ChatGPT for Financial Services combines licensed financial data, finance-specific interfaces, and firm templates for investment banking and equity research.
2
The capability overhang exists because model progress has moved faster than the value people get from models in daily work.
3
Joseph Kim expects persistent agents and finance-specific training to push more financial tasks past the threshold where delegation saves time.
Summary
Joseph Kim describes how OpenAI built ChatGPT for Financial Services with design partners Morgan Stanley and Evercore. The product targets investment banking and equity research, combining licensed financial data, custom search indices, citation viewers, interactive charts, and firm-approved templates for spreadsheets and presentations. Kim argues that capable models alone do not change how people work. Products must connect models to reliable data and familiar workflows. He calls the gap between model capability and practical workplace value the capability overhang. The talk then turns to future model progress, including scaling, algorithmic improvements, reasoning, and what Kim calls unhobbling. He explains the threshold of usefulness, when a model becomes reliable enough that reviewing its output takes less time than doing the task directly. For finance, OpenAI plans more finance-specific post-training and expects persistent agents to work in the background. Kim's 2027 bets include AI analysts that remember context, take ownership, and outperform leading experts on some financial judgments.
ChatGPT for Financial Services is built around the daily work of investment bankers and equity researchers
Joseph Kim introduces ChatGPT for Financial Services as a tailored version of ChatGPT for financial professionals, beginning with investment banking and equity research. OpenAI built it with design partners Morgan Stanley and Evercore. The product focuses on three parts of financial work: access to financial data, an interface that lets users inspect sources and explore charts, and firm-specific formats. Its outputs include Excel models, PowerPoint presentations, and written analysis. Firm templates help users create spreadsheets and presentations that follow approved organizational standards instead of requiring the same formatting work for every task.
The capability overhang comes from the gap between model ability and useful workplace output
Kim calls the distance between what models can do and the value people get from them the capability overhang. Models have moved from solving high school mathematics problems to solving much harder mathematical problems, while many financial professionals still work in much the same way. Closing the gap requires more than a capable model. Users need suitable data, tools, and an experience that fits their workflow. Kim presents the product as an attempt to turn existing model intelligence into work that financial professionals can actually use.
A better product experience can bring non-engineering teams into AI use
Kim uses OpenAI's internal Codex adoption as an example of the effect product design can have. Engineers began using Codex early, while finance, legal, and recruiting started with almost none of their AI output generated in it. By June, those functions were generating around 90% of their AI output in Codex. Kim attributes the sharp change around February and March to the release of the Codex desktop app, rather than to a new model. His conclusion is that making an experience easier to use can help people outside engineering apply capabilities that were already available.
OpenAI licensed and indexed financial data to control quality and reliability
The financial product includes sources such as financial statements, earnings transcripts, market data, and private-company data. Kim says OpenAI evaluated existing MCP connectors and found problems with latency, errors, or search quality. The team therefore licensed much of the data directly and built its own hosting and search indices using the same infrastructure as the web index. The model still accesses the data through the MCP protocol. Hosting and indexing the data internally gives OpenAI more control over the experience and lets it post-train models to work more closely with those datasets.
Finance workflows need reviewable outputs that fit a firm's existing process
OpenAI added firm-approved templates for spreadsheets and presentations, interactive charts for stock prices and financial metrics, and citation viewers for the source passages and tables behind an answer. Kim says financial users need to check where a number came from and understand its surrounding context. These features make review part of the interface. He also explains that the product was built with the ChatGPT plugin platform available to developers. MCP tools and connectors provide data, citation metadata and MCP apps support source inspection, and skills with shared artifact tools guide Excel and PowerPoint workflows.
Model progress comes from scaling, algorithms, and unhobbling
Kim describes three overlapping drivers of model improvement: scaling compute, improving algorithms, and unhobbling, which means finding ways to get more useful work from models. He discusses pre-training, scaling laws, algorithmic efficiency, and reasoning through reinforcement learning and post-training. He says OpenAI has addressed concerns about limits such as the supply of human-generated training data and context length through synthetic data, verifiable rewards, algorithmic improvements, sparse attention, and compaction. He expects larger, more efficient, and more capable models in 2027, while acknowledging unresolved alignment and safety problems.
A task becomes valuable when the model crosses its threshold of usefulness
For each task, Kim describes a threshold where a model becomes capable and reliable enough to save time. Below that point, better answers may not matter if checking and fixing them takes as long as doing the work manually. Once the model crosses the threshold, review takes much less time than the task itself, so people begin using it regularly and delegating the work. The change can feel sudden even when model improvements have been gradual. Kim asks builders to design now for capabilities that may become reliable enough to delegate later, including what users will still need to control, review, or approve.
Finance will get both stronger general models and finance-specific post-training
Kim describes general model progress as a rising tide that improves financial work through transferable abilities in mathematics, coding, and reasoning. OpenAI's finance team is also investing more directly in finance-specific post-training. Over the next three months, the focus is general financial reasoning, financial analysis, Excel models, and PowerPoint presentations. The team also plans to work on senior-level judgment, such as deciding what matters in large information sets, forecasting earnings, assessing commodity prices, and evaluating investments. Kim says the goal is to match the best human experts on these tasks sometime in 2027.
Persistent agents could work more like human coworkers
Kim describes proactive and persistent agents as a further form of unhobbling. Instead of waiting for each prompt, a persistent agent can hold a goal, remember relevant context, and continue working in the background. This also changes how users interact with an agent. Rather than learning a special prompting workflow, a user could give the agent a project, share context, and ask for an update. Kim says this is closer to working with another person and could make AI more accessible to people who have not developed prompting or tool-use skills.
"The analogy I like to use is MCP and skills are like handing an instruction manual to a junior analyst versus post training, where the model's like having a senior analyst who's been working with the data for over two decades and knows all of its quirks like the back of their hand."06:49
Who should watch
You are building an AI product for a professional workflow and need to connect model capability to reliable data, review tools, and familiar outputs.
Your finance team is evaluating how to use AI for financial analysis, Excel models, presentations, or source checking.
You are designing agents and want a concrete explanation of when a model becomes useful enough for people to delegate work to it.