# Enterprise Aligned AI, full text A podcast featuring real conversations with enterprise leaders implementing AI, exploring ROI, reliability, security, organization, and real-world solutions for moving from pilots to production. Hosted by Aparna Sinha. Every document below is also available on its own. Episodes and blog posts live at https://enterprisealignedai.com/episodes/ and https://enterprisealignedai.com/blog/, with markdown at the same URL plus `.md`. The index is at https://enterprisealignedai.com/llms.txt. Generated from 6 episodes and 10 blog posts. --- # Episodes ## Ep. 1: How Wayfair uses Arize for Reliable Agentic AI - URL: https://enterprisealignedai.com/episodes/how-wayfair-uses-arize-for-reliable-agentic-ai - Published: 2026-03-03 - Duration: 50 min - Video: https://www.youtube.com/watch?v=mKcRisRl_eM - Guests: Aman Khan, Victor Sulaiman - Tags: agentic AI, evals, retail, logistics, Wayfair, Arize AI **Summary:** Host Aparna Sinha discusses the current landscape of enterprise AI with speakers Victor Sulaiman from Wayfair and Aman Khan from Arize AI. They explore the integration of AI in retail, the importance of evaluating AI systems, and the emerging field of Agentic Commerce. The conversation delves into optimizing logistics and support with AI, the necessity of human oversight on autonomous agents, and techniques for measuring ROI of enterprise AI investments. **Why listen:** If you're trying to figure out where the ROI actually lives in Agentic AI, this is the episode for you. Learn how one of the world's largest retailers moved beyond the pilot phase into production-grade agentic systems — from customer support automation to freight logistics optimization. ### Key takeaways - Agentic Commerce: the future of retail isn't a search bar — it's an agent that finds, visualizes, and recommends products in new ways - Wayfair reduced ticket resolution times from 7 days to 2 days by automating low-lift tickets - LLM reasoning can optimize freight capacity beyond traditional algorithms - Using multiple LLMs as a jury to deliberate on decisions avoids costly errors - Sometimes a simple reflex agent beats a complex hierarchical one — avoid over-engineering ### Chapters - 0:00: Introduction to Enterprise AI and Speakers - 2:16: AI in Retail: Enhancing Customer and Associate Experience - 6:00: Evaluating AI: Arize's Unique Approach - 10:36: Open Source: Power of Community Engagement in AI - 13:41: Adapting to Generative AI: Evolving Use Cases - 15:31: Agentic Commerce: The Future of Retail - 19:08: Agentic Logistics: Can Reasoning Optimize Delivery? - 22:24: Measuring Reliability and Determinism in AI Systems - 25:52: Launching AI Agents: The Importance of Real-World Feedback - 28:52: LLM as a Jury with Human Oversight in AI Systems - 33:25: Navigating Autonomous Agents in Business - 37:43: Understanding ROI in AI Investments - 42:58: Build vs. Buy: Strategies for AI Deployment - 45:44: Scaling AI from Pilot to Production ### Clips - How Wayfair Makes Its Agents Reliable - URL: https://www.youtube.com/watch?v=VSd1AupMyWo - Speaker: Victor Sulaiman - Occurs at: 22:24 - Claim: Multiple LLMs as a jury catch costly errors. - Agentic Commerce at Wayfair: Agents That Find and Recommend Products - URL: https://www.youtube.com/watch?v=9YHk25KM48E - Speaker: Victor Sulaiman - Occurs at: 15:31 - Claim: Agents find, visualize, and recommend products for shoppers. - Inside Arize's Alyx Agent CLI - URL: https://www.youtube.com/watch?v=06WIXbFnBoA - Speaker: Aman Khan - Occurs at: 6:00 - Claim: Arize's Alyx runs as an agent from your CLI. In the debut episode of Enterprise Aligned AI, host Aparna Sinha sits down with **Victor Sulaiman**, Senior Product Manager at Wayfair, and **Aman Khan**, Head of Product at Arize AI, to get into the weeds of how one of the world's largest retailers moved beyond the "pilot phase" into production-grade agentic AI systems. ## AI in Retail: More Than Chatbots Victor walks through how Wayfair is deploying AI across the business — from customer support automation that reduced ticket resolution times from 7 days to just 2 days, to agentic commerce experiences that help customers find and visualize products in entirely new ways. ## The Evaluation Problem Aman explains why building an AI agent is easy but making it reliable is the real challenge. Arize's approach to scalable AI evaluation and governance — including open-source tools — gives teams the confidence to ship agents to production. The key insight: you need to evaluate agents in the real world, not just in test environments. ## Agentic Logistics One of the most fascinating segments explores how Wayfair uses LLM reasoning — not just traditional algorithms — to optimize freight capacity and container volume. This is agentic AI applied to hard operational problems where the ROI is immediately measurable. ## LLM as a Jury Why does Wayfair use multiple LLMs to "deliberate" on decisions like furniture translations across languages? Because a single model makes hilarious (and costly) errors. The jury approach with human oversight creates a reliability layer that makes autonomous decisions safe for production. ## Build vs. Buy and Scaling Up The episode concludes with practical insights on the build vs. buy dilemma for AI tooling and strategies for scaling AI from pilot projects to full production deployment. --- ## Ep. 2: Royal Bank of Canada's $1B AI Transformation - URL: https://enterprisealignedai.com/episodes/rbc-1b-ai-transformation - Published: 2026-04-21 - Duration: 51 min - Video: https://www.youtube.com/watch?v=pVWtq0yCrEA - Guests: Vinh Tran - Tags: agentic AI, financial services, MCP, OpenTelemetry, governance, platform engineering, RBC **Summary:** Host Aparna Sinha sits down with Vinh Tran, Vice President of Data & AI Platforms and RBC Fellow at Royal Bank of Canada, to unpack the $700M to $1B AI value commitment RBC's CEO made to the street at the 2025 investor day, and how Vinh's team is delivering it by the end of 2027. They get into which use cases move first (back-office and call-center insight, not direct-to-customer), how RBC is measuring ROI across nine flagship projects and five lines of business, why Vinh favors build over buy in a regulated industry, the platform pattern behind RBC Assist and Self-Serve Agents, how the Lumina data platform and a data-product mindset feed the agents, the control plane of guardrails and judges that keeps agentic AI reliable, and why 'AI for All', 96,000 employees and not just 8,000 developers, is the real transformation. **Why listen:** If you're building AI inside a regulated enterprise, this is a rare, specific look at what it actually takes to move from pilots to production at the scale of Canada's largest bank. Vinh walks through the platform, the standards (MCP, OpenTelemetry), the control plane, and the workforce story, and shares his mental model on Building vs. Buying: 'I really feel that we have to have the courage to build. We're not going to write every line of code. We're going to use the tools. We're going to use the cloud. We're going to use the community.' ### Key takeaways - RBC's CEO committed $700M to $1B of AI-driven value at the 2025 investor day, to be delivered by the end of 2027 across nine flagship projects spanning all five lines of business plus technology and operations - The real transformation isn't the 8,000 developers and data scientists who already used AI, it's getting AI into the hands of all 96,000 staff: mortgage brokers, branch managers, account managers, engineers - RBC is starting with back-office and call-center insight use cases (not direct-to-customer) while the technology and the organization build the muscle; fully autonomous customer-facing AI is 'possibly, one day' - In a regulated industry, 'build' beats 'buy' not because it's cheaper but because it gives you control of your destiny: your policies, your patches, your compliance posture, plus a stronger engineering culture - Three pillars scale AI across the enterprise: a foundation platform (gateway-fronted approved models), universal access and literacy, and reusable scaffolding so builders don't start from an empty folder - Reliability for agentic workflows comes from a control plane: LLM-as-judge guardrails, PII and content filters, a gated MCP server registry, human-in-the-loop on material changes, and continuous evaluation from production telemetry, not once-a-week batch evals - Sometimes you don't need an agent at all, deterministic workflows should stay as deterministic code; agents are for complex, changing, human-interaction-heavy work like intent classification and summarization - Engineers become builders: coding tools are giving 10x productivity, but 'just because Vin can vibe code a system, we should not put it in production and run our banking systems against that' - Don't fall in love with today's tools or paradigms, 'we're in chapter one or two of a six-chapter book.' Build AI literacy, get started, iterate ### Chapters - 0:00: Introduction to Vinh Tran - 3:34: The Scope of RBC's $1B AI Transformation - 6:10: Transforming Customer Service - 8:47: Measuring ROI of AI Across Business Units - 14:23: Build vs. Buy: Why Build in Regulated Enterprises - 18:37: Self-Service Agents at RBC - 22:25: Setting Platform Standards: MCP and OpenTelemetry - 26:00: Preparing Data for AI: Best Practices - 30:33: Overcoming Data Readiness Challenges - 33:16: RBC's Control Plane for Reliable Agentic AI - 37:10: Use Cases for LLMs in Banking - 40:37: RBC Assist: Productivity Chatbot or Superpower? - 43:13: Workforce Transformation through 'AI for All' - 45:58: Hiring the Next Generation of Builders - 49:25: Advice for Enterprise AI Leaders ### Clips - How RBC Scales AI With Self-Serve Agents on Its Platform - URL: https://www.youtube.com/watch?v=eEATehEnnSs - Speaker: Vinh Tran - Occurs at: 18:37 - Claim: RBC is putting AI in the hands of all 96,000 staff. - RBC's AI Transformation Spans All Five Lines of Business - URL: https://www.youtube.com/watch?v=bLe0xFnVVQE - Speaker: Vinh Tran - Occurs at: 3:34 - Claim: Nine flagship AI projects across all five lines of business. - RBC's CEO Committed $700M to $1B of AI Value by 2027 - URL: https://www.youtube.com/watch?v=FqNUauiuXsQ - Speaker: Vinh Tran - Occurs at: 3:34 - Claim: RBC committed $700M to $1B of AI value by 2027. - Why a Regulated Bank Builds Its AI In-House - URL: https://www.youtube.com/watch?v=BnpfmiOi_h4 - Speaker: Vinh Tran - Occurs at: 14:23 - Claim: In a regulated bank, building gives you control. - Coding Tools Are Turning RBC Engineers Into Builders - URL: https://www.youtube.com/watch?v=QuvREwmW36s - Speaker: Vinh Tran - Occurs at: 45:58 - Claim: Coding tools give engineers 10x productivity. - Why RBC Standardized on MCP and OpenTelemetry - URL: https://www.youtube.com/watch?v=OmvAcBNiiOk - Speaker: Vinh Tran - Occurs at: 22:25 - Claim: RBC standardized its platform on MCP and OpenTelemetry. - RBC Assist: From Chatbot to Self-Service Agent Platform - URL: https://www.youtube.com/watch?v=L7of3L9bdoM - Speaker: Vinh Tran - Occurs at: 40:37 - Claim: RBC Assist grew from a chatbot into a self-service platform. - Everyone at RBC Can Build an Agent - URL: https://www.youtube.com/watch?v=mdyEH17tsVE - Speaker: Vinh Tran - Occurs at: 43:13 - Claim: Every employee can build a secure agent, no code required. - RBC's Call Center Summarization Use Case - URL: https://www.youtube.com/watch?v=gs3_RB_UNgg - Speaker: Vinh Tran - Occurs at: 6:10 - Claim: Back-office and call-center insight use cases came first. In the second episode of Enterprise Aligned AI, host Aparna Sinha sits down with **Vinh Tran**, Vice President of Data & AI Platforms and RBC Fellow at Royal Bank of Canada, ranked #1 in Canada and #3 globally for AI maturity on the 2025 Evident AI Index, to get into what an enterprise-wide AI transformation actually looks like from the inside. ## The Scope of RBC's $1B AI Transformation At RBC's 2025 investor day, the CEO committed $700M to $1B of AI-driven value, to be delivered by the end of 2027. Vinh joined the Borealis team shortly after to lead the platforms that make that possible. Nine flagship projects, with hard KPIs tracked by finance, span all five lines of business plus technology and operations. But the number Vinh keeps coming back to isn't the dollar figure, it's **96,000**. Historically AI at the bank served 6 to 8 thousand developers and data scientists. The transformation lifts everyone with a computer: mortgage brokers, branch managers, account managers, wealth advisors. ## Measuring ROI of AI Across Business Units When the CEO goes to the street and commits a number, that number needs to be measurable. RBC tracks ROI across those nine projects with clear KPIs, split between revenue uplift and cost reduction, with finance and accounting teams doing the quantification. A concrete example: call-center productivity. AI drafts the post-call notes and summaries that agents used to write by hand, shaving meaningful time off every call and materially reducing the cognitive load on staff who take dozens or hundreds of calls a day. ## Build vs. Buy: Why Build in Regulated Enterprises Vinh is unapologetic about building. Three things changed the math: open-source ecosystems are now production-grade, coding tools give 10x to 100x engineering leverage, and the cloud democratizes services that used to be the domain of large vendors. But cost savings aren't the main reason. In a regulated industry, building gives you **control of your destiny**, your policies, your patches, your compliance posture, and it builds a culture of engineering excellence that you can't buy. ## Self-Service Agents and Platform Standards RBC's platform scales AI through three pillars: a foundation with pre-approved models behind an LLM gateway (with model risk, security, and bias reviews done up-front), universal access and literacy, and reusable scaffolding with LangGraph references, MCP tool registration, and knowledge connectors, so builders aren't starting from an empty folder. Standards matter too: RBC is standardizing on **Model Context Protocol (MCP)**, agent-to-agent, and **OpenTelemetry** to stay pluggable in a market that mostly isn't standardizing. ## Preparing Data for AI: The Lumina Platform Data is the fuel. Vinh credits his predecessor for starting the Lumina data platform five to six years ago, data lake, warehouse, semantic layers, metadata, and for moving the bank to a **data-product mindset** with curated data hubs (retail credit, mortgage) instead of everyone copying data everywhere. For other enterprises: the blocker isn't the tool. It's the endless architecture meetings. Start, classify, iterate, and use AI itself to generate metadata and tags. ## RBC's Control Plane for Reliable Agentic AI Reliability in banking workflows comes from a control plane: LLM-as-judge guardrails where one model reviews another's intent classification, PII and sensitive-content filters, an MCP gateway that enforces human-in-the-loop for material actions, and an AI architecture review that asks the most important question first: *do you actually need an agent for this?* Deterministic workflows should stay deterministic code. Evaluation is continuous, pulled from production telemetry, not a weekly batch of 1,000 prompts. ## RBC Assist and 'AI for All' What started in May 2024 as an internal chatbot has evolved into a no-code agentic productivity platform now used across the bank every day, product managers synthesize Jira boards, staff analyze uploaded files alongside MCP-served enterprise data, and anyone can build an agent without writing code. The workforce story is the one Vinh is proudest of: RBC serves 17 million clients with 96,000 staff today and wants to serve 25 million with the same team, which means making every employee more effective, not cutting. ## Hiring the Next Generation of Builders Engineers become **builders**: fewer ceremonies, faster delivery, but also a much harder peer-review problem when AI is writing thousands of lines of code. *"Just because Vin can vibe code a system, we should not put that in production and run our banking systems against that."* The next generation of RBC engineers needs AI fluency, orchestration instinct, and the discipline to make sure the code the AI writes is the right code, with the right tests, for the right use case. ## Advice for Enterprise AI Leaders "We're in chapter one or two of a six-chapter book." Don't fall in love with today's tools or paradigms. Get started, build the AI literacy, pivot with the industry. --- ## Ep. 3: Logitech's Award-Winning AI Share of Voice, Mastering Agentic Commerce with Typesense - URL: https://enterprisealignedai.com/episodes/agentic-commerce-with-logitech-and-typesense - Published: 2026-05-24 - Duration: 43 min - Video: https://www.youtube.com/watch?v=vFNzd3rh3wY - Guests: Deepa Shekhar, Jason Bosco - Tags: agentic commerce, e-commerce, search, AI visibility, composable architecture, Logitech, Typesense, MCP **Summary:** Host Aparna Sinha sits down with Deepa Shekhar, Director of E-commerce Technologies at Logitech, and Jason Bosco, CEO and Co-founder of Typesense, for a two-sided look at agentic commerce. Deepa lays out her framework of three tracks of agentic commerce, distinguished by where the transaction happens and who owns the consumer experience, and walks through the composable architecture and content strategy that helped Logitech.com win the 'Growth Engine' award in Semrush's inaugural 2025 AI Visibility Awards for Consumer Electronics, measured across thousands of real prompts in ChatGPT and Google AI Mode. Jason brings the view from underneath the stack: why the search and retrieval layer is the most underestimated piece of the agentic commerce stack, what most teams get wrong when they reach for a vector database before they've gotten lexical retrieval right, and how Typesense scales to 10B+ queries a month as the open-source alternative to Algolia and Pinecone. The conversation gets into the fragmented product catalog and commerce protocols (ACP, UCP, MCP), real-time inventory and post-purchase gaps, legacy OMS and WMS constraints, emerging efforts like OnX, and the financial guardrails, auditability, observability, and verification work that agentic commerce still has ahead of it. **Why listen:** If you sell anything online, agentic commerce is your new top of funnel. This episode pairs a brand-side operator who has already rebuilt her stack for the agent era with the infrastructure founder whose search engine runs underneath some of the largest commerce stacks in the world. Deepa's three tracks give you a way to think about where to invest. Jason's view on retrieval gives you a way to think about whether your current stack can actually serve agents at the latency they expect. ### Key takeaways - Three Tracks of Agentic Commerce: commerce happening inside LLM platforms, brand-owned agents across web/app/messaging, and AI agents acting as autonomous buyers on behalf of consumers - Every agentic system is, at its core, a search system, retrieval decides whether agents can find the right product, in the right context, at the right latency - There is no universal real-time product catalog interface for LLMs yet, and protocols like ACP, UCP, MCP, and OnX are emerging to fill the gap, each with different tradeoffs - Content strategy inverts in the agentic era, from brand storytelling to answering customer questions, structured for machine readability with frameworks like llms.txt - Composable architecture is the prerequisite, not the strategy, when the front end, search, cart, and content can be independently composed and called by an agent, you can show up wherever the customer is - Logitech won the 'Growth Engine' award in Semrush's 2025 AI Visibility Awards for Consumer Electronics, measured across thousands of real prompts in ChatGPT and Google AI Mode, demonstrating that AI visibility is already a measurable funnel - Brand agents need shared intelligence, not isolated chatbots, one shared AI layer should power web, app, WhatsApp, and RCS, not five fragmented experiences - Most teams reach for a vector database before they have lexical retrieval right, and pay for it in relevance and latency - Legacy ERP, WMS, and OMS systems were architected for human web traffic and overnight batches; supporting agent traffic requires event-driven, sub-second translation layers - Agentic commerce introduces new risk surface area, prompt injection, agent manipulation, and material financial actions, that requires LLM-as-judge evaluators, audit trails, and human-in-the-loop guardrails ### Chapters - 0:00: Agents Disrupt Shopping - 0:21: Podcast and Guests - 2:34: Three Tracks of Commerce - 5:49: Search for Agentic Retrieval - 8:14: Catalog Data and Protocols - 10:55: Unifying Data with Search - 13:49: Content Strategy for LLMs - 17:09: Modern Retrieval Stack - 22:08: Building Brand Agents - 26:32: Payments, Inventory, and Ops - 33:16: Track Three: Agent Buyers - 35:58: Agents Evaluate Vendors - 40:12: Security, Guardrails, and Wrap ### Clips - The Three Tracks of Agentic Commerce - URL: https://www.youtube.com/watch?v=xajQ6vN98MA - Speaker: Deepa Shekhar - Occurs at: 2:34 - Claim: Commerce in LLMs, brand-owned agents, and agent buyers. - Every Agentic System Is a Search System - URL: https://www.youtube.com/watch?v=ywLIxiDqWhk - Speaker: Jason Bosco - Occurs at: 5:49 - Claim: Every agentic system is, at its core, a search system. - Agentic Commerce Needs Markdown and llms.txt - URL: https://www.youtube.com/watch?v=lwH6LQQ2Npw - Speaker: Jason Bosco - Occurs at: 13:49 - Claim: Content strategy inverts to machine-readable answers. - One Shared AI Layer Should Power Web, App, and WhatsApp - URL: https://www.youtube.com/watch?v=Ut5I8kVnU-Q - Speaker: Deepa Shekhar - Occurs at: 22:08 - Claim: One shared AI layer powers web, app, and WhatsApp. - Logitech Won Semrush's 2025 AI Visibility Growth Engine Award - URL: https://www.youtube.com/watch?v=8egd70l-DGk - Speaker: Deepa Shekhar - Occurs at: 13:49 - Claim: Logitech won Semrush's AI Visibility award for consumer electronics. In the third episode of Enterprise Aligned AI, host Aparna Sinha sits down with **Deepa Shekhar**, Director of E-commerce Technologies at Logitech, and **Jason Bosco**, CEO and Co-founder of Typesense, for a two-sided look at agentic commerce, the brand-side operator and the infrastructure founder. ## Deepa's Three Tracks of Agentic Commerce Deepa offers a clean way to think about where this is heading. The three tracks are distinguished by two questions: where does the transaction actually happen, and who owns the consumer experience? > "I see agentic commerce evolving across three tracks, distinguished by where the commerce transaction takes place and who owns the consumer experience. We must execute on all three tracks to be successful." — Deepa Shekhar - **Track 1: LLM Platforms as Strategic Engines of Commerce.** Discovery and transaction journeys are moving from retailer websites to third-party LLMs like ChatGPT, Perplexity, Gemini, and Claude. The new KPIs are visibility and conversion inside LLM platforms, not website click-through rates. - **Track 2: Proprietary Brand Agents Grounded in Your Data.** Building your own AI agents (in your app, your website, messaging channels) is more important than ever, in your unique brand voice enriched with your user and product data. - **Track 3: Pure Algorithmic Buying.