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

Summary

• ChatGPT for Financial Services makes reliable data access a central part of its enterprise offering. Joseph Kim says some big-name MCP connectors showed a “40%-plus error rate,” justifying bundled data despite customers potentially paying twice for existing licenses. Morgan Stanley and Evercore are design partners; Kim says data is customers’ number-one positive callout.
• Kim predicts a billion-dollar ChatGPT-plugin business because domain specialization supplies the decisive last 20% of a product. Codex gets roughly 80% of the way, but an unchanged finance launch “would fail”; nevertheless, “95% of what we built” was something a startup or another company could absolutely do. The “secret sauce” might be data, UI changes or integrations, while financial training enters the main model for outside builders too.
• OpenAI’s internal adoption jump shows how product packaging can unlock existing intelligence. Finance, legal and sales adoption of Codex went from roughly zero to 90% around February’s desktop-app release, while engineering adoption climbed steadily. Kim’s product mandate is to make “model intelligence accessible to as many users as possible.”
• The current product’s safety model depends on humans reviewing its work. It might complete 30% or 80% of a task, with auditing remaining the analyst’s responsibility. As autonomy increases into ’27 and ’28, Kim wants agents’ permissions and supervision “to look like a person,” drawing on existing workplace controls and banking’s review hierarchy.
• Kim expects human taste to become a substantially stronger model capability in 2027. His training mechanism is pairwise human preference: people can judge slides within seconds, while models may reason extensively and still choose incorrectly. Coding supplies the precedent—verifiable correctness combined with judgments about clean design—and financial models have a comparable “feel.”
• Moving into investment ideas creates a harder learning problem because the outcome may take three years to judge. Richard challenges how models will develop intuition that runs against prevailing wisdom; Kim’s answer centers on continued scaling, research focus and P99.9-type expert feedback. He expects strides in 2027, while acknowledging that these objectives become fuzzier as work moves toward associate and MD responsibilities.
• Kim says flawed financial benchmarks can obscure real capability gains. OpenAI stopped tracking GDPval because it was saturated or broken, and flat GPT-5.2 scores on financial evals prompted investigations that uncovered faulty answer keys and prompts. His team wants an investment-banking eval that predicts how much junior-analyst work can actually be automated.

Deep dive

1. The pivotal finance layer is within reach of startups

• Kim describes ChatGPT for Financial Services as an enterprise product combining consumer ChatGPT’s tools with firm-specific templates and sources including Daloop and S&P. Morgan Stanley and Evercore are design partners. Financial training goes upstream into the “Big Run,” the main run that becomes something like 6-Ash or 6-1-Ash, reaching the main model and remaining available to startups and financial institutions building their own products.
• The adoption evidence came from inside OpenAI: engineering’s Codex usage climbed steadily, while finance, legal and sales sat near zero before jumping to 90% around February’s Codex desktop-app release. Kim describes OpenAI as a “research and deployment company,” deliberately putting research first; product development then makes that intelligence accessible to people who otherwise would not use it.
• Kim says this build was unusual: product, model harness and research were all under one roof, giving the team flexibility. His thesis has two halves: Codex’s memory, skills and persistence provide about 80% of a competitive finance product, but launching it unchanged “would fail.” The missing 20% is pivotal. Nevertheless, “95% of what we built was something a startup or another company could absolutely do,” without needing to change the underlying model.
• Most of the product uses ChatGPT’s plugin architecture and the same primitives available to other builders. Kim likens his team to a startup building on OpenAI and predicts “a billion-dollar business built on top of a ChatGPT plugin.” The valuable addition might be data, UI changes or an integration: “that secret sauce is really valuable.”

