How Financial Authorities Define and Track AI Use

Financial authorities use different definitions of AI. What the FSB's 2025 report found about definitions, indicators, and four monitored vulnerabilities.

Last verified: October 20266 min read

Financial authorities still have a partial view of how much AI the financial sector uses, and one reason is that they do not share a definition of it. In a survey the Financial Stability Board (FSB) ran in early 2025, answers from 19 jurisdictions and one international organization showed that some supervisors adopt the OECD or EU AI Act definition, others write their own, and a few use none. The FSB's report of October 10, 2025 lists that gap among the obstacles to tracking AI in banks, insurers, and markets.

Key takeaways

  • The FSB survey drew 28 responses from supervisors and other authorities, and the report describes monitoring of AI in finance as still at an early stage.
  • Four vulnerabilities from the FSB's 2024 report remain the framework: third-party dependence, market correlations, cyber risk, and model risk with data quality and governance.
  • Firms classify the same model differently, for example by calling a vendor model that was modified in-house either third-party or internally developed.
  • Fewer than half of the respondents collect monitoring data every year, and none collect it more often.
  • Many indicators are proxies, such as patent filings, job postings, and survey answers, not direct measures of risk.

Why one word covers so many systems

A traditional credit scorecard, a fraud model retrained every week, and a chatbot that drafts customer letters can all be described as AI, and each carries a different kind of risk. The FSB's report on monitoring AI adoption and related vulnerabilities notes that definitions differ across jurisdictions. Some authorities use the OECD definition or the one in the EU AI Act, which is itself based on the OECD's 2023 update. Others rely on definitions specific to their own jurisdiction.

The survey respondents also said they prefer technology-neutral rules. The authorities interviewed plan to rely on existing policy frameworks unless AI-specific ones prove necessary. That choice keeps one rulebook for every kind of model, but it also means that a count of "AI users" depends on whichever definition the counting authority picked.

What the survey counted

The FSB asked its members how they track AI adoption. Responses came from 19 jurisdictions and one international organization, mostly from authorities with supervisory duties such as prudential bank supervision or market conduct oversight. The most common method was a survey of financial institutions, followed by research on public data. Over a third of the respondents with supervisory powers had AI-related supervisory reporting in place. Just over half of the authorities that collect AI data publish aggregate results, and participation in most surveys is voluntary for firms.

Voluntary participation brings selection bias, and the report lists it among the limits of surveys. Cost is the other limit, for firms and for authorities. Less than half of the respondents collect monitoring data annually.

Four vulnerabilities and how each is watched

The FSB's 2024 report on financial stability implications of AI, published on November 14, 2024, named the four vulnerabilities below. The 2025 report examined how far authorities can see each one. The table condenses the survey findings; it is a summary, not a legal classification.

VulnerabilityMeaning in plain termsWhat the 2025 survey found
Third-party dependencies and service provider concentrationMany institutions depend on a few outside providers of models, data, or computingMany respondents ask about third-party AI in surveys; few ask how critical those applications are
Market correlationsSimilar models and data could push many firms to act alikeA limited number of authorities collect anything on it; the report says there is still little empirical evidence of an effect on market outcomes
Cyber risksAI can be attacked, or used to attackMany respondents collect general cyber incident data that includes AI-related cases
Model risk, data quality, and governanceErrors, poor input data, and weak oversight of how models are built and usedMany ask about governance and explainability; a limited number have issued AI-specific model governance guidance

Direct measures and stand-ins

The report separates direct indicators from proxies. A direct indicator would be an inventory of AI use cases sorted by financial activity (trading, lending, insurance, payments), by type of AI (generative, agentic, or other), and by materiality, meaning whether the use sits in a core business line, a critical operation, or a low-risk internal process. A proxy would be the number of AI-related patents a bank files, its AI-related job postings, or its technology spending.

The report gives two examples of proxies. The US Census Bureau's Business Trends and Outlook Survey asks firms two questions about AI and is published every two weeks. It shows securities and investment firms using AI more often than lenders and insurers, but coverage of the largest institutions is uneven. In patent data, US global systemically important banks filed more than 1,400 AI-related applications over ten years, a share of total filings that grew for much of that period and then leveled off in the last three years. Neither figure says how much a bank relies on AI, which is why the report advises combining several indicators.

Questions to ask about a statement such as "this bank uses AI"

The checklist below is an illustrative reading aid built from the report's own distinctions. It is not an official test.

  • Which definition is in use, and does it include older statistical models?
  • Is the figure a share of firms, a share of activities, or a count of projects?
  • Does the use sit in a core function or in a low-risk internal process?
  • Was the model built in-house, bought from a vendor, or bought and then modified?
  • Does the number come from a voluntary survey, and how many firms answered?

This article gives general information about how financial authorities monitor AI. It is not financial, investment, or legal advice, and it does not rate the safety or conduct of any bank, insurer, or fund. Questions about a specific institution belong with its regulator, and questions about legal duties belong with a qualified professional.

FAQ

Does the FSB set binding rules for banks?

No. The FSB coordinates national financial authorities and standard-setting bodies, and its charter says its activities do not create legal rights or obligations. National authorities decide what applies in their own jurisdictions.

Which definition of AI does the EU AI Act use?

The FSB report says the EU AI Act contains a regulatory definition of AI based on the OECD's updated 2023 definition. Authorities outside the EU may use that definition, a variant of it, or their own.

Has AI been shown to cause correlated market moves?

The FSB says there is still little empirical evidence that AI-driven market correlations affect market outcomes. It lists the vulnerability as one that may change as the technology develops.

Why do authorities monitor AI if they prefer technology-neutral rules?

Some risks, such as explainability problems or limited transparency in pre-trained models, may be heightened by AI. Monitoring shows whether existing frameworks are sufficient or whether AI-specific ones become necessary.

Last verified: October 2026

Sources

  • Monitoring Adoption of Artificial Intelligence and Related Vulnerabilities in the Financial Sector, Financial Stability Board, October 10, 2025
  • The Financial Stability Implications of Artificial Intelligence, Financial Stability Board, November 14, 2024

New to artificial intelligence?

Start with the basics, then follow the guides to tools, rules, and effects on society.

Start with AI Basics