** Soon, personal consumer AI agents will shop on behalf of humans, and enterprise systems need to be ready. Agents rely on structured data and select products objectively based on operational parameters: granular specifications, shipping velocity, and historical fulfillment reliability. ## AI Visibility Is the New Top of Funnel Logitech won the "Growth Engine" award in Semrush's inaugural 2025 AI Visibility Awards for Consumer Electronics, measured across thousands of real prompts in ChatGPT and Google AI Mode. Logitech.com was also named PCMag Reader's Choice 2025 Best Manufacturer Online Store. The work to show up inside agentic surfaces is already producing a measurable funnel. This isn't future state. To win mindshare inside LLM answer windows, Deepa explains that marketing and technology teams must partner closely and jettison the keyword-based SEO mindset. Shoppers in LLMs ask intent-driven questions: "What is a good mouse for long working hours?" "What is a good mouse that works cleanly on a glass surface?" Logitech rewrote their content for those questions, and engineered it for machine readability with standards like `llms.txt`. ## Composable Architecture Is the Prerequisite Logitech's shift to composable architecture wasn't an AI play. It's the reason agentic commerce is available to Deepa's team today. When the front end, the search layer, the cart, and the content can each be composed and called by an agent, you can show up wherever the customer is, whether that's the brand site, an LLM, or something that doesn't exist yet. The infrastructure decisions made to "modernize" the stack now look like AI-readiness decisions. The work compounds. ## Jason's View from Underneath the Stack Jason has a slightly counter-narrative point of view. The conversation has gravitated to LLMs and vector databases, but whether an agent can find the right product, in the right context, at low latency, is decided in the search and retrieval layer. Typesense ships as a single C++ binary, open-source, revenue-funded, with 25,000+ GitHub stars and 25M+ Docker pulls behind it, and runs 10B+ queries a month on Typesense Cloud as the open-source alternative to Algolia and Pinecone. Jason explains what most teams get wrong when they reach for a vector database before they've gotten lexical retrieval right. In Deepa and Jason's recommended architecture, the LLM does intent detection (translating a query about "repetitive strain injury" into an ergonomic design match), then passes the extracted intent to a high-performance search engine that runs dozens of query combinations across millions of records in milliseconds. The result is more queries per second, better answers, and higher AI visibility. ## The Missing Product Catalog Interface One roadblock facing enterprises trying to plug into Agentic Commerce is that there is no unified mechanism to share product catalogs with LLMs. Deepa walks through the fragmented protocol landscape: - **OpenAI's Agentic Commerce Protocol (ACP)** provides structured capabilities for syncing catalogs, pricing models, and active promotions, but lacks native payment capabilities. - **Universal Commerce Protocol (UCP)** defines standards for checkout, order updates, and payment handoffs, but only checks inventory at the final moment of checkout, which forces merchants to build parallel low-latency inventory APIs to avoid cart abandonment. - **Model Context Protocol (MCP)** is not yet intuitive for scaled retail and functions more like a sandboxed app extension inside the LLM container. - **Order Exchange Protocol (OnX)**, championed by a consortium of modern OMS and e-commerce vendors, is aiming to introduce standardized, real-time eventing layers over legacy backends. The industry needs a unified, cross-platform protocol that binds real-time inventory, batch product attributes, and multi-vendor financial transactions into a single standard. Until then, engineering teams face significant architectural friction. ## Agents Are Already Making Product Decisions Perhaps the best moment of the discussion: Jason recounts an autonomous agent (built to construct a personal knowledge repository) that was given autonomy to choose its own technology stack. It rejected a vector-only database, discovered Typesense, systematically tested its hybrid search and typo tolerance, and deployed it into production, documenting the whole evaluation in its own log files. After seeing this, Jason updated how Typesense exposes its technical assets, serving documentation as pure markdown alongside HTML, and embedding metadata that points crawling AI bots at the markdown versions so they don't waste tokens parsing decorative web formatting. ## Guardrails for the Buy Click Agents executing financial transactions require financial integrity and system auditability. The primary defense against AI-driven transactional risk is automated, real-time evaluation layers, specialized LLM evaluators in production to dynamically audit conversation flows, monitor financial thresholds, and catch prompt injection vectors before anomalies register on the financial ledger. This is a domain that's still early; the conversation barely scratches the surface of what's coming. ## Why This Matters Now If you're running e-commerce, the question isn't "should we have an agentic strategy?" anymore. The question is which of Deepa's three tracks you're investing in, and whether the layer underneath your store (the part Jason cares about) can actually serve those agents at the latency and relevance bar they expect. --- ## Ep. 4: How Coursera Became an AI-Native Enterprise (with Mint MCP) - URL: https://enterprisealignedai.com/episodes/coursera-becoming-ai-native - Published: 2026-07-06 - Duration: 48 min - Video: https://www.youtube.com/watch?v=hcKmyD27mms - Guests: Mustafa Furniturewala, Jiquan Ngiam - Tags: AI-native, MCP, enablement, enterprise AI, agentic engineering, Coursera, Mint MCP **Summary:** Host Aparna Sinha sits down with Mustafa Furniturewala, CTO of Coursera, and Jiquan Ngiam, co-founder and CEO of Mint MCP, on what it takes to become AI-native inside a fourteen-year-old public company, in the one industry everyone assumes AI will replace. Mustafa walks through the playbook: encourage AI rather than police it, then add the enablement and guardrails to make it safe, an MCP gateway that governs data access, tools bundled by role, and deterministic policies IT can stand behind. Jiquan explains why MCP exists and what it does that APIs and CLIs can't: tools built for the model to call, with a security boundary an API key cannot express. The conversation gets into agentic code migration across 200+ repos, managing context with nested CLAUDE.md files instead of graph-database indexing, background agents that act like coworkers, a supply-chain attack that turned a compromised npm package against developer machines, and how Coursera's business users now write their own AI skills, driving more than a quarter of the company's Claude usage. **Why listen:** If you are trying to move an enterprise from AI experiments to AI-native, this is the operator's account of how one CTO did it without slowing down. Mustafa's lean-in-then-guardrail sequence is a repeatable playbook, and his candor about what fails (point an agent at a decade-old codebase cold and it hallucinates) is as useful as the wins. Jiquan gives the clearest explanation of why MCP matters that cuts through the hype, and both guests show how the same enablement layer that transforms engineering also unlocks business users. ### Key takeaways - Blocking AI is a security risk: block a useful tool and people route around it with unsafe copy-paste and temporary data stores, so enablement is the safer path - MCP tools are built for the model to call, not for developers; one MCP tool call can wrap several API calls with validation, and expose a boundary an API key cannot (read and write, but never delete) - An MCP gateway governs MCP sprawl: Coursera bundles tools by role, hosts its own servers, and sets deterministic policies like no deleting and read only public Slack channels - Agentic migration turned weeks of work into two to three hours across 200+ repos and millions of lines, but only with the system around the agent; point Claude Code at a codebase that size cold and it hallucinates - For large codebases, manage context with inline search and nested CLAUDE.md files rather than reaching for graph-database indexing - Background agents act like coworkers: a sandboxed loop where Claude plans and codes and Codex reviews with fresh eyes, triggered from Slack, running async and in parallel - Business users write their own AI skills: Coursera forked Anthropic's knowledge-work skills, and more than a quarter of Claude usage now comes from people who don't code, via Claude Cowork - Becoming AI-native is a practice you keep improving, sustained by an AI council, a plugins repo, monthly awards, and internal sessions where people show how they use AI - Security has to be designed in, not bolted on: a compromised npm package used Claude Code on developer machines to find GitHub secrets, making the case for secure-by-design MCP deployments ### Chapters - 0:00: Migrating a Decade of Code in Hours at Coursera - 2:03: AI in Education: Block or Embrace? - 5:46: Personalization at Scale: Adaptive Assessments and AI Tutors - 10:02: MCP vs APIs and CLIs: Secure Data Access for Agents - 14:07: Skills: Teaching AI your SOPs as Markdown - 18:03: Why Coursera Needed an MCP Gateway - 21:53: How Coursera Became an AI Native Enterprise - 24:37: AI Does the Grunt Work, Engineers Do the Rest - 27:24: GraphRAG vs Inline Context for Large Codebases - 30:37: Building a Culture of AI Adoption - 32:35: The New Agentic Engineering Lifecycle - 35:28: Background Agents That Act Like Coworkers - 39:41: When a Supply-Chain Attack Turns Your AI Against You - 43:28: Unlocking Business Users with Claude Cowork - 47:42: The Future of Personalized Learning ### Clips - Business Users at Coursera Write Their Own AI Skills - URL: https://www.youtube.com/watch?v=I09YzKAi36M - Speaker: Mustafa Furniturewala - Occurs at: 43:28 - Claim: A quarter of Coursera's Claude usage comes from business users. - Why Coursera Needed an MCP Gateway - URL: https://www.youtube.com/watch?v=ThTBBL6eqjc - Speaker: Mustafa Furniturewala - Occurs at: 18:03 - Claim: Coursera bundles MCP tools by role. - AI Is Exposing That Multiple-Choice Tests Do Not Teach Well - URL: https://www.youtube.com/watch?v=FiaJ1iIcRXI - Speaker: Mustafa Furniturewala - Occurs at: 2:03 - Claim: Multiple-choice questions are not the most effective way to learn. - How Coursera Sustains AI Adoption: AI Council, Plugins Repo, Monthly Awards - URL: https://www.youtube.com/watch?v=XQ-cWd5ZhkA - Speaker: Mustafa Furniturewala - Occurs at: 30:37 - Claim: Coursera runs an AI council and a plugins repo. - Claude Cowork Edits Your Files Directly - URL: https://www.youtube.com/watch?v=0WTlUM7KL78 - Speaker: Jiquan Ngiam - Occurs at: 43:28 - Claim: Cowork edits your files directly. - A Background Agent Where Claude Codes and Codex Reviews - URL: https://www.youtube.com/watch?v=pCxDyiztEDg - Speaker: Jiquan Ngiam - Occurs at: 35:28 - Claim: Claude writes the code. Codex reviews it. - How Mint MCP Enforces Enterprise AI Security & Policy Controls - URL: https://www.youtube.com/watch?v=7MVL4XSBNOY - Speaker: Jiquan Ngiam - Occurs at: 18:03 - Claim: The agent can read and write. It cannot delete. - Why an API Key Cannot Express an Agent's Security Boundary - URL: https://www.youtube.com/watch?v=MC6J-tGgiBE - Speaker: Jiquan Ngiam - Occurs at: 10:02 - Claim: MCP tools are built for the model to call. In the fourth episode of Enterprise Aligned AI, host Aparna Sinha sits down with **Mustafa Furniturewala**, CTO of Coursera, and **Jiquan Ngiam**, co-founder and CEO of Mint MCP. Coursera is fourteen years old and public, in the one industry everyone assumes AI will replace. Instead of blocking AI, Mustafa leaned in, then built the structure to make leaning in safe. ## Lean in, because blocking is worse Mustafa's starting point was to encourage AI use, in the product and across the company. On education, he reframes the smarter-or-dumber debate: the problem AI surfaced was already there. > "The assessment types, like multiple-choice questions, are not the most effective way to learn, and AI is exposing that faster than we expected." — Mustafa Furniturewala The case for enablement is a security case. Block a useful tool and people route around it, copying data into temporary stores you cannot see. Blocking relocates the risk instead of removing it. ## Why MCP, and why a gateway Jiquan corrects a debate the internet keeps garbling. APIs were designed for developers and are CRUD-based. MCP tools are designed for the model to call, so reliability is higher, and one tool call can wrap several API calls with validation. The security difference is concrete: an API key can read, write, and delete, while the MCP layer exposes a tool abstraction that never calls the delete API. Coursera hit close to 100% AI adoption before it had enablement. API tokens were scattered, business users wanted access, and nobody could answer which MCP tools were approved. An MCP gateway closed that gap, bundling tools by role and letting IT set deterministic boundaries. > "It cannot be prompt the bot to say please don't do this, pray and hope. Put a sandbox around it, and the only way for data to leave and enter the sandbox is through the MCP gateway." — Jiquan Ngiam ## Transforming engineering, then everyone Coursera migrated code that took engineers weeks in two to three hours, across 200+ repos and millions of lines. The honest caveat matters as much as the number: point Claude Code at a codebase that size cold and it hallucinates. The setup around the agent, documented architecture and nested CLAUDE.md files, does the work. Jiquan runs review as a background agent, a sandboxed loop where Claude plans and Codex reviews with fresh eyes, triggered from Slack. The demand that surprised Mustafa came from business users. Coursera forked Anthropic's knowledge-work skills, tuned them to how the company runs, and now more than a quarter of its Claude usage comes from people who don't write code. ## What becoming AI-native takes Becoming AI-native is a practice you keep improving, sustained by an AI council, a plugins repo, monthly awards, and sessions where people show how they use AI. > "I don't think we are fully there either. We're getting there, but the key is to have a system, keep improving it, and create feedback loops where the value compounds over time." — Mustafa Furniturewala --- ## Ep. 5: Who Owns the Legal AI Harness? - URL: https://enterprisealignedai.com/episodes/who-owns-the-legal-ai-harness - Published: 2026-07-22 - Duration: 48 min - Video: https://www.youtube.com/watch?v=voxL8usOV-A - Guests: Danielle Benecke, Max Junestrand - Tags: Legal AI, Harnesses, Build vs Buy, Agentic AI, Compliance, Pricing **Summary:** Host Aparna Sinha sits down with Danielle Benecke, who founded and leads Baker McKenzie's Applied AI practice, and Max Junestrand, CEO and co-founder of Legora, on the question the legal industry has quietly moved to. The conversation starts with what a harness is and ends with who should own one. Danielle argues the real opportunity is not accelerating the work lawyers already do but building services that could not exist before, and that the third wave of AI in legal will be runtime compliance steering for enterprises running agents across the business. Max explains what a legal harness decides that a general one cannot: which documents the model reads, what counts as authoritative between a court case and an internal precedent, which models run where, and where sub-agents are allowed. Along the way they get into why generating legal analysis is now cheap while judgment stays scarce, why consumption pricing arrived in legal AI, why one general counsel's AI-assisted first pass made the work more expensive rather than cheaper, and how a global firm decides which capabilities to buy and which to keep inside. **Why listen:** Most AI conversations in the enterprise are still about which tool to roll out. This one is about what happens after that question is settled: which parts of the AI stack a firm should own, and which it should rent. Danielle's generalizable-to-custom and strategic-criticality axes give you a way to make that call. Max gives the clearest working definition of a harness available anywhere, using Claude Code and Cursor as the reference points. Both are operators describing decisions they have already made, with the commercial consequences attached. ### Key takeaways - The industry question has moved from which tool to roll out to which harnesses a firm should own rather than rent - A harness decides what the model reads, what is authoritative between a case and an internal precedent, which models run where, and whether sub-agents are allowed; the simplest harness is a folder of markdown files - Own the harness where the work turns on your expert steering, internal playbooks, and know-how; rent it where the capability is generalized and not strategically critical - Enterprises run multiple harnesses at once: direct model access, productivity-layer harnesses, vendor tooling, and their own bespoke systems per service line - Generating content and legal analysis is cheap now; judgment and accountability are scarce, so the goal is steering agents upstream rather than reviewing more output downstream - The three waves in legal AI: tools for lawyers, then AI-native legal and compliance businesses, then compliance systems that steer agents and business behavior in real time - The bigger opportunity is complex work humans could not do before, not accelerating the lower-level work the market is focused on - Marketing was first to roll out AI at many enterprises, moving legal review from tens of pieces of content to tens of thousands across the long tail of countries - Expertise now sits upstream in designing the system; the better the design, the less downstream validation is needed, which changes what a firm sells - Consumption pricing came to legal AI because these businesses do not have software margins, while law firms move the other way toward fixed fees to absorb that uncertainty for clients - One general counsel asked firms to quote both validating her team's AI-generated first pass and doing the deal end to end; validating the AI first pass came back more expensive - A more expensive model can be the cheaper choice, because token efficiency varies and a cheaper model can use more of them ### Chapters - 0:00: Legal AI: Whose Harness Is It Anyway? - 0:33: Meet the Guests: a Global Law Firm and a $5.6B Legal AI Startup - 1:45: Why a Law Firm Started Building Its Own AI - 3:10: Legora: How This High-Growth AI Startup Was Founded - 4:57: Legora's Vision: Making Law More Accessible and Creative - 6:57: Baker McKenzie on the New Legal Work AI Enables - 9:24: Baker McKenzie: The Three Waves of AI in Law - 10:57: Legora's "Legal Engineer": A Lawyer Who Builds AI - 12:50: Legora's Real Moat: Deep Enterprise Integration - 17:42: What Legora's Legal Harness Does That Claude Code Can't - 21:05: Baker McKenzie: When to Own vs Rent Your AI Harness - 22:26: Baker McKenzie: Use Human Judgment Upstream to Design the System - 25:05: Baker McKenzie: From Billable Hours to Fixed Fees - 26:25: Legora: Why It Moved to Consumption-Based Pricing - 28:01: How Legora Scaled From $1M to $100M in Under 18 Months - 34:03: Baker McKenzie: Same-Day Legal Work in a Cyber Attack - 38:15: When a Local Lawyer Beats a Frontier AI Model - 39:46: When Using AI Costs the Client More - 40:37: Build vs Buy for Legal AI - 44:12: How Legora Stands Out in a Crowded Vendor Market - 47:10: The Future of Law and Lawyers ### Clips - Legora: Making Law More Accessible and Creative - URL: https://www.youtube.com/watch?v=SSDURRKRfvQ - Speaker: Max Junestrand - Occurs at: 4:57 - Three Waves of Legal AI: Compliance Is Next - URL: https://www.youtube.com/watch?v=122-zVuVrKo - Speaker: Danielle Benecke - Occurs at: 9:24 - What Legora's Legal Harness Does That Claude Code Can't - URL: https://www.youtube.com/watch?v=ASbV3-js0Ww - Speaker: Max Junestrand - Occurs at: 17:42 - Legal AI: Use Human Judgment Upstream to Design the System - URL: https://www.youtube.com/watch?v=ZdYKJ-VVF-4 - Speaker: Danielle Benecke - Occurs at: 22:26 - Baker McKenzie: Same-Day Legal Work in a Cyber Attack - URL: https://www.youtube.com/watch?v=9rmd1r-V_2Q - Speaker: Danielle Benecke - Occurs at: 34:03 - From Billable Hours to Fixed Fees - URL: https://www.youtube.com/watch?v=gGlYAohxJxM - Speaker: Danielle Benecke - Occurs at: 25:05 Danielle Benecke founded and leads Baker McKenzie's Applied AI practice. Max Junestrand is CEO and co-founder of Legora. They come at enterprise AI from opposite sides of the table, one buying and building inside a global law firm, the other selling the platform. They agree on the question that matters now: which parts of a firm's AI it should own, and which it should rent. ## The question underneath the tooling Danielle describes the shift in how her clients think about their AI stack: > "They're starting to think about what harnesses they should own rather than rent." Her framework uses two axes. One runs from generalizable to custom, the other from strategically critical to not. Generalized capability that is not strategic can safely run on a platform you do not own. Work that turns on your expert steering, your internal playbooks, and your own know-how should stay in-house. ## What a harness is Max gives the working definition, using coding tools as the reference: > "Claude Code is optimized for coding. It goes back to how do you let the model read the documents, make sense of documents, make sense of context? What's authoritative between that case over there and some internal precedent that the firm might have? Which models do we use where? Where do we allow it to launch sub-agents? Where do we not allow it to launch sub-agents?" Danielle adds that enterprises are not dealing with one harness but several at once: direct model access, productivity-layer harnesses, vendor tooling, and the bespoke systems a team builds per service line. The simplest harness, she notes, is a folder of markdown files. ## Cheap analysis, scarce judgment > "Generating content is cheap, generating legal analysis is cheap. True judgment and accountability is scarce." That imbalance is why Danielle expects the third wave of legal AI to be compliance systems that steer agents in real time. The first wave was tools for lawyers. The second, where most investment is now, is AI-native legal and compliance businesses. The third moves upstream, into designing the framework agents operate within so the output is compliant in the first place. ## The counter economics The most surprising moment comes from a general counsel Danielle spoke with, who sent an RFP for an M&A deal asking every firm for two quotes. One to validate the first-pass diligence report her team had generated with internal AI. One to do the whole thing end to end. > "Every one of those firms came back. It was more expensive to do the validation on the first pass that the client did internally versus doing it end to end." Which brings both guests to the same conclusion, from different directions: the scarce input is the expert who can steer the system, and where that expert sits determines where the work should go. ## Transcript Full transcript of the episode. Timestamps match the published video. **[0:00]** *Cold open.