Chat and citations won't save your vertical AI
summary
Filed builds AI data entry for tax firms and hits 80%+ vs 50–60% baseline yet users still audit every result; the team discovers chat and citation UIs push verification back onto people. The fix is concrete: implement long-running background agents with a planner pattern, automatic skill capture so edge cases become encoded, full traces plus pause/resume control, and measure success by sessions executed by agents rather than raw accuracy percentages.
AI-written summary · about
chapters (12)
- 0:00Vertical AI challenges
- 1:24Citations and chat limits
- 3:07Product evolution layers
- 5:13Agentic conveyor belts
- 6:35Four delegation features
- 7:16Delegatable tasks
- 8:22User-specific agent skills
- 9:46Monitoring agent tasks
- 10:50User control and trust
- 12:12Retaining trust features
- 13:17Value in weekly sessions
- 14:41Vertical AI takeaways
related talks
26mFrom autocomplete to agents: AI coding assistance state of play
49mHow to be an AI First company
46mIntelligence as Infrastructure: What Developers Must Learn to Build with AI
26mAI, You complete me
39mHow Chatty Leads Builds AI Agents Businesses Actually Want
20mFrom Copilot to Colleague: Trustworthy Agents for High-Stakes
44mPannel Discussion | Agentic Automation: Are AI Agents The Future of Digital Work
46mRecalibrate - How AI shakes up software team