⚡ Building ChatGPT for Finance — Joseph Kim, OpenAI's Head of Financial Services

Why OpenAI's finance product hinged on owning its own data

OpenAI built ChatGPT for Financial Services, with Morgan Stanley and Evercore as design partners, mostly from plugin primitives any company can use. The hard part was adapting to how finance actually works. Kim's lesson is that enterprise products must be built around an industry's realities and competitive forces. That meant owning reliable data and fitting review and evaluation to the standards banks already apply to people.

Takeaways

• Owning the data beat third-party MCP connectors. The team bundled quarterly earnings transcripts even though MCP connectors already existed. Testing several MCPs, including big-name ones, showed error rates of 40% or more. Bringing data in house was the only way to fix that at scale, and customers now value it most.
• Startups could build 95% of this product. Most of ChatGPT for Financial Services runs on ChatGPT's plugin architecture and primitives, open to any builder. Kim says Codex gets a product about 80% of the way to market success, but a team still has to build the final 20% itself.
• Financial evals are largely broken today. OpenAI considers GDPval saturated and no longer reports it. GPT-6 barely moved financial eval scores, and a deep dive found the benchmarks saturated or badly built, with wrong answer keys and vague prompts. Kim's team is building an investment banking eval predicting automatable junior analyst work.
• Agent safety mirrors how banks review juniors. Today a human always reviews the output, which shifts responsibility to users. As agents gain autonomy, OpenAI will shape model security like a new employee's, reusing checks banks apply to analysts, associates, VPs and managing directors. Kim argues humans err as often as models.
• Taste and judgement now lag behind accuracy. Hallucination is much rarer, and the career-ending error is a wrong input that flips a conclusion. Humans still judge slides far better, so OpenAI trains on pairwise slide comparisons. Kim likens finance to coding, where verifiable checks sit beside judgement of clean design.
• Research comes first, with finance data in training. OpenAI calls itself a research and deployment company. Financial training data goes into the main training run, so outside developers benefit too. Internally, Codex adoption among staff outside engineering, including finance and legal, jumped from near zero to about 90% after the February desktop app.

What comes next for finance AI

• OpenAI plans to move from junior analyst tasks to associate and managing director work in 2027. Objectives there are fuzzier, and judging an investment idea can take three years of waiting on the market.