* Danielle Benecke: "As an industry, we've gone from a what tool do we roll out conversation to where does our harness live conversation, and I think most organizations are wising up to that right now." Max Junestrand: "Take the entire service market, it's roughly 5% software, 95% service, and that is about to change. It will make the supply of legal services much more widely available. It will make it possible to pursue strategies in litigation that we otherwise may not have found or not have thought of." **[0:41]** Welcome to "Enterprise Aligned AI," the podcast where we interview enterprise practitioners on the reality of agentic AI adoption. What gets deployed in regulated security-conscious enterprises? How is success measured? And how are the best leaders navigating the changes? My guests today are Danielle Benecke, who leads Baker McKenzie's Applied AI practice a firm with roughly 4,600 lawyers more than 70 offices worldwide, consistently ranked the world's number one legal brand. And Max Junestrand, CEO and co-founder of Legora, the Stockholm-born legal AI platform founded in 2023, now valued at over $5.6 billion and used by more than 1,200 legal teams worldwide. **[1:29]** It's truly a rocket ship. Thank you both for joining the podcast. Thank you so much for having us, Aparna. It's great to be here. Pleasure to be here. I'm glad we could make this happen. This all came out of Future Law at Stanford, so shout out to Roland Vogel and the team for bringing us together. Totally. Legal is a hot vertical right now for AI, but both of you have been working on AI for many years before danielle, let's start with you. People picture a global law firm as being a buyer of technology, not necessarily building AI capabilities, and yet you've been leading the Applied AI team at Baker McKenzie. **[2:03]** You built it up from scratch. What led to that, and what unique capabilities does your team bring ? So we've been, scaling our Applied AI practice since 2020, 2021, and it builds on a much bigger and longer AI program that dates back to the 2010s. But the real inflection point we saw back in 2020, 2021 started with the GPT, 3 API, and we started experimenting with that and bringing it into different service lines. And that, for us, was the, moment where we realized we needed a forward deployed capability. That's the capability that I founded, grew, continue to grow and lead now. **[2:40]** And we are a group of lawyers and AI technologists that specialize in designing, calibrating, and steering AI systems that do complex legal work. The companies that we work with are large global multinationals that work across many jurisdictions and have really complex legal and regulatory landscapes. And so what we do is we design systems that can undertake certain types of legal and compliance work to support those businesses. Max, Legora also started in this field before we really had the foundation models, so there's a lot of vision that went into the company. **[3:18]** Tell us, as a co-founder of Legora, what made you start this company? I don't think I get to take credit for those, first gnarly years pre-GPT, because I jumped on the ship in 2023, and when things really started taking off. But with the early BERT models, and in particular there was a Swedish version called SweBERT, that my co-founders were, utilizing to summarize and work with court cases. It was really hard. These models were not very intelligent. They were very hard to work with, enormous context issues. And then suddenly we get to GPT-3.5, right? **[3:53]** And things just, cascade from there. And so we took the bet and we said, "Hey, let's, quit our jobs. Let's, let's drop out. Let's pursue this full time." And about one month later, we started working with the biggest firm in the Nordics. A month after that, we got accepted into Y Combinator, and then we were off. And since then, not only has the underlying technology continued to develop, but the way that the scaffolding, the tool calling, the harnesses, the amount of performance that we can squeeze out of these models, and the different user experiences that we can build around them, that has all flourished. **[4:29]** It feels like we're running a marathon, but that we're still in mile one in many ways. Was the original vision to, summarize and automate legal work, but then in 2023 there was a breakout moment and you realized that it could be more than that? Well, back in 2020 it was more of a fun side project, I'd say. They were building their own text editor. They built their own semantic search layer, and all of this we just put on the shelf as soon as GPT-3.5 came. And if you think about w- where the vision for the company and the impact that we want to have in the world continues to develop, it's that, the legal service market is so people-dependent, right? **[5:10]** Take the entire service market. It's roughly 5% software, 95% service, and that is about to change. And I think the impact that that will have in society will be very positive. It will make the supply of legal services much more, widely available. It will make it possible to pursue strategies in litigation that we otherwise may not have found or not have thought of. It will allow us to, level up from gruesome, doc review to the more interesting and maybe creative aspects of the work. And to Danielle's point, it actually allows us to solve some of the challenges that your clients and large organizations are facing, because at the end of the day, we're building this technology, we're applying the technology, and we're innovating because we wanna go solve problems. **[5:57]** It's so interesting that you mentioned BERT, Max. So one of the first AI native services that we deployed to clients back in 2021 was we were taking the GPT-3 API and combining it with BERT and some other capabilities to do some scaled supply chain risk- monitoring for clients, right? Obviously a very primitive capability to what we're doing right now. But I think for us, that was a really early signal about the commercial model- that was possible. To your point, Max, this is fundamentally about new services that solve unmet needs. I think one of the challenges we have as an industry right now is most of the market is focused on accelerating existing services, existing ways that individual lawyers work at the task level. **[6:44]** The real opportunity that certainly we see for our clients, and that we see for the industry at large, is around entirely net new categories of services, and that has been our focus from day one of our program. Danielle, you've been at the forefront of both, the legal challenges that AI cr- creates, as well as the types of legal problems that AI can solve. Most people think there's potential for AI to do the rote work of, the legal profession. But you're saying there's new types of problems and new types of work that maybe we haven't seen. **[7:17]** Can you give us a few examples? Yes. And somewhat counterintuitively, I think the true opportunity is around those more complex categories of work. A lot of the market sees the opportunity more around accelerating lower level work. I think the real opportunity is around complex work that humans were simply not able to do without this toolbox that we have now. The other big thing that we're seeing in our practice is a lot of the work we're doing with clients stems from them rolling out AI across their businesses. In the very early innings of the shift, we had a number of clients come to us and say, "Marketing and comms have been the first to roll out AI across the business." **[7:59]** They're now creating massive, mountains of legal work that need to be done because in the old days, they partnered with the advertising agencies, and they would produce tens of pieces of content that our legal and compliance r- teams would review. They're now cutting deals with the frontier labs, and they're producing tens of thousands of content that is highly customized down the long tail of countries that we operate across. So one of the other early service lines that we got into is AI-enabled triage for global marketing and comms content for a lot of our global multinational clients. **[8:33]** And we would be analyzing this content across Product regulation across IP risk, across ambush marketing, and all sorts of other legal and regulatory verticals. But we said to the clients very early on, "This is not a scalable way to solve this problem. What we need to do is move upstream with you and start to design compliance systems that can help steer your marketing and comms agents and other agents operating across the business." So there is compliant possible in the first instance, right? One of the big challenges we have right now is generating content is cheap, generating legal analysis is cheap. **[9:11]** True judgment and accountability is scarce, right? And so we don't wanna get in the situation where we're in this AI arms race, where we're just generating more and more content that the legal and compliance teams need to deal with downstream. We want to improve the quality of action across the company, and ultimately, I think the third order effect or the third wave of AI in our industry is going to be all about runtime compliance steering for companies that, have agents operating across their business. The first wave was obviously tools for lawyers. **[9:44]** The second wave is now AI native legal and compliance businesses. That's where a lot of the investment activity is now. I think the third wave is going to be about designing compliance systems that can help steer agents and business behavior in real time. That's fascinating. Designing, compliance agents that are, steering, output in real time, sounds like it requires a collaboration between legal practitioners who, understand regulation, and all the nuances, and AI engineers. How do you make that collaboration happen? We are seeing the early sparks of this right now in our industry, and to take that example earlier of the marketing comms compliance, where the original service was the downstream triage- the next iteration of that service is around designing a framework for the agents to operate within so that the content is as compliant as possible in the first place. **[10:41]** And that starts to bring the walls down between legal risk compliance silos in the business as they historically have operated into, a much more collaborative en- environment. We're already seeing that happen across a lot of big global enterprises. Does Legora, which is a technology provider for the legal space, versus Baker McKenzie, which is a law firm do you, work very closely with your customers to start to invent the new aspects of the application? I think Legora is one of the only technology companies where we have more lawyers than software engineers. **[11:16]** It's very, very rare, but we do. And we've actually, I don't know who, who invented the term first, but we've certainly adopted it, and it's the legal engineer. And the legal engineer is a very tech-savvy lawyer who both understands the problem that we're solving in terms of the use case, but also is adept enough to actually help drive change management and the transformation of how legal teams operate. Because if you just take this technology and you go, just as Danielle said, "How can we improve, the way that we already do things?" **[11:50]** you'll get one type of result. But if you Say, well, now we have this technology, how should we reinvent the process all to begin with? Then you can get something much more, a- aligned with the capabilities that we now have. And I think in terms of new things coming, I love where this is going because effectively, all of the agents operating within an enterprise are going to have to, in real time, be able to query, is this plan or is this thing that I'm now going to do compliant with our policies or compliant with regulation? **[12:21]** And the regulatory, body that you actually need to take into, account is enormous. It's too big for one single person to actually keep in their brain at all times. There's just too much context. And as we start to build these layers for agents that can operate more autonomously, then I think that's certainly a kind of cool path to keep going down. There are all these new roles, forward deployed engineers, which have been around for some time, and now legal engineers who are lawyers and also engineers. But I wanna ask you, Max, about differentiation and moat. **[12:53]** This is a really important question, especially in the legal profession, because there are many different types of players. There's all these incumbent companies which have distribution like LexisNexis and Thomson Reuters. They can obviously pull in AI capabilities into their existing services. And then there's a range of startups. Legal is a very hot area for, funding startups- all going after different parts of the market, and certainly law firms are a big part of that. And then there's the question of can foundation models themselves start to, increase what they offer by, post-training, for example For Legora, what are the true sources of moat? **[13:31]** Well, the space is developing at an exponential pace. And I like to say that velocity is a moat because you have to be at the forefront, and you have to be at the frontier at all times. If you look at what our product was back in 2023, it would have looked like a total joke today. If you brought out the product from 2024, it would have looked like, a really poor executed legal AI platform, and maybe even towards the mid of 2025. And with the things that we've just come out with, we are, clearly moving ahead of what I'd say the most of the pack is doing. **[14:10]** And as you said, everybody's thinking about how can we win our, slice of the cake here, and we're still in mile one, right? Technology and product tends to compound. So if you make a lot of really, correct strategic decisions about where should we train models, where should we not train models, where should we build, where should we partner, where should we buy? How do we build a user experience at the end of the day, and a product that delivers, more value than anything else can do on the market? And that's what Legora is obsessed about. **[14:42]** There's, I'd say, different strategies at play. Yes, you have some of the big incumbents who are saying, "Oh, we're gonna, leverage this piece of the data and our big distribution muscles, and that will be our way of winning." And, fine, that's one strategy, and we see what's happening with those, public stocks. Um, you know, they're, they're maybe not doing, super hot right now. And then on the other hand, you have private companies like Legora, who is in, enormous demand, , in terms of our growth rate and where we're going. And I think that the old sort of SaaS world is also really getting disrupted by these AI native businesses. **[15:18]** If you look at the growth rate of some of these new players entering on the market, it's just, it's throwing all of the old metrics and the even venture capital playbook, out of the window. Because the net revenue retention, the DAU over MAU, growth in revenue, this is just, unheard of. We get to experience that in legal, but the same thing is happening in coding, right? You've got Cursor on one end who just exited to, SpaceX and, you have, all of these different, verticals that are playing out, and we're certainly not at the finish line, for the legal one. **[15:50]** Aside from velocity, what are the other moats? Well, I mean, the- there's the degree of em- embeddedness with a customer, when you work with a firm like Baker McKenzie, and you roll out to thousands of practitioners, and they start to adopt Legora in terms of how they do work and deliver to their clients, that becomes one type of moat. When you,, build products like our portal, that actually puts both the clients and the law firm- In one single software experience, that becomes a way of driving both, new types of delivery models, but which of course are quite sticky once you get them right. **[16:25]** The way that you integrate with the rest of the ecosystem and that you build functionality that frankly is, quite far away from what a foundation model would naturally go after. A lot of the software that we're building now has nothing to do with LLMs whatsoever. , The amount of audit logs, permission management, trails, scheme. The amount of just things you need to solve in order to deploy within an organization like Baker or a big enterprise for that matter, in the legal, department specifically, are quite extensive towards what just the, horizontal models are doing. **[17:01]** It's maybe not so clear at the surface, but the minute you step your foot in it, you realize that, oh, this pool was really shallow. If you wanna upload 1,000 documents from a VDR or 100,000 documents from a e-discovery exercise, and it just like blatantly fails in some of these frontier, models, like in the way that those are packaged in their, very general harness. And you go to Legora and it's just like night and day. We're not positioned as we wanna be the legal AI assistant for everyone. We're a agentic, infrastructure layer that works with some of the biggest enterprise and legal firms in the world. **[17:37]** I think we're naturally going after quite different segments of the market. We've all worked with very large enterprises. There's a huge compliance burden would you say that Legora is a legal specific harness? We certainly have a harness that's tuned to legal, but there's many layers in the product. , We articulated this quite well, I think, with our agentic operating system that we released a few weeks back. It starts with the foundation models, and then we put those foun- foundation models into work within our legal harness, and then we give that harness access to legal specific tools, legal specific context, legal specific surface areas. **[18:11]** Because not everything should be served in a chat, right? Every use case, the UX is not, a chat is not optimized for everything. And then all of that has to be wrapped in, a secure and enterprise grade system, and it has to be combined with the legal engineers. We should not forget the human aspect of this. Even if you have a tool like, Claude, implementing that within an organization to drive the outcomes you want is quite extensive. There's a lot of DIYing you have to do if you take a general harness. **[18:45]** Probably, no organizations in the world can take that to the extent that we've taken Legora. And that's because we're 700 people who wake up every day and obsess about our particular problem. Whereas in these, very general tools, there's maybe one PM and a few engineers. But what is the harness? What is different about your harness than, let's say, Claude Code? Claude Code is optimized for coding. I-i-it goes back to how do you let the model read the documents, make sense of documents, make sense of context? How does it tie together? **[19:20]** What's authoritative between that case over there and some internal precedent that the firm might have? How do you actually weigh those two together? Which models do we use where? Where do we allow it to launch sub-agents? Where do we not allow it to launch sub-agents? How do we instruct the agent to actually plan? We recently, introduced a new agent, and that is tailored to create a very detailed plan for the user before it sets off to do the work. Cursor does the same, right? When you're coding with it, presents a plan before you actually a-allow it to go out and execute. **[19:54]** But then the way that it plans is, of course, a function of the tools at its disposal and, subject to its own capabilities. I think you can go very, very deep, and I could probably talk for an hour about, where we wanna take it Just to jump in on harnesses, it's important to distinguish between inner harnesses and outer harnesses. One of the things we see in the narrative in the market right now is harnesses becoming a monolithic concept, when in reality we're actually dealing with multiple harnesses if you're a big enterprise business. **[20:26]** So we, for example, have direct, access to a range of models through our cloud AI environment. We work with open AI models, we work with Anthropic capabilities, we have access to Google capabilities, et cetera. We then have access to productivity layer harnesses and Anthropic Cowork is a great example of that. You see similar capability filtering up through Microsoft environment as well, and then you get into legal specific tooling. Most large scale law firms and other actors in the knowledge industry have access to multiple tools with different types of harnesses. As an organization, you also are starting to create your own harnesses. **[21:08]** Every service line, every bespoke system that we design and deploy for our clients, that has its own unique harness. The most simple harness is probably markdown, a folder of markdown files, right? A big piece of work, that my team and I do is working with chief legal officers, chief compliance officers, general counsel, other leaders on the client side, and help them design their stack, their strategy around this technology. They're starting to think about the strategic dependency of having their harness live inside another platform, they're starting to think about, what harnesses should they own rather than rent. **[21:45]** One way that we think about it is there's a scale from generalizable to custom. There's a scale from extremely strategic to not so strategic. And I think for the stuff that is super generalized, not particularly strategic or important to your organization, totally fine to rely on a harness that lives outside of your environment. But for the very strategic stuff that ultimately turns on your expert steering or internal playbooks and internal know-how, you need to be very careful about where that harness lives. As an industry, we've gone from a what tool do we roll out conversation to where does our harness live conversation, and I think most organizations are wising up to that right now. **[22:26]** Does Baker McKenzie think about, AI adoption as a top-down mandate or a bottom-up mandate? You're using many different tools, many different models. Is that , by choice from, lawyers within the firm, or has it been something that's top-down? We're a services business, so everything starts with what our clients need. And in this era, it really comes down to three things. Our clients need strategic clarity, decision velocity, and risk mitigation if you look at every piece of legal or compliance work, it's ultimately those three things. As an orchestrator of capability that serves that ultimate need, what do we need? **[23:03]** That really drives every question about the capabilities that we plug in, whether that is our baseline cloud AI capability and multi model access, whether that is bringing in best-in-class vendors, whether that is Legora or other players in the ecosystem. The market is moving so quickly, we need to be very agile, not bet the farm on any one thing, but just c constantly thinking about what is ultimately going to serve the client need. Many players in the industry still think of this as a primarily a technology shift. Technology is obviously a big part of it, but the commercial model transformation part is arguably much more profound than the technology shift. **[23:44]** When I look to what we build and sell to our clients, the technology bit in many ways is the easy bit. The hard bit is how do you sustain that capability in a way that is commercially viable for the business and for the client, and that also meets the governance needs of sophisticated global companies, when my team is, building out this set of primitives that ultimately drive all of our service lines, the commercial and the governance pieces are equally, if not more important than the technology pieces that we bring in. **[24:19]** I see. So depending on client needs, you'll bring in the relevant technology. But the technology shift is actually leading to a commercial model shift. Tell us more about the commercial model shift so historically, knowledge