School Parent Survey: Can AI Summarise Hundreds of Comments?

Yes. A school can write a short parent survey with two or three open questions, strip names from the export, and ask a chat assistant to group the comments into themes in batches of about 100. The AI sorts and counts; a person reads a sample of each theme against the raw comments before anything reaches governors.

The catch is that AI summaries sound confident even when they drift. A theme called "communication concerns" can quietly swallow forty comments about homework and ten about a bus route. The method below is built around catching that: a coding frame you control, a spot check with a pass mark, and a rule for the comments no tool should decide on.

Follow me on Instagram@sagnikteaches

Why free-text comments are worth the effort

Tick-box answers might tell you that 62% of parents agree the school communicates well. They do not tell you that half of the disagreeing parents mean "the weekly email arrives after the club has already started". Free text is where the fix lives. It is also where a head teacher's evening disappears, because 400 comments of two or three sentences each is roughly 25,000 words.

Connect on LinkedInSagnik Bhattacharya

That is the real job for AI here: reading volume, not making judgements. The same pattern appears in analysing customer feedback surveys with AI; this tutorial adapts it to the particular care a school needs with children's families.

Subscribe on YouTube@sagnikteaches

Writing questions that give you codable comments

AI cannot rescue a survey that asks vague questions. "Any other comments?" produces a pile of one-word replies and a few essays. Better questions point parents at something specific and ask for one example.

Three open questions are enough

Say a primary school with 420 pupils wants feedback after the autumn term. Keep the survey under five minutes: six rating questions, then three open ones.

• "What is one thing the school does well that we should keep?" Gets strengths you must not break while fixing other things.
• "What is one thing that made your child's term harder than it needed to be?" Gets specific frictions, not general moods.
• "If you could change one thing about how we contact you, what would it be?" Communication is nearly always the biggest theme, so give it its own question and the AI will not have to disentangle it.

Avoid double-barrelled questions such as "How do you find homework and reading books?". The answers mix two topics and no theme will be clean.

Add one routing field

Ask which year group the child is in, as a drop-down with bands (for example Reception to Year 2, Years 3 and 4, Years 5 and 6, or the equivalent grades). Bands rather than single years keep each group large enough that no parent can be identified. You can then ask the AI to compare themes by band without ever seeing a child's name.

Before and after: a rewritten question

Before: "Please share any feedback about the school." After: "Think about the last time you contacted the school. What happened, and what would have made it easier?" The first version tends to return "all good, thanks" and a few venting paragraphs; the second returns concrete accounts, such as "phoned at 8:40, left a message, nobody rang back by pick-up". That is an illustration, but the pattern is reliable: specific prompts produce comments an assistant can code.

Anonymity: what to settle before you open the survey

Parents are more candid when they trust that comments cannot be traced. Decide these four things in advance and tell parents the answers on the cover page.

1. Anonymous or confidential? Anonymous means you cannot identify the respondent at all. Confidential means you can, but promise not to share it. Choose anonymous for a general satisfaction survey and say so plainly.
2. Who sees raw comments? Name the roles, for instance the head teacher and one governor, not "staff".
3. Where the AI step fits. One sentence: "Comments are anonymised and may be grouped using an approved AI tool; no names are passed to it."
4. How long you keep the data. Pick a date, such as the end of the next academic year, and write it down.

Your data-protection lead should approve the tool and the wording. Free-text comments about children and families are personal data under most data-protection rules, even in an "anonymous" survey, so treat that conversation as part of the setup, not an afterthought.

Cleaning the export before the AI sees it

Download the responses as a spreadsheet. Work in a copy and follow this order.

1. Delete identifying columns. Remove email, IP address, timestamps and any name field. Keep a response ID (R001, R002) so you can trace a comment back later.
2. Search for names. Filter the comment columns for staff names, pupil first names and "Miss", "Mr", "Mrs". Replace with "[teacher]" or "[pupil]". A quick find-and-replace handles the staff names; pupils need eyes on the page.
3. Flag the red comments. Add a column called FLAG and type "S" next to anything about safety, bullying, discrimination, a named adult's behaviour or a medical matter. Move those to a separate tab. They go to the right person by a human route, whatever the survey promised, and they are not part of the AI summary.
4. Count what you have. Write down the number of rows per question. This is your check figure for later.

