How OpenAI Built ChatGPT for Bankers and Analysts — Joseph Kim, OpenAI

Software engineers were already using tools such as Codex to change their daily work, while many financial professionals had yet to feel a comparable shift. Joseph Kim, OpenAI’s lead for ChatGPT for Financial Services, starts with that discrepancy: model intelligence was improving quickly, but turning it into useful financial work still required products that fit people’s daily responsibilities.

ChatGPT for Financial Services, launched a month before the talk, is a tailored version of ChatGPT Work. Its initial focus is investment banking and equity research at large financial institutions. OpenAI developed it with design partners Morgan Stanley and Evercore, concentrating on the requirements that surround an analyst’s work as well as the model’s ability to answer questions.

The initial design addresses three practical needs:

• Financial sources: Give people access to the information their work depends on, directly within ChatGPT. The quality of an analysis is limited by the quality of its sources.

• An interface for investigation and review: Let analysts inspect source material, explore financial charts and control which sources inform an analysis.

• The firm’s format: Help generated models and presentations follow the organization’s formats and standards through firm-specific templates. A deliverable has to work for the team that will review and share it.

The distance between what a model can do and the value people obtain from it is the capability overhang. A model can improve substantially while a professional’s working day barely changes. Closing that gap requires usable access to data and tools, together with an experience that fits the work. This gives product design a direct role in making existing intelligence useful.

An internal Codex usage chart addresses a concrete question: when did colleagues outside engineering begin putting the tool to work? The chart tracks the share of AI output generated in Codex across job functions. Engineers started early and their share grew steadily. For finance, legal and recruiting, almost none of their AI output came from Codex at the beginning of 2026; by June, the share was around 90%. The pronounced inflection came around February and March.

Kim initially went looking for a model release to explain that turn. What he found was the February release of the Codex desktop app. His interpretation is that a more accessible product helped people outside engineering use capabilities already available to them. The timing alone does not isolate the app’s causal effect, but it supplies a concrete reason to take interface changes seriously. The finance product pursues the same opportunity: make capable models easier to put to work in a particular profession.

The data foundation had to cover sources professionals already rely on: financial statements, earnings transcripts, market data and private company data. Existing MCP connectors offered a way to reach that information, but many of the connectors the team examined had problems with latency, errors or search quality. A model’s ability to call a tool did not resolve those problems in the information it received.

OpenAI licensed much of the data directly and built its own hosting and search indices using the same infrastructure as its web index. MCP remained the protocol through which the model accessed the data. The change was ownership of the infrastructure behind that access. Taking on licensing, hosting and indexing gave the team more control over search quality and the reliability of the experience, while also making it responsible for those parts of the product.

That investment also creates an option for post-training models to work with the datasets. MCP and skills provide access and instructions; post-training can develop familiarity with the data itself. Kim’s analogy is a junior analyst with an instruction manual compared with a senior analyst who has spent more than two decades learning a dataset’s quirks. The useful distinction is between being told how to reach and use information and having learned its peculiarities through training.

Financial work often ends in an Excel model, a PowerPoint presentation or a written analysis that other people must review and share. The template system makes the firm’s approved templates available when generating spreadsheets and presentations. Its intended saving is the preparation work between obtaining an output and making it ready for a team or client.

Interactive financial charts support the investigation that happens along the way. Analysts can explore stock prices and financial metrics directly inside ChatGPT while working through an analysis. The interface therefore supports examining the information as well as producing a final document from it.

For a financial figure, the review question is concrete: where did this number come from, and what context surrounds it? Citation viewers let people inspect the source passages and tables behind an answer. That connects the generated analysis to material the analyst can examine.

The connection makes source review a normal part of working through an answer. The reader can examine whether the cited material supports the figure and its interpretation. Much of this experience uses the same ChatGPT plugin platform offered to developers, with different components serving different jobs:

• Data access: MCP tools and connectors give the model the financial information it needs.

• Source inspection and charts: Citation metadata and MCP apps support inspecting sources. MCP apps also power the interactive financial charts.

• Workflows and deliverables: Skills work with shared artifact tools to guide financial workflows and produce Excel models and PowerPoint presentations.

• Onboarding: The product uses the platform’s native plugin onboarding experience for financial onboarding.

For builders, the valuable contribution can be unique data, an existing tool or a deep understanding of a particular workflow. The platform supplies building blocks around which that expertise can become a usable product. Kim is bullish enough on this approach to predict multiple billion-dollar businesses built as ChatGPT plugins, with the opportunity expanding as models become more capable.

The forward-looking half of the talk begins with three overlapping drivers of model improvement: scaling compute, improving algorithms and unhobbling, the changes that let people get more useful work out of models. Looking backward at those drivers is Kim’s way of building an intuition for what might become possible next.

Pre-training scaled capability by making models larger and adding more data. Before GPT-4 finished training, OpenAI researchers could predict its final loss and its performance on a coding evaluation. Algorithmic improvements supplied another kind of gain: Kim cites a study in which the compute required to train a language model from scratch to a fixed performance level fell by roughly half every eight months. That is an efficiency trend at a given target, even though the timing of individual breakthroughs remains hard to predict.

