AI Audit for Schools: A Checklist to Find Which Tools Staff Use

Run a seven-step audit: tell staff it is an amnesty, send a short anonymous survey, check apps and browser extensions, map where data goes, rate each tool for risk, publish a short approved list, and set a review date. Do it in about a week and expect to find more tools than leaders guess.

The finding that tends to surprise leaders is not that staff use AI but how ordinary the uses are. Lesson-plan drafting, parent email wording, translated letters, marking comments, a transcription app for meetings. Almost none of it is reckless. The risk sits in a few details: which account they used, whether pupil names went in, and whether the vendor trains on what was typed.

Follow me on Instagram@sagnikteaches

Step 1: Set the tone with an amnesty before you ask anything

An audit that feels like an investigation gets silence. Before the survey goes out, send a short note from the head or principal. It should say three things plainly: you are looking for tools in use so that the school can support the good ones; nobody will be disciplined for past use that they disclose in this window; and a short approved list is the outcome.

Connect on LinkedInSagnik Bhattacharya

Illustrative wording:

Subscribe on YouTube@sagnikteaches

We are mapping which AI tools colleagues already use for school work. This is not a disciplinary exercise. Anything you tell us between now and [date] will be treated as an amnesty: no action will be taken over past use. What we want is a short list of tools we can support, with clear rules on what information may go into each. The survey takes four minutes and is anonymous unless you choose to add your name.

Keep the promise. If one disclosure leads to a sanction, the next audit gets nothing. The exception is where a disclosure reveals a serious safeguarding or data breach. Say that in the note too, in one sentence, so nobody feels tricked: "If something you tell us shows a risk to a child's safety, we have a duty to act on it."

Step 2: A four-minute staff survey that gets real answers

Keep it short. Eight questions is the limit. Use tick-box answers where you can, and always add an "other" box. Anonymous responses will be more honest, but ask for role group (teaching, support, admin, leadership) so you can see patterns.

1. Which of these have you used for school work in the last term? (A list: ChatGPT, Gemini, Claude, Microsoft Copilot, an AI feature inside a tool we already use, a transcription or meeting app, an image generator, a translation tool, a grammar or writing checker, other.)
2. Which account did you use? (School account, personal account, no account.)
3. What did you use it for? (Lesson planning, resources, marking or feedback, parent communication, reports, admin or emails, translation, meeting notes, other.)
4. Did you enter any of the following? (Pupil names, pupil work, reports or grades, parent details, staff details, none of these, not sure.)
5. Do you pay for it yourself?
6. Did it save you time? (None, under 30 minutes a week, 30 minutes to two hours, more.)
7. What would you want from an approved tool?
8. Is there anything you are unsure about using?

Question 4 is the one that matters. Questions 5 and 6 are quiet gold: personal subscriptions show where staff found value worth paying for, and time saved shows which tools deserve support. Question 8 gives you the training agenda for free.

Reading the results of the survey

Illustrative. Say a school of 45 staff gets 31 replies. Twenty-two have used at least one tool. Of those, 18 used a personal account. Nine entered pupil names or work. Four paid for a subscription. The top uses are lesson planning (17), parent emails (11), and report comments (7). Don't draw conclusions from the headline number. The two figures that matter are 18 using personal accounts and nine entering pupil data, because those are where data leaves your control.

Step 3: Check apps, browser extensions and accounts you can see

Surveys capture what people remember and admit. A technical check finds what they forgot. Ask your IT lead, or whoever manages the school's accounts, for five things. None of these requires special software.

• Sign-in with the school account. In Google Workspace, the Admin console's Accessed apps list (under Security, then Access and data control, then API controls) shows third-party apps that have used school accounts, and it downloads as a CSV file. In Microsoft 365, the Enterprise apps list in Microsoft Entra does a similar job. Export the list and look for AI names.
• Browser extensions. On managed devices, list installed extensions. AI writing assistants and meeting recorders often arrive as extensions.
• Installed apps on school laptops and tablets.
• Web filter or firewall logs. Look at categories or domains, not individuals. A count of visits to a few AI sites per week tells you about scale without identifying anyone.
• Existing licences. Many tools you already own have added AI features. Check your learning platform, your management information system, your meeting software and your office suite for AI settings that are on by default.

Report what you find as numbers and tool names, not as names of staff. If a log shows one person making 600 visits to an AI site in a week, do not open it unless there is a specific concern. The audit's trust depends on you not using the amnesty as cover for surveillance.