The data-handling habits here match those in running staff surveys and exit interviews with AI analysis, where the same identifying-column cleanup applies.

Building a coding frame before you ask for themes

If you simply paste comments and ask "what are the themes?", the assistant invents categories, and each batch gets different ones. A coding frame fixes that. Write it yourself from a first read of about 40 comments, or ask the AI to propose one from those 40 and then edit it.

A frame for a primary school might look like this (illustrative):

The "does not include" column matters more than it looks. It stops one theme growing into a dumping ground. Keep the frame to seven to ten codes. More than that and the counts become too thin to report.

Running the theme coding in a chat assistant

Use an account your school has approved, such as a business or education plan of ChatGPT, Claude or Gemini. OpenAI says ChatGPT Business and Edu content is not used for training by default; Anthropic says the same of Claude's business plans; Google says Gemini in qualifying Workspace editions, education editions included, does not use customer data for training without permission. Settings still differ by plan, so confirm on the vendor's own page before you start. One wrinkle: rating a reply with a thumbs up or down can send that chat to the vendor as feedback, so ask staff not to rate chats that contain comments. A free personal account is the wrong place for parent comments.

Paste the coding frame first, then 80 to 100 comments with their IDs. Use a prompt like this:

You are helping a primary school analyse anonymous parent survey comments.
Use ONLY the coding frame below. Do not invent new codes. If a comment fits two codes, give the main one first and the second after it. If nothing fits, use OTH.

CODING FRAME:
[paste table]

For each comment return one row: ID | code(s) | a 6-word paraphrase | sentiment (positive / mixed / negative).
Never rewrite the comment itself. Do not merge or drop any comment. At the end, state how many comments you received and how many rows you returned.

COMMENTS:
R001: ...
R002: ...

The "how many received, how many returned" line is your first safety net. If you pasted 94 and it reports 91, three comments went missing and you re-run the batch.

Sample output and what to fix

A realistic reply (illustrative) looks like this:

R017 | COM | emails arrive too late | negative
R018 | HWK, COM | maths homework instructions unclear | mixed
R019 | PRS | reception staff warm and helpful | positive
R020 | OTH | wants lunch menu changes | negative

R020 deserves a look. Lunch is a logistics matter in your frame (BLD includes "lunches"), so the assistant has misfiled it. Fix by tightening the frame ("BLD includes lunch menus and queues") and re-running the batch, not by hand-editing a single row, because the same slip will repeat across batches.

Counting, merging and comparing batches

Paste each batch's output into one sheet with three columns: ID, code, sentiment. A pivot table then gives theme counts by question and year band. For help with the sheet side, the ChatGPT Excel guide covers pivots and formulas step by step.

Let's run the numbers on the illustrative 420-pupil school. Suppose 214 parents respond (51%) and 187 write something on question two. The pivot shows:

• COM: 61 comments (33%)
• BLD: 38 (20%)
• HWK: 29 (16%)
• CLB: 24 (13%)
• SUP: 17 (9%)
• OTH: 18 (10%)

These sum to 187. Always check this. If your theme counts do not add up to the comment count, something was dropped or double counted. Note that a comment given two codes will count twice, so decide in advance whether you report "mentions" (can exceed 100%) or "main theme" (sums to 100%). For governors, main theme is easier to read.

Spot checking against the raw comments

This is the step that turns "the AI said so" into evidence. Do it for every theme you plan to report.

1. Sample. For each theme, pull 10 comments at random (the RAND function in a spare column, then sort). For a small theme with fewer than 10, read all of them.
2. Judge. For each sampled comment, mark Agree or Disagree with the code the AI gave. Be strict.
3. Score. If 9 or more of 10 agree, accept the theme. If 7 or 8 agree, tighten the frame and re-run that batch. If 6 or fewer agree, the code is not usable. Rebuild it.
4. Read the tails. Read every OTH comment and every comment tagged "negative" in the praise theme. These are where misreads hide.