work has been about, execution, human workers executing on certain tasks. There's no secret we're five years into the shift at this moment. More and more knowledge work is about designing systems that execute rather than the downstream execution. Many lawyers and compliance professionals still are thinking of their role as shifting downstream. A lot of firms and other players initially thought, "Well, we'll have the AI generate output, and then we'll validate downstream." **[24:58]** The most important expression of expertise at this point in time is actually upstream in designing the system. The better the design, the better the output is going to be, the less downstream validation you're going to have to do. That ultimately changes what we're selling completely, and that ultimately creates an en- entirely new model, commercial model for delivery of work. So for example, my team, we operate off an entirely fixed fee model. Every engagement, we will design and price that around what the client needs. What kind of risk are they trying to mitigate? **[25:31]** What kind of commercial opportunity are they trying to unlock? What is the ROI they're looking for? And then we design the service around that. Very different, pricing model to the way that historically legal work has been delivered. So it's not about the billable hour and the expertise. The expertise is, like you said, better applied in the design of the solution, and so you scope out based on a problem what the solution is, and then it's a fixed fee, which is more outcomes based. Exactly. And working within a large incumbent law firm, moving the ship takes time, right? **[26:06]** My team operates in this way. That does not mean that every team across my firm is operating off of the same billing and pricing model because it takes time for a wider industry to shift. But for the AI native services that we sell, that is the way that we think the commercial model, is best shaped. If we look at the legal industry and the transformation that it's going through many of the new YC startups are totally AI native, where they're providing a legal outcome, and, it is a black box in some sense, whether it's done by AI or it's done by lawyers or who it's done by. **[26:39]** Do, do you also, price based on outcomes? First off, we're not an AI native law firm, we're a software company. And I wouldn't be, I wouldn't wanna compete with these guys. They're very competitive. I think that the AI native l- law firm model, is tricky because you're actually competing with the full-time software builders, and you're competing with the full-time experts who have built their entire, expertise, customer base, processes around delivering legal services. So I think it's a hard market. How do we price? Well, we've priced primarily on a seat and enterprise basis, but the move to introducing more and more agentic functionality, has actually made us introduce consumption-based pricing. **[27:26]** Most AI tools since the, beginning of Cursor and Lovable and Clay and also the big frontier labs have been on a consumption basis because that's the only thing that aligns the value delivered with, the cost basis. We don't have software margins. These AI businesses, are fundamentally built in a different way, and this is the only way to make sure that if one user is leveraging, hundreds of dollars worth of tokens or even thousands of dollars worth of tokens, that has to be priced differently to a user who's doing much more, foundational and basic things. **[28:01]** We spoke about moats earlier, and you mentioned speed legora has grown exceedingly fast, you mentioned that there are certain important, decisions, and pricing being one of them, but there are other important decisions. So tell us what helped you grow? Thank you. We have grown very fast. We formed the company, in the late spring of 2023. We were accepted into Y Combinator in '24, and then from 1 million to 100 million, I do believe that we, beat the record for the fastest enterprise company, to go from 1 to 100 million in less than 18 months. **[28:31]** We were, like, very on, on a knife's edge to beating Wiz. But I think we didn't with about a week or two to go. And I think early, we made some tactical decisions that we were going to focus all of our attention on the application layer, and that we were going to build with our customers. So I'm a software engineer. The entire founding team were software engineers. We didn't hire a lawyer at Legora until six months into the journey, and he also happened to be a software engineer and a lawyer, so, he's a little special. **[28:58]** We have three cultural values at Legora, and the last one is grow together because it's so clear to us that we only win if our clients win, and we need them emotionally and commercially invested in how this technology will be brought to, the field. And so every new piece of functionality that we build, we co-design and build with a group of design partners. We run a really tight customer advisory board. We iterate a lot, and we have a saying that when some zig, we wanna zag, and that's because you can't just look at what the competition or what others are doing. **[29:43]** I think there's a lot of copying And if you're swimming a race and you're looking sideways all the time, you're gonna be slower. So you need to look down. You need to do your own thing, and you just need to move really fast. And one of those decisions about being one of the first major players to come out with consumption-based pricing now is also because it allows us to continue being the product and innovator in the space. What's actually quite good about it too is maybe in comparison to some of the, general and big tools, we can work with a much, finer, group of models. **[30:16]** We don't have to put Opus 4.8 into everything we can use open source models. We can use, older generational models because once something is good enough at a specific task, you don't actually have to push the frontier models anymore. So we really wanna align our commercial interests with the commercial interests of our clients, and then be very transparent about how it's going to continue developing forward. So you can potentially be more efficient. Token usage is right now, quite a premium, especially for the larger- Yeah. We don't want people token maxing in Legora, right? **[30:46]** We want them to, get the results they need. And as a company, we started in, in Sweden because that's where we happen to be born. But very, very quickly we've become a global, company. Maybe not as global as Baker, yet. Not, not yet. But we're getting there. We have something like 15 offices up and running all the way from San Francisco to Japan and in between. It's also because, different markets needs different things. The way that you work in Singapore is different from Germany, which is different from the US, and we've really, recognized the need to be local, in these markets. **[31:19]** So what led to Baker McKenzie choosing Legora as one of your, partners? We can talk about that in a sec, but I actually wanted to jump on, the tokens point because I think it's a really important one when it comes to the commercial model. And to Max's point, we are well out of the AI subsidy era, right? We're past the token maxing phase where we all had our Apple Minis running all weekend on- all sorts of, even just our personal projects. We have started to see that, come up as an important factor in pricing services. **[31:50]** So to take one example, we do a lot of AI-driven global compliance audits for clients. Those can generate billions of tokens of analysis once you're working across multiple verticals across many countries. That's becoming a meaningful part of the cost equation for delivering this work. So that's just another reason why the pricing model needs to change. Ironically, I think, Max's announcement about, the consumption-based pricing, which is, very expected, it takes us out of the earlier narrative about legal, AI because I think a lot of the clients thought, "Oh, we're gonna get out of hourly-based pricing, which is super unpredictable." **[32:26]** Yeah. Now the vendors are coming up with a even more unpredictable pricing model, which is tokens. For now, that's why we are very focused on fixed pricing. We do a lot of work to model out what we think those cost inputs are going to be before we scope the work and deliver it to clients. I think this is a kind of natural part of the shift that we're going through, and most enterprises are going to have to figure it out. The other thing that's a little bit counterintuitive is when you're looking at the relative cost of different models Different models have different token efficiency. **[32:59]** And so sometimes you might be looking at a more expensive model, and the instinct is we'll go with the cheaper model that's most performant, but actually the expensive model may be much more cost efficient. It's like a great shampoo. Exactly. Sometimes it's better to get the slightly more expensive shampoo because you end up using less. You use a little less of it. I agree that introducing consumption-based pricing introduces another, metric of uncertainty. However, once you have the data, it's actually very predictable. If you're running a big, say, tabular review on, a thousand documents x 100, queries, we know what that's gonna be. **[33:33]** And similarly, if you kick off a big agent workflow- we can predict with pretty good certainty what that's gonna be. Now, some firms like Baker have a really strong pricing, team, and you have a lot of data in the firm to know how to price stuff, but not every firm has that. And we already have, certain litigation tools that get passed, right? Like The cost of certain discovery tools gets passed down to the end invoice, I think some firms might experiment with, doing the same for their AI usage. I think that's right. **[34:03]** Then going back to your question about how do we orchestrate different vendors in our stack, it's really about what do we think is best in class capability to bring in. As I mentioned before, we're not gonna build something that we think is generalized capability that is best served up by another player. Openly, we have our cloud AI layer, which plugs into all sorts of capability, whether that is Google, Anthropic, OpenAI, and other models. Like many in legal, we are primarily an Azure shop, but we do have other cloud AI environments. **[34:33]** I am a technology lawyer by background, so a lot of the work that I do actually is with, the legal teams of Frontier and other tech driven firms. Sometimes those clients wanna use their own models for obvious reasons. We can meet those clients there. And then as we get up the stack into different capabilities, we use Legora for certain things. Certainly in the transactional realm, we're seeing some really, useful capability to plug in. We use other specific tools, whether that is Relativity Air for very domain specific things. My team will often layer A lot of capabilities within the same service line. **[35:07]** To give you one example, we support a lot of cyber incident, global cyber incident response. So one of the services that my team delivers is rapid same-day, analysis of reporting obligations for very large cyber incidents. Historically, it's taken days or weeks to do that kind of analysis, triangulating across different regimes, especially if a client is dealing with an incident that affects millions of consumers across multiple countries. So we have our own internal harness, which is built up of all sorts of logic and playbooks, that drive that service. We use foundation models directly from some of the labs to power that service, but equally as an input if we wanna run a forensic on the incident, we might use a tool like a Relativity, and then we will take that forensic report, and we will feed it into our system, and we will analyze that report against our playbooks. **[36:02]** We see the future of these kinds of AI-driven services as just increasingly complex stacks of best-in-class capabilities orchestrated, with the right commercial model around it, the right governance around it to give the clients what they need, which is strategic clarity, decision velocity, and risk mitigation. What you're delivering there is, fundamentally a new value function. Yes. Because if you have a cyber incident, you are willing to pay more if you can get the results tomorrow than if you have to wait three weeks. The technology savviness of the, firms is increasingly becoming a competitive advantage, I appreciate that, Max. **[36:44]** The other thing is the service completely changes. So in the old days, the service would be you do the analysis once. Today, you can't do that because most of the cyber incident reporting regimes range between six hours deadline to 48 hours deadline, and you are getting new facts, new assumptions. You know, the boom happens, and most organizations get the information in piecemeal. What we're able to do is just constantly model the reporting obligations as the incident develops. So we might run the analysis many times over the course of 48 hours, so the client is getting the best real time analysis of what they should do. **[37:24]** What I'm hearing from you, Danielle, is that, there's a range of problems that were difficult to solve, previously. Now that you have this, riches of tools, including the foundation models themselves, you're able to build solutions that didn't exist before, you're pricing for those solutions based on outcomes, which bakes in the consumption piece. What you're surfacing here is a tension with AI that, all of us feel, which is that because token consumption is such a big part of the cost, you have to pass through some amount of consumption on the other hand, you're providing new kinds of capabilities and you want to price for that. **[37:58]** Once you see some of the hourly rates, you realize that the tokens, in a lot of legal work delivered, the tokens are a very small piece compared to what some of the hourly rates are. That's a really interesting point as well because we operate across 45 markets that have very different historical hourly rates and very different economics around legal work. I was in India recently for the AI summit connecting with a number of our clients out there interestingly when you talk to the chief legal officers of those companies, they will say it's actually more efficient for us to get our local expert to weigh in on something than to have a expensive US commercial reasoning model mull over it you have to be really thoughtful, particularly in this AI scarcity era, about the kinds of problems that are ripe for this approach. **[38:47]** I was talking to, a chief legal officer of a technology company recently, and he framed the calculus around whether he will send work outside the organization or keep it in around three things. He said The primary question is, do I have the right expert to steer the AI system to the right conclusion? If I am not confident that I have the right expert inside of my in-house team, I am likely to get a lot of slop and a lot of token spend that is just not efficient versus sending it to an expert that is proficient at steering the AI system to get the work done. **[39:25]** Secondly, even if they think they have the right expert, do they really want to hold on to the risk of that work? The accountability question, do they want to send that risk outside because it's a kind of decision point that they need, somebody to be accountable for. And thirdly, what is a overall kind of cost efficient way of delivering this work, taking those two things into account? And I thought that was a really interesting way of framing it. The other thing I've seen recently, this is another general counselor who said that to me. **[39:53]** She said she had sent out an RFP for an M&A deal, and she said to all of her firms, "Give me two quotes." So one quote is for you to validate the first pass diligence report that the client generated using their internal AI. Second quote they wanted was just to do the whole thing end to end. Every one of those firms came back. It was more expensive to do the validation on the first pass that the client did internally versus doing it end to end, I think it signals that where we're at right now is it's all about do you have the right expert to steer the system to get the work done efficiently because we're seeing all sorts of counter economics. **[40:36]** Wow. My question was going to be about build versus buy. It's not about build versus buy anymore because, you're choosing the right tools, you're creating custom solutions. You're seeing a whole new set of problems that you can solve that you couldn't solve before. The question is really about where is that boundary? This is a question to both of you, because you have to decide what your roadmap is, what you're going to build, and what you're not going to build. Yep. You've mentioned that, you are working very closely with customers and, you're charting your own path. **[41:01]** Not looking at other swim lanes. At the same time, every client is different you have to say no to things. Yeah, my main job is saying no, actually. So where is that boundary between- what you would build into Legora versus what you would build a custom harness for- I think the first thing to mention is, we are very transparent with, our clients about what we are going to build and what we are not going to build. We actually recently just got sent an Excel from one of our close partners with, "Here's all the software tools we have. **[41:31]** Mark an X where you think, over the next three to five years you might do something in this space." So we can, plan our software stack, because, it's, clear that, Legora is growing to be something quite big. The part around, what do we decide to build and what do we not decide to build, the most important thing is that on top of Legora, you need to be able to configure and build the differentiation. So if Danielle and team wants to bring their own MCP server into Legora, they can do that today. **[41:57]** But they probably don't want to build the UI- UX and routing and orchestration of a Microsoft Word add-in, 'cause that's, really annoying and very gnarly, and we get a lot of scale on building that. But building a custom cyber solution for a very specific type of cyber case, that's maybe something more up your alley. It's also a timing question, right? Because, our clients can get arbitrage by having certain capabilities now versus having them a year or two years from now. Because technology is becoming an increasing competitive advantage in how we pitch and how we win work, buy versus build for me is more of a how do we partner and how do we make each other as successful as possible? **[42:41]** Anything to add to that? We do a lot with our clients on this, help them think about where to build, because even the most sophisticated technology companies in the world, their legal and risk and compliance arms, it does not make sense for them to roll their own for everything, even though when they have the frontier models inside of their ecosystem. One of the ways I think about it is on the spectrum from generalizable to customizable, where is that capability? Typically, the more generalized it is, the more that this is something that other firms, other players in the ecosystem would benefit from that kind of capability, it makes sense to leave that to the wider market and to buy that in. **[43:20]** The more bespoke it is, the more it turns on our unique expertise, our unique playbooks, things that others do not have in the market. That's going to make us more inclined to keep that capability internal and to have our own harness, our harness around that. The other axis would be around strategic criticality. How important is this to our business versus is this a fairly commoditized capability that others can replicate in the market? So the quadrant where my team typically tends to focus is complex, high-value, strategically critical services that turn on our expertise, our playbooks, things that nobody else in the market has access to. **[44:01]** Generalized capabilities we would typically buy in, but everybody's stack is going to be different. Ultimately, I think that's where most enterprises will land. Great. That makes sense, and I think you're both aligned on that. One last question, There's so many companies offering all kinds of AI services, of course, legal services these days when I talk to CXOs, they're like, "Well, I don't know who to go with because there's so many vendors. I can't tell the difference." You've obviously been able to cross that hurdle. What do you do? What's your secret sauce for winning over clients? **[44:32]** Except for Dude Law, of course. It all started at a very, from a very honest place, we wanna go build delightful software with the best lawyers in the world, and that's the pitch, and that is continuing now as we scale. Of course, you cross these critical thresholds, you get to one million revenue, you get to 10 million revenue, you get to 100 million revenue, and then the next one, and that builds your brand, and that builds, interest. But ultimately, buyers are comparing many solutions. And so the only way that we actually continue to stay relevant is by being the absolute best product on the market, when we go into a competitive pilot, we have to out-compete other providers to win the trust our clients. **[45:19]** And that comes down to product, that comes down to the partnership, that comes down to the roadmap. And as Danielle said, you have to build, trust, and you have to think about how this is a critical piece of infrastructure. And so the company that you decide to work with, you need to make sure that they have their heart in the right place. I think time and time again, you have to prove that and build up a reputation. Given the trust aspect, it's very important to actually check the work of the AI, make sure that it's correct, right? **[45:44]** Yeah. You've mentioned before that, having the right expert, anywhere in the world. There's a very human aspect to that, at least for Baker McKenzie. Would you say that that's part of the edge? Absolutely. I think most of the market is very much focused on the downstream validation. For us, it is equally, if not more important, to have the experts designing the system in the first place. That is where you get the more performant systems when you have the right experts designing the logic, designing the playbooks. Because many of the playbooks that power our services, it's not pure black letter law. **[46:18]** It's a pragmatic, risk-calibrated approach to the law. When we're advising our clients, we're not just saying what the law is, we're advising the clients what to do in the context of not just the legal landscape, but the market, their product roadmap, the political, geopolitical landscape and society. And those are two very, very different products. Validation is important, but upstream and downstream validation need to happen. I think also AI actually gives the first opportunity to, scale this across bigger firms too. 