--- title: "Expedia's AI Agents and Accessibility | Expedia SVP and Evinced CEO" episode_number: 7 canonical_url: https://enterprisealignedai.com/episodes/ai-agents-with-expedia-and-evinced md_url: https://enterprisealignedai.com/episodes/ai-agents-with-expedia-and-evinced.md last_updated: 2026-10-30 duration: "42 min" youtube_url: https://www.youtube.com/watch?v=gjYamaZQVhA spotify_url: https://open.spotify.com/episode/6urKBk9zWoaSP8232jA4us apple_url: https://podcasts.apple.com/us/podcast/expedias-ai-agents-and-accessibility-expedia-svp/id1890781384?i=1000798858311 guests: ["Reena Patil", "Navin Thadani"] tags: ["Agentic AI", "Travel", "Accessibility", "Reliability and Security of AI", "Deterministic AI", "Browser Agents"] --- # Ep. 7: Expedia's AI Agents and Accessibility | Expedia SVP and Evinced CEO **Summary:** Host Aparna Sinha talks with Reena Patil, SVP of Product Management at Expedia Group, and Navin Thadani, co-founder and CEO of Evinced, about how Expedia deploys AI agents and why accessible code makes agents work better. Evinced makes software that finds and fixes accessibility defects in websites and mobile apps. Reena explains how travelers now ask detailed questions in plain language, and how a content agent collects the answers from partners. Expedia's onboarding agent cut partner onboarding from up to 14 days to under a day. Expedia lets agents act on their own only inside hard limits, and it checks their work against its own data before it reaches a partner. Navin explains why Claude Code can't make a site accessible on its own. Pages it builds average 45 accessibility issues, and Evinced's deterministic test, fix, retest loop brings that under one. The same approach took a large bank's forecast accessibility spend from $27 million a year to less than $3 million. They also cover the AI trust gap in travel booking, agents that hire humans to pass CAPTCHA, and how accessible sites can make browser agents 30X to 100X cheaper. ## Why listen Reena gives a working model for putting agents into production at a large marketplace: start with one measurable use case, set hard limits, and check the output against your own data. She is specific about what went wrong early and how Expedia now verifies agent output. Navin brings numbers from Evinced's own experiments on Claude-generated code, and explains why a deterministic check has to sit inside the AI coding loop. If you run agents against websites, or ship code that coding agents write, both halves of this conversation apply. ## Key takeaways - Travelers now ask detailed questions in plain language, such as whether the pool is heated and when it closes, and without real answers the AI makes things up - Expedia's content agent tells partners which traveler questions went unanswered, such as 67 queries about a pool - Expedia's autonomous supply agent cut hotel onboarding from up to 14 days to under a day - Expedia uses bounded autonomy: agents cannot cross price floors or recommend outside a threshold, and must check inventory in back-end systems - Expedia cross-checks agent output against its own database and the property's website, and holds back properties it cannot verify - Agent-found properties go into the same leads database as human-found ones, and Expedia's regional sales teams take them from there - Pages Claude or Codex build average 45 accessibility issues, and telling the model to follow WCAG leaves 44 - Evinced runs deterministic tests inside the AI coding loop to bring pages under one issue on average - Evinced fixed 350 accessibility issues for one customer in a little over an hour, at about 20 cents per issue - A browser agent that sends every screenshot to an LLM can spend $3.50 to $4 to book one simple United flight - Accessible sites let browser agents skip the LLM for known controls, which Evinced benchmarks at 30X to 100X better price performance - About 68% of travelers still feel comfortable booking with a trusted brand, which Reena calls the AI trust gap ## Timestamps - 0:00: Welcome and Guests - 1:25: Reena Career Journey - 2:01: Navin Founding Story - 3:16: Why Accessibility Is Hard - 4:22: Expedia Marketplace Overview - 5:34: Agentic AI Use Cases at Expedia - 7:07: Measuring AI ROI - 9:58: Expedia Partner Adoption - 11:25: Automating Accessibility Fixes - 14:25: What Accessibility Issues Look Like - 18:39: Why Claude Code Can't Fix Accessibility - 20:24: Deterministic AI is 45X better: Test, Fix, Retest - 23:22: Unharnessed AI Burns Tokens and Time - 26:24: Bounded Autonomy at Expedia - 27:43: How Expedia Verifies AI Agents - 30:44: Accessible Sites Help AI Agents - 34:55: The AI Trust Gap - 37:00: Fraud and Agents That Hire Humans - 38:04: Booking on Expedia From Claude or Gemini - 40:03: Looking Ahead and Closing ## Clips Short segments published from this episode. Each is citable on its own. - **Expedia's AI Agents Cut Onboarding From 14 Days to Under a Day** - URL: https://www.youtube.com/watch?v=cPT8TTjtMMI - Speaker: Reena Patil - Occurs at: 6:35 - Duration: 0:29 - Claim: Partner onboarding went from 14 days to under a day. - **AI Travel Agents and Swimming Pools** - URL: https://www.youtube.com/watch?v=_VkdNUYevX4 - Speaker: Reena Patil - Occurs at: 5:43 - Duration: 0:45 - Claim: Travelers now ask whether the pool is heated. - **The AI Trust Gap** - URL: https://www.youtube.com/watch?v=06s4mqXS5vA - Speaker: Reena Patil - Occurs at: 35:22 - Duration: 0:25 - Claim: 68% of travelers still book with a trusted brand. - **How to Harness Claude: Test, Fix, Retest** - URL: https://www.youtube.com/watch?v=zqUAIfayBB4 - Speaker: Navin Thadani - Occurs at: 20:23 - Duration: 0:42 - Claim: Test, fix, retest brings pages under one issue. - **Accessible Sites Help AI Agents Too** - URL: https://www.youtube.com/watch?v=T9IH1Kb0fq4 - Speaker: Navin Thadani - Occurs at: 34:44 - Duration: 0:10 - Claim: Accessibility