https://video.ut0pia.org/videos/watch/1f049aa3-ba4c-41ad-b910-eed082b6d452 Most vertical SaaS teams are doing the same things: chasing higher accuracy, building better model harnesses, shipping more features. And their customers are saying the same things: the AI got this wrong, it hallucinated, the accuracy is not good enough. So teams go back and push the numbers higher. We did the same at Filed. We built AI data entry for tax firms and hit 80%+ accuracy against an industry baseline of 50-60%. Many users still complained. Same model, same stack, different outcomes. So we dug in. The unhappy customers were not experiencing worse AI. They were reverse-engineering everything we produced. We had not removed work from their day. We had just changed its shape. Chat interfaces and citation trails feel like the fix. They are not. They hand the verification burden back to the user with extra steps. Accuracy %s are the score you get after the game is already over. The complaints, the hallucination reports, all of it: symptoms of the same underlying problem. Users are still holding the bag, and when they are, every error is catastrophic. When we started building the real fix, we realised the coding world had already been here. Early coding AI dumped a full function and asked engineers to review 200 lines. Same problem. The fix was not a better model. It was Copilot in the editor, not a separate tab. The planner pattern instead of dumping full outputs. Skills and memory that compound with every use. We reached the same conclusion independently, from taxes. This talk is those three patterns and what they look like in a vertical SaaS product. Go where the work is. Most users will try a new feature. Almost none will adopt a new platform. AI has to live inside existing workflows, not alongside them. 1000 feet first. The right unit of work matters more than accuracy on any given unit. Start at the macro level, let users orient, then drill down. Each level is small enough to verify fast. Users stop auditing and start deciding. Skills over models. Every edge case is a skill waiting to be encoded, not a model failure. Turn real usage into institutional knowledge that makes every future user better off. The specific lessons are from taxes. The pattern is universal. Speakers: Atul Ramachandran (Filed Inc): Atul has cofounded multiple startups and is currently CTO of Filed, which has raised over $17M to build AI infrastructure for tax firms. He is an active open source contributor in the JavaScript ecosystem, with projects like NodeGui. He is currently based out of Stockholm, Sweden. X/Twitter: https://x.com/a7ulr LinkedIn: https://www.linkedin.com/in/atulanand94/ GitHub: https://github.com/a7ul Sat, 11 Jul 2026 20:57:20 GMT https://validator.w3.org/feed/docs/rss2.html PeerTube - https://peerate.fr https://peerate.fr/lazy-static/avatars/d7b83a50-fe89-4e9f-8234-2a5bb1c39ab5.jpg https://video.ut0pia.org/videos/watch/1f049aa3-ba4c-41ad-b910-eed082b6d452 All rights reserved, unless otherwise specified in the terms specified at https://peerate.fr/about and potential licenses granted by each content's rightholder.