Stanford Bootcamp · April 2026
Building AI Agents in Law
Aparna Sinha
Host · EnterpriseAlignedAI.com
What We'll Cover
- 01 · The AI Evolution — models, RAG, agents, where we are in 2026
- 02 · Legal Impact — how AI is reshaping legal work today
- 03 · Architecture of Agents — the enterprise tech stack, Skills, MCP
- 04 · Live Demo — Claude → Cowork → Code → Integrate → Build → Deploy
- 05 · Conclusions — what this means for law students, business students, and engineers
Key insight
Whether you write code or not — what I'm going to show you matters for how you work.
Section 01
The AI Evolution
AI models, agents, and where we are today
The AI Evolution: 2022 → 2026
Key Trends Shaping 2026
Technical
- Scaling laws still work, but... new scaling = post-training + inference time
- Specialized models — long thinking, research, coding, role play, tool calling
- Agent Harness Engineering — sandbox, memories, secure tool use, context and token efficiency
- Multimodality — text, vision, audio, docs in one context
Industry
- Open models proliferating — more in 2026 than 2025
- AI as collaborator — partner, not a tool you query
- The governance gap — tech outpaces organizations
- Cost inversion — inference is the real cost now
Source: Lex Fridman #490 — Nathan Lambert & Sebastian Raschka, Feb 2026
My Take
Governance, training, and workflow integration are.
We're in the electricity moment — everyone knows AI matters, most orgs still haven't rewired how they work.
Section 02
Legal Impact
How AI is reshaping legal work today
AI's Impact on Legal Work
2024–2026
or exploring AI
market by 2030
Key players
- Harvey — $11B · $1B raised · Sequoia + GIC
- Legora — $5.55B · $816M raised · Accel
- EvenUp — $2B+ · $385M raised · Bessemer
- Eve — $1B · $164M raised · Spark Capital
- Spellbook — $350M · $80M+ raised · Khosla
Thesis shift: "AI for legal research" → AI as operating system for legal work
The incumbent counterpunch · Thomson Reuters
- ~$11B AI capital capacity through 2028
- $200M+/yr productized AI spend
- Proprietary legal LLM — claims to beat GPT-5 & Opus 4.5
- CoCounsel at 1M users · Casetext acquired for $650M
Funding: Crunchbase, PitchBook · Market: MarketsandMarkets, Lucintel · Survey: ILTA 2025 · Company data: press releases (Apr 2026)
What's Actually Getting Automated
| Task | Time Reduction | Status |
|---|---|---|
| Contract review & due diligence | 30–80% | High-volume ceiling up to 80%; ~30% firm-wide on mixed workload (A&O / Harvey) |
| Legal research | 30–50% | Still needs human verification |
| Initial drafting | ~30% | First drafts only; peer-reviewed RCT (Choi et al. 2024) |
| Compliance monitoring | Ongoing | Only ~13% of orgs deployed today (KPMG 2025) |
| Document classification | 50–80% | Intake, categorization, routing |
Sources: Choi et al. 2024 (Minn. L. Rev.); Thomson Reuters Future of Professionals 2025; Harvey AI; Everlaw Am Law 100 case study; KPMG 2025.
The billable hour paradox
AI makes tasks faster → less billable time → but remaining human time is more valuable. Value-based pricing captures this. Hourly billing doesn't.
The Autonomy Spectrum
| Level | How It Works | Legal Example | Risk |
|---|---|---|---|
| Scripted | Predefined steps, constrained tools | Clause review against checklist | Low |
| Guided | AI chooses strategy, human approves | Research agent proposes, lawyer confirms | Medium |
| Autonomous | AI plans and executes independently | Full due diligence from data room | High |
The reliability math
95% reliable per step × 10 steps = ~60% end-to-end. This is why human-in-the-loop matters.
Sweet spot today
Guided agents for enterprise legal work. Enough autonomy to be useful, enough checkpoints to be governed.
