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

Joseph Kim opens by describing ChatGPT for Financial Services, a new enterprise product designed to combine ChatGPT’s consumer-facing tools with firm-specific enterprise capabilities. It includes templates tailored to individual firms and licensed sources such as Dealogic and S&P. Morgan Stanley and Evercore served as design partners, and Kim says customer reception has been “incredible.” The first half of his keynote covered this product; the second looked ahead to what improving models might do by mid-to-late 2027 and what builders should do now to prepare for that intelligence. Kim frames OpenAI as a “research and deployment company,” deliberately putting research before deployment. The company first asks what new model capabilities are becoming possible, then builds product features and interfaces that make those capabilities usable. He points to Codex adoption as evidence that product design matters: software engineering usage climbed steadily, while finance, legal, and sales usage stayed near zero until the Codex desktop app launched in February, after which adoption jumped to roughly 90%. His conclusion is that model capability alone is insufficient; products must be co-developed around how professionals actually work.

Asked what financial reasoning data looks like and how it affects model training, Kim says financial analysis data is fed upstream into OpenAI’s main training run—the model line he refers to as GPT-6 Astra and GPT-6.1 Astra—rather than being reserved only for the financial-services product. That means improvements from financial data should be accessible to people building startups, businesses, and internal tools on top of OpenAI models, not just to OpenAI’s first-party applications. Kim says this openness is intentional: OpenAI wants its intelligence available both through products it builds internally and through models that outside companies and financial firms can integrate directly. He also explains that ChatGPT for Financial Services was unusual inside OpenAI because the product, model harness, and research work were housed together. That arrangement gave the team flexibility to test how model behavior, product design, and enterprise requirements interacted. The main lesson from the first six months was that enterprise products cannot be designed from first principles alone; they must account for the existing incentives, constraints, competitors, licenses, and workflows of the industry they enter.

Data became a concrete example of the difference between an elegant architecture and a reliable enterprise product. Kim says the team initially faced an obvious objection to bundling earnings-call transcripts and other financial data: customers might already license the information, and an MCP connector for a major provider might already exist in ChatGPT’s store. On paper, bundling seemed redundant and potentially like charging customers twice. The team tested available MCP connectors and found that some prominent providers had error rates above 40%. Kim’s conclusion was that MCP connectors looked attractive in theory but were not reliable enough for a financial product where data accuracy matters. To fix the problem at scale, OpenAI brought the data in-house and controlled the experience end to end, accepting the additional cost of bundling it. Customers have since consistently identified the data as the product’s most valuable feature, which Kim presents as evidence that the industry-specific decision was correct. For builders in legal, sales, or other domains, his broader advice is to inspect real performance rather than assuming an existing integration works because it is technically available.

Kim pushes back on the idea that OpenAI’s success came mainly from secret model modifications. He says 95% of what the team built could have been built by a startup or another company. Most of the product sits on top of ChatGPT’s plugin architecture and primitives, which OpenAI has been investing in so internal teams can build on the base model just as external companies do. His team’s working thesis was that ChatGPT or Codex could get a product roughly 80% of the way to market, but the missing 20% would determine whether the product actually won. Simply launching Codex as a finance product would fail because it would lack that domain-specific layer. The valuable additions might be proprietary data, changes to the user interface, integrations, firm-specific workflows, or other “secret sauce.” Kim goes so far as to predict that a billion-dollar business will be built on top of a ChatGPT plugin by combining OpenAI’s horizontal capabilities with a focused, high-value specialization. He describes his own team as something like a startup building on OpenAI, since even internal groups need scalable ways to “stand on the shoulders of giants.” That makes the platform architecture strategically important both inside and outside the company.

Kim says safety is especially important as agents become more autonomous through 2027 and 2028. The currently launched financial product is primarily a productivity tool for knowledge workers, with a built-in assumption that a human remains in the loop. A financial analyst can ask the system to tackle an item on a daily task list; it may complete 30% or 80% of the work, but the analyst is still expected to review and audit it before relying on the result. For future products that operate with less continuous supervision, Kim wants their security profile to resemble how organizations already manage human coworkers. A new engineer may need “readability” or equivalent permissions before touching production systems, and work is checked through existing organizational controls. In banking, junior analysts’ work is reviewed by associates, then passed through vice presidents, directors, and managing directors. Kim argues that these hierarchical checks already exist to catch human mistakes, and can also help supervise models. He expects future agents to behave more like junior colleagues or interns: they should have scoped read/write permissions, produce auditable work, and fit into established review chains. This is not a claim that models are risk-free; it is a proposal to use familiar institutional controls rather than inventing an entirely separate safety system.

In response to the fear that one AI mistake could end an analyst’s career, Kim first says that premise is too absolute, then considers it as a worst-case assumption. Not all errors have equal consequences, and the risk depends on the employee’s level. A mistake reaching a managing director is more serious because multiple review layers should already have caught it. The most dangerous error is a factually incorrect piece of information that changes the conclusion of an analysis. Kim says models have made major progress on factual reliability. “Hallucination” was a dominant concern in 2022 and 2023 but comes up less often now, and he expects continued improvements plus existing review systems to make a career-ending factual error reaching senior leadership unlikely. The more persistent weakness is taste: a slide can be factually correct but still look sloppy, use poor colors, or fail to meet professional standards. He expects models to improve substantially at these softer, comparative judgments by 2027. He explains that taste can be trained through reward models and human preference data. Humans compare two outputs and identify which is better; for slides, people can make an aesthetic judgment in seconds, often more reliably than a model that spends many reasoning tokens analyzing pixels. Coding provides a precedent: good code is not only code that compiles, but code that is clean, maintainable, and likely to be merged.

Kim says most current AI applications in finance focus on junior-analyst tasks because today’s models are good at concrete instructions and because people dislike tedious work such as building Excel models, assembling slides, copying logos, and positioning elements precisely. The product direction has therefore followed both current model capability and human preferences about which tasks to delegate. He expects OpenAI to move up the banking hierarchy in 2027, from junior analysts to associates and eventually toward managing-director work such as generating investment theses and ideas. That work is harder because the objective is fuzzy: whether an investment idea was good may not be clear for three years, unlike whether a slide follows a template or whether an Excel formula has a reference error. Kim says solving it will require continued base-model scaling, focused research, high-quality data, and feedback from exceptional domain experts—what he calls “P99.9” feedback—but he does not view it as impossible. The hosts ask about benchmarks, and Kim warns that financial evals are exceptionally difficult. An eval is only a proxy for the skill it measures, answer keys are tedious to construct, and flawed grading can make systems reward-hack the test. OpenAI found that some GPT-6 financial evals had barely changed because the tests were saturated or broken by bad answer keys and poorly specified prompts. Kim’s team is building profession-specific evals intended to predict whether a model can automate substantial portions of work such as junior investment banking analysis.