On a trial with 7 codes, 10 comments each, that is 70 comments read: about 20 minutes. It is the cheapest insurance you will buy. For a deeper checklist of this kind, see the checklist for checking AI-written reports before sending.

A mistake that shows up in the check

Say the sample of COM comments contains "my son says the teacher shouts". A pattern-matching model can file it under communication, because the word "says" is a communication word. It is not a communication theme. It is a comment about a named adult's conduct, which belongs in the red tab. This is why flagging by hand comes before the AI step, and why you read the tails. Add a sentence to your prompt: "Any comment about an adult's behaviour towards a child: return FLAG and nothing else."

Pulling quotes without inventing them

Governors like a quote or two per theme. Ask the assistant to select comments by ID, never to write or tidy them:

For the COM theme, list the IDs of the 5 clearest comments that show different problems (timing, channel, length, reply speed, access). Return IDs only.

You then copy the quotes from the sheet yourself. Trim identifying details, mark any cut with "[...]", and never merge two comments into a composite. A paraphrase written by AI and displayed in quotation marks is a made-up quote, and parents who recognise their own words will notice.

The one-page report for governors

Governors want to know what parents said, how sure you are, and what you will do. A page is plenty. Ask the AI to draft it from your verified counts and chosen quotes, then edit:

Draft a one-page summary for school governors from the verified figures below. Use only these figures and quotes. Structure: response rate and method (2 sentences); top three themes with counts and one quote each; one thing parents praised; limits of the survey (one paragraph); three proposed actions as questions for governors to decide. Plain British English, no jargon, under 400 words.

FIGURES: [paste counts]
QUOTES: [paste]

Check the draft line by line against the sheet. Two limits belong on the page every time: the response rate (214 of roughly 420 families is a bit over half, and the silent half may think differently), and the fact that themes were grouped with AI and checked on a sample of 70 comments. That honesty is worth more than a smooth paragraph.

Turning themes into decisions

A list of themes is not a plan. For each top theme, write a one-line owner and date. For instance: COM, 61 comments, mostly about email timing; action: move the weekly email to Thursday morning from January; owner: office manager; review: a three-question pulse survey in February. Governors can then close the loop with a short "you said, we did" note to parents, which is what makes the next survey worth answering.

Variations: smaller and larger surveys

Not every survey is 400 comments. Adjust the method to the size.

• Under 60 comments. Skip the AI. Read them, code them in a sheet, and finish in an hour. The AI adds risk without saving time.
• A group of schools with 3,000 comments. Batch by school and year band, keep one shared coding frame, and sample 15 per theme per school. Assign one person to own the frame so codes stay consistent.
• A nursery or early-years setting. Answers are often short and warm. Use fewer codes (five or six) and treat everything about a key person's relationship with a child as a human read, in the spirit of how nurseries use AI to write daily reports for parents.
• A music or after-school provider. Parent comments tend to be about scheduling and progress. AI for a music school shows how parent-facing wording can be handled in a similar setting.

What goes wrong, and how to see it early

• Dropped comments. Counts do not match. Fix by always asking for received and returned totals.
• Theme drift between batches. Batch 2 uses "Communication" and batch 3 "Contact". Fix by forbidding new codes and pasting the frame every time, not relying on chat memory.
• Soft sentiment. Sarcasm ("lovely to hear about the trip the day after it happened") is read as positive. Spot check sentiment on a sample, and report themes rather than sentiment scores if the sample disagrees.
• Over-weighting loud voices. A 600-word complaint is one response. Counts treat it as one. Say so in the report.
• Re-identification. A comment such as "as the only parent governor in Year 6" is identifying even without a name. Read the cleaned file once more for roles and unusual details.