What a typical technical check turns up

Illustrative: the admin console shows 11 third-party apps with AI features linked to school logins, of which the school formally bought two. Three browser extensions on staff laptops summarise web pages, and one records meetings. The management system has a new "generate report comment" button that was switched on by an update. That last one is the most important finding in the example, because it is inside an approved system and nobody decided to enable it.

Step 4: Map where the data goes

For each tool found, fill in one row of a data-flow table. This turns a vague worry into a specific question you can answer from the vendor's own pages.

The training default matters. ChatGPT, Claude and Gemini all let personal-account users stop their chats being used for training (ChatGPT's personal plans share data until the user turns "Improve the model for everyone" off), while ChatGPT Business, Claude's Team and Enterprise plans, Gemini in Workspace and Microsoft's work Copilot say they do not train on business content by default. Describe it as a default to be checked, not a guarantee, because settings change. Fill "Unknown" honestly. An unknown is a finding in its own right and the next step handles it.

For a tool that sees pupil data, a data-protection impact assessment is usually appropriate. The worked impact assessment example shows the layout; you can adapt it for a school tool.

Step 5: Rate each tool with a simple risk score

You are not writing a thesis. A three-factor score is enough to sort the list into "fine", "fix" and "stop".

Add the three scores. A total of 3 to 4 is low: allow, perhaps with guidance. A total of 5 to 6 is medium: allow only with a fix, such as moving to a school account or banning pupil data. A total of 7 to 9 is high: stop until resolved, and offer an alternative.

Apply it to the table above. The personal free chat assistant used with pupil first names scores 3 + 3 + 3 = 9, which is high. The same assistant on a school business plan used for lesson ideas scores 1 + 1 + 1 = 3, which is low. The meeting recorder scores 3 (pupil discussion) + 3 + 3 = 9. The report-comment button scores 3 + 1 + 2 = 6, which is medium: probably fine once the vendor's addendum is read, but not before.

That produces a useful result: the same underlying product can be low or high risk depending on how it is used. This is why banning a tool by name is usually less effective than setting rules on accounts and data.

Three special cases to treat as high risk until proven otherwise

• Anything that records audio or video of children, including meeting recorders left running during pupil discussion.
• Tools that take special educational needs, welfare or safeguarding information, which is highly sensitive whatever the vendor says.
• Image generators and "AI tutors" used directly by pupils, which raise age-limit and content questions beyond this staff audit and should be handled separately.

Step 6: Publish an approved list that people will use

An approved list is the part staff will actually read. Keep it to one page. For each tool give its name, what it is approved for, which account to use, and what must never go in.

Give a route for new requests: a one-page form or a named person, with a promised reply time such as ten working days. Without a route, staff will go back to personal accounts as soon as they need something not on the list.

For a tool that the audit showed was valuable but risky, offer a fix rather than a ban. If 17 people use a free chat assistant for planning, the answer is probably to buy the business plan and move them over, not to block the site. The shadow AI tutorial on stopping data leaks goes into how to make the safer route the easier one.

Step 7: Set a follow-up date and a small review routine

An audit goes stale within a term because the tools change and so do the staff. Put a date in the diary before you finish. Three months is realistic for the first review, then every six months or on any significant change, such as a new contract or a vendor announcing an AI feature.

A review needs four things only: re-run the survey with the same eight questions so you can compare; re-check the admin console for new apps; walk through the approved list and delete anything unused; and add anything newly approved. Record the date, the lead and the outcome in a single page that governors or trustees can read. A short regular check is far better than a one-off thorough one.

Alongside the review, write or update a short usage policy so the rules in the list have a home. The tutorial on writing an AI usage policy gives a template that you can shorten for a school.

Five-minute conversations that fill the gaps

A survey and a console export leave blind spots, so add a few short conversations. Pick four or five people across roles: a head of department, a teaching assistant, someone in the school office, the person who handles data requests, and a teacher in their first year. Ask the same three questions and write the answers on one sheet.

1. "What is the most useful thing you have done with an AI tool this term?" This opens the door without any suspicion.
2. "Which tool, and how did you get access to it?" The second half reveals personal accounts, trials that quietly became paid, and tools recommended by a colleague at another school.
3. "Is there anything you stopped doing because you were not sure it was allowed?" This tells you what the future approved list should allow.

The office conversation is often the richest. Office staff handle parent letters, attendance chasing, translated notices and minutes, all jobs where AI helps and where personal details are close to hand. If a tool is in use anywhere with real names in it, it will probably be there.

Deciding what to do with each finding

Once every tool has a score, give each one a decision. Four outcomes are enough, and each should have a named action and date.

