Video summary
How OpenAI Built ChatGPT for Bankers and Analysts — Joseph Kim, OpenAI
Main summary
Key takeaways
Summary
Joseph Kim, who leads ChatGPT for Financial Services at OpenAI, describes how the team is adapting ChatGPT for investment banking and equity research—and what the build process suggests for other industry-specific AI products.
Product strategy and design
ChatGPT for Financial Services is an adapted version of ChatGPT for financial professionals at large institutions. It launched about a month before the talk, with Morgan Stanley and Evercore as design partners.
The team focused on three areas:
- Trusted financial data: Provide access to financial statements, market data, and private-company data within ChatGPT.
- Finance-oriented workflows: Make it easier to inspect financial charts, verify sources, and control which sources inform an analysis.
- Firm-specific outputs: Help users create Excel models and PowerPoint presentations using company-approved templates and standards.
The product also includes interactive financial charts and citation viewers that show source excerpts and tables, making review and verification part of the workflow.
Frameworks and product-building playbook
- Bridge the “capability overhang.” This is the gap between what models can do and the value people can reliably get from them at work. Closing it requires more than a capable model; it also requires relevant data, useful tools, and an experience that fits existing workflows.
- Build around the user’s work product. Identify the artifacts employees must produce, such as spreadsheets, presentations, and written analysis, then make them easier to create in the organization’s established formats.
- Make verification native to the experience. In high-stakes work, users need to check where figures came from and understand their context. Source review should be integrated, not bolted on.
- Use specialized data infrastructure when quality demands it. The team found problems with some financial-data connectors, including latency, errors, and search quality. It licensed data directly and built its own hosting and search indexes to improve control and reliability, while still using MCP to provide data access to the model.
- Separate access and instructions from domain familiarity. Kim compares MCP tools and skills to an operating manual for a junior analyst. Post-training can make the model more familiar with the data and how it is used.
- Build ahead of capability thresholds. For each task, ask what model capability is not yet useful enough to delegate, what might change within a year, and what users will still need to review, control, or approve once it becomes reliable.
Concrete examples and operating lessons
- Codex adoption as a product lesson: Kim says engineers began using Codex earlier, while finance, legal, and recruiting had little use by early 2026. He attributes a major turning point around February–March to the Codex desktop app, rather than to a new model. In the chart he cites, usage in those non-engineering functions reached about 90% by June. His takeaway: a better product experience can unlock existing model capabilities for users who are not technical specialists.
- Developer opportunity: The financial-services features were built using the ChatGPT plugin platform and components including MCP tools, citation metadata, MCP apps, skills, and artifact tools. Kim recommends bringing unique data, existing tools, or deep workflow expertise into ChatGPT using these building blocks. He predicts that plugins could support many $1 billion businesses.
- Operational priority: Reduce the work required to turn a model’s output into something ready to share with a team or client—not just improve the model’s raw answer quality.
Targets, timelines, and predictions
- Next three months: OpenAI’s finance team plans to improve financial reasoning and the quality of work on financial analysis, Excel modeling, and PowerPoint creation, with an emphasis on accuracy, structure, and usefulness in real workflows.
- By 2027: The team aims to improve performance on more judgment-intensive tasks associated with senior bankers, analysts, and investors, such as interpreting large volumes of information, forecasting company earnings or commodity prices, and evaluating businesses.
Kim’s three predictions for finance in 2027:
- Professionals will rarely rely on Excel and PowerPoint to create financial models or presentations.
- AI financial analysts will feel more like human colleagues—remembering context and preferences, taking ownership of tasks, tracking work proactively, and continuing in the background.
- Models will begin to outperform top human experts on important financial tasks, including profit forecasting and investment-idea development.
Models and organizational change
Kim groups model progress into three drivers: more computation and data, algorithmic improvements, and “unlocking”—finding ways to get more useful work from models through methods such as post-training and tools.
He expects finance-specific training to build on general improvements in reasoning, math, and coding.
He also sees proactive, persistent agents as a major next step. Instead of requiring users to prompt the model at every stage, an agent could retain relevant context, work toward an assigned goal in the background, and provide updates. This could make AI more accessible by letting people interact with it more like a colleague.
Presenter/source: Joseph Kim, OpenAI.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.