How can I find out how staff are using AI without spying on them?

31 October 2026

Find the tools and risks, not a minute-by-minute account of people

Start with a declared discovery exercise: which AI tools are being used, for what work, with which accounts and which categories of business information? You do not need to read every prompt to answer those questions. You need enough visibility to decide what is safe, what needs approval and where staff need a better alternative.

Separate three purposes before collecting anything. A tool inventory identifies suppliers and owners. A security investigation examines a specific concern. Performance management evaluates someone's work. Combining all three into one vague promise to “understand AI usage” creates unnecessary risk. Information collected to secure business accounts should not quietly become a league table of supposedly productive employees.

The UK Government's Cyber Security Breaches Survey 2025 reported that 43% of businesses identified a cyber security breach or attack in the previous 12 months. That is a dated, general cyber security finding, not an estimate of AI misuse. It explains why business oversight matters; it does not justify blanket surveillance.

Write the actual question you need answered before selecting software. “Are customer files going into unapproved services?” is a useful question. “What is everyone doing on their laptop?” is too broad. The first points towards approved accounts, clear data rules and limited checks. The second encourages you to collect material you neither need nor want to hold.

Ask staff first, without demanding their private conversations

Use a short questionnaire or team discussion. Ask for the tool name, whether the account belongs to the business, the task it supports, the type of information entered, any connections to email or files, and whether the output is checked. Include built-in AI features, browser extensions, meeting assistants and mobile apps. People may not describe an automatic meeting summary as “using AI”.

Explain what happens next. A useful message is: “We want to support useful tools and stop confidential information going somewhere unsuitable. We are collecting tool names and work purposes, not asking to inspect your personal account history.” Offer a private route for questions and accidental disclosures. Do not promise immunity from disciplinary action unless your actual employment policy authorises that promise.

Anonymous feedback can reveal blocked workflows or missing training, but it cannot reliably identify every supplier that holds company information. In a five-person team, even supposedly anonymous comments may identify the writer. Avoid collecting department, seniority and detailed examples together if that combination makes people recognisable. Ask staff not to include client names, health details or prompt transcripts in the survey.

Managers should participate too. If a director uses a personal chatbot to summarise contracts, the same approval standard should apply. Otherwise staff learn that policy is for junior people. The most useful outcome is often discovering why an unapproved tool is attractive: it fixes a genuine problem faster than the approved route. Address that problem rather than treating the disclosure itself as evidence of wrongdoing.

Compare answers with business records you already control

Check company-card subscriptions, expense claims, procurement records and approved software lists. Review work-account sign-ins, installed extensions and connected-app permissions where you have a lawful, explained reason and authorised access. Record what each source can actually tell you. A payment proves a subscription exists, not what someone uploaded. A sign-in proves access, not that a customer file was shared.

Microsoft 365 and Google Workspace administration can help identify work accounts and integrations within their respective environments. Available records depend on your licence, configuration, permissions and retention. Neither platform gives you a complete view of personal accounts, personal phones or every standalone AI service. Do not buy an upgrade on the assumption that it solves the whole discovery problem.

A simple spreadsheet or your existing service register is often enough for the first pass. Record the supplier, business owner, purpose, account type, data categories, connected systems, approval status and next review date. Our guide to the risks of an incomplete AI register covers the record itself. This exercise is about gathering its inputs without crossing unnecessary privacy boundaries.

Use discrepancies as questions, not verdicts. If a card payment is missing from the questionnaire, ask whether it was a trial, a team subscription or an abandoned purchase. If an integration has broad permissions, establish whether it is still needed before changing access. Disabling an unknown connection without checking its owner can interrupt legitimate work. Contain urgent risks, but do not confuse a discovery exercise with indiscriminate shutdown.

Where does proportionate oversight become intrusive monitoring?

Reading every conversation, taking continuous screenshots or capturing keystrokes is a very different intervention from listing approved suppliers. Personal accounts may contain family matters, job applications, medical information or confidential material from someone else. Owning the laptop does not make every possible inspection necessary or lawful.

The ICO's guidance on monitoring workers says employers must identify a lawful basis, be transparent and consider necessity and proportionality. Its discussion of consent highlights the imbalance of power in employment. A signed policy or a manager's request to “agree” is not a shortcut around those requirements.

Where planned monitoring is likely to result in high risk to people's rights and freedoms, a data protection impact assessment is required. Assess the actual proposal rather than assuming every small survey needs one or that small employers are exempt. Check current ICO guidance when implementing: relevant guidance may be updated following changes to UK data protection law.