Hours and subscription cost for 187 comments

For 187 comments, expect roughly this (illustrative): an hour to write questions and set up the survey, 30 minutes to clean the export, 40 minutes running three batches, 20 minutes for the spot check, 20 minutes for the pivot, and 45 minutes to draft and edit the governor page. That is under four hours, against a day or more for reading and hand-coding every response. The subscription cost is the plan your school may already hold; otherwise a business seat on ChatGPT Business or Claude Team is $25 a user a month billed monthly ($20 billed annually), with at least two seats on both. Education pricing is set separately, so ask the vendor. The survey tool itself can be a free form builder.

If you want to map this into how your school gathers feedback across the year, from parent evenings to exit questionnaires, my AI implementation consultation covers exactly that kind of set-up; you can book a call when you are ready. The method above works without it.

A reusable checklist for the next survey

1. Three open questions, each pointing at something specific.
2. Year band as a drop-down; no names collected.
3. Cover note states the AI step and the retention date.
4. Export copy: identifying columns deleted, names replaced, red comments moved out.
5. Coding frame of seven to ten codes with "does not include" notes.
6. Batches of 80 to 100; received and returned counts agree.
7. Ten random comments per theme read; pass mark nine of ten.
8. Quotes copied by ID from the sheet, never rewritten.
9. Report states response rate, method and limits.
10. Actions with an owner and a date; a "you said, we did" note to parents.

Keep the coding frame and prompts with the survey file. Next year's survey will reuse them, and being able to compare this term's 33% communication figure with next year's is what turns a one-off exercise into a measure of whether the changes worked.

Questions school leaders ask before pasting parent comments into AI

Is it safe to paste parent comments into ChatGPT or Claude?

Only after removing names, pupil details and anything identifying, and only on an account that your school's data-protection lead has approved. Business and education plans usually do not train on your content by default, but check the plan's own terms. If a comment names a child, redact it first or leave it out of the AI step.

How many comments can an AI assistant handle at once?

A few hundred short comments fit in one chat on current assistants, but accuracy drops on very long pastes. Work in batches of 80 to 120 comments, ask for themes per batch, then merge. Batching also makes spot checks easier because each batch has its own count.

Should we tell parents that AI will read their comments?

Yes. One sentence on the survey cover note is enough: comments are read by staff, anonymised, and may be summarised using an approved AI tool, with no names passed on. Being open usually costs nothing and protects trust if a parent asks.

Can AI tell us which comments matter most?

It can count and group, but it cannot know which theme matters to your school. A single comment about a safeguarding worry outweighs fifty about the car park. Read every comment that mentions safety, bullying, or a named adult yourself, whatever theme the AI gave it.

Further reads

• Event Feedback Survey: Write It and Summarise the Answers With AI — The same write-then-summarise method for a feedback form after an event.
• AI Chatbot for a School Website: Parent FAQs From Your Handbook — Answer the questions parents keep asking with a handbook-based chatbot.
• How to Give AI Your Business Context So Answers Fit — Set up a reusable context block so theme coding stays consistent.
• Hotel Guest Feedback: Turn Post-Stay Surveys Into Fixes With AI — See how another sector turns survey comments into fixes.
• Using AI for Childcare Observations and Learning Journals — Handling children's information with care when using AI.
• School Trip Letter to Parents: The Whole Pack, Drafted by AI — One trip brief in, four documents out: parent letter, consent slip, medical form and kit list, with voluntary-contribution wording and a leader checklist.
• School Trip Risk Assessment: A Template Drafted With AI, Signed by You — A hazard-by-hazard template with a worked coach-trip example, a leader walk-through and a rule that AI never signs it off.
• Friends and Family Test Comments: Summarise Them With AI — A GP practice routine for coding Friends and Family Test free text into themes, a one-page report and a you-said-we-did poster.
• 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 page and ChatGPT Business pricing; Anthropic privacy centre (commercial products) and Claude pricing; Google Workspace generative AI privacy hub (all checked 4 November 2026); general survey-design practice.