Shadow AI in financial services: how to find it and what to do about it

Shadow AI is the use of artificial intelligence tools that an organisation has not approved: a public chatbot in a browser tab, an AI note-taker joined to a client call, a translation extension, an AI feature switched on in a personal app. It is the AI equivalent of shadow IT, and in most firms it is already widespread.

For a regulated firm, shadow AI is a governance problem rather than a technology one. This article explains the risks, how to find out what is happening, and how to bring it under control without driving it further out of sight.

Why shadow AI happens

People rarely use unapproved AI to break rules. They use it because it helps them: a faster first draft, a quicker summary, a second opinion. Shadow AI grows where three things meet:

• AI tools are free or cheap and easy to reach.
• The firm has no approved alternative, or the approved one is hard to use.
• The policy is unclear, unknown, or seen as unrealistic.

Understanding this matters because the response that feels safest — banning AI outright — often makes the problem worse.

The risks for regulated firms

Client and confidential data leaving the firm

The most immediate risk is information being pasted into tools that store, process or reuse it outside your control. Depending on the provider’s terms, data may be retained, reviewed or used to improve services. For personal data, this raises UK GDPR questions about lawful basis, transfers and security. See protecting client data.

Unchecked output reaching customers

AI output can be wrong in convincing ways. If a client letter, complaint response or promotion is drafted with an unapproved tool and sent without proper review, the firm carries the consequences.

No record

When AI use is invisible, there is no record of what was used, for what, or with which data. That makes incidents harder to investigate and makes it difficult to show your arrangements are effective.

Dependence you have not assessed

Teams can come to rely on a tool the firm has never assessed, with no contingency if it changes or disappears.

How to find shadow AI

No single method gives a complete picture. Combine several:

1. Ask. A short, non-judgemental survey or team conversations surface far more than you expect. Make it clear the purpose is to understand, not to punish.
2. Attestations with exceptions. A periodic attestation that asks “have you used only approved AI tools?” — and lets people declare an exception with an explanation — turns hidden use into recorded, manageable information.
3. Existing technical data. Network, web-filtering, browser-extension and expense data can show which AI services are being reached or paid for.
4. Incident and speak-up routes. Make it easy to report AI concerns, including confidentially. Firms within scope of SYSC 18 already have whistleblowing arrangements that can cover AI concerns.

Record what you find in your AI use case register, including uses you decide to stop.

What to do about it

Provide an approved route

The most effective control is an approved AI tool that is genuinely useful: access to capable models, a familiar interface, and clear answers when something is not allowed. People switch when the approved route is easier than the workaround.

Set clear, realistic rules

Publish an AI acceptable use policy that says which tools are approved and what data may be used with them, with practical examples.

Put controls in the path of the request

Within the approved tool, controls should check each request before it reaches a model: is this model approved for this person, does the request contain restricted information, does it need human approval? Controls that act automatically prevent accidental breaches.

Use network controls where appropriate

Blocking known public AI services at the network or endpoint level can support the policy, but on its own it tends to move use to personal devices. Treat it as a complement to an approved route, not a substitute.

Treat declared exceptions as information

When someone declares that they used an unapproved tool, follow up — understand what was used and why, assess any data exposure, and fix the gap that caused it. If people are punished for honesty, declarations stop.

Measuring progress

Useful indicators include the number of declared exceptions over time, the share of AI activity going through approved tools, the number of requests where sensitive information was masked or blocked, and time to resolve AI incidents. None of these needs a single “score” — trends and exceptions tell you more.

How Dutira helps

Dutira gives employees an approved AI workspace with access to OpenAI, Anthropic and Google models, checks every request against your policies, and runs attestations that let people declare exceptions. Monitoring covers AI activity that runs through the platform and the tools you connect to it; it does not automatically see use of other third-party tools, which is why attestations and incident reports matter. See the AI Workspace.