From Copilot to Colleague: Trustworthy Agents for High-Stakes
summary
Claims a move from helpful assistants to agentic systems that actually produce judgments and end to end work using legacy tools, backed by 4,500 domain experts and 1.5 TB of proprietary content and $200M a year in AI R&D. Describes concrete builds: an AI pipeline that extracts W2/1099 data, maps it into a tax engine with a validation module to produce tax returns, and a legal research agent that fetches hard citations. Notes practical limits: evals are hardest, expert accuracy swings 10%+, agents drift and require rubrics and coordination controls like an autonomy dial and persistent memory
AI-written summary · about
chapters (8)
- 0:00Introduction to agentic AI
- 1:32Context on Thomson Reuters
- 3:13The evolution of AI capability
- 3:59Defining the agentic spectrum
- 7:21Lessons learned in development
- 12:01Demos of agentic applications
- 16:08Summary of key takeaways
- 17:34Q&A session
related talks
27mLarge-scale Enterprise Co-pilots - Making AI Useful for Business
21mAI Copilot - Enterprise-wide and Heterogeneous
26mFrom autocomplete to agents: AI coding assistance state of play
26mAI Agents as Financial Co-Pilots | AI-Driven Strategy & Decision Making
41mGoing beyond copilot with AI Agents
29mBuilding Trust in AI Agents
26mAI Agents as Your Co-Developer
44mPannel Discussion | Agentic Automation: Are AI Agents The Future of Digital Work