Why find existing AI use before writing rules?
A policy written without knowing current use either bans tools that staff depend on or misses the risky ones. The National AI Centre recommends an AI register that captures all AI systems, including AI features embedded in common software, and notes that a shared register helps reduce shadow AI use. The OAIC adds a privacy reason: obligations apply to personal information entered into an AI system and to outputs that contain it, and as best practice it recommends that organisations do not enter personal information, particularly sensitive information, into publicly available generative AI tools. Neither piece of guidance can be applied until you know where the data is going.
Where does unapproved AI use show up?
Shadow AI is rarely one tool. In most businesses it is spread across personal accounts, add-ons and features that arrived in a software update. Look in each of the places below and record what you find as you go. Include contractors and casual staff, who often bring their own tools, and include phones as well as computers. Each place needs a different check, so a single question at a staff meeting will not find them all.
- Chat assistants used on personal or free accounts in a browser or phone app.
- Browser extensions that read pages, rewrite text or summarise email.
- Meeting note-takers and transcription bots that join video calls.
- Third-party apps staff have connected to their work email, calendar or files.
- AI features switched on inside accounting, customer management, HR, design or helpdesk software.
- Paid AI subscriptions on company cards or claimed as expenses.
How do you ask staff without driving use underground?
Say why you are asking before you ask: to find out what helps, choose approved tools and protect client information. Make it clear that answers will not be used for discipline. Use a short form, named or anonymous to suit your workplace, with the questions below. Ask for a description of the task and leave the content out, so nobody pastes client data into the survey. Share a summary of the results and what will change, because staff who see their answers lead to a better tool are more likely to report honestly next time.
- Which AI tools or features have you used for work in the last month?
- Which tasks did you use them for?
- Was it a personal account, a free account or a work account?
- What kind of information went in: public, internal, client or personal details?
- What would you lose if the tool were unavailable tomorrow?
What can the Microsoft 365 and Google Workspace consoles show?
Both platforms record third-party apps that staff have allowed to access company data. In Microsoft Entra ID, all users can by default consent to apps for permissions that do not need administrator consent, such as access to their own mailbox. An administrator can review the user consent settings under Enterprise apps, Consent and permissions, and Microsoft recommends allowing user consent only for apps from verified publishers. In the Google Admin console, Security, Access and data control, API controls, Manage App Access lists configured apps, apps that have accessed Google data and apps pending review. Google notes that app details typically appear 24 to 48 hours after authorisation.
How do you check AI features in software you already pay for?
Built-in features are the easiest to miss because nobody installed anything. List your main business systems, then check each vendor's admin settings and recent release notes for AI features, whether they are on by default and who can switch them off. Ask each vendor in writing what data the feature sends and where it is processed, whether your data is used to train models, and whether the feature can be disabled for a user or a team. The OAIC lists similar due diligence points, including data flows, third-party access and whether features can be disabled.
What should you do with what you find?
Record every use in the AI register with the tool, the task, the account type, the data involved and an owner. Then sort each one into one of the decisions below. Where you stop a tool, offer an approved alternative for the same task on the same day if you can, otherwise the use tends to move to a personal phone where you cannot see it. Repeat the discovery at a set interval and whenever a major software update adds AI features.
- Approve: low-risk use on a business account with acceptable data terms.
- Approve with conditions: allowed for named tasks or data types only, with a review date.
- Replace: the task is legitimate but the tool or account type is not, so move it to an approved tool.
- Stop: client or personal information is going into a public tool, or the terms are unacceptable.