Can AI Draft a Headteacher Report to Governors From Your Data?

Yes. Export attendance, progress and budget figures as anonymised tables, give an AI assistant a fixed section layout, and ask it to turn the numbers into short paragraphs. You then check every figure against the source, add your own judgement on each priority, and send it. AI drafts the prose; you own the claims.

The catch is that governors do not want a tidy summary, they want a report they can question. A draft that says "attendance is broadly in line" helps nobody. So the work is in the instructions you give: ask for movement against last term, against target and against the same term last year, and make the AI flag anything it cannot support from your tables.

Follow me on Instagram@sagnikteaches

What goes in, what stays out

The rule is simple: the AI sees aggregates, never people. Governors scrutinise patterns, so the report never needs a pupil's name, date of birth or a free-text case note. That also keeps you well clear of the awkward question of whose data you have pasted where.

Connect on LinkedInSagnik Bhattacharya

Before you paste anything, check which assistant you are using. Consumer plans of ChatGPT and Claude let you switch off the use of your chats for model training in privacy settings. Business and education plans do not train on your content by default: that covers ChatGPT Business and Edu, Claude Team, the work version of Microsoft Copilot (formerly Microsoft 365 Copilot) and Gemini for Education, which Google includes free in every Workspace for Education edition. Whichever you use, confirm what your school's data-protection lead has approved. If you are unsure, read which free AI tools suit a school and what data they keep first.

Subscribe on YouTube@sagnikteaches

Here is a practical split.

Small numbers are a trap. A table saying "Year 6, disadvantaged, persistent absence: 1" identifies a child to anyone who knows the cohort. Suppress or merge groups under five before the table leaves your management system, and tell governors you have done so.

Building the data pack in one spreadsheet

Most of the time saved comes from having the same one-page data pack every term. Set up a single workbook with one tab per section, each with the same column pattern: measure, this term, last term, same term last year, target. The AI needs those four comparison points to say anything useful.

Say a primary school with 210 pupils builds this attendance tab.

All figures here are invented for illustration. The point is the shape: with five columns, the AI can write "attendance is 0.5 points below last term but 0.4 above the same term last year, and 1.4 points off target", and you can verify that sentence in ten seconds.

Build similar tabs for progress (percentage at or above expected standard by subject and year), behaviour and exclusions (counts only), staffing (vacancies, absence days, no names), safeguarding (training completion, referrals counted, single-central-record audit date) and finance (budget, spend to date, forecast outturn, variance by line). Keep every number as a number in its own cell, not buried in a text note. If you export from your management information system, paste values into the pack rather than sending the live file.

The prompt that sets the layout

Write the layout once and reuse it. Governors learn where things are, and you stop reinventing headings each term. This prompt is deliberately strict about what the AI may and may not do.

You are helping a headteacher draft the termly report to the governing body.

Use ONLY the tables below. Do not add figures, causes, or events that are
not in them. If a comparison cannot be made from the tables, write
"[data not supplied]" instead of guessing.

Sections, in this order, each under 150 words plus a short list of
questions governors might ask:
1. Attendance
2. Progress and standards
3. Behaviour and exclusions
4. Safeguarding (counts, training and audit dates only)
5. Staffing
6. Finance
7. Priorities for next term (use my notes)

For every measure, say: this term vs last term, vs same term last year,
vs target, and whether the movement is good, neutral or concerning.
Write in plain British English. No marketing language. Do not praise
the school. Mark anything you are unsure about with [CHECK].

TABLES:
[paste anonymised tables]

MY NOTES ON PRIORITIES:
[paste your own bullet points]

Two instructions do most of the work. "Do not praise the school" strips the cheerful filler that makes governors distrust a report. And "[CHECK]" gives you a quick way to find every spot where the model felt unsure, rather than reading line by line for hidden errors.

What comes back, and what you fix

Here is a realistic output for the attendance tab above. It is illustrative, so you can see the pattern.

Overall attendance was 94.6%, down 0.5 points on last term (95.1%) but up 0.4 points on the same term last year (94.2%). It remains 1.4 points below the 96.0% target. Unauthorised absence rose to 1.9% from 1.6% last term, still better than last year (2.1%) but 0.4 points above target. Pupils below 90% attendance increased from 24 to 27 (12.9% of the roll), against a target of under 10%. Reception is the weakest year group at 93.1%, 1.9 points below target [CHECK: cause not supplied].

That is good raw material. Now apply your judgement, because a draft like this has three weaknesses.

• It reports, it does not explain. The cause is yours to add: say a cluster of illness in January, or two families on extended leave. The AI correctly refused to invent one.
• It treats all movement as equal. A 0.5-point fall across a whole school is larger than it sounds, because it equals roughly one extra absence day per pupil over a full year. Only you know whether that is a blip.
• It has no position. Add a sentence beginning "My concern is…" or "I am not worried about…". Governors want your view to challenge.

