How OpenAI's Products Are Used in Financial Services

Finance is the sector where AI use is talked about most and audited hardest. This article brings together the customer stories OpenAI has published on its own site and what OpenAI's finance team has told the press. The rule is simple: no number or claim appears here unless its source is an OpenAI page or a named publication. In the last section I look separately at what these bank-sized examples mean for a mid-sized manufacturing or trading company.

In the banks: governance first, roll-out second

The big-bank stories share one sequence: usage rules and review boards first, then mandatory training, and only then a company-wide roll-out.

BNY. According to OpenAI's account, the bank set up a central AI hub and let employees build their own agents on an internal platform called Eliza. The platform supports more than 125 live use cases and 20,000 employees are actively building agents. Use cases pass through a data-use review board, an AI launch board and an enterprise AI council; employees complete mandatory training before using the platform. The most-quoted result comes from the legal team: a contract review assistant covering more than 3,000 vendor contracts a year cut review time by 75 percent, from four hours to one.

BBVA. The collaboration started in 2024 with 3,300 accounts, grew to 11,000 and in December 2025 became a multi-year programme covering all 120,000 employees. The June 2026 story says more than 100,000 employees now use ChatGPT Enterprise, with about three hours saved per employee per week and productivity gains of up to 80 percent in selected workflows. The roll-out model is described too: a network of AI ambassadors, power users, and dedicated training for 250 leaders including the CEO and chairman. On the customer side there is a virtual assistant called Blue, built on OpenAI models.

Commonwealth Bank of Australia. ChatGPT Enterprise for about 50,000 employees; the aim is fluency in daily workflows first, then agent-supported scenarios such as customer service and fraud and scam response. The story highlights connectors, training programmes and leaders acting as role models.

MUFG. Mitsubishi UFJ Bank is rolling ChatGPT Enterprise out in stages to about 35,000 employees. Employees complete e-learning before they can use the tool, and AI champions have been appointed in every department. The clearest line in the story comes from a bank executive: the obstacle was not the technology but people not knowing what they were allowed to use it for.

Morgan Stanley. The oldest and most-cited example. The internal assistant built for financial advisors answers questions from a library of 100,000 documents; 98 percent of advisor teams use it and document access rose from 20 to 80 percent. A meeting summary tool turns consent-recorded calls into CRM notes and draft follow-ups, which the advisor corrects before sending. The real lesson is in the method: every use case passes an evaluation framework before release, a regression suite of sample questions runs every day, and a zero data retention policy answered the security question.

In investment research: agents and evaluation discipline

Balyasny Asset Management. According to OpenAI's March 2026 story, about 95 percent of investment teams use the firm's own AI research platform, and deep research tasks that used to take days now take hours. The most concrete example is an agent that analyses central bank speeches: macroeconomic scenario analysis went from two days to about thirty minutes. What stands out is not the model but what surrounds it: before putting models into production the firm built an evaluation pipeline measuring more than 12 dimensions, it selects models task by task, and agents are built centrally and deployed to teams with scoped access.

Model ML. OpenAI's August 2026 startup story is about the "last mile" of finance work: reconciling evidence, formatting the file, tying every number to its source. Model ML's agent produces an editable PowerPoint or Excel file from a brief and source material; the finance professional checks assumptions, sources and message before sharing. The story reports that at a global asset manager a bespoke summary report that took an analyst about an hour now takes about five minutes, and that in another workflow data rooms with more than 100,000 rows were processed in a single pass.

In the finance team's own work: ChatGPT Work, Excel and Codex

OpenAI offers finance teams two different things, and they are often confused.

