OpenAI tells banks to put AI agents in ChatGPT, not new apps

Two engineers at OpenAI told a workshop on AI agents that a bank gets to production sooner with its data and rules in skills and plugins. These go in tools that the bank approved, on the web and on the desktop. The finance tool uses the same model as standard ChatGPT, but the data around it is new.

How it works

The model is the same as in standard ChatGPT. The data around it is new: company reports, market data from 15 minutes before, news and earnings call transcripts. Each dataset was prepared for the model with a custom tool call and not with embeddings.

A skill is a text package with prompts or knowledge that persons can use again. A plugin is one zip package with skills, a connection server and screens.

The numbers

OpenAI gave no test results for bank work. Engineer Stefano Fabbri estimates that smaller models in the sub-agents give about 95% of the result in a team of agents.

Fabbri also reports that clients put the cost of use first for about three months.

What is not known

At this time there is no command line or API access to the financial data. Fabbri reports that this can change. Plugin screens work only on the desktop, and the hosts plan the web next. The hosts did not name the banks that tested the plugins.

Insight: AI platforms become the gate for company agents

On 4 November 2026, Adobe put its agent for marketing in ChatGPT, Claude and Copilot. OpenAI tells banks to do the same, and the workshop shows a research app rebuilt as a plugin. This looks like a race to own the surface where agents live, and the owner can set the rules for data and access. Leaders must find out who sets those rules.