Building ChatGPT for Finance — Joseph Kim, OpenAI's Head of Financial Services
Joseph Kim argues that building finance AI means building around industry realities, shipping on plugin primitives, training for human taste, moving up the analyst hierarchy by 2027, and fixing broken evals.
Watch on YouTube ↗Charlie's summary
- ▶ 2:30
Research comes first, deployment second: Codex adoption in finance, legal and sales jumped from near zero to 90% after the desktop app
Joseph says OpenAI calls itself a “research and deployment company”, in that order. Financial training data goes into the main big run, so every builder can use it. Codex adoption among finance, legal and sales staff sat near zero, then jumped to 90% when the Codex desktop app shipped in February. Model and product have to be developed together.
- ▶ 4:34
In enterprise, build around the industry's reality: bundled data beat MCP connections with 40%+ error rates
Joseph says first-principles logic can mislead you. Bundling earnings transcripts looked hard to defend when MCP connections already exist and customers may already hold licenses. But the team tested the MCPs and found some big names had “40% plus error rate”. So they brought the data in-house to control the experience end to end. Customers now name the data as the thing they like most.
- ▶ 7:04
95% of the product could be built by a startup, and a billion-dollar business will be built on a ChatGPT plugin
Joseph says most of ChatGPT for financial services used the same plugin architecture and primitives open to any builder, not model changes. His team's thesis was that Codex gets you about 80% of the way, but launched as a finance product unchanged it would fail, so the missing 20% is pivotal. That 20% can be data, UI changes or an integration. He believes there is a billion dollars of value in that secret sauce.
- ▶ 10:39
Make agent safety look like working with people: human review, read-write permissions and the banking hierarchy
The current product assumes a human in the loop, so auditing stays with the analyst. For more autonomous agents, Joseph wants the security profile to resemble how you work with coworkers and interns. Humans already have checks like Google's readability before touching systems. Banking has junior analysts, associates, VPs and MDs checking work, and he notes humans “make just as many mistakes as models”. Those systems can be reused.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.