The Claude-Native Law Firm
Zack Shapiro's Thesis
- Solo firm, ~200 startups at half BigLaw rates
- 2,000-word prompts — treats Claude like a senior associate
- Skills = "your judgment encoded, scaled like software"
- General-purpose AI beats specialized tools
- 7.5M views — hit a nerve
The Debate
- Supporters: Practical roadmap for solo/small firms
- Critics: Only works for low-complexity work
- U.S. v. Heppner (Rakoff, J., SDNY, Feb 2026): Defendant used consumer Claude without counsel's direction — not privileged, not work product. Enterprise, attorney-directed use analyzed differently.
Section 03
Architecture of Agents
Building agentic AI for legal work
Agent Tech Stack
What We Built at Vercel
- AI Gateway — centralized routing, model selection, cost tracking
- AI Sandbox — safe isolated runtimes for AI-generated code
Key insight
The "magic" isn't just the model — it's the infrastructure around it.
Agents, Skills & the Evolution
MCP
Connect agents to the world
- Databases, APIs, doc stores
- Westlaw, Lexis, DMS
- Standardized protocol
Skills / Plugins
Reusable encoded expertise
- "How a partner thinks about M&A diligence"
- Lightweight vs. fine-tuning
- One expert's judgment → scales to the firm
Agents
Autonomous loops with model tool calls
- Plan and execute tasks
- Use tools: search, code, APIs
- Iterate and recover from errors
Section 04
Live Demo
Question → Analyze → Build → Integrate → Deploy
Question → Analyze → Build → Integrate → Deploy
| Step | Mode | What You'll See | Tool |
|---|---|---|---|
| 1 | Chat | Ask a legal question, get an instant answer | Claude.ai |
| 2 | Cowork | Contract review — GREEN / YELLOW / RED flags | /review-contract |
| 3 | Code | Run a custom contract review agent | Anthropic SDK |
| 4 | Integrate | Draft an email, check calendar | Google Workspace MCP |
| 5 | Deploy | Push a live dashboard to Cloudflare | Frontend Skill + Wrangler |
From question to deployed app in 6 minutes
Claude writes the code · You provide the judgmentStep 2 · Cowork — The Legal Plugin
The Claude Legal Plugin
- /review-contract — clause-by-clause review
- /triage-nda — NDA pre-screening
- /vendor-check — compliance tracking
- Plus /brief, /respond, more
Market signal
Launched Feb 3, 2026. RELX (LexisNexis parent) had its steepest single-day drop since 1988. Anthropic moved from model supplier to workflow owner.
How the Agent Works
Agent Tools
- read_contract — load the contract text
- read_playbook — load acceptable terms
- flag_clause — flag a risky clause
- generate_report — produce the final summary
Key insight
Claude decides which tools to call and in what order. You define the tools and the judgment criteria.
Step 4 · Connect to Your Workflow
The Integration Story
Same Claude. Same context. Now operating across calendar, email, documents — through MCP.
Claude reads my calendar…
Found · Google Calendar MCP
TechVentures × CodeCraft
Contract Review Call
Today · 4:00 – 4:30 PM PT
Counsel: [email protected]
Discuss outstanding redlines on the software development agreement — focus on IP ownership, liability cap, and indemnification.
…and drafts the email
Gmail · Draft
Subject: TechVentures × CodeCraft — Top 3 redlines for 4pm call
Hi counsel,
Ahead of this afternoon's call, the three issues we need to resolve in the SDA:
1. §3.1 — IP Ownership. Developer retains all work product despite Client's $450K investment. We need work-for-hire with full assignment to Client.
2. §7.2 — Indemnification. Client cannot indemnify Developer for Developer's own negligence. This clause must be deleted and replaced with mutual indemnification.
3. §6.1 — Liability Cap. A 3-month-fees cap is too low for a mission-critical ERP build. We need total fees paid, with carve-outs for IP infringement.
Talk soon,
Aparna
Safety
Draft saved. Not sent. Human-in-the-loop before anything leaves your inbox.