2. Data reliability and familiar controls determine deployment

• Bundling earnings transcripts looked difficult to defend: customers might already have licenses and MCP connectors, making another charge appear duplicative. Testing changed the decision. Kim says some big-name MCPs had a “40%-plus error rate,” so the team brought data in-house to control the experience end to end. Customers’ number-one positive callout, he says, is now the data.
• Richard presses on increasingly autonomous agents’ financial risks. Kim says the product launched last month assumes “there’s always a human in the loop”: an analyst might get 30% or 80% of a task completed, depending on the work, but retains responsibility for auditing it. He expects autonomy to increase into ’27 and ’28.
• For what comes next, Kim wants the security profile “to look like a person.” Google’s readability requirement before engineers change systems illustrates permission checks; banking supplies an analyst–associate–VP–director–managing-director review chain. Kim argues that “humans make just as many mistakes as models,” making existing systems for checking people’s work a template for supervising agents.

3. Taste becomes a trainable part of financial work

• swyx challenges the fear that “one mistake, my career is over.” Kim rejects the premise, then considers it conditionally: a factual error that changes the conclusion is dangerous, particularly if it reaches an MD after multiple reviews. He feels confident that improving factual reliability, combined with existing checks, will make such propagation “very low.”
• The remaining weakness, in Kim’s framing, often involves judgment: “This slide looks ugly,” although its facts are correct. Formatting, colors and perceived sloppiness still determine usefulness. He expects substantial improvement in 2027 on these “softer objective functions,” where the question is whether another version is better.
• Asked how to train taste, Kim returns to early RLHF: show humans pairs, ask which is better, then train a reward model. OpenAI did this for slides. Humans could choose within seconds, while a grading model might spend many reasoning tokens dissecting pixels and still choose incorrectly. The technical answer is simple, he says; the practical answer is more complicated.
• Coding supplies Kim’s precedent: he estimates it is “80% of the way there,” combining verifiable rewards with judgments about maintainability and clean design. Excel similarly has sums and reference errors to check, but also a model’s “feel”: an experienced professional perturbs inputs and watches the outputs. swyx connects this to Cognition’s Frontier Code and SWE-bench: correctness is not the same as deciding which of several valid PRs you would merge.

4. Investment ideas make the objective harder to define

• Richard’s challenge is that experienced investors’ intuition often runs against prevailing wisdom, potentially enabling a great return. Kim explains today’s junior-analyst focus through two forces: models execute concrete instructions well, and people dislike tedious Excel and slide work, such as precisely placing logos. Looking into 2027, he expects movement into some associate and MD tasks, where objectives become fuzzier.
• An investment thesis might require “three years of waiting” for a clear market verdict. Kim calls improvement difficult, but “not impossible”: it needs wonderful data, P99.9-type expert feedback, continued base-model scaling and research priority. He feels confident about strides in 2027. In the opening clip, he also says pre-training problems constrained scaling and that “In 2026, we’ve been able to solve many of these issues.”

5. Better benchmarks and domain expertise remain bottlenecks

• Kim’s warning on evals is causal: a benchmark is a proxy for a skill; faulty answer keys or grading enable reward hacking and high scores despite poor underlying performance. “Building a good eval is incredibly hard.” His assessment of finance: “we don’t think that there’s a really good eval today.”
• OpenAI no longer tracks or reports GDPval: Kim says it is saturated, with the remaining pieces effectively broken. For GPT-6 Astra, he says OpenAI did not discuss its internal financial evals much. When GPT-5.2 scored little differently on a set of financial evals, investigation found incorrect answer keys, poorly specified prompts and other flaws. He flags researchers’ tendency to investigate disappointing scores more readily than flattering ones.
• The replacement ambition is an investment-banking eval predicting whether a strong score corresponds to automating substantial junior-analyst work. Kim says internal financial evals are coming. His team is also trying to broaden the enterprise product to Pro-plan prosumers: “hopefully we’ll have some cool announcements there.”
• Domain expertise remains an organizational constraint. Kim calls great product management “the main bottleneck” in winning this market, highlighting investment banking, equity research and asset management. Subject-matter experts are needed because finance lacks coding’s built-in advantage: the people training coding models already understood what good code looked like.