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

## Executive summary

Joseph Kim details the development of ChatGPT for Financial Services, a product designed for investment banking and equity research. The core challenge addressed is the 'capability overhang'—the gap between model potential and practical user value. The solution involves building a highly tailored experience that integrates proprietary financial data, firm-specific templates, and interactive viewing tools. Future development is predicted to center on proactive, persistent agents and the crossing of the 'threshold of usefulness,' leading to AI systems that function as true co-workers, eventually outperforming human experts on complex financial tasks.

## Key takeaways

- The Capability Overhang: Model capabilities are advancing rapidly, but the value derived by financial professionals has not kept pace. Closing this gap requires more than a capable model; it demands the right data, tools, and a workflow-fitting experience.
- Product Focus for Finance: ChatGPT for Financial Services focuses on three core areas: integrating financial data sources, building an interface familiar to financial professionals (with source inspection and charting), and adhering to the firm's specific format using templates.
- The Threshold of Usefulness: Meaningful model progress is felt when a task crosses a 'threshold of usefulness'—the point where the time saved by the model significantly outweighs the time required for human review and correction. This transition can feel like a sudden leap, even if the underlying improvement was gradual.

## Technical details

- Data Indexing and Control: Instead of relying solely on existing MCP connectors, the team licensed and indexed financial data directly. While the MCP protocol is still used for model access, self-hosting and indexing provide greater control over data quality, reliability, and the ability to post-train models on specific datasets.
- Product Architecture: The product was built using the native ChatGPT plugin platform. Functionality relies on MCP tools and connectors for data access, citation metadata for source inspection, and 'skills' combined with shared artifact tools to generate structured outputs like Excel models and PowerPoint presentations.
- Future Agent Architecture: The next form of AI utility is predicted to be proactive and persistent agents. Unlike current workflows that require explicit prompting, a persistent agent can hold a goal, remember context, and continue working in the background, making the AI more accessible and workflow-integrated.

## Practical implications

- When building AI products, focus on the entire workflow experience (the 'product') rather than just model capability. The integration of structured outputs (templates, charts) is key to adoption.
- For data-intensive applications, consider the limitations of existing connectors and evaluate the need to control the data pipeline (indexing, hosting) for reliability and specialized post-training.
- Design products anticipating the 'threshold of usefulness'—the point where the AI output saves more time than it takes to review. Build the interface to manage this review process naturally.

## Topics

Generative AI, Financial Technology (FinTech), Product Design, Agentic Systems, Large Language Models (LLMs), Joseph Kim, OpenAI, AIE CODE summit

Source: https://www.youtube.com/watch?v=opHmX7qClzE