helps AI agents cost less. - **AI Learned to Code From an Inaccessible Web** - URL: https://www.youtube.com/watch?v=dCsOzahqLFU - Speaker: Navin Thadani - Occurs at: 18:51 - Duration: 0:28 - Claim: The average homepage has 70 accessibility issues. - **Claude Can't Make Your Site Accessible** - URL: https://www.youtube.com/watch?v=XoQX1buGeyQ - Speaker: Navin Thadani - Occurs at: 19:38 - Duration: 0:44 - Claim: Telling Claude to follow WCAG leaves 44 issues. - **Expedia Handles It When Your Trip Goes Wrong** - URL: https://www.youtube.com/watch?v=ShW2Q-8nfjo - Speaker: Reena Patil - Duration: 0:21 - Claim: Someone has to handle a cancelled flight at 11pm. - **Expedia's AI Agents Tie Together Its Older ML Models** - URL: https://www.youtube.com/watch?v=r0edNtTLCfA - Speaker: Reena Patil - Occurs at: 8:23 - Duration: 0:34 - Claim: Agents connect models Expedia built years ago. Reena Patil is SVP of Product Management at Expedia Group. Navin Thadani is co-founder and CEO of Evinced, which makes software that finds and fixes accessibility defects. Expedia is an Evinced customer. The conversation covers how Expedia puts agents into production, and why accessible code makes those agents cheaper and more reliable. ## Travelers ask in plain language now Reena opens with a hotel pool: > "You had a simple question: Does the hotel have a pool? Simple structured content, yes or no, we had that, our hotels provided to us. Now, the traveler expectations have changed. So for example, is the pool heated? To what temperature is the pool heated? Is the pool open up to this time?" Expedia found that the AI made things up when it had no answer. So Expedia built a content agent that collects the answers from partners and tells them which traveler questions went unanswered. ## The AI trust gap An Accenture survey this year found that 9 in 10 travelers would use AI to plan a trip. Only 7% would use it to complete the purchase on their own. Reena sees the same gap in Expedia's data: > "It's called the AI trust gap, and our data shows about 68% of the travelers still feel comfortable booking with a trusted brand. ... All of those beginning phases, the decide moment, is won by AI, and the buy moment is still with people who can handle when things go wrong and who can book." Travel is a large purchase that travelers can't see before they buy, and there is no industry standard yet for agents making those purchases. ## Expedia's high-ROI agent use cases On the partner side, an onboarding agent scans ZIP code by ZIP code for properties missing from Expedia's database. That cut hotel onboarding from up to 14 days to under a day. Agents also connect machine learning models that Expedia built before generative AI: > "I have seen that Expedia has had many models, like seven, eight years old, but I think they were not being fully utilized. The power of these models wasn't being harnessed. Now with agentic AI, we are able to stitch it together." Expedia also limits sprawl. Teams were building their own MCP servers with no shared standard, so Expedia set guidelines and a central repository. ## Why Claude Code can't fix accessibility Navin explains that models learned to code from a web that is mostly inaccessible. Evinced tested it: > "We've done an experiment where you can ask Claude or Codex or any one of these models, 'Here's a page, go build it,' and you'll see 45 issues on average. ... 'Oh Claude, go look at WCAG and follow all the guidelines and make sure you make it accessible.' How many issues do you think you're gonna get? It turns out 44." ## A deterministic check inside the coding loop Evinced runs deterministic tests inside the coding loop: > "We test it to find any accessibility defect, then we go to Claude and say, 'Hey, you made these three mistakes for accessibility. Go fix them like this,' in machine-readable language. Claude will go about fixing it, then we test it again and make sure that there are no issues, otherwise we give it back." That brings pages under one issue on average. One customer had 350 issues fixed in a little over an hour, at about 20 cents each. A large bank forecast $27 million a year in accessibility spend and now spends less than $3 million. ## Accessible sites make agents cheaper A browser agent today sends a screenshot to the LLM for every click. Navin puts the cost of booking one United flight this way at $3.50 to $4. Evinced built an agent that reads a site's accessibility traits instead, so it can operate known controls without calling the LLM. > "Some of our initial benchmarking based on the current state of the internet is that we should see a 30X price performance improvement out of the box. ... If you did make changes to your site to make it slightly more accessible, like we fix five, six small little things that we'll tell you to do, you could see up to 100X improvement in price performance." ## Bounded autonomy Reena describes Expedia's approach as bounded autonomy: > "AI agents cannot cross certain price floors. It cannot make recommendations outside a threshold. It has to check across inventory and other systems in the back end." ## How Expedia verifies AI agents Expedia's onboarding agent made mistakes early, when several new properties shared one building address. Expedia now cross-references agent output against its own database and the property's website, and holds off when a hotel has no website. For the most important properties, the agent's listings go into the same leads database that Expedia's sales teams call from. > "We are having a higher bar of accuracy when something comes through an agent rather than a human." ## Agents hire humans Reena on what agents mean for fraud: > "Agents can now hire humans to bypass CAPTCHA. ... It's a fraud nightmare. As we used to say in Google Security, it's an arms race. They just have to win once, and we have to win every single day."