ChatGPT Work and the Excel add-in. OpenAI's finance page describes a tool that connects to business applications through plugins (Excel, PowerPoint, Snowflake, Salesforce, Databricks and others) and hands multi-step jobs to agents: ready-made flows such as budget variance, cash forecast and revenue dashboard; every output traceable back to its source; sensitive actions kept subject to human approval; usage and spend controls. ChatGPT for Excel, released in beta in March 2026, builds and updates models inside the workbook, runs scenarios, follows links across sheets and can detect and fix errors; data integrations from FactSet, Dow Jones Factiva, LSEG and S&P Global were added alongside it. A Google Sheets version went into beta in April 2026.

Codex. The coding agent; for a finance team it means having the scripts that build the data flow written without waiting for a programmer. I covered this step by step in a separate article: How to automate financial reporting with Codex

OpenAI's own finance team. According to Fortune's August 2026 report, the team audits quarter-end materials with ChatGPT Work: it catches mismatches between the board deck and actuals, pulls the monthly close into one validated dashboard, and a feature called "CFO spotlight" surfaces the month's priorities and biggest variances. The director of product finance says tasks are completed two to three times faster; the same report states plainly that the tool does not replace the accounting system or the controller's sign-off. CFO Dive's September 2026 report adds that a manager on the same team cut computing-cost reporting from five days to five hours with Codex, noting that this is his own estimate. In a June 2026 conversation the CFO described custom GPTs used for investor relations and audit support.

What a mid-sized company can take from this

All of these stories come from organisations with tens of thousands of employees. The lesson for a 200-person manufacturer in Turkey is not the budget but the order of steps:

  • Rules before tools. Every story has a written usage rule and a review step; banks do it with boards, a small company does it with a one-page policy and a single owner. "Which data may leave, which may not" is the first line of that page.
  • Training is mandatory, access comes after. BNY and MUFG do not give employees the tool until training is done. For a ten-person finance team that means a two-hour workshop and three concrete jobs.
  • Do not go live without an evaluation set. Morgan Stanley's daily question suite and Balyasny's 12-dimension measurement are not a matter of scale but of discipline. A 30-question checklist built from your own data tells you whether an agent is reliable before an expensive mistake does.
  • The agent produces, a person approves. At Morgan Stanley the advisor, at Model ML the finance professional, at OpenAI the controller has the last word. In none of these examples is the agent given write access to the ERP; that is the principle we argue for on our security page.
  • Start with one workflow. Even the banks started with a single job such as contract review or meeting summaries. For you that is usually the monthly management report or supplier invoice checking.

Most of the tools described here run in the cloud. In the banks in this article, where the data goes was settled by contracts and zero-retention commitments. In a mid-sized company the same question is answered with a local-model option or a written scope; we describe both routes on our AI page.

Sources

The OpenAI pages below were read for this article on 8 October 2026; figures are as published on that day.

  1. OpenAI, BNY builds "AI for everyone, everywhere" with OpenAI, December 2025.
  2. OpenAI, BBVA and OpenAI collaborate to transform global banking, December 2025, and BBVA puts AI at the core of banking with OpenAI, June 2026.
  3. OpenAI, Commonwealth Bank of Australia builds AI fluency at scale, December 2025.
  4. OpenAI, MUFG aims to become AI-native with OpenAI, July 2026.
  5. OpenAI, Morgan Stanley uses AI evals to shape the future of financial services.
  6. OpenAI, How Balyasny Asset Management built an AI research engine, March 2026.
  7. OpenAI, Model ML gets finance work done efficiently with GPT-5.6 Sol, August 2026.
  8. OpenAI, ChatGPT Work for finance teams and Introducing ChatGPT for Excel, March 2026.
  9. Fortune, What OpenAI's finance team is becoming in the age of the AI CFO, August 2026.
  10. CFO Dive, Inside OpenAI's experiment with AI coding in finance, September 2026.
  11. CFO.com, OpenAI CFO Sarah Friar offers a look inside the company's finance function, June 2026.

ChatGPT, Codex and the GPT model names belong to OpenAI; other company and product names belong to their respective owners. This article is an independent compilation; figures are reported as stated in the sources and are the organisations' own accounts.