OpenAI’s Joseph Kim says most of ChatGPT for Financial Services could be built with standard engineering, while the harder-to-copy value lies in licensed data, interface design and integrations. His account makes a useful case for where specialist AI businesses might win: not necessarily by building a better model, but by making it work for a particular industry.
OpenAI Watch analysis
What happened
Kim, who leads product and engineering for ChatGPT for Financial Services, discussed the product on the Latent Space podcast, according to BigGo Finance. He described 95% of the product as work a startup could do using standard engineering and OpenAI’s available tools. For a general coding model turned into a finance product, he said, the remaining 20% of domain-specific “secret sauce” could be pivotal.
That sauce, in Kim’s account, includes financial data, tailored interfaces and integrations. He said OpenAI brought earnings-call transcripts and other data into the product rather than relying on third-party connectors. Kim said some connectors from major providers had error rates above 40%, while customers named the bundled data as the product’s most-liked feature.
Why it matters
This is a practical argument about where value can sit when companies build on shared AI models. For financial services, a capable model is only part of the product: the data and workflow around it can determine whether professionals trust it enough to use.
The product is currently enterprise-only, though Kim said the team is working towards a Pro-plan version. He also described a longer-term push towards more senior financial work, where judging whether an investment idea is good is much harder than checking whether a spreadsheet formula is correct.
Our read
The most persuasive detail is not the percentage split, which is Kim’s estimate, but the connector example: a polished model still needs reliable data and a workflow people will use. For customers, the useful question is whether a bundled product solves a real problem better than their existing tools, not whether it wears an AI badge.
There is a catch. Kim said there is no good evaluation for financial work today, and that building one is difficult because benchmarks can measure the wrong thing. That makes claims about progress into more senior tasks harder to assess from headline scores alone.
What to watch
- Whether OpenAI brings the product to Pro-plan users, and what access that includes.
- Whether it publishes a usable evaluation for financial work.
- Whether customers continue to prefer bundled data over third-party connectors.
Discussion spark: For specialist AI products, which is the more defensible advantage: access to a powerful model, or the data and workflow built around it?
Sources and evidence
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.