Step 5 · Deploy to the World
Contract Review · Live Dashboard
TechVentures × CodeCraft
Software Development Agreement · $450K
Verdict
Very High Risk
$450K
Value
8
Issues
4
Critical
3
High
The Frontend Design Skill
- 277K+ installs — the most popular Claude skill
- Bold aesthetics, distinctive typography
- Production-grade UIs, not generic templates
- Auto-activates in Claude Code for frontend tasks
Live now · built with the Frontend Design Skill
main.techventures-review.pages.dev →Section 05
Conclusions
Law students · Business students · Engineers
What This Means For You
Law Students
- Most valuable skill: directing AI, not competing with it
- Encode your judgment — it becomes leverage
- AI fluency is now a core competency
Business Students
- The billing model transition is the business opportunity
- Vertical AI wins — go deep, not wide
- Governance is a product, not a policy
Engineers
- Legal AI: massive market, unsolved infra problems
- Agent orchestration + evaluation are open frontiers
- Model + gateway + RAG + agents + review
That's what's possible.
Now it's your turn.
What's next today
- Next up: Deep dive on Claude — hands-on workshop
- Then: You build — today and tomorrow
- Goal: Hackathon — build something real
Q&A — 15 minutes
References & Further Reading
Papers & research
- Choi, Monahan & Schwarcz, "Lawyering in the Age of AI" — Minn. L. Rev. 109 (Nov 2024); peer-reviewed RCT, 12–32% speed gains
- Generative AI in the Legal Landscape — Stanford Law (2024 & 2026 updates, co-authored)
- Thomson Reuters Institute — 2025 Future of Professionals Report (n=2,275)
- LexisNexis — 2025 Future of Work Report (n>1,800)
- Clio — 2025 Legal Trends Report
- McKinsey — "Legal Innovation and Generative AI"
- Goldman Sachs — "How will AI affect the global workforce?" (2025)
- KPMG — How AI is Poised to Reshape Compliance Functions (Jul 2025)
- White & Case — 2025 Global Compliance Risk Benchmarking Survey
- ILTA 2025 Technology Survey — law firm AI adoption
- MarketsandMarkets & Lucintel — legal AI market forecasts through 2030
- Wolters Kluwer Future Ready Lawyer 2026
Cases & legal analysis
- U.S. v. Heppner, S.D.N.Y. (Rakoff, J., Feb 10, 2026) — AI privilege & work product
- Heppner client alerts: Gibson Dunn, Debevoise, Akin, McDermott, Proskauer (Feb–Mar 2026)
- Harvard Law Review — United States v. Heppner note (Mar 2026)
Company press & primary sources
- Harvey — harvey.ai/blog (methodology: "How Harvey Saves Lawyers Time"; "Contract Intelligence Benchmark"); CNBC Mar 2026 ($11B)
- Everlaw — Am Law 100 doc-review case study (50–67% reduction, 90%+ accuracy)
- Legora — legora.com, TechCrunch Mar 2026 ($5.55B Series D)
- EvenUp — Fortune Oct 2025 ($2B Series E)
- Eve — eve.legal/press (Sept 2025, $1B)
- Spellbook — Law.com (Oct 2025, $350M)
- Rains LLP — rains.law (Claude-native boutique)
- Thomson Reuters / CoCounsel — LawSites "1M users" (Feb 2026)
Podcasts & commentary
- Lex Fridman #490 — Nathan Lambert & Sebastian Raschka (Feb 2026)
- Zack Shapiro — "Claude-native law firm" X thread (~7.7M views)
- agents.law — "Inside the Claude-Native Law Firm" (Matt Pollins)
- LawDroid Manifesto — "The Claude-Native Lawyer: Zack Shapiro" (Tom Martin)
- Artificial Lawyer · LawSites (Ambrogi) — ongoing coverage
- EnterpriseAligned AI podcast — enterprisealignedai.com
Funding, market & technical docs
- Crunchbase News & PitchBook — funding rounds & valuations
- Fortune, Bloomberg, CNBC, TechCrunch — startup coverage
- Claude Agent SDK — docs.anthropic.com
- Model Context Protocol — modelcontextprotocol.io
- Claude Skills & Legal Plugin documentation
Appendix
Appendix
Additional reference material