'Cause if you have an amazing partner in one of the offices write a skill for how an agent should do a specific piece of task, that can now be scaled across the entire firm. **[47:00]** So actually expertise and precedent and knowledge and knowhow, can in a way get scaled across the business in a much more pragmatic way than ever before. This was a very energizing and rich discussion. Anything else you'd like to say? Anything about the future of law? Internally at Legora, we like to say that it is the best time in history to be a lawyer because never before has this many hardworking, devoted, and talented people been focused on building great software to make it more fun, energizing, and deliver better results for the clients. **[47:35]** So I'll leave all the listeners with that. I would agree with that. I think it is the most consequential moment to be a lawyer or a compliance professional. I would encourage the industry to focus more on what are the services that don't exist now that could be created with this capability. We have spent the last four or five years as an industry focusing on accelerating existing services. There is value in that, but I think far and above the bigger value that we can deliver to the clients, to the companies, is around unlocking some of those existential risks that keep the chief compliance officers up at night, that keep the CLOs up at night, which are things we haven't solved for quite yet. **[48:19]** Thank you so much. Thank you. Thank you. It's been fun. --- ## Ep. 6: AI Agents Need Clean Data | Cohesity CMO and Kaggle Founder, Sumble CEO - URL: https://enterprisealignedai.com/episodes/marketing-ai-with-cohesity-and-sumble - Published: 2026-08-19 - Duration: 43 min - Video: https://www.youtube.com/watch?v=JVO3V2rORhg - Guests: Carol Carpenter, Anthony Goldbloom - Tags: Marketing AI, Go-to-Market, Knowledge Graphs, ROI & Measurement, Data Privacy, Organization and Talent **Summary:** Host Aparna Sinha sits down with Carol Carpenter, Chief Marketing Officer of Cohesity, and Anthony Goldbloom, CEO and co-founder of Sumble, on the function that adopted AI first and is furthest along with it. Carol runs a marketing org that is a heavy user of enterprise LLMs, a builder of custom agents for translation and brand compliance, and a demanding buyer of third-party AI tools. She gives the numbers behind each: a 60% target reduction in translation cost, lead follow-up moved from 20% to 80% of MQLs, and top-of-funnel gains in the 10 to 20% range. She is equally direct about where AI has not delivered, and about how much human design work sits behind every agent that works. Anthony sold Kaggle to Google in 2017 and now builds the grounded data layer that go-to-market agents run on. He explains why a language model cannot tell you which team inside an account is ready to buy, what a knowledge graph adds at the team level, and why the work is mostly cleaning data. They disagree productively on whether an enterprise should build its own knowledge graph, and Carol is unmovable on one point: her proprietary data does not leave the company. **Why listen:** Marketing was first to adopt AI and is now far enough in to have real numbers rather than pilots. Carol gives hers with the caveats attached, including where the gains are only 10 to 20% and where the setup work was harder than she expected. Anthony gives the clearest available account of why grounding matters: language models handle the reasoning, but the facts about what companies are doing have to come from somewhere. If you are deciding what to buy, what to build, and what data to hand a vendor, both sides of that call are argued here by people who have made it. ### Key takeaways - Enterprise B2B took 57 touch points to close a decade ago and now takes 152, which is what makes orchestration a data problem rather than a volume problem - Lead follow-up moved from 20% to 80% of MQLs through a combination of a few more SDRs and better tooling, not headcount alone - The AI translation tool is targeting a 60% cut in translation cost, built on OpenAI with brand guidelines, tone, cultural nuance, and a memory of prior translations - Image generation and the automated customer journey are the two places AI is not yet a game changer in marketing - Agents do not automate the customer journey on their own; a person has to design the sequences and set the baseline before an agent can learn from it - A language model is good for the thinking, but the thinking has to be applied to a hard-grounded corpus that knows what the world's companies are doing - Sumble maps tech stacks and buying windows to individual teams inside an account rather than to the company, which is the entity CRMs are missing - The moat is curation and depth: 70 million companies in cold storage, 3 million active, and roughly 40,000 that buyers ever look up, where the data is kept pristine - Building the product is 80% cleaning data, including disambiguating job titles that drift from SalesOps to RevOps to GTM engineering - Putting all go-to-market data into a warehouse gets roughly 70% of the value of a knowledge graph, because language models are good at SQL - Proprietary first-party data does not leave the enterprise, even if sharing it would make a vendor's agents smarter - Hiring for AI means asking about the candidate's latest prompt and the problem behind it, not whether they can write one ### Chapters - 0:00: 152 Touch Points to Close One Enterprise Deal - 1:04: Meet the Guests: Cohesity's CMO and Sumble's CEO - 2:49: How Cohesity Uses AI - 4:09: Founding Kaggle and Sumble - 5:50: Where CMOs See Real ROI from AI - 7:33: Busting the Hype of AI in Marketing - 9:04: Agents Require Humans: Don't Underestimate the Work - 11:23: Cutting Translation Cost 60% and Moving MQLs 20% to 80% - 14:55: Sumble Provides Team-Level Tech Stacks and Buying Windows - 17:43: Sumble's Moat: Pristine Data on 40,000 Companies - 21:57: Data Privacy Is Paramount for Enterprises - 24:06: Building Sumble Is 80% Cleaning Data - 27:02: Build vs Buy: Should Cohesity Build Its Own Knowledge Graph - 31:22: Cohesity's Brandi Agent and the Pre-Event ROI Calculator - 36:51: Organization and Talent: Defining the GTM Engineer - 38:37: Hiring for AI: "Tell Me About Your Latest Prompt" - 40:30: Sumble: Staying Out of the LLM Kill Zone - 42:13: Closing Thoughts ### Clips - LLMs Do the Thinking. The Corpus Does the Grounding. - URL: https://www.youtube.com/watch?v=apwxgKCvH5E - Speaker: Anthony Goldbloom - Occurs at: 20:21 - Claim: LLMs think, the corpus grounds - AI Agents Can 10X Your Outreach. A Human Still Designs It - URL: https://www.youtube.com/watch?v=1pcBUeYq9hM - Speaker: Carol Carpenter - Occurs at: 8:31 - Claim: Agents scale outreach, humans design it - Greg Brockman Learned AI on Kaggle: Why Data Beats Models - URL: https://www.youtube.com/watch?v=_VGUnbTUudM - Speaker: Anthony Goldbloom - Occurs at: 4:23 - Claim: Data beat models, so he built Sumble - Claude and ChatGPT Can't Prioritize Your Funnel - URL: https://www.youtube.com/watch?v=SqDU3bFFrYU - Speaker: Carol Carpenter - Occurs at: 9:04 - Claim: 57 touch points became 152 - Claude Is Good at SQL: You Don't Need a Knowledge Graph - URL: https://www.youtube.com/watch?v=lfJKhZGH7TQ - Speaker: Anthony Goldbloom - Occurs at: 29:07 - Claim: A warehouse gets you 70% of the value - How an AI Data Company Uses AI - URL: https://www.youtube.com/watch?v=OK8jpGISwuE - Speaker: Carol Carpenter - Occurs at: 2:56 - Claim: Three ways Cohesity uses AI - Building AI Is 80% Cleaning Data - URL: https://www.youtube.com/watch?v=_MaS1lrbia8 - Speaker: Anthony Goldbloom - Occurs at: 24:13 - Claim: 80% cleaning, 20% complaining - AI Agents for Salespeople - URL: https://www.youtube.com/watch?v=s9TqgqtOIA0 - Speaker: Carol Carpenter - Occurs at: 28:06 - Claim: AI agents for salespeople - The LLM Kill Zone: Kaggle Founder on Staying Out of It - URL: https://www.youtube.com/watch?v=EXNmrV7Es2w - Speaker: Anthony Goldbloom - Occurs at: 41:08 - Claim: Staying out of the LLM kill zone Carol Carpenter is Chief Marketing Officer of Cohesity. Anthony Goldbloom is CEO and co-founder of Sumble, and before that co-founder and CEO of Kaggle. One runs a marketing organization that buys, builds and uses AI daily. The other builds the data layer underneath it. The conversation is about what has paid back so far, and what the agents still cannot do without a person. ## The number that reframes the problem Carol opens with a figure that explains why marketing is a data problem now: > "10 years ago, we used to say you have to touch a customer 57 times in enterprise B2B. I just heard the other day, it's 152 touch points." Those touches are spread across a website visit, a webinar, a Gartner enquiry, G2, Reddit. Finding and sequencing them is what AI is being asked to do. ## Where the returns are real The translation work is the clearest win. Cohesity does business in 140 countries with 12 prioritized languages, and the cost of translation is significant. The team built on OpenAI, loading in brand guidelines, tone, documentation standards, cultural nuance, and a memory of everything already translated. The goal is a 60% reduction in cost, and Carol believes they have already shown it is reachable. Further down the funnel, SDR follow-up went from touching about 20% of marketing qualified leads to close to 80%. Carol is specific that this was not solved by hiring 200 more SDRs; it was a few more people plus the tooling. ## Where it has not delivered > "Do not underestimate, we underestimated the level of work. It is far more painful than I anticipated." Carol names two areas where AI is not yet a game changer. Image generation still reads as AI-generated. And the automated customer journey still requires a person to think through why someone buys and what the moments of truth are, before an agent has anything to learn from. ## Why the data matters more than the model Anthony's argument is that the reasoning is the easy part now: > "You could think of the LLM as quite good for the thinking portion, but then you need to apply the thinking to a hard-grounded corpus that has the world's companies in it, and knowledge of what the world's companies are doing." Sumble's answer is a knowledge graph built at the team level. A conventional data vendor tells you a bank uses a particular backup product. Sumble tells you which teams inside it use what, and which of those teams are in a buying window because their architecture is changing. The unglamorous half of that is data cleaning, which Anthony puts at 80% of the work, including keeping up with job titles that drift from SalesOps to RevOps to GTM engineering. ## The disagreement worth listening to Aparna pushes on whether Cohesity should build its own knowledge graph over its first-party data. Anthony's answer is that a warehouse gets most of the way there, because language models are good at SQL, and that a private graph is still too manual to be worth it. Carol's position on the underlying question is not negotiable: > "There's all this proprietary data that we would never ship externally to share, even if it would make some agents smarter." --- # Blog ## What is Enterprise Aligned AI? - URL: https://enterprisealignedai.com/blog/what-is-enterprise-aligned-ai - Author: Aparna Sinha - Published: 2026-03-08 **Excerpt:** Practical strategies for implementing AI at scale. Deep dives with real leaders implementing change — covering the reality of AI adoption inside companies. ![Enterprise Aligned AI](/images/blog/enterprise-aligned-ai.jpeg) ### Deep dives with real leaders implementing change There is a lot of hype in AI. The technology is real but it is over hyped by the scale of investment and the pressure for growth. That makes it doubly important to get honest insights about what's working and what's not working. We cover the reality of AI adoption inside companies. Not only startups, but non-AI companies who are cautious about adoption. Tune in for the mainstream customer perspective critical to your success with AI. We will cover this topic through a series of in-depth interviews with enterprise leaders, diving into the messy reality of adoption and value generation. AI is a statistically variable technology that is quite new, not fully reliable or secure and largely unintegrated into enterprise systems. We also cover the human aspect of the organizational and up-skilling challenge AI poses. In addition, we feature short pieces on the technology itself and promising trends we see, as we discuss AI with the most active enterprise change makers. ## 1. Who is this series for? This series is for anyone interested in **how to adopt the latest AI technology to extract business value in the work setting**. That means if you are a business leader who is tasked with using AI to move your company or division forward, there will be useful reality checks and experienced best practices here for you. If you are a founder of an AI company targeting enterprises, you will find real requirements and stories of how to help your customers succeed. ## 2. What can you expect here? Succinct writing, and practical tools and suggestions. Often deeply technical, but with a business perspective and accessible to AI practitioners in a business setting. There will also be interactive content with questions and polls because this space is fluid. Expect an in-depth **blog post every week and a podcast every month** with an interesting enterprise leader in the trenches of AI adoption. ## 3. Why the name Enterprise Aligned? *Alignment* in AI model development is the work done to make systems more reliable, secure and coherent with corporate and human values. Adopting AI successfully to achieve business outcomes requires stakeholder and employee *alignment* within an enterprise. This work is intellectually interesting and essential, hence the name! --- ## The Hidden Power Dynamic - URL: https://enterprisealignedai.com/blog/the-hidden-power-dynamic - Author: Aparna Sinha - Published: 2026-03-09 **Excerpt:** How AI coding assistants are quietly shaping technology choices. As AI coding agents become indispensable, they're influencing which technologies we choose — and their recommendations aren't neutral. ![Small figure stands at the entrance of a large maze](/images/blog/hidden-power-dynamic.jpg) *Photo by Imkara Visual on Unsplash* As AI coding agents become indispensable to software development, they're quietly influencing more than just how we write code. They are shaping which technologies we choose. From programming languages to databases, frontend frameworks to cloud services, these agents guide our tech stack decisions. Unfortunately we may not realize that their influence isn't neutral. The recommendations come from training data heavily weighted toward open-source repositories and startup preferences. **AI is potentially steering the entire industry toward a narrower set of solutions** that may not be optimal for every use case. **Savvy startups are optimizing their documentation and online presence to gain preferential AIO (vs. SEO) placement**. ## The Illusion of Neutral Guidance Here's how this plays out in practice: in building an enterprise financial management system I mentioned S3 as an example for enterprise data storage, Claude started building around it without exploring alternatives. Only when I specifically asked about Box did Claude acknowledge it might be a better functional fit given the compliance requirements. This pattern repeated when Claude recommended Langfuse for LLM observability and Tavily for search APIs, presenting these choices as obvious defaults rather than deliberate selections from a range of options. My guess is these are not accidental suggestions. Tavily, Langfuse, and others may be putting effort into achieving this outcome, whereas others are not (yet). Claude's suggestions are not wrong, they're often solid choices. The issue is that Claude presents one solution as the solution, without revealing tradeoffs and alternatives. Unless users ask for analysis, AI makes a number of implicit assumptions, but as a product owner I want every choice to be intentional and steeped in customer requirements. ## The Data Behind AI's Technology Preferences The pattern is measurable. According to SitePoint's analysis of Claude Code's technology choices, React dominated frontend recommendations in 85% of trials, frequently paired with Next.js. On the backend, Node.js with Express appeared in roughly 80% of cases, while PostgreSQL was chosen 70% of the time for databases. These statistics reveal more than preference, they suggest a default tech stack. The projects in Claude's training data are heavily weighted toward those with extensive public documentation and tutorials, so the playbook for getting preferential treatment by Claude is starting to emerge. But these technologies are not necessarily what companies use in production, which is not documented in an AI friendly way and not marketed to AI crawlers. This disconnect between AI recommendations and optimal technology choices suggests we're seeing those targeting AIO winners, not the best solutions for every use case. **Reference:** "[What Claude Code Actually Chooses: Research Reveals AI Tool Preferences](https://www.sitepoint.com/claude-code-research-tool-preferences/)," SitePoint, February 2026. ## Remember to Challenge Your AI As more tech vendors wake up to the need for optimizing their documentation and tutorials for coding agents (rather than humans only), the range of AI suggested technologies should expand. For the time being, its important for users to develop practices for maintaining agency in technical decisions. Here's how: ### 1. Use Assumption-Forcing Prompts Instead of asking "What database should I use?" try: "Before recommending a database, please: - List your key assumptions about my requirements - Provide 3 different database options with explicit tradeoffs - Explain why you're excluding other alternatives - Identify any limitations in your knowledge" ### 2. Align Architecture Decisions to Business Requirements Architecture decisions, vendor selections, and framework choices have long-term implications. Sharpen your requirements here because these decisions: - Create vendor lock-in or technical debt - check against your broader architecture - Impact performance, scalability, and cost at scale - set explicit envelopes for these - Influence the entire application's development patterns ### 3. Develop Technical Breadth Be Claude's equal in technical judgment by: - Understanding the categories of solutions available - Knowing the key tradeoffs - Maintaining awareness of emerging alternatives ## The Path Forward The next time Claude confidently recommends a solution, remember: it's showing you what it knows best, not what's best for you. As we integrate AI more deeply, smart developers are becoming excellent AI partners, knowing when to trust, when to question, and when to dig deeper. The quality of your software, and the diversity of our technology ecosystem, depends on getting this balance right. --- ## How Wayfair Scaled Agentic AI - URL: https://enterprisealignedai.com/blog/how-wayfair-scaled-agentic-ai - Author: Aparna Sinha - Published: 2026-03-10 - Related episodes: how-wayfair-uses-arize-for-reliable-agentic-ai **Excerpt:** Why evals are the secret to reliable agents. Building an AI agent is relatively easy; making it reliable enough to handle mission-critical customer support and million-dollar freight logistics is the challenge. [![How Wayfair Scaled Agentic AI](/images/blog/wayfair-agentic-ai.png)](/episodes/how-wayfair-uses-arize-for-reliable-agentic-ai/) Building an AI agent is relatively easy; making it reliable enough to handle mission-critical customer support and million-dollar freight logistics is the challenge. In this debut episode of our podcast, I sat down with [Victor Sulaiman](https://www.linkedin.com/in/victor-a-s/) (Senior PM at Wayfair) and [Aman Khan](https://www.linkedin.com/in/amanberkeley) (Head of Product at Arize AI) to get into the weeds of how one of the world's largest retailers moved beyond the "pilot phase". Wayfair is one of the few retailers shipping Agentic AI at scale, and making it reliable enough to run mission-critical applications in customer support, freight logistics, and a globalized product catalog. If you're trying to figure out where the ROI actually lives in Agentic AI, this one is for you. ## Practical Insights from the Episode: **Agentic Commerce:** The new SEO! The future of retail is not search, but an agent that finds, visualizes, and recommends products in entirely new ways [15:31] The big shift in retail is the chat interfaces customers are using to ask what to buy. ChatGPT, Gemini, Claude, and Anthropic-powered agents are becoming the new top of funnel. Wayfair has chosen to lean in: partnering with all three foundation labs to make sure their catalog is surfaced when shoppers ask AI agents about home decor. Victor draws an explicit parallel to the early SEO era. Retailers who refused to optimize for Google because they had enough foot traffic eventually found themselves invisible! > "Imagine you're at a party and you love the table. You can circle to search it and say 'where did we buy this table?' That gets surfaced from a catalog within Wayfair. Discovery is going to start changing beyond just SEO and keywords — into image catalog and image classification." — Victor Sulaiman **Customer Service Win:** Wayfair reduced ticket resolution times from 7 days to 2 days by automating low-lift tickets, which surprisingly boosted employee satisfaction by letting staff focus on high-impact work [03:08] **Agentic Logistics:** A fascinating look at how Wayfair is using LLM reasoning—not just traditional algorithms—to optimize freight capacity and container volume [19:41] Wayfair's freight problem has historically been an optimization-algorithm problem. Containers were getting shipped at ~10% capacity because the system was rigidly optimizing for speed. The team turned to LLM reasoning to dynamically decide what to add to a container, where to route it, and how to maximize volume without slipping delivery promises. Early results show container-volume gains and meaningful savings. **LLM as a Jury:** Wayfair uses multiple LLMs to "deliberate" on decisions like furniture translations to avoid hilarious (and costly) errors [31:10] Wayfair sells globally, translating product catalogs into many languages at scale. They started with an LLM-as-a-judge approach: a single model deciding whether a translation was correct. > "'Sofa' in Polish would be 'rug.' The judge LLM said: 'Yes, this is from English. Yes, this word is in Polish. Yes, this is home decor. Therefore, this translation makes sense.'" — Victor Sulaiman The fix: replace the single-judge with a jury of multiple LLMs that surface a *set* of recommendations, then escalate ambiguous cases to a human. The jury catches the 1% that breaks the customer experience. Aman has seen Arize customers adopt LLMs as a Jury across regulated and customer-facing use cases. It's how you get reliability without bottlenecking on humans. **Build vs. Buy is Becoming "Buy and Build"** — Victor's framework: plot every system on a spectrum of *control needed* versus *iteration speed needed*, and let that drive the choice. > "Right now it's 'I have a problem, do I build or buy?' I think that's going to change to 'buy, and then build on top of it.'" — Victor Sulaiman **The Pilot Trap:** Choosing the right agent for the right job is key. Sometimes a simple "reflex agent" beats a complex hierarchical one [47:20] **📖** [Full episode](https://enterprisealignedai.com/episodes/how-wayfair-uses-arize-for-reliable-agentic-ai/) with timestamps, key takeaways, and show notes --- ## MCP is the Fastest-Growing Protocol in AI History - URL: https://enterprisealignedai.com/blog/mcp-is-the-fastest-growing-protocol - Author: Aparna Sinha - Published: 2026-03-30 **Excerpt:** Despite reports of its untimely death, MCP has grown 5,000% in 16 months with over 100 million monthly SDK downloads. Here's why the protocol is thriving and what it means for enterprise AI. ![MCP Protocol](/images/blog/mcp-protocol.png) In February, Eric Holmes published ["MCP is Dead. Long Live the CLI."](https://ejholmes.github.io/2026/02/28/mcp-is-dead-long-live-the-cli.html) and the tech community on Twitter piled on. At that time, OpenClaw didn't use MCP and was actively promoting CLIs and Skills as more context efficient, composable, debuggable alternatives to MCP. But as of April 2026 MCP is looking pretty alive. It's the fastest-growing open-source AI project in history, growing 5,000% in 16 months. Here's the current state in April: - **over 100 million monthly [SDK downloads](https://npm-stat.com/charts.html?package=%40modelcontextprotocol%2Fsdk)** and over 10,000 active servers - **[First-class client support](https://modelcontextprotocol.io/clients)** across ChatGPT, Claude, Gemini, Cursor, Microsoft Copilot, and VS Code - **Anthropic** now has 75+ connectors powered by MCP, and launched Tool Search for production-scale deployments - **OpenAI** shipped [dynamic tool search for MCP](https://openai.com/index/introducing-gpt-5-4) in GPT-5.4, reducing token overhead by 47% - **The [Agentic AI Foundation](https://aaif.io/)** co-founded by Anthropic, OpenAI, and Block, with platinum members AWS, Google, Microsoft, Bloomberg, and Cloudflare now stewards MCP under the Linux Foundation - **The MCP Dev Summit** in NYC (April 2-3), featuring 95+ sessions from Anthropic, Datadog, Hugging Face, Microsoft, and others, is oversubscribed So what is going on here? ## CLIs Are Great — Especially for Autonomous Agents I love CLIs. Most developers do. LLMs have been trained on CLI docs and are therefore experts at using 'grep', 'awk', and other Bash commands. AI workflows often require multiple tool calls chained together (e.g., you might call your email server and then your calendar etc.). If every tool had a good CLI, LLM would write a long string of commands that executes in a single leap for any given workflow. CLIs are also self documenting, so LLMs can discover features on the fly (using help/man). And, CLIs can be more context efficient than MCPs - describing tools in fewer tokens than large MCP schemas. Having said that, none of these advantages are permanent. MCP is a new spec and very much a work in progress. Still when comparing CLIs to local MCPs, CLIs are a clear winner because the security value proposition of local MCPs is not much better than CLIs. However, remote or gateway managed MCPs are a totally different story. At the root of the issue is a tradeoff between Agent Autonomy and Enterprise control. A consequence of this is that **what an individual developer is comfortable trading off for the benefit of greater Agent autonomy, is not always what is best for teams, enterprises, and the broader ecosystem** of non-technical builders. ## Agent Autonomy Needs Robust Controls The deeper tension beneath the MCP vs. CLI debate is worth naming: **how much autonomy should we give agents?