"Move" is usually the biggest group and the most valuable. It converts hidden use into supported use at modest cost, and it gives you the admin controls you lacked. If you buy, work through a short vendor question list first; the questions to ask an AI vendor tutorial has one that applies equally to schools.

A worked decision: the free assistant used by 18 staff

Illustrative figures. Eighteen staff use a free chat assistant on personal accounts, mostly for lesson planning and parent emails; nine of them have typed pupil first names. The score for the current setup is 9. The options are to ban it, ignore it, or move users to a school plan. A business plan at $25 per user per month (the monthly list price of a standard ChatGPT Business or Claude Team seat) for the 18 heaviest users would cost $450 a month, or $5,400 a year; annual billing at $20 a seat brings that to $4,320. Both plans have a two-seat minimum and a seat cap, and education pricing may differ, so check current terms. Set against that, the survey says those 18 save between 30 minutes and two hours a week each. The school decides to move them, with a rule that pupil names never go in, and checks the result at the three-month review. That is a more defensible decision than either a ban or a shrug, and it has numbers to back it.

Where audits go wrong

The survey that nobody finishes. If it takes over five minutes, response rates fall. Test it with two colleagues and cut anything not needed to make a decision.

The audit used as discipline. A leader spots a named account in a log and raises it in a one-to-one. The next survey gets 12 replies instead of 31. Protect the amnesty.

The ban without an alternative. A school blocks all AI sites on the network. Staff use phones on mobile data, and the data goes to the same tools with less visibility. A block should come with an approved route.

The list nobody maintains. The approved list says a tool is approved for six months after the vendor changed its terms. Give each entry a review date and an owner.

The feature nobody chose. A management system switches on an AI function in an update. Subscribe to vendor release notes for your main systems, or ask your IT lead to check settings each term.

A one-page audit report for governors

When the audit is done, summarise it in one page. Include the response rate, the number of tools found, how many scored high, medium and low, the actions taken, and the next review date. Illustrative:

AI tools audit, autumn term
Survey: 31 of 45 staff replied (69%). Amnesty in place.
Tools found: 14 (6 via survey only, 5 via admin console only, 3 in both).
Risk: 3 high, 5 medium, 6 low.
Actions: personal-account use of chat assistant A to move to school business plan by mid-term; meeting recorder extension removed from managed laptops; report-comment feature reviewed with vendor, settings confirmed; approved list published (6 tools).
Next review: end of spring term.

That is a complete, honest record. It shows leaders what staff use, what you did about it, and when you will look again. If you want help turning a messy set of findings into a list like this, that is what my AI implementation consultation covers, though most schools can run this checklist themselves.

Questions school leaders ask before an AI audit

Will staff admit to using tools if it might get them in trouble?

Far more will if you announce an amnesty first. State in writing that the audit looks for tools, not culprits, that nobody will be disciplined for past use disclosed during the window, and that the aim is a short approved list people can actually use.

Do we need to check pupils' use as well?

This checklist covers staff. Pupil use needs its own conversation with different questions about age limits, homework rules and safeguarding. Run the staff audit first, because what staff already do shapes what you can reasonably ask of pupils.

How long does an audit take?

For a school of around 40 staff, plan about a week of elapsed time and six to eight hours of the lead's work: one survey window, a few conversations, a short technical check and a one-page report. Most of the time is waiting for replies.

Should the approved list be long?

No. Five to eight tools is plenty. A long list is hard to keep checked, and staff will not read it. Each tool should have a named purpose and a clear line on what data may go in.

Further reads

• How to Pilot AI in Shadow Mode Before Customers See It — Trial an approved tool quietly before staff rely on it.
• What to Ask an AI Chatbot Vendor Before You Sign Up — Vendor questions if a chat tool scores medium or high.
• AI Answering Service for a School Office: Absence Calls and Cost — An example of an approved tool: AI for the school office phone.
• AI Automation for Education: Five Workflows, Tools and Costs — Which education workflows are worth automating first.
• Questions to Ask Before Buying AI That Touches Client Data — A vendor checklist for any AI tool that touches personal data.
• Free AI Tools for Schools: Who Qualifies and What Data They Keep — A school leader's data-protection checklist for six free AI offers: who qualifies, what each says about training data, and what to ask before staff sign up.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Data Controls FAQ and business pricing, Anthropic privacy centre, Gemini Apps Privacy Hub and the Workspace generative AI privacy hub, Microsoft Learn on Copilot data and Entra enterprise apps, Google Workspace Admin help on accessed apps, checked 4 November 2026.