Rewrite the paragraph with your additions and the list of questions the AI suggests, then decide which questions you want governors to ask. Some will be helpful ("What is the attendance of pupils with a special educational needs plan?"), others you can drop.

The accuracy check before it leaves your desk

Models get arithmetic wrong, drop a decimal, swap this term and last term, and occasionally invent a plausible number. A school report goes to people who will quote it at a parents' meeting, so treat the check as non-negotiable. It takes about 20 minutes for a three-page report.

1. Highlight every number in the draft. Use a highlighter in your word processor. If a number is not highlighted afterwards, it was not checked.
2. Tick each against the pack. Find the cell. Check the term, the group and the decimal.
3. Recompute every difference. "Down 0.5 points" means 95.1 minus 94.6. Do it on a calculator, not in your head.
4. Check the direction words. "Improved", "fell", "rose" are where models flip meaning, especially for measures where lower is better, such as absence and exclusions.
5. Search for causes. Any sentence with "due to", "because" or "as a result of" must come from you, not the model.
6. Search for [CHECK] and [data not supplied]. Resolve or delete each.

A second pair of eyes helps. Ask your school business manager to check the finance section alone; they will catch a wrong line in a minute because they know the budget.

You can also ask the AI to audit itself, which catches some errors and misses others, so use it as an extra, not the only check:

Here is my draft report and the source tables. List every number in
the draft that does not exactly match a table, and every comparison
(up, down, above, below) that is arithmetically wrong. Show the
calculation. Do not rewrite the draft.

A finance section that governors can actually challenge

Finance is where AI helps least by itself and most with a clear table, so give it the variance column and ask for flags rather than commentary. Say a primary school's finance tab looks like this (illustrative):

Ask: "List lines where forecast exceeds budget by more than 5% and write one neutral sentence on each. Do not suggest causes." The model will pick out supply cover (up 54.5%) and teaching staff (up 1.3%, below the threshold). You add the cause: perhaps long-term sickness cover in the autumn. That is a sentence only you can write, and it is the one governors will ask about.

Watch for a classic fault: the AI adds the variance column wrongly when some variances are negative. Add it yourself. Here, +8,000 + 12,000 − 1,500 − 1,000 = +17,500, so the total matches.

Turning the draft into a safe safeguarding section

Treat this section as a data table with a sentence under it, nothing more. Give the AI these items: staff with up-to-date training as a proportion of all staff, the date of the last single-central-record check, the number of referrals made this term as a count, the number of open cases as a count, and any actions from the last review.

A good output reads like this (illustrative): "All 31 staff completed their annual safeguarding update this term (previous term: 29 of 31). The single central record was last checked on [date]. Referrals this term: 3 (autumn term: 5). Open actions from the last governor review: 1 [CHECK: status not supplied]."

Notice what is missing: any story. The designated safeguarding lead owns the narrative and gives it to the safeguarding governor through whatever route your school uses. If a governor asks for detail, the answer is a conversation, not a pasted paragraph.

Progress data: where a table needs a sentence of context

Progress figures are the hardest to draft well, because a percentage means little without the cohort size. Give the AI three columns, not one: the percentage, the number of pupils, and the same measure last year. Say a Year 4 class of 28 shows these reading results (illustrative): 20 pupils at or above the expected standard this term (71%), against 17 of 29 in the same class a year ago (59%), with a target of 75%.

Ask the model to state the pupil count as well as the percentage. "71%" sounds like a big gain; "three more pupils than last year's class" is the honest version, and governors should see both. A good prompt line is: "For any cohort under 30 pupils, give the count alongside the percentage and say that one pupil equals more than three percentage points."

Here is how the output might look after your edit: "20 of 28 Year 4 pupils (71%) are at or above the expected standard in reading, up from 17 of 29 (59%) in the same class a year ago. The target is 75%, so one more pupil (21 of 28) would meet it. One pupil is worth 3.6 points, so small changes here move the percentage a lot." Check the sum yourself: one divided by 28 is 3.57 per cent. The model usually gets such sums right, but this is exactly the sort of line the highlighter check exists for.

Finally, add what the numbers cannot show. If the class had three arrivals mid-year, say so. If the gains came from a new phonics programme, say that too, because governors will want to decide whether to keep funding it.

Priorities and the "so what" paragraph

This is the part governors read first. Give the AI your own bullet notes, not data. For example:

• Reception attendance: letters sent, family conversations, breakfast club offered to three families.
• Writing in Year 4: moderated sample showed weak sentence structure; training day booked.
• Supply cost: appointing a permanent cover supervisor to cut daily agency use.

Ask it to expand each into three lines: the issue, the action, how you will know it worked. Then edit hard. The "how you will know" line must be measurable. If the AI writes "improved outcomes", replace it with "Reception attendance at or above 95% by the summer term".

When attendance drives a priority, link it to what actually happens in school. The three-stage attendance letter sequence is one such action, and you can cite the number of letters sent at each stage as a count in your report.