** This question is actually fundamental to Enterprise AI adoption and the current state of the AI bubble (massive investment and usage in SF/Silicon Valley circles but hesitance, fear and slower adoption in Enterprises where AI could have significant impact). The CLI-maximalist position is synonymous with more autonomy. Give the agent shell access, let it figure out which commands to run, let it chain tools together however it sees fit. The more access the better - because that makes the agent more capable and the user much more productive. How could this be bad? In fact, it is even more appealing as models get smarter. And they are getting smarter, fast. But of course the reality is when you are inside a company, autonomy without guardrails is a recipe for all sorts of problems. We don't have to look farther than the OpenClaw security crisis to get a sense for what could go wrong. ## Requirements CLIs Alone Don't Solve MintMCP founder [Jiquan Ngiam](https://jngiam.bearblog.dev/mcps-clis-and-skills-when-to-use-what/) uses CLIs for side projects, but MCPs at work to run background agents, each with scoped access. He will be on my podcast next month. The tradeoffs he notes: - **Access control**: CLIs are hard to scope per session. MCPs have built-in `allowedTools` per session. - **Auth**: CLIs require on-device tokens and terminal commands. MCPs offer OAuth with single-click UIs. The CLI **Credential UX** is painful for non-developers. MCPs connect in one click. - **Observability**: CLIs produce ad-hoc output. MCP servers enable standardized OpenTelemetry metrics across teams. I found Charles Chen's article, ["MCP is Dead; Long Live MCP!"](https://chrlschn.dev/blog/2026/03/mcp-is-dead-long-live-mcp/), to be the first super sensible take that busted the CLI hype-train on X. It was a widely-read take ranked highly on Hacker News. His core argument is that MCP enables organizational-scale agentic engineering with security, observability, and governance built in. ### 1. Security In an enterprise, you need OAuth-based authentication with secrets managed server-side. You need to revoke access when someone leaves without worrying about API keys sitting in dotfiles. You need to know for sure that an agent running a background task can read call transcripts but *can't* send emails. Remote MCPs accessed via Gateways are the pattern enterprises are using to secure Agent tool use. This isn't just an enterprise problem. *Anyone* running AI agents needs security. The individual developer vibing in their terminal is one malicious dependency away from a compromised environment. OpenClaw illustrates what happens when security is an afterthought. Researchers found over 135,000 OpenClaw instances exposed to the public internet, with [15,000 vulnerable to remote code execution](https://www.darkreading.com/application-security/critical-openclaw-vulnerability-ai-agent-risks). Would MCP have prevented this? MCP's remote server model means there is no listening port on the user's machine to expose. The 135,000+ exposed instances are architecturally impossible with remote MCP. MCP's OAuth 2.1 model keeps credentials server-side with scoped, short-lived tokens, not stored in local dotfiles waiting to be exfiltrated. Of course MCP is not a silver bullet. It had 30+ CVEs of its own in early 2026, but its *design* makes secure deployment the path of least resistance. ### 2. Observability When an agent uses a CLI, what gets logged? Whatever the agent decides to capture. When an agent uses an MCP server, you get structured, standardized telemetry: what was requested, what was executed, what was returned, and how long it took. The [2026 MCP roadmap](https://workos.com/blog/2026-mcp-roadmap-enterprise-readiness) makes this explicit: end-to-end audit trails enterprises can feed into *existing* logging and compliance pipelines. For regulated industries, and many of the large US banks and financial institutions I speak to, this is required. ### 3. Governance Who can use which tools? What versions are deployed? How do you roll out an upgrade across 200 agents? How do you ensure consistent behavior? Charles Chen jokes that the CLI-only argument is "cowboy vibe-coding culture." Every developer installs their own version, configures their own flags, manages their own upgrades. As Ngiam points out, there's no "consistent interface across CLIs: credential management, flag conventions, error handling all work differently." MCP servers provide centralized configuration, version control, and dynamic content delivery, what Chen calls "server-delivered SKILL.md" that auto-updates across all tools without manual synchronization. This is well-managed. It means your OpenClaw agent, isn't auto-installing skills and CLIs from the web on its own! The tradeoff of agent autonomy for security and governance is clear here. Lastly, there is the sobering practical consideration that many enterprise backends don't have CLIs. They may have APIs — often REST, sometimes GraphQL, but wrapping them in a CLI means custom tooling for each system. MCP provides a standard protocol for wrapping these APIs once and making them available to any MCP-compatible client. ## What About Skills and Code Mode? **Skills** are process documentation for agents, encoding domain-specific know-how, team conventions, and multi-step workflows. They're complementary to both CLIs and MCPs. In fact skills + MCP is often what Anthropic uses to post specific vertical plug-ins. (As an aside **agents sometimes don't pickup the right skill.** Vercel's evaluation of their Next.js agent found that in the majority of eval cases, the [agent never invoked the available skill](https://vercel.com/blog/agents-md-outperforms-skills-in-our-agent-evals) at all). **Code Mode** gives the agent an API spec and lets it write custom code in a sandbox for each request - this is of course maximally flexible and best for intricate use cases, but hard to audit, govern, or make repeatable. ## The Path to Agent Autonomy The path forward, which every AI model company is building towards is to create the infrastructure that lets autonomy scale safely. MCP's gateway pattern, scoped tokens, and structured telemetry make sense as steps in that direction, at least for now. As models improve and earn more trust, the guardrails can loosen. In a notable turn, OpenClaw's founder Peter Steinberger shared that the next version [will adopt MCP](https://www.rootdata.com/news/592688), replacing its proprietary messaging channel with the standardized protocol. The enterprises I speak to — large US banks, financial institutions, technology companies are already using MCP or actively adopting it, and almost always with Gateways to secure the MCPs. The ecosystem has grown to over 5,800 community and enterprise servers spanning databases, CRMs, cloud providers, and developer tools. I expect the future to look like smarter models that use MCP, Skills, CLIs, and Code Mode fluidly, with perhaps many more custom plugins available and better automatic Skills recognition by agents. With the right guardrails, there's no doubt we will get to Agent autonomy one day soon - the benefits are too good to ignore. *Aparna Sinha is the host of the [EnterpriseAligned AI](https://enterprisealignedai.com/) podcast, where she speaks with enterprise leaders about AI adoption in practice. Upcoming episodes feature conversations with RBC and Coursera about their own MCP adoption journeys.* --- **Sources:** - [Eric Holmes, "MCP is Dead. Long Live the CLI"](https://ejholmes.github.io/2026/02/28/mcp-is-dead-long-live-the-cli.html) - [Charles Chen, "MCP is Dead; Long Live MCP!"](https://chrlschn.dev/blog/2026/03/mcp-is-dead-long-live-mcp/) - [Jiquan Ngiam, "MCPs, CLIs, and Skills: When to Use What"](https://jngiam.bearblog.dev/mcps-clis-and-skills-when-to-use-what/) - [Vercel, "AGENTS.md Outperforms Skills in Our Agent Evals"](https://vercel.com/blog/agents-md-outperforms-skills-in-our-agent-evals) - [The New Stack, "MCP Roadmap 2026"](https://thenewstack.io/model-context-protocol-roadmap-2026/) - [WorkOS, "MCP's 2026 Roadmap Makes Enterprise Readiness a Top Priority"](https://workos.com/blog/2026-mcp-roadmap-enterprise-readiness) - [Linux Foundation, "Agentic AI Foundation (AAIF) Announcement"](https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation) - [Anthropic, "Donating MCP and Establishing the AAIF"](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation) - [OpenAI, "Agentic AI Foundation"](https://openai.com/index/agentic-ai-foundation/) - [Dark Reading, "Critical OpenClaw Vulnerability"](https://www.darkreading.com/application-security/critical-openclaw-vulnerability-ai-agent-risks) - [Sangfor, "OpenClaw Security Risks"](https://www.sangfor.com/blog/cybersecurity/openclaw-ai-agent-security-risks-2026) - [Microsoft, "Running OpenClaw Safely"](https://www.microsoft.com/en-us/security/blog/2026/02/19/running-openclaw-safely-identity-isolation-runtime-risk/) - [MCP Dev Summit North America](https://events.linuxfoundation.org/mcp-dev-summit-north-america/) - [Google Cloud, "Official MCP Support for Google Services"](https://cloud.google.com/blog/products/ai-machine-learning/announcing-official-mcp-support-for-google-services) --- ## Royal Bank of Canada's AI Platform is Scaling Agentic AI to 96,000 Employees - URL: https://enterprisealignedai.com/blog/rbc-scaling-agentic-ai-to-96000-employees - Author: Aparna Sinha - Published: 2026-04-23 - Related episodes: rbc-1b-ai-transformation **Excerpt:** The 'AI for All' Blueprint. How RBC built a self-service platform that turns 96,000 employees into builders, backed by MCP, OTel, and the Lumina data foundation. [![Royal Bank of Canada's AI Platform is Scaling Agentic AI to 96,000 Employees](/images/blog/rbc-ai-for-all.png)](/episodes/rbc-1b-ai-transformation/) In Episode 2 of [Enterprise Aligned AI](https://enterprisealignedai.com/), I sat down with [Vinh Tran](https://www.linkedin.com/in/vinhbtran/), VP of Data and AI Platforms and RBC Fellow at Royal Bank of Canada, to discuss one of the most ambitious AI journeys in global finance. While many organizations are stuck in the "pilot" phase, RBC has made a public commitment to drive **$700M to $1B in value** through AI by 2027. The key insight is the platform they have built to scale their human transformation. RBC is delivering on nine elite AI projects and also lifting every employee in the bank with AI. > "We're driving the value not just from the six or seven or eight thousand developers or data scientists, which is historically where AI was being utilized. We're now looking at transforming the 96,000… everybody with a computer: every mortgage broker, every engineer, every branch manager." — Vinh Tran ## The Power of "Self-Service" Agents RBC's core strategy is to move beyond a central "experts-only" model. They have built a self-service platform, headlined by **RBC Assist**, that empowers everyone from mortgage brokers to branch managers to build their own agents without writing code. What this unlocks goes beyond productivity. > "We have an opportunity to change how everybody works, how everybody does their job, drive more efficiency and productivity, but also be able to do things that people were never able to do before." — Vinh Tran Vinh notes the platform tools allow staff to explore complex data across multiple sources using natural language. In doing so, RBC is turning their workforce into a massive engine of builders. ## Standardizing for Security, Reliability and Speed To make "AI for All" a reality in regulated industries is no small feat. RBC focused on standardizing the "connective tissue" of their stack: - **The Data Foundation:** This transformation is fueled by **Lumina**, RBC's modern data platform. Vinh credits his predecessor for starting to build a robust semantic layer. The organization is now adopting a **data-product mindset** with curated data hubs (retail credit, mortgage) that give agents the semantic layer and high-quality metadata they need to understand banking fields. > "A lot of that work was underway well before our agentic journey. Now that we're moving quickly on this agentic journey, we're seeing that that semantic layer, that data classification, that metadata is very important to the agents." — Vinh Tran - **MCP Gateway:** RBC uses the Model Context Protocol (MCP) to standardize how agents access tools and data. By placing MCP servers behind a secure gateway, they ensure every agentic action is approved, compliant, and has a "human in the loop" for material system changes. > "We've put our MCP servers behind a gateway. That gateway has metadata on the tools and the MCP servers behind it, so we know that if you're calling this tool, you better have some human in the loop and a plan that's approved before you can do that." — Vinh Tran - **Observability with OTEL:** Standardization on **OpenTelemetry (OTel)** ensures continuous evaluation of agent plans and actions from production telemetry, not weekly batch runs. ## Build vs. Buy Calculus has Shifted **It's a better time than ever to build**, and Vinh is passionate about it. He argued that three shifts have lowered the bar to build inside enterprises: open source that's now unambiguously production-grade (the old "you can't run open source in production" fear is dead), new **AI coding tools are giving engineers 10x to 100x leverage,** and cloud services that used to be the domain of large vendors are now deployable by anyone. You don't have to write every line of code to build, you can be an integrator composing with what's already out there. In regulated enterprises, building is of even greater importance. When compliance and policy enforcement are non-negotiable, building provides greater control over policy guardrails and timelines. Vinh also argued that **building is a culture-forging function for the engineering org**, something you can't buy. > "You're generally going to save money by building, but I don't think that's the most important thing. When you're in a regulated industry like we are, having control of your destiny, being able to incorporate your policies, your guidelines, patch your code, do everything that we need to do to meet regulatory requirements, it's powerful for us to be able to have that control in our own hands." — Vinh Tran ## From Engineers to "Builders" Vinh described the elevation of the engineering role. In this new world, engineers are no longer just writing code; they are becoming **Uber Builders** who orchestrate AI tools, audit model outputs, and ensure compliance. > "I think we'll still need engineers, and I think they'll become builders... the journey right now is what's important: building those muscles, knowing how to use those tools." — Vinh Tran RBC's $1B transformation proves that the most powerful accelerator in enterprise AI isn't the model you buy. It's the platform you build to empower your people. My favorite quote from Vinh in this episode: > "I really feel that we have to have the courage to build. We're not going to write every line of code. We're going to use the tools. We're going to use the cloud. We're going to use the community, but I think that allows us to really drive benefits for the organization and our clients." — Vinh Tran **📖** [Full episode](https://enterprisealignedai.com/episodes/rbc-1b-ai-transformation/) with timestamps, key takeaways, and show notes --- ## Logitech's Award-Winning AI Share-of-Voice - URL: https://enterprisealignedai.com/blog/logitechs-award-winning-ai-share - Author: Aparna Sinha - Published: 2026-05-24 - Related episodes: agentic-commerce-with-logitech-and-typesense **Excerpt:** How the hardware giant mastered Agentic Commerce using Typesense's low-latency search engine. [![Logitech's Award-Winning AI Share-of-Voice](/images/blog/logitech-ai-share-of-voice.jpg)](/episodes/agentic-commerce-with-logitech-and-typesense/) Brand storytelling, creative visual merchandising, and emotional triggers have been the basis of e-commerce targeting humans. But Agents don't respond to these. AI without eyes, emotions, or patience for advertisements requires different inputs. In this episode of [Enterprise Aligned AI](https://enterprisealignedai.com/), I sat down with [Deepa Shekhar](https://www.linkedin.com/in/deepa-shekhar-677779/), Director of E-Commerce Technologies at Logitech, and [Jason Bosco](https://www.linkedin.com/in/jasonbosco/), CEO and Co-Founder of Typesense, to map out the technical, semantic, and structural plumbing required for Agentic Commerce. Logitech recently won a prestigious accolade **in Semrush's 2025 AI Visibility Awards for Consumer Electronics**, as a highly discoverable and trusted brand in the AI ecosystem. The secret to their win was a radically new content strategy, and their investment in a low latency, high accuracy enterprise data architecture. This is the operational blueprint of how Logitech.com transformed to optimize share of voice in AI, enabled by the underlying search infrastructure from Typesense. ## Semrush AI Visibility Award: Navigating the Great Content Shift To win users inside LLM answer windows, Deepa explains that marketing and technology teams must partner closely, and jettison the keyword-based SEO mindset. When a consumer shifts their shopping journey to platforms like ChatGPT, Perplexity, or Gemini, they are no longer browsing catalogs; they are asking highly contextual, intent-driven questions about problems they want to solve. Deepa notes the queries that are winning Logitech AI market share today: > _"What is a good mouse for long working hours?"_ > > _"What is a good mouse that works cleanly on a glass surface?"_ Logitech brought together marketers from various business units, channels, and the core brand team and fully updated their content strategy. By optimizing content for machine readability and semantic relevance, Logitech ensured their products surfaced as the top trusted answers. Here's their playbook: 1. **Write Semantic Content:** Marketers stopped writing copy about abstract product features and began authoring content to answer problems customers were trying to solve. 2. **Engineer Agent Readability:** The technology team took this problem-solving content and published it in highly structured formats that LLMs can digest quickly. They deployed modern AI-readability standards like **llms.txt**, which acts as an explicit AI sitemap instructing LLMs on the navigation paths to find authoritative brand data. > **💡 Actionable Advice for Leaders:** **Update your marketing content to mirror customer questions.** Audit top-performing pages and anchor text around contextual customer problems that semantic search algorithms look for. ## The Three Tracks of Agentic Commerce > "I see agentic commerce evolving across three tracks, distinguished by where the commerce transaction takes place and who owns the consumer experience… We must execute on all three tracks to be successful." — Deepa Shekhar - **Track 1: LLM Platforms as Strategic Engines of Commerce.** Discovery and transaction journeys are moving from retailer websites to third-party LLMs like ChatGPT and Gemini. The **new KPIs are visibility and conversion inside LLM platforms**, not website click-through rates. - **Track 2: Proprietary Brand Agents Grounded in Your Data.** Building your own AI Agents (in your app, your website, messaging channels etc.) is more important than ever in your unique brand voice enriched with your user and product data. - **Track 3: Pure Algorithmic Buying.** Soon, personal consumer AI agents will shop on behalf of humans, and enterprise systems need to be ready. Agents rely on structured data and select products objectively based on operational parameters, e.g., granular specifications, shipping velocity, and historical fulfillment reliability. ## Preparing Data for Agentic Commerce All three tracks require data preparation, integration and LLM optimized search. Commerce data is typically fragmented across separate backends including pricing engines, real-time inventory systems, content databases, and Product Information Management (PIM) software. Instead of forcing an Agent to query multiple backends in real time, which causes extreme latency and wastes tokens, enterprises must assemble this multi-source data into a single, unified database layer, indexed by a specialized, low latency search engine. Deepa recommends an architecture that uses a high-performance search engine like **Typesense as the semantic layer and central orchestration hub**. In this architecture, the Agent LLM does intent detection: it parses natural language variations, captures customer synonyms, and maps user context (like translating a query about a "repetitive strain injury" into an ergonomic design match). Intent is then passed to a search engine like Typesense allowing the Agent LLM to run more complex queries in a shorter timeframe, resulting in better answers, and higher AI visibility. **Typesense can run dozens of query combinations across millions of records in milliseconds. It takes the fragmented, slow data from your core systems, indexes it, and exposes it as a blazing-fast, high-availability API.