Shadow AI: how to find it and decide what to do next

A list of purchased licences misses free accounts, browser extensions, personally paid subscriptions and new features inside existing software. Start with business teams: ask them to walk through a recent task, including the steps between receiving information and sending the finished result.

Explain what you are trying to establish and how the findings will be handled. Invite people to show the kind of information they use without asking them to paste confidential content into a discovery form. You need to know whether a prompt contains names, contractual information or commercially sensitive material; you usually do not need a copy of that material.

Compare those accounts with supplier records, approved application lists and available administrative information, using the organisation’s established privacy and security arrangements. A technical signal is a lead to investigate: access to a website does not establish the purpose, data or impact of the use.

Capture a specific use rather than a product name. “Drafting public event descriptions from an approved brief” is different from “summarising attendee accessibility requests”, even if both involve the same assistant.

If a discovery suggests confidential information has been shared or a consequential answer is reaching people unchecked, contain the activity and involve the responsible security, privacy or business lead. Preserve the facts needed to assess what happened. A possible incident should follow the organisation’s incident process; completing a register entry is not a substitute.

For a use without an immediate exposure, investigate why it exists. Is the approved tool unsuitable? Is access too slow? Has the employee misunderstood what an existing licence permits? Removing the tool without addressing the task can leave the work and the incentive unchanged.

Choose an explicit outcome. Stop the use if it cannot be supported responsibly. Modify it if a narrower input, approved account or review step addresses the concern. Approve a trial when there is enough evidence to test the bounded workflow. Set an owner and a date for unresolved questions instead of letting “under review” become an indefinite permission.

In this fictional example, an events team uses a personal assistant account to draft event descriptions. Discovery establishes that the current inputs are public venue information and an approved programme. Staff were considering adding attendee requests to personalise the text, but have not done so. That proposed extension changes what must be assessed.

This is not a claim that an enterprise licence makes any input safe. The decision still depends on actual settings, terms, data, purpose and controls. If personal data is involved, assess the relevant data protection obligations separately.

Publish a usable route for proposing an AI use and finding its decision. Tell the team which account to use, what information is permitted, who checks the output and where to raise a problem. Explain a refusal in terms of the work and risk so the same request does not return under a different product name.

Revisit the picture when teams change workflows, suppliers enable features or new accounts appear. Ask managers about changed uses during existing operational reviews. Discovery is not complete because every department has returned a spreadsheet once.

For the next step, take one discovered use through the worked AI risk assessment. Keep unknowns visible until somebody resolves them.

When the decision is to support a better way of doing the task, AI at Work, Acuity’s practical adoption resource, covers the skills and workflows that help people use AI in everyday work.