A term-by-term routine that takes under two hours

After the first term, the work shrinks. Here is a realistic rhythm for a head with a part-time business manager.

1. A week before the clerk's deadline for papers, 30 minutes: the office exports the data into the pack and suppresses small numbers.
2. Same day, 10 minutes: you paste the tables and your notes into the saved prompt.
3. That evening, 40 minutes: you rewrite, add causes and positions, remove generic lines.
4. Next day, 20 minutes: you run the accuracy check; the business manager checks finance.
5. By the clerk's deadline: send the report and appendix so they go out with the agenda. Your board's procedures set how far ahead governors must receive papers, so work back from that date, not from the meeting.

Keep the prompt and the empty pack in a shared folder. If you use a shareable Project in ChatGPT or Claude on a business plan, save the layout instructions there so a deputy can run it when you are away. OpenAI has scheduled custom GPTs to retire on 11 December 2026, so do not build the routine on one; a Project holds the same instructions without that end date.

Four ways an AI-drafted governors' report goes wrong

A few failure patterns turn up repeatedly with tools like this.

The confident invention. You omit last year's column for one measure, and the model produces "up 0.3 on last year" anyway. The instruction to write "[data not supplied]" reduces this but does not abolish it. The highlighter check catches it.

The softened bad news. Models drift towards reassuring language. "A slight dip" for a 2.4-point fall in Year 1 reading is not acceptable. Ask for "concerning" labels with a threshold you set, such as "any fall of more than 2 points".

The sameness. Three terms running, the report reads identically, and governors stop reading. Vary what you emphasise. Add one short "focus this term" box, such as a closer look at one subject, written by you.

The pasted table that was not anonymised. A pupil-level export has a name column you did not notice. Open the file, scan the headers, and delete the column before copying. If a pupil-level paste does happen, tell your data-protection lead promptly and delete the chat according to the tool's own controls.

What to ask governors to do with the report

End each section with two or three questions you want challenged. This changes the meeting from a presentation to a discussion. Ask the AI for ten candidate questions per section, then choose the three that are hard and fair. Include at least one that you cannot easily answer. A board that never struggles with a question is not doing its job, and a head who invites one is more likely to be trusted.

If you want to go further, the same data pack can feed a one-page parent summary or a staff briefing. Keep those separate: a parent summary has different words, different caveats and no finance. Run a different prompt on the same tables rather than pasting the governors' report to a wider audience. For broader ideas on the office side, five school office jobs and what each costs shows where else the time goes, and a one-term plan for a small school puts reporting in sequence with everything else.

The idea to hold on to: AI is a fast way to turn a clean table into a first draft. The cleanliness of the table, the cause behind each number and the judgement about what it means are still yours.

Questions heads ask before trying this

Should I paste pupil names into ChatGPT or Claude?

No. Governors need patterns, not names. Use counts, percentages and group labels such as Year 4 or disadvantaged pupils. If a case needs discussing, give it a code and keep the key in your own system, in a part of the meeting governors handle confidentially.

How long should the report be?

Many governing boards prefer a short written report they can read in ten minutes: one to three pages plus a data appendix. Ask your chair what they want. A tight narrative with tables behind it is easier to challenge than a long one.

Can AI write the safeguarding section?

It can format counts and training dates you supply. It should never summarise individual cases or make judgements about them. Keep that section to numbers, dates and actions, and discuss anything sensitive verbally or through your designated lead.

Will governors know it was AI-drafted?

Possibly, if you send it unedited. The giveaway is bland, even-handed prose with no opinion. Add your own judgement on each priority and say plainly what worries you. A report in your voice, with the AI doing the table-to-sentence work, reads as yours.

Further reads

• Is Gemini Free for Schools? What Is Included and What Costs Extra — Whether a free Gemini plan covers a school.
• AI Training for School Staff: Session Plan, Tasks and Handout — Teach staff what data never goes into a chatbot.
• AI Answering Service for a School Office: Absence Calls and Cost — Absence calls feed your attendance figures.
• AI Consultant for Schools: Role, Cost and How to Brief One — What to ask if you bring in outside help.
• AI Chatbot for a School Website: Parent FAQs From Your Handbook — Parent FAQs answered from your handbook.
• Sports Club Treasurer's Report: A Plain Summary of the Books With AI — How a volunteer treasurer builds a one-page table, has AI draft a plain summary, three talking points and member answers, then checks every figure.
• How Can AI Help School Administrators? Ten Jobs, With Prompts — Ten school office jobs AI handles well, each with a copyable prompt, a sample output and a check, plus the placeholder rule and where AI must not be used.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI enterprise privacy and data-controls pages, Anthropic privacy centre and pricing page, Microsoft Learn Copilot privacy page, Google Gemini for Education page and OpenAI custom GPT retirement FAQ (all checked 4 November 2026). Report structure is general practice, not statutory guidance, so check your governing board's own expectations.