Limit who can see identifiable findings and set a justified retention period. A small business might choose to delete raw survey responses after checking its inventory, while keeping supplier approvals for longer; those are operational choices to document, not statutory deadlines. If a particular incident warrants reviewing content, scope that review separately, restrict access and involve the appropriate privacy, security or employment adviser. Do not widen routine monitoring because a dashboard makes it easy.

What should we measure instead of counting prompts?

Measure whether you can account for the tools handling company information. Useful checks include whether each active service has an owner, whether its access is understood, whether high-risk uses have been reviewed and whether former staff still have access. None requires ranking employees by how often they speak to a chatbot.

For effectiveness, measure the work rather than the typing. Does the drafted reply require less correction? Are customer details accurate? Does the workflow reduce rework after checking time is included? A high prompt count could mean experimentation, repeated failures or a task that should not use AI at all. A low count could mean one carefully designed workflow is doing the job well.

The same government survey reported that 19% of businesses undertook staff cyber security training or awareness raising in 2025, compared with 76% of large businesses. These are historical survey results, not AI-specific training figures. They are a useful reminder that buying oversight software does not establish whether people understand safe working practices.

Here is illustrative budgeting, not a market quote: if six people each spend 20 minutes answering a survey and the owner spends two hours reviewing it, the exercise takes four staff-hours. At an assumed internal cost of £30 per hour, that is £120 before any follow-up. Compare that visible effort with buying another subscription, managing its alerts and reviewing information you may never have needed. Keep the first exercise small enough that you can repeat it when the tools or risks change.

When this approach is NOT right for you

A light-touch inventory is not an incident response plan. If you have credible evidence that sensitive information has been exposed, preserve relevant evidence, contain access and assess the incident promptly. Do not wait for a quarterly questionnaire. Reporting obligations depend on the facts, so obtain appropriate advice rather than assuming every mistake is either reportable or harmless.

Nor is a general staff survey sufficient for regulated work, sensitive health information or systems making consequential decisions about people. Those situations need specific checks of purpose, data flows, permissions and accountability. If you are considering detailed employee monitoring, get specialist advice before switching it on. A supplier's claim that its product is “GDPR compliant” does not decide whether your use is lawful.

At the other end, a business with two approved tools, business-owned accounts and no connected client records probably does not need a dedicated AI surveillance platform. Review existing controls and have an honest conversation first. A spreadsheet, clear approval route and usable alternative may solve the immediate problem better than an enterprise discovery product.

End discovery with decisions people can understand: approved uses, uses needing further checks, data that must not be entered, and who can help. Give useful tools a route to approval rather than leaving them permanently in limbo. If you need help distinguishing an account-security gap from a monitoring proposal, bring a list of tools and work purposes to a conversation with an adviser. There is no need to share staff prompt histories to begin.

Is This Right For You?

This approach suits UK small businesses trying to identify AI suppliers, unsafe data sharing and missing approvals. It is not a substitute for incident response or specialist assessment of intrusive monitoring. Start with tool names and purposes, not people's private conversations.

Frequently Asked Questions

Can I ask staff to show me their personal ChatGPT history?

Do not make this your routine discovery method. Ask for tools, work purposes and data categories first. Personal histories may contain unrelated sensitive information. Any targeted inspection needs a specific justification, appropriate authority and a proportionate scope.

Do I need employee consent to check business AI use?

Consent is not automatically the right lawful basis in employment because of the imbalance of power. Identify an appropriate lawful basis, explain the processing and assess necessity and proportionality. Consult current ICO guidance for your proposal.

Will anonymous surveys find every unapproved AI tool?

No. They can expose useful patterns and barriers, but they are not a complete inventory. Compare the findings with authorised procurement, subscription and account records, without assuming missing information proves misconduct.

Should I monitor personal phones used for work?

Do not assume you can inspect personal devices or private apps. Set clear work-data and approved-account rules, consider less intrusive controls and take advice before introducing device monitoring.

Can Microsoft 365 or Google Workspace show all our AI use?

No. Their administrative records can help within the environments you control, subject to licensing and configuration. They will not reliably reveal every personal account or separate service.

How long should we keep discovery information?

Choose a period justified by the purpose and document it. Raw identifiable answers may be needed for less time than supplier approval records. Avoid keeping prompt transcripts or personal information just because storage is inexpensive.

What if someone admits sharing confidential information?

Move it into your incident process. Restrict further disclosure, preserve relevant facts and assess containment and any reporting duties. Do not copy the sensitive material into a team-wide survey report.

Should AI usage figures be used in performance reviews?

Prompt counts are not reliable measures of performance. Assess work quality and outcomes, and do not quietly repurpose security or discovery information for employee scoring.