** > **💡 Actionable Advice for Leaders:** **Create a fast semantic search layer and use LLMs to augment your data.** Decouple extraction from storage. Use LLMs offline to auto-tag product catalogs with synonyms, and real-world context so the underlying data store is pre-optimized for semantic queries. ## Agents Already Make Product Decisions Perhaps the best moment of the discussion was Jason's recent Agentic Commerce encounter. After this, Jason updated Typesense's technical assets to ensure full machine-readability; so that documentation is served as markdown alongside HTML. Typesense also now embeds metadata instructions in their pages pointing to markdown versions of each page, ensuring agents don't waste tokens parsing decorative web formatting. > **💡 Actionable Advice for Leaders:** **Create markdown-native layouts for your documentation and assets.** AI crawlers and buying agents prefer markdown over complex HTML. ## Product Catalog Integration One roadblock facing enterprises trying to plug into Agentic Commerce is the lack of a unified mechanism to share product catalogs with LLMs. Deepa summarizes the fragmented mosaic of evolving, incomplete protocols engineering teams face: - **SEO and Structured Feeds:** traditional semantic tagging and canonical metadata feeds remain essential. However, the penalty for data discrepancies is now severe. If your feed frequency falls behind, AI (e.g., ACP) will penalize your visibility. Feeds must now be refreshed in minutes rather than hours. - **OpenAI's Agentic Commerce Protocol (ACP):** provides structured capabilities for syncing catalogs, pricing models, and active promotions. It lacks native payment capabilities, requiring a secondary marketplace model or merchant-of-record partnerships with Stripe. ACP uses inventory data for _discovery ranking_, making a fast backend more important for AI visibility. - **Universal Commerce Protocol (UCP):** defines standards for checkout, order updates, and payment handoffs. As buying agents execute these transactions asynchronously, inventory validation often happens in a single flash-point at checkout. To survive this without systemic cart abandonment, merchants must construct high-availability, low-latency APIs running in parallel to feed real-time inventory counts during protocol handoff. - **Model Context Protocol (MCP):** While highly anticipated, MCP is not yet intuitive for scaled retail. It functions more like a sandboxed app extension inside the LLM container (similar to custom apps within ChatGPT) rather than a dynamic, automated communication framework. The industry needs a unified, cross-platform protocol that seamlessly binds real-time inventory, batch product attributes, and multi-vendor financial transactions into a single standard. ## Modernizing Legacy Backends An interesting barrier to scalable agentic commerce is that traditional enterprise backends (ERP, WMS, and OMS systems) are architected for batch processing, syncing data overnight or in hours-long intervals rather than milliseconds. They assume human web traffic, which naturally throttles itself, and are structurally unequipped to handle high-frequency, low-latency read/write loops from automated software. Emerging unified frameworks like the **Order Exchange Protocol (OnX)**, championed by a consortium of modern OMS and e-commerce vendors, are aiming to solve this by introducing standardized, real-time eventing layers. Startups that successfully build real-time, ultra-low-latency translation layers over legacy, batch-processed systems will unlock an immense enterprise market opportunity. Lastly, we barely scratched the surface on how to manage financial risk in Agentic Commerce, noting the significant developments still to come in this area. ## Bottom Line Agentic commerce requires significant re-architecture across functions. The brands that win the next decade will build their data, their content, and their infrastructure for machines and humans. Logitech is showing that work is already paying off in measurable share of voice; Typesense is showing the stack underneath has to be rebuilt for agents querying in milliseconds, not humans browsing in seconds. 📖 [Full episode](https://enterprisealignedai.com/episodes/agentic-commerce-with-logitech-and-typesense/) with timestamps, key takeaways, and show notes. --- ## Defensible Moats for Vertical AI Companies - URL: https://enterprisealignedai.com/blog/defensible-moats-for-vertical-ai - Author: Aparna Sinha - Published: 2026-06-18 **Excerpt:** Foundation models keep improving and getting cheaper. A new Stanford CodeX paper, co-authored by Aparna Sinha and Jay Mandal, ranks five product moats vertical AI companies can build to compete, from workflows and UX to embedded judgment. Foundation models improve every quarter and get cheaper. For anyone building a vertical AI application in legal, accounting, financial services, or consulting, that poses a question: as model capability rises, what durable moats can a vertical company build to compete? Jay Mandal and I wrote a paper on it, published this week by Stanford Law's CodeX (the Stanford Center for Legal Informatics): "Defensible Moats for Vertical AI Application Companies in a New Competitive Landscape." ## The competitive landscape is shifting rapidly Vertical software was dominated by SaaS companies for two decades. Foundation models added a new set of competitors at once: AI-native startups, AI-enabled SaaS businesses, AI-native and AI-enabled services firms, open-source applications, and in some cases the model providers. In Legal the competitive landscape has broadened. Incumbents (Thomson Reuters, LexisNexis, Wolters Kluwer) still hold exclusive access to primary law and regulatory content and are adding AI products on top. AI-native startups (Harvey, Legora, Eudia) have gained distribution quickly and are carving out the market. Law firms are also building their own applications and open-source projects are replicating the leaders cheaply. Finally the model providers have also to an extent entered the market with Anthropic's Claude for Legal and OpenAI's announced legal vertical, and their recently announced partnerships with system integrators. The same shift is starting across accounting, financial services, banking, healthcare, and life sciences. ## Five product moats for competing at this layer We rank five product moats in ascending order of strength. The weaker ones are easier for a competitor to replicate. ![The five moats of vertical AI applications, in ascending order of strength](/images/blog/moats-vertical-ai.png) 1. **Workflows and UX.** Encode a customer's operating procedures as skills, an expert-in-a-box. This has value, but a competitor can build the same, so it is the lowest moat. Decagon does this for banking customer service, with agents that field more than one million calls a month for Chime. 2. **The vertical harness and custom tools.** Build the tools and integrate the in-house and legacy systems horizontal players will not touch, then orchestrate tools, context, memory, and security around one industry. Harvey and Legora compete on legal-native integrations: iManage, document management, and research connections. The underlying model is interchangeable. 3. **Built-in compliance.** Meet regulatory and policy requirements with the determinism, explainability, and auditability that raw models lack, and keep the UX clean. Jump does this for wealth management and commands a price premium over horizontal recording tools. It reached roughly 27,000 advisors in under two years. 4. **The Brain, a data-driven operating system.** Encapsulate hard-to-obtain proprietary data into a vertical-specific system: curated public authority, private client context, real-time data, and physical data, plus operating insights from the combination. Accordance built this for tax and accounting on a curated corpus of statutes, standards, and precedents fused with each firm's own context. Switching costs rise the longer the system runs. 5. **Embedded judgment.** Encode the judgment and taste of expert practitioners as business logic that guides decisions at key points in complex workflows. The system learns from the decisions the best practitioners make, and the gap widens with every customer. This is the hardest moat to codify and the most exposed to model progress. Three operational moats (distribution, engineering, and operational excellence) strengthen the product moats but do not stand on their own. ## The takeaway As model capability rises, the value of any single workflow, tool, or product shipped alone falls. The durable value is the system: a combination of the five product moats, strengthened by the operational moats. The paper has the full framework, the company case studies, and open questions on how the hierarchy shifts as models grow more autonomous. Read the paper: https://law.stanford.edu/publications/defensible-moats-for-vertical-ai-application-companies-in-a-new-competitive-landscape/ Stanford CodeX shared on LinkedIn: https://www.linkedin.com/feed/update/urn:li:activity:7473473710163795969/ --- ## How Coursera Became an AI-Native Enterprise - URL: https://enterprisealignedai.com/blog/coursera-becoming-ai-native - Author: Aparna Sinha - Published: 2026-07-08 - Related episodes: coursera-becoming-ai-native **Excerpt:** Coursera is a fourteen-year-old public company that transformed itself to the leading edge of AI. CTO Mustafa Furniturewala and Mint MCP founder Jiquan Ngiam break down how an established enterprise becomes AI-native: enablement, culture, and automation. What's an AI-native enterprise? Can a decades-old public company become one? Coursera is a fourteen-year-old public company that has transformed itself to the leading edge of AI, and in doing so redefined how education and engineering can be done better with AI: personalized assessments and AI tutors for learners, a Course Builder for its three-sided marketplace, and a custom reinvention of the agentic engineering SDLC inside Mustafa's team. Coursera is turning AI into a growth engine with courses on exactly the skills companies are looking for in new hires. In May 2026 it completed a $2.5 billion combination with Udemy, and the two platforms now reach more than 200 million learners. I sat down with Mustafa Furniturewala, CTO at Coursera, and his MCP Gateway technology partner Jiquan Ngiam, CEO of MintMCP, to discuss how they did it. It comes down to three things: enablement, culture, and automation. ## 1. Enablement Being AI-native means empowering the entire company, and the benefits of AI come with risk. Engineering is hard enough to transform when you have decades of technical debt. But being truly AI-native means getting everyone in the company to become an expert user, not at all easy to do securely in a large enterprise. > "The primary goal is to enable everyone. My first instinct is not to block access, but to see what is the secure, right way to enable people with the tools." Before MintMCP, API tokens were scattered across laptops, business users wanted Claude Code access, and nobody could say which MCP tools were approved. MintMCP is a gateway that governs AI's access to data. It centralizes tokens so no keys sit on laptops, bundles tools by role, scans inline for sensitive data, and keeps audit logs. IT sets the policies once and they work across Claude, Cursor, and ChatGPT etc. with a unified view of every agent. Enablement is also teaching AI how to work. A skill is a markdown file of standard operating procedures. Coursera forked Anthropic's knowledge-work skills repo and built many skills. For example, an onboarding skill for new engineers replacing the Confluence pages that used to hold that process. ## 2. Culture Coursera built an inclusive AI culture. There is an AI council, a group of engineers that meets to discuss the company's AI frameworks and how to improve them. There is a plugins repo everyone contributes to, an award for the best plugin of the month, and Spark sessions where people show each other how they use AI, which sparks more plugins in turn. A developer experience team runs much of this. More than a quarter of Coursera's Claude usage now comes from people who do not write code. Cowork edits the document, the spreadsheet, and the deck directly and pulls the data itself. Those business users are contributing skills back. Mustafa put the mindset this way: > "I don't think we are fully there... But the key is to have a system, keep improving it, and create feedback loops where the value compounds over time." ## 3. Automation Mustafa's team is creating their new agentic SDLC. He says it is a great time to be an engineer, because AI takes the grunt work across more than 200 repos and millions of lines of tech debt built up over a decade, leaving design and product taste to humans. A Scala-to-Java migration that took engineers weeks now runs in two to three hours, deployment included. A separate migration of 500 tests ran in weeks instead of quarters. It doesn't just work by default though: > "If you just spin up Claude Code and just do this long-running thing, it's likely going to fail, or it will hallucinate, it will create bugs. So the setup and the system around it becomes very critical." The setup around the agent for Coursera means nested CLAUDE.md files with the architecture and patterns written down, and a model that retrieves the right context itself rather than through a separate index. At MintMCP, Jiquan goes further. A review loop has Claude write a plan and the code, Codex review it with fresh eyes, and the two go back and forth until the code is simpler and secure; sometimes the pull request gets smaller. This "Do-Anything" loop or Dan-loop takes a while, so JQ moved it into a sandbox triggered from Slack. Jiquan runs a whole team of agents: one recreates his exec team to debate a document, and another, called Charlie, reads the week's sales calls to tell the team what is and is not selling. The agents do real work across functions, each scoped and read-only where it should be. ## The tech: API, CLI, or MCP Jiquan explains why MCP, when APIs, CLIs, and browser use exist. APIs aren't designed for agents and may not allow the level of granular access controls or have the usability needed. MCP moves that work to the server, so a business user never touches a command line. Notion's first MCP copied its API one-to-one; the second was rebuilt for agents, with a new language, more explicit granular scopes, and works far better. Ultimately agents are powerful only when they have the right tools. MCP Gateways are a secure way to enable agents for all sorts of users, not only engineers. Clips from the episode: [YouTube playlist](https://www.youtube.com/playlist?list=PLORpLzeUmxyE) 📖 [Full episode](https://enterprisealignedai.com/episodes/coursera-becoming-ai-native/) with timestamps, key takeaways, and show notes. --- ## Who Owns the Legal AI Harness? - URL: https://enterprisealignedai.com/blog/who-owns-the-legal-ai-harness - Author: Aparna Sinha - Published: 2026-07-26 **Excerpt:** Baker McKenzie & Legora on Build vs. Buy, Pricing, and the future of Legal AI [![Who Owns the Legal AI Harness? Baker McKenzie & Legora](/images/episodes/ep5-legora-bakermckenzie-og.jpg)](/episodes/who-owns-the-legal-ai-harness/) Who owns the AI harness inside a law firm? And what happens to the billable hour when AI enables everyone? Enterprise AI usage for legal work is surging with **79% of lawyers now using AI** (up from 19% two years ago)[^1]. Legal is not just an early adopter, it’s a bellwether for all knowledge work. A sampling of the questions now facing the legal profession: - Does AI change the billable hour pricing model for Law Firms? - What is the role of legal expertise and judgement and how best to apply it? - How can legal tech startups differentiate themselves from model providers? - Where does it not make sense to use AI in legal? - Should lawyers build technical skills? In this episode of **Enterprise Aligned AI**, I sat down with two incredible leaders tackling these questions: **[Max Junestrand](/speakers/#max-junestrand)**, CEO and Co-founder of Legora, and **[Danielle Benecke](/speakers/#danielle-benecke)**, Founder and Global Head of Baker McKenzie’s Applied AI practice. - **Legora:** Founded in Stockholm in 2023, Legora is one of the fastest enterprise software companies in history to reach **$100 million in ARR**, hitting that milestone in just 18 months. Valued at $5.6 billion this spring, Legora now serves over 1,200 legal teams. - **Baker McKenzie:** A 77-year-old firm and more global than most peers, with 70+ offices worldwide. Danielle leads a specialized unit of lawyers and AI technologists building custom AI systems directly for multinational clients. ## Key Takeaways from the Conversation - **The Business Model Shift:** Legal are 95% human delivered services and 5% software today. Max and Danielle explain that this ratio is shifting rapidly and has broad implications. They explain why Legora moved to consumption-based pricing, while Baker McKenzie is leaning into fixed-fee models over billable hours. - **Expertise should move upstream:** Because generating a first-draft legal analysis is now cheap and fast, human judgement and accountability are the scarce assets. Danielle’s team pulls human judgement into designing AI architecture and guardrails upfront rather than mechanically validating AI output line-by-line downstream. [Clip](https://youtu.be/ZdYKJ-VVF-4) - **Work that could not be done before:** The prize is not doing today’s work faster. Danielle’s team now delivers same-day analysis of reporting obligations during a live cyber incident, work that used to take days or weeks before AI. She argues the industry should focus on the services that do not exist yet rather than accelerating the ones that do. As AI generated output proliferates this becomes urgent. [Clip](https://youtu.be/9rmd1r-V_2Q) - **Differentiating via the Harness and Operating System:** Max gives his insights on building a winning product: the harness is the orchestration layer that determines which documents the model reads, resolves conflicts between court precedent and internal firm playbooks, chooses which models run where, and manages sub-specialization. [Clip](https://youtu.be/ASbV3-js0Ww) - **When to Own vs. Rent:** Danielle shares a rule that generalizes beyond law: **Own the harness** where the workflow hinges on your unique institutional expertise, playbooks, and strategic edge. **Rent it** where the capabilities are general or standardized. I ranked the five moats vertical AI companies can build in a [Stanford CodeX paper](/blog/defensible-moats-for-vertical-ai/) with Jay Mandal. This episode adds a great deal of depth to that work. ## Join the Conversation Whether you are an enterprise executive, in-house counsel, or AI practitioner, this debate provides a practical blueprint for navigating build vs. buy decisions and structural pricing shifts in agentic AI. *If you know a legal or enterprise leader currently deciding what to build and what to buy, pass this post along to them!* [^1]: Source: [Clio Legal Trends Report](https://www.clio.com/about/press/clio-latest-legal-trends-report/) 📖 [Full episode](https://enterprisealignedai.com/episodes/who-owns-the-legal-ai-harness/) with timestamps, key takeaways, and show notes. --- ## Marketing's Claude Code Moment - Coming Soon? - URL: https://enterprisealignedai.com/blog/marketings-claude-code-moment - Author: Aparna Sinha - Published: 2026-08-20 **Excerpt:** Cohesity CMO Carol Carpenter and Kaggle Founder, Sumble CEO Anthony Goldbloom on Data Enrichment, Go-to-Market Agents, and Using AI at the Bottom of the Funnel [![Marketing's Claude Code Moment - Cohesity & Sumble](/images/episodes/ep6-cohesity-sumble-og.jpg)](/episodes/marketing-ai-with-cohesity-and-sumble/) Carol Carpenter, Chief Marketing Officer of Cohesity, is inundated with AI offerings for marketing. Almost all of them want the same thing from her, and her answer is no. > "The challenge in the market is literally I get 10 solicitations a day. There's a new AI tool every day." > > — Carol Carpenter, [26:34](https://youtu.be/JVO3V2rORhg?t=1594) > "If I have another vendor come to me and say, 'Yeah, you just have to feed your data into our…' No. I don't need 10 different agents… We want to have a way to keep our data proprietary, maybe build a knowledge graph, and then feed all of that into workflow tools." > > — Carol Carpenter, [36:14](https://youtu.be/JVO3V2rORhg?t=2174) ## Carol's team Builds and Buys AI Cohesity has adopted AI top down and bottom up, internally and externally for their customers. In marketing, Carol's team built its own agents for translation and brand compliance for example. They use Copilot, ChatGPT, Claude and Gemini across the org, and outreach.ai on the GTM pipeline. Tokens are their internal currency now. The ROI per project is higher but tangible gains are in the 10-20% range overall. > "Depends on the project, but anything that saves is good so it's 10 to 20%." > > — Carol Carpenter, [13:30](https://youtu.be/JVO3V2rORhg?t=810) The most promising projects are improving lead conversions at the bottom of the funnel, but these require pristine data, domain knowledge and human creativity. A person has to know the ten reasons a customer buys and the moments of truth for a buying committee that spans CIOs, IT ops, CISOs and security ops. > "So now, you look down the funnel where the rubber meets the road, are these converting? It just is requiring a lot more work than we anticipated in terms of training and the setup." > > — Carol Carpenter, [14:10](https://youtu.be/JVO3V2rORhg?t=850) ## Anthony Goldbloom thinks the missing piece is data Anthony founded Kaggle, the world's largest data science community, and sold it to Google in 2017. Kaggle has had tremendous impact on Machine Learning with its famous competitions - Greg Brockman, now president of OpenAI, learned machine learning competing in Kaggle competitions. Kaggle is also a public data platform, which is what sowed the seed for Anthony's latest startup - Sumble. > "My co-founder and I left with this vision that we were gonna start the world's best data vendor, an AI native data vendor. We think that hard grounded facts are going to be a huge part of the LLM ecosystem, and we are starting off with a focus on data for go-to-market teams." > > — Anthony Goldbloom, [04:55](https://youtu.be/JVO3V2rORhg?t=295) Sumble curates a pristine dataset about all the companies in the world, and provides it to Enterprises so their go-to-market agents can have a hard-grounded corpus with deep information about their customers. ## His bet is that this solves Carol's problem Anthony is betting the pipeline agents fail because of what they (don't) know, and that better data can supercharge the creativity of Carol's team. > "We're there to put the creativity into practice based on real robust knowledge of what's happening into the company. You can come up with a very creative strategy. We can help you actually execute on it." > > — Anthony Goldbloom, [21:31](https://youtu.be/JVO3V2rORhg?t=1291) So what's different about Sumble? A conventional tool could tell Carol that her customer prospect uses a competing backup product. Sumble would say 400 teams at Wells Fargo care about backup, 20 use Commvault, 50 use Cohesity, and which of them is in a buying window! > "It's one thing to know the footprint of a company's tech stack, but even better if you can know they're in a buying window. If they're using Commvault and that team is also doing a cloud migration at the moment, their architecture is changing, that is an amazing moment to rip out a legacy player." > > — Anthony Goldbloom, [16:11](https://youtu.be/JVO3V2rORhg?t=971) ## Sumble's Moat: is difficult to build knowledge graph The work is mining public sources and then cleaning them without stopping. > "We clean data all day long. My dream is that one day this will be the most pristine data set. It's quite hard to do." > > — Anthony Goldbloom, [24:25](https://youtu.be/JVO3V2rORhg?t=1465) > "We will never miss a job post. Because we just think job posts are gold. Other crawlers, if I miss 20% of job posts, so be it. We have a rule, we do not miss job posts." > > — Anthony Goldbloom, [19:46](https://youtu.be/JVO3V2rORhg?t=1186) Job titles keep drifting, from SalesOps to RevOps to GTM engineering, and every drift has to be resolved before a fact is usable. ## And they never take the private data Sumble does not ask Carol for her customer records. The clean public corpus goes into the warehouse the enterprise already owns and enriches what is there. > "The approach that Cohesity and others are taking where they put everything go-to-market related into a data warehouse is probably 70% of the value you would get if it was structured as a knowledge graph." > > — Anthony Goldbloom, [29:07](https://youtu.be/JVO3V2rORhg?t=1747) This is an exciting new approach to making agents more effective and more efficient and most importantly amplifying the creativity of your marketers. Carol argues that only her team has the domain expertise and creativity required to achieve the results Cohesity needs, no AI can replace that. But with the constantly updated, deep knowledge graph from Sumble, Marketing may soon be reaching its Claude Code moment. Watch the episode for much more on data, knowledge graphs, agents, and how to build an enterprise grade system for go-to-market engineering. 