Reasoning added another route in 2024 and 2025. Reinforcement learning and post-training taught models to work through problems more effectively, while additional compute at answer time gave them more opportunity to think. These changes operated alongside pre-training. Kim also reports that pre-training problems had made scaling difficult at OpenAI, and that resolving many of those problems in 2026 enabled larger, more capable models.

Potential barriers had prompted concern that this progress might stop. The feared data wall was a shortage of human-generated training text. In Kim’s assessment, synthetic data generation, learning from verifiable rewards and learning more from each token had largely addressed that constraint. He similarly points to sparse attention and compaction as answers to context-length constraints. His expectation is continued progress toward larger, more efficient and more capable models in 2027, while alignment and safety remain difficult problems.

Mathematics supplies the concrete trajectory: GPT-4 scored 700 out of 800 on SAT math in 2023, followed later by an OpenAI model reaching gold-medal-level performance at the International Mathematical Olympiad. Kim then reports internal research work: disproving a long-standing conjecture connected to a problem studied since 1946, proposing a solution to the Navier–Stokes Millennium Prize Problem, and sharing 719 mathematical research manuscripts across 372 families of results. These later examples are research claims; a proposed solution and a collection of manuscripts do not by themselves establish accepted mathematical results.

The extrapolation is conditional on another year of progress roughly in line with the recent pace. Under that condition, Kim expects models by the end of 2027 to outperform leading human experts on a growing range of difficult tasks across fields and specialties. Mathematics is his example of how a capability can move from familiar tests toward work at a field’s frontier.

A gradual capability curve can produce a sudden change in working habits. Kim calls this the threshold of usefulness: the point at which a model becomes capable and reliable enough to save time after the user has checked and corrected its work. The threshold belongs to a particular task and depends on how much review that task requires.

What changes on either side of the threshold is the practical value of the same task:

Several similar increments in capability can therefore feel very different to the user. Early improvements reduce errors without changing the decision to do the work manually. A later improvement makes the output usable with a manageable review cost, and the decision changes. The underlying gain need not be unusually large for the change in behavior to feel dramatic. Different tasks have different thresholds, so these transitions will arrive at different times.

This creates product work that can begin before the next capability gain arrives. Identify a task whose model performance is easy to underestimate because it is almost useful, then consider the experience needed once delegation becomes worthwhile. How should the interface change? What should the user still control, review or approve? Having those decisions worked out helps a new capability become useful as soon as it crosses the threshold.

Financial tasks have two routes toward that threshold. Kim describes them as a rising tide and the waves on top:

• The rising tide: Larger, more capable general models bring transferable skills from mathematics, coding and reasoning into finance. Historically, OpenAI had devoted relatively little training time or compute specifically to finance, so much of the financial capability came from those broader improvements.

• The waves: Finance-specific post-training builds on the general model and concentrates learning on the work financial professionals do. It can create stronger performance on particular tasks above the improving general baseline.

At the time of the talk, the team’s next three months of work focused on general financial reasoning and the quality of common analyst outputs: analyzing financial information, building Excel models and creating PowerPoint presentations. These tasks occupy a significant part of a junior analyst’s day. The training objective is work that is accurate, well structured and useful within an actual workflow.

The longer-term target moves into the judgment of senior bankers, analysts and investors. That work involves taking in large amounts of information, deciding what matters and making forward-looking assessments: forecasting a company’s earnings, assessing where commodity prices may go or evaluating which businesses make attractive investments. A model has to select the important information and use it to reason about an uncertain future.

Kim’s stated goal is to match the best human experts on these financial tasks sometime in 2027. He expects both forces to contribute: continued improvement in the underlying model and more direct training on financial work. The broad model supplies transferable capability; the specialized training concentrates it on the tasks the product needs to perform.

Unhobbling returns as the third driver of useful progress. Earlier examples include RLHF helping models follow instructions and hold conversations, and tools letting them search for information, write and run code, and interact with software. Each change gives model capability a more useful way to operate in the world.

The next form Kim expects is a proactive, persistent agent. A persistent agent would retain a goal, remember relevant context and continue working in the background after receiving direction. Progress would no longer depend on the user being present to supply the next prompt. The assignment could remain active while the person turns to other work.

Persistence also changes how people cooperate with the agent. Getting useful work from today’s agents often requires learning how to prompt them and operate their tools. With continuity across an assignment, the interaction can become more familiar: give the agent a project, share context and ask for an update. The accessibility benefit comes from organizing the relationship around ongoing work, much as people coordinate with colleagues.

Kim closes with three explicit predictions for 2027:

• Less reliance on Excel and PowerPoint to make the work: People will rarely, if ever, rely on those applications to create financial models or pitch decks.

• AI analysts that behave more like coworkers: They will remember preferences and context from previous days, take ownership of assignments, proactively follow up and keep working while the person focuses elsewhere.

• Better forward-looking judgments than leading human experts: Models will begin to outperform the best people on critical financial tasks, including forecasting earnings and developing investment ideas.

The product already aims to produce Excel models and presentations in a firm’s format. These bets extend that direction into how assignments are managed and the quality of the judgments inside the deliverables. The design question is what a user will still need to control, review and approve when delegation becomes worthwhile—and what builders should start making now so that experience is ready.