Joseph Kim on Building AI Tools for Financial Work

At AI Engineer New York, Joseph Kim joins swyx and Richard to discuss ChatGPT for Financial Services. Kim describes it as an enterprise product combining the general ChatGPT experience with firm-specific templates and financial data. He says OpenAI worked with design partners Morgan Stanley and Evercore.

Kim also looks ahead to what improvements in models could mean for people building products. His comments about capabilities in 2027 are forecasts, and he encourages builders to consider how their products might work as models change.

Asked about financial reasoning, Kim says training data for financial analysis is incorporated into OpenAI’s main model training rather than reserved for a separate first-party model. He says the aim is to make those capabilities available to firms and startups building on the shared model as well.

He describes OpenAI as a research and deployment company: his team first considers model capabilities and where they are headed, then develops product features that make those capabilities accessible to more people.

Kim cites an internal Codex adoption chart. He says engineering use rose steadily, while finance, legal and sales adoption jumped after a desktop app became available. In his view, the example illustrates why model capabilities and their user interface must be developed together.

Building for an enterprise industry requires more than an argument from first principles, Kim says. Product teams also need to understand existing licenses, industry practices, competitors and the incentives of the organizations they serve.

He gives data as an example. At first, bundling earnings transcripts seemed hard to justify when customers might already have licenses or MCP connectors. Kim says the team tested connectors and found high error rates in some cases, including rates above 40% in its tests; he does not provide enough detail here to generalize that number to all connectors.

Kim says his team chose to bring data into the product to control the experience end to end, even though bundling adds cost. He reports that customers have responded positively to the data component; that is the team’s account rather than an independently measured result presented in the interview.

Asked what other domain builders can learn, Kim estimates that 95% of what his team built could also be built by a startup or another company without changing the underlying model. He says the product relies heavily on the same plugin architecture and primitives available to outside builders.

He describes Codex as providing much of the general functionality, including persistent workflows, memory and skills. His team’s thesis was that a financial product still needed a crucial domain-specific layer: launching a generic tool unchanged would not, in his view, meet the market’s needs.

Kim predicts that a major business could be built on a ChatGPT plugin by adding distinctive data, an interface or an integration to horizontal capabilities. That is his prediction about an opportunity, not evidence that such a business has already been built.

The hosts ask about safety as agents become more autonomous. Kim says the current product is designed as a productivity tool with a human in the loop: an analyst can delegate a task, but remains responsible for reviewing and auditing the resulting work.

Depending on the assignment, he says, the system may complete only part of the work or much more of it. The existing design assumes a person is supervising; that assumption made it easier for the team to explain who is responsible for checking the result.

Kim says the team is also thinking about future systems that may not be monitored continuously. One idea is to make their permissions and behavior understandable through familiar workplace practices, such as the checks placed on a new employee before that person can change a system.

He points to banking’s review chain: junior analysts prepare work, associates review and add judgment, and material passes upward to more senior staff. Kim believes established checks on human work can inform the design of AI-assisted workflows; the analogy itself does not prove the controls are sufficient.

A host asks how analysts should begin using AI when a consequential error can damage a career. Kim argues that mistakes differ in severity and that a wrong fact that changes a conclusion can be especially serious, particularly if it passes through several layers of review.

Kim says he sees improvement in model reliability and expects better models combined with review to reduce the chance that such an error reaches a senior decision-maker. This is his assessment and forecast; the interview provides no quantitative error rate for that claim.

He distinguishes factual accuracy from judgment and taste. A presentation may be factually correct but look sloppy or use an awkward format. He expects models to improve at these comparative, less easily specified qualities, while acknowledging that they matter to the usefulness of financial work.

Asked how to train taste, Kim points to reinforcement learning from human feedback and pairwise comparisons. The technical idea is to have people compare outputs and use their preferences to train a reward model, though he notes that practical implementation is harder than the brief description suggests.

As an example, he says people can compare pairs of slides and quickly choose the better-looking one. In the team’s experience, a grading model can spend considerable effort analyzing visual details and still choose poorly, so human preferences remain useful training information.

Kim draws an analogy with software engineering. Some properties are verifiable, such as whether code compiles or a function produces the expected output; others involve judgments about clean design, maintainability and testing. He argues that both types of feedback matter.

Financial models have a similar mixture, he says. A spreadsheet can be checked for arithmetic or reference errors, but an experienced professional may also vary inputs and judge whether the resulting behavior feels sensible. Kim sees that judgment as analogous to assessing code quality.

The conversation moves to investment ideas. A host notes that strong investment judgments can run against prevailing opinion and asks how an agent could be trained for work where the right answer is hard to establish immediately.

Kim says many current finance applications focus on junior-analyst tasks for two reasons: models can follow concrete instructions relatively well, and people often dislike repetitive work such as formatting slides or building detailed spreadsheets. Product priorities reflect both present capabilities and user preferences.

He expects future systems to move toward tasks performed by more senior professionals. But he says objectives become less clear as the work changes: judging an investment thesis may require years of market outcomes, making training and evaluation harder than checking whether a slide follows a template.

Kim says progress will require research, high-quality data and feedback from experienced experts, along with continued improvement in the base model. His expectation of substantial progress in 2027 is a forecast, not a guarantee of investment performance.

On evaluations, he warns that a benchmark is only a proxy for the capability it seeks to measure. Creating a good answer key and grading method is difficult; errors in either can make a test easy to exploit, allowing a model to score well without being good at the underlying work.

Kim says his team has found limitations in existing financial and broader knowledge-work evaluations. He describes a case where unchanged scores led researchers to investigate and find saturated questions, incorrect answer keys or poorly specified prompts. These are the team’s findings as he reports them.

His team is working on evaluations intended to predict performance on a specific profession’s work, such as junior investment-banking tasks. He also says the finance team is hiring engineers, product managers and subject-matter experts so financial practitioners can work with researchers and product teams.

Kim says the product is enterprise-focused and that the team is trying to make it available to more individual professional users. The hosts close by referring listeners to his keynote; any broader availability remains an intention discussed in the interview.