📖 [Full episode](https://enterprisealignedai.com/episodes/marketing-ai-with-cohesity-and-sumble/) with timestamps, key takeaways, and show notes. --- # Glossary ## Model Context Protocol (MCP) - URL: https://enterprisealignedai.com/playbook/glossary/model-context-protocol **Definition:** An open standard introduced by Anthropic in late 2024 that lets AI agents securely access tools and data across multiple sources. MCP servers wrap underlying APIs and present only the specific tools and operations an agent should be allowed to call. **Importance:** Without MCP, every connection between an agent and a tool is a custom integration, and each AI client needs its own wrapper for each API. MCP standardizes that interface. A tool is wrapped once as an MCP server, and any MCP-compatible client (Claude, ChatGPT, Gemini, Cursor, VS Code, Microsoft Copilot) can call it. This standardization makes downstream patterns like gateways, tool registries, and centralized audit trails possible. ## Designed for AI consumption APIs were designed for humans and developers, with CRUD-based create/update/delete operations that don't map to how a model would call a tool. MCP servers redesign that surface for agent use. Jiquan Ngiam, founder of MintMCP, gave the Notion example in the Coursera episode: the v1 MCP server was a one-to-one wrapper around the existing API; the v2 server was rebuilt from scratch with a different vocabulary designed for AI. In Salesforce, a useful MCP tool exposes a single business-level operation. It wraps multiple SOQL queries, validates the results, and returns a clean response to the model. An underlying API key can read and write records, and there's no clean way to tell it "you can read but you cannot delete." An MCP server exposes only the operations you want (read, search, lookup), while the underlying credentials retain full access on the server side. The boundary is enforced at the abstraction layer. Credentials stay on the server. ## How enterprises are deploying MCP at scale At Coursera, MCP usage grew ad hoc before any central oversight. Mustafa Furniturewala, CTO, described the state that pushed him to adopt a gateway: > "MCP usage was going up. My worry was that this was happening in a very ad hoc way across the company. There were a lot of API tokens everywhere. If someone asked what MCP tools are available, there wasn't a clear answer. I had my own experience of using some AI IDEs to add MCPs and there would be 50 of them and they would not connect." > > — Mustafa Furniturewala, CTO, Coursera He also flagged that blocking MCP entirely backfires: > "When you block certain tools, you see people use patterns that are very unsafe. If you block MCP access, they'll start copying data from one place to another, or creating some temporary data store, which is worse from a security perspective." > > — Mustafa Furniturewala, CTO, Coursera At RBC, Vinh Tran's platform team uses the same gateway pattern with metadata on every tool: each MCP call carries the tool's risk profile, and material actions require a human-in-the-loop plan approval before the agent can proceed. Most enterprises run multiple AI clients at once (Claude for engineering, ChatGPT for research, Cursor for IDE work). Without a gateway, IT installs MCPs in five separate places. With a gateway, IT installs each MCP once and every AI client gets a single click into the same set, with single-pane visibility across all of them. Coursera bundles MCP installations by role. A developer installs the developer package and gets the right MCPs configured. A product manager installs the productivity package. Mustafa reports that more than 25% of Claude usage at Coursera now comes from non-engineering co-work, enabled by giving business users the right MCP tools through the gateway. ## MCP in commerce In e-commerce, MCP is one of four mechanisms brands use to expose product catalogs to LLM platforms, alongside SEO/structured feeds, ACP, and UCP. Deepa Shekhar, Director of E-commerce Technologies at Logitech, sees today's deployment as in-LLM apps (a Target MCP inside ChatGPT, for example), and expects the next phase to be agent-to-agent commerce: LLMs invoking brand-owned agents to retrieve real-time product and inventory information. (Covered in the Logitech episode.) ## Supply-chain risk MCP servers are software packages, which means they inherit the supply-chain risks of their package ecosystems. In March 2026, the popular Python package `litellm` was compromised on PyPI (versions 1.82.7 and 1.82.8 were live for about 40 minutes before quarantine, downloaded over 119,000 times). Any developer with an MCP server that pulled `litellm` in could be attacked simply by opening Cursor. The malicious payload installed a `.pth` file that triggered on every Python interpreter startup. The `mcp-slack` package on PyPI is community-built and not official Slack code, though the naming suggests otherwise. A gateway gives IT a single place to vet which MCP servers are installed across the organization and to remove unofficial ones. ## llms.txt - URL: https://enterprisealignedai.com/playbook/glossary/llms-txt **Definition:** A Markdown file a publisher places at `/llms.txt` on their website to tell LLMs where to find the site's authoritative content. **Importance:** llms.txt acts as a sitemap for LLMs. Without one, the LLM answers from whatever pages it has indexed, which may not be the publisher's canonical pages. ## In commerce In the Logitech episode, Deepa Shekhar, Director of E-commerce Technologies, identified llms.txt as one of the methodologies Logitech's technology team used to make brand content discoverable by LLMs: > "There are methodologies like llms.txt, which serve as a site map on the website, which instructs LLMs that this is the navigation part to find this kind of content." > > — Deepa Shekhar, Director of E-commerce Technologies, Logitech She described it as one element of a broader effort: marketers writing content around the questions customers were asking LLMs, technology teams publishing that content in structured formats LLMs could understand. The combined work contributed to Logitech earning the "Growth Engine" award in Semrush's 2025 AI Visibility Awards for Consumer Electronics. --- # Decision matrices ## Build vs Buy for Enterprise AI - URL: https://enterprisealignedai.com/playbook/decision-matrices/build-vs-buy-enterprise-ai **Problem:** You're an enterprise leader deciding whether a piece of your AI stack should be built in-house or bought from a vendor. ## Start with two questions Danielle Benecke, who founded and leads Baker McKenzie's Applied AI practice, uses two scales to decide what her firm owns and what it rents: > "One way that we think about it is there's a scale from generalizable to custom, there's a scale > from extremely strategic to not so strategic." Plot the capability on both, then apply her rule: | | **Not strategic** | **Strategic** | |---|---|---| | **Generalizable** | Rent. "For the stuff that is super generalized, not particularly strategic or important to your organization, totally fine to rely on a harness that lives outside of your environment." | Rent, but avoid lock-in. Watch the dependency of running on someone else's platform. | | **Custom** | Buy, or build only if it is cheap to maintain. | **Own it.** "For the very strategic stuff that ultimately turns on your expert steering or internal playbooks and internal know-how, you need to be very careful about where that harness lives." | The test is whether the work turns on your expert steering, your internal playbooks, and your know-how. When it does, owning that layer is the point. When it does not, renting is fine. ## You are already running several of these Danielle's description of the layers a large organization ends up with, all at once: 1. **Direct model access** through cloud AI environments. 2. **Productivity-layer harnesses**, with Anthropic's Cowork and the equivalent capability in Microsoft environments as examples. 3. **Domain-specific vendor tooling**, where most large firms hold several tools with different harnesses. 4. **Your own bespoke harnesses.** At Baker McKenzie every service line and every system they design for a client carries its own. Her note on how simple this can start: "The most simple harness is probably a folder of markdown files." Her caution on the portfolio: be agile and "not bet the farm on any one thing." ## The framework | Criterion | Build when... | Buy when... | |---|---|---| | **Speed to value** | You have an internal team that can ship within weeks, AND no vendor currently nails the problem | A vendor already solves it well and you can integrate in days | | **Control & policy** | You're in a regulated industry where policies, patches, and compliance posture must be yours (Vinh's framing at RBC) | The use case is horizontal and the vendor's controls already match your bar | | **Differentiation** | The capability IS your differentiation, or the way you operationalize it is | The capability is generic and customers won't care who built it | | **Data moat** | You have proprietary data competitors can't replicate | The vendor's moat is data aggregated across many customers, something you can't realistically reproduce alone (Carol on Sumble) | | **Team capability** | You have AI builders embedded in the function that owns the problem | You have functional experts but not engineering bandwidth | | **Total cost of ownership** | The build will produce reusable infrastructure (a platform, not a feature) | The buy avoids long-term maintenance you don't want to own | | **Risk tolerance** | Mistakes are tolerable, learning is part of the goal | You need production-grade reliability from day one | ## Examples from the show **Build calls that worked:** - **RBC's Lumina data platform and Agentic Control Plane** (Ep 2). Vinh's argument: in a regulated industry, building gives you "control of your destiny, your policies, your patches, your compliance posture." Open source has matured, AI coding tools give 10-100x engineering leverage, and cloud democratizes services that used to require a large vendor. Building costs more and gives control. - **Logitech's composable commerce architecture** (Ep 3). Deepa's reframing: composable architecture wasn't an AI play. But it's the reason agentic commerce is available to her team today. The platform decisions made to "modernize" the stack now look like AI-readiness decisions. **Baker McKenzie runs both at once** (Ep 5). The firm reaches frontier models directly, buys best-in-class vendor tooling including Legora, and builds its own harnesses per service line. Danielle starts from what clients need rather than from the technology: "Our clients need strategic clarity, decision velocity, and risk mitigation." That need decides which capability gets plugged in where. She also flags what most buyers underweight: "Many players in the industry still think of this as primarily a technology shift. Technology is obviously a big part of it, but the commercial model transformation part is" the larger change. **Buy calls that worked:** - **Cohesity using Sumble for top-of-funnel ABM** (Ep 6, on the recording schedule). Carol's framing: "Same old decision tree of build vs. buy. What do you want to continue to maintain? In this case, a company like Sumble, let somebody else collect all the buying signals. We don't need to." - **Logitech using Typesense for the agentic commerce search layer** (Ep 3). Jason's pitch to a customer who could conceivably build it: "Sure, you could build this with LLMs, but it would take you a long, long time." The proprietary knowledge graph is the moat. ## Warning signs you picked the wrong side - **You're building** but every sprint feels like you're rebuilding something a vendor already does. The build was about pride, not strategy. - **You're buying** but the vendor's roadmap is taking them away from your use case. The moat you bought into is eroding. - **You're building** but the team isn't shipping. The build was a real strategic call, but you don't have the engineering bandwidth to back it. - **You're buying** but the data your vendor's AI is being trained on is your data. The integration is more strategic than you priced it. - **You're renting a harness** that steers work turning on your own playbooks and know-how. By Danielle's test, that one belongs inside your environment. ## Where this came from Every quote above is from an episode of Enterprise Aligned AI. The build-vs-buy discussion with Baker McKenzie and Legora is in [Who Owns the Legal AI Harness?](/episodes/who-owns-the-legal-ai-harness/). Vinh Tran's case for building inside a regulated bank is in [Royal Bank of Canada's $1B AI Transformation](/episodes/rbc-1b-ai-transformation/). Real conversations with enterprise leaders implementing AI. New episodes by email: [enterprisealignedai.com/subscribe](/subscribe) --- # Playbook ## Building AI Agents in Law - URL: https://enterprisealignedai.com/playbook/stanford-llmxlaw-april-2026 - Venue: Stanford Law School A walk through how AI is moving from retrieval-augmented generation to agentic systems in legal practice, with a hands-on demo of a contract clause review agent. # Speakers ## Anthony Goldbloom - Role: CEO and Co-Founder, Sumble - LinkedIn: https://www.linkedin.com/in/anthonygoldbloom/ - Episodes: marketing-ai-with-cohesity-and-sumble Anthony Goldbloom is CEO and co-founder of Sumble, an AI-native data vendor for go-to-market teams. Before Sumble he was co-founder and CEO of Kaggle, the world's largest data science community, which he sold to Google in 2017 and continued to run inside Google until 2022. Kaggle's public dataset platform is what convinced him that models matter less than the data underneath them, and Sumble is built on that premise: a knowledge graph of what companies are doing, assembled so that language models have hard facts to reason over. ## Carol Carpenter - Role: Chief Marketing Officer, Cohesity - LinkedIn: https://www.linkedin.com/in/carolwcarpenter/ - Episodes: marketing-ai-with-cohesity-and-sumble Carol Carpenter is Chief Marketing Officer of Cohesity, the enterprise data security and management company. She has spent her career building category-defining technology brands. Before Cohesity she was Chief Marketing Officer of Unity Software, and before that Chief Marketing Officer of VMware, where her team led the company's pivot to multi-cloud. She has also held senior leadership roles at Google Cloud and Trend Micro. At Cohesity she runs a marketing organization that is at once a heavy user of enterprise AI tools, a builder of custom agents for translation and brand compliance, and a demanding buyer of third-party AI vendors. ## Aman Khan - Role: Head of Product, Arize AI - LinkedIn: https://www.linkedin.com/in/amanberkeley - Episodes: how-wayfair-uses-arize-for-reliable-agentic-ai Aman is Head of Product at Arize AI, an AI Development and Evaluation platform used by companies like Uber, Duolingo, Reddit, Instacart, Booking.com and Wayfair. At Arize, Aman helps teams launch and improve their AI agents and systems. He recently led a popular deeplearning.ai course on Evaluating AI Agents, and has been featured by Lenny's Newsletter to cover AI Product Management and Evals a number of times. Prior to Arize, Aman led products at Spotify, Cruise and Apple. ## Danielle Benecke - Role: Founder and Global Head, Applied AI Practice, Baker McKenzie - LinkedIn: https://www.linkedin.com/in/daniellebenecke/ - Episodes: who-owns-the-legal-ai-harness Danielle Benecke is the founder and global head of Baker McKenzie's Applied AI practice, an award-winning team of lawyers and technologists delivering AI-native services for complex, high-stakes legal and regulatory work. She founded the practice in 2021, before most of today's legal AI companies existed; it now operates across 45 markets, pricing by outcome rather than hours. A technology lawyer by background, Danielle has spent her career advising global tech leaders on AI development, commercialization, and responsible use; her focus is unlocking legal capabilities that were not previously achievable. She is recognized in Chambers Global, Lawdragon Global 100, and Forbes Top Lawyers, has received honors including Law.com Innovator of the Year and Financial Times Changemaker of the Year, holds a Master's from Stanford, and serves as a venture coach with Stanford Law School's CodeX. ## Deepa Shekhar - Role: Director of E-commerce Technologies, Logitech - Episodes: agentic-commerce-with-logitech-and-typesense Deepa is Director of E-commerce Technologies at Logitech, where she leads the transformation of the company's global digital commerce platform. Her shift to composable architecture and high-performance experiences materially improved conversion and scalability on Logitech.com, and contributed to the site being named PCMag Reader's Choice 2025 Best Manufacturer Online Store and Logitech earning the "Growth Engine" award in Semrush's inaugural 2025 AI Visibility Awards for Consumer Electronics, measured across thousands of real prompts in ChatGPT and Google AI Mode. She is now focused on the next wave — agentic commerce — where she has outlined the three tracks reshaping how people discover, evaluate, and buy products. ## Jason Bosco - Role: CEO and Co-founder, Typesense - Episodes: agentic-commerce-with-logitech-and-typesense Jason is the CEO and co-founder of Typesense, the open-source search engine that powers more than 10 billion searches per month on Typesense Cloud, with 25,000+ GitHub stars and 25M+ Docker pulls. Built as a single C++ binary delivering lightning-fast search across large datasets, Typesense is the open-source alternative to Algolia and Pinecone — and is revenue-focused, not VC-backed. Before Typesense, Jason was VP of Engineering at Dollar Shave Club and VP of Technology at Verishop, so he has lived the e-commerce journey from multiple sides. He has a strong point of view on why the search and retrieval layer is the most underestimated part of the agentic commerce stack. ## Jiquan Ngiam - Role: Co-Founder & CEO, Mint MCP - LinkedIn: https://www.linkedin.com/in/jngiam/ - Episodes: coursera-becoming-ai-native Jiquan Ngiam is the co-founder and CEO of Mint MCP, an MCP gateway that gives enterprises secure, governed data access for AI agents. Before Mint MCP, he worked on Google Brain and Waymo, and did his graduate work under Andrew Ng at Stanford, where he contributed to early deep learning research and helped build the first online machine learning courses that became Coursera. His thesis is that as agents become coworkers, enterprises need a security boundary around what they can access, and Mint MCP provides that layer. ## Max Junestrand - Role: CEO and Co-Founder, Legora - LinkedIn: https://www.linkedin.com/in/maxjunestrand/ - Episodes: who-owns-the-legal-ai-harness Max Junestrand is the co-founder and CEO of Legora, the AI-native legal workspace he started in Stockholm in 2023 and which is now used by law firms and in-house legal teams worldwide. Legora's stated mission is to empower exceptional lawyers by building the world's first truly collaborative AI for lawyers, reducing friction and automating repetition so legal teams can focus on the highest-impact work. Before founding the company, Max worked at McKinsey and in venture capital while completing a double degree. He holds a Master of Science in Computer Science with a specialization in Machine Learning from KTH Royal Institute of Technology, a Bachelor of Science in Business Administration from the Stockholm School of Economics, and a Master of Engineering in Computer Science from the University of Illinois Urbana-Champaign. Legora is a Y Combinator alum. ## Victor Sulaiman - Role: Senior Product Manager, Wayfair - LinkedIn: https://www.linkedin.com/in/victor-a-s/ - Episodes: how-wayfair-uses-arize-for-reliable-agentic-ai Victor is a Senior Product Manager at Wayfair, where he leads AI and Agentic workflows for retail and marketplace platforms. He has over a decade of experience driving enterprise AI adoption, from Uber to Meta and even worked on foundational model research, he specializes in building scalable, reliable agent systems that put the user at the center. Victor is passionate about the intersection of product strategy, marketplace economics, and responsible AI deployment. ## Mustafa Furniturewala - Role: Chief Technology Officer, Coursera - LinkedIn: https://www.linkedin.com/in/mustafaf/ - Episodes: coursera-becoming-ai-native Mustafa Furniturewala is the Chief Technology Officer at Coursera, where he has spent more than eleven years and now leads product and technology strategy across engineering, infrastructure, data, security, and IT. He is spearheading Coursera's generative AI strategy, both in the products Coursera builds for learners and in how the engineering team itself works, from agentic code migration to background agents to an MCP gateway for secure data access. Before Coursera, he held engineering roles at Twitter, Evernote, and Citrix. ## Vinh Tran - Role: Vice President, Data & AI Platforms, Royal Bank of Canada - LinkedIn: https://www.linkedin.com/in/vinhbtran/ - Episodes: rbc-1b-ai-transformation Vinh Tran is the Vice President of Data & AI Platforms at Royal Bank of Canada and an RBC Fellow at RBC Borealis. He leads the transformation of Canada's largest bank into an AI-powered enterprise. With nearly 10 years architecting cloud, data, and AI platforms across major Canadian banks, Vinh has built enterprise-scale infrastructure on Azure, AWS, and GCP serving thousands of developers and millions of customers. His work spans from pioneering multi-cloud Kubernetes environments to now spearheading generative AI, autonomous agentic systems, and enterprise data ecosystems, all while maintaining the rigorous security and compliance standards required in financial services.