AI for Nurseries: Drafting Risk Assessments Managers Can Sign Off

Yes. A childcare nursery can use AI to draft risk assessments for rooms, outdoor areas and outings, provided the draft starts from your own written description of the setting and the manager then walks the space, checks the rules and signs it. The AI saves typing time. It cannot see your building.

This page is about childcare nurseries, not plant nurseries. The risk is not that the draft looks bad, because AI drafts look tidy. The risk is that a tidy page gets signed without anyone testing it against the real room, the real stairs and the real children in the group. Everything below is built around that gap.

Follow me on Instagram@sagnikteaches

What the AI is good at, and what it cannot know

A chatbot is a fast first-draft writer for structured documents. Give it a description of a baby room and it will produce a hazard list, a table of controls and a review line in about a minute. That is worth having, because a blank page is the slowest part of the job for most managers.

Connect on LinkedInSagnik Bhattacharya

What it cannot know is everything you did not tell it. It has not seen the radiator cover that came loose in October, the gate that sticks, the gap between the garden fence and the shed, or the child who has just started climbing. It also cannot tell you which of its controls you do not actually have. If it writes "stair gate fitted at all times" for a room with no stair gate, the sentence reads exactly as confidently as a true one.

Subscribe on YouTube@sagnikteaches

Write the setting description first, in your own words

The quality of the draft depends almost entirely on the description you feed in. Spend ten minutes on it and the draft needs light edits. Spend thirty seconds and you get a generic page that fits no nursery in particular.

A useful description covers the room's age group, its size and layout, the fixed equipment, the door and gate arrangements, the sleeping set-up, what happens at transitions such as arrival, meals and pick-up, and what is outside the door. It does not include a single child's name, photograph, medical detail or family circumstance. The assessment is about the environment and the activity, so it never needs them.

Here is an illustrative description for a toddler room, the sort you might keep as a saved text file and reuse each term.

Setting: day nursery, 52 places, converted house, two floors.
Room: Toddler room (ages 2 to 3), ground floor, up to 12 children, 3 staff.
Layout: rectangular, one door to a corridor shared with the cloakroom,
one patio door to the garden, a low sink area, a book corner with a
carpet, a role-play kitchen, an open shelf unit and a small climbing
frame with a soft mat.
Doors: corridor door has a high handle and a finger-guard; patio door
is locked during indoor play and opened by staff only.
Routine: arrival 7:30 to 9:00, snack 9:30, garden 10:00, lunch 11:45,
sleep 12:30 to 14:00, pick-up from 16:00 in the corridor.
Notes: radiator under the window with a cover; heavy shelf unit fixed
to the wall; nappy changing in the room next door.
Do not include any child's name or personal details.

A prompt that produces something a manager can mark up

Ask for a table with blank columns for the parts only you can complete. That one instruction prevents most of the false confidence, because the AI cannot fill in a column it has been told to leave empty.

You are helping a nursery manager draft a room risk assessment for review.
Use only the setting description below. Do not invent equipment, rooms
or procedures that are not in it. If something is missing, write
"ask manager" rather than guessing.

Produce a table with these columns:
1. Hazard
2. Who could be harmed (age group or staff role, no names)
3. What could go wrong, in one plain sentence
4. Controls the description says are in place
5. Further controls to consider
6. Checked on walk-round (leave blank)
7. Requirement reference (leave blank)
8. Action, owner, date (leave blank)

Rank hazards from most to least likely to cause harm for this age group.
Finish with a list of questions you could not answer from the description.

[paste setting description]

Notice the last line of the prompt. The "questions I could not answer" list is often the most useful output, because it shows you where your own description was thin.

Reading a sample output the way a manager should

This is an illustrative excerpt of what the prompt above might return for the toddler room. Reading it critically is the skill.

Four things a manager should notice. First, the "controls in place" column quotes only what the description said, which is the point of the instruction. Second, "ask manager" appears where the description was silent. Third, the choke-hazard row is a good prompt but needs a decision: does your room have a choke tester, and who owns the check? Fourth, "a door alarm if not fitted" is a question dressed as a control. If the answer is no, that line becomes an action with an owner and a date, not a sentence in the finished document.

The walk-round that turns a draft into an assessment

This is the step a chatbot cannot do and the one inspectors and insurers care about. Print the draft, take a pen and a phone for photographs of anything you change, and walk the room at child height. Crouch. A hazard that is invisible from standing is obvious from a toddler's eye line.

1. Go down the table row by row. For each hazard, mark the "checked" column only when you have looked at it yourself.
2. Test each control. Pull the gate, try the door, tug the shelf, feel the radiator cover. If a control is described but absent, strike the sentence and write an action.
3. Add hazards the draft missed. There are always two or three: a rug edge, a cable, a lunchtime routine that crowds the corridor.
4. Delete hazards that are not there. A draft pads itself with generic items, and a long list full of irrelevant rows hides the three that matter.
5. Fill the requirement column from the source, not from the chatbot (see the next section).
6. Sign, date and set a review date. Tell the room team what changed and where the file lives.

A realistic mistake shows why step 2 matters. Say the draft states that wall fixings on the shelf unit are checked monthly. Nobody has ever recorded such a check. If you sign without asking, you have signed a procedure that does not exist. Turning that line into an action ("start a monthly fixings log, deputy, first entry Friday") makes the assessment honest and the nursery safer.

Verifying the rules: the early years framework and local guidance

Group childcare settings work to the statutory framework for the early years foundation stage that applies to them. The government's publications page says new versions of the frameworks came into force on 1 September 2026, with the requirements unchanged since they were first published for information in July 2026. In the version for group and school-based providers, paragraph 3.93 says you must be able to show how you are managing risks and must decide where written assessments on specific issues are helpful, and that an assessment should identify which parts of the environment are checked, when and by whom, and how each risk is removed or reduced. A footnote adds that health and safety rules may require written assessments once five or more staff are employed.

Paragraph 3.91 covers outings: the provider must assess the risks and hazards, the assessment must consider staff-to-child ratios, and whether it is written down is left to the provider. Treat that flexibility as no reason to skip paper. A signed outing assessment is the thing you can show a parent or an inspector who asks, and it is the version the staff on the trip actually read. Check how your inspector and local authority want it done, and read the current text yourself, because a chatbot will paraphrase from memory and may quote an older version.

Treat these as sources you open, not facts you ask for:

• The current statutory framework for your type of setting (group and school-based, or childminders), including the safeguarding and welfare section on staff ratios, premises and outings.
• Your local authority's early years team guidance, which often includes outing and visitor templates.
• Your insurer's conditions, especially for trips and any water, animal or transport element.
• Your own policies on lost children, medicines, allergies and emergency evacuation. A risk assessment should point to them, not contradict them.

Use the requirement column as a check on the AI, not a record of what it said. If the draft said a ratio or a rule and the source says something different, the source wins, and you note the difference in your own files so the next draft does not repeat it.

Outings: the draft that needs the most human input

A room assessment changes slowly. An outing assessment depends on the destination, the date, the group, the route and the weather, so the draft is only a skeleton. The same prompt pattern works, with a few additions: transport, crossing points, toilet access, head counts, the adult-to-child arrangement you plan to use, and what happens if a child becomes unwell or lost.

Say a nursery plans a morning visit to a local farm park with eight pre-school children and four adults. An illustrative outline of the manager's extra inputs:

• Recce: someone visits first, or phones the venue and asks for its own risk assessment and its hand-washing and animal-contact arrangements. The draft cannot do this.
• Route and transport: the walking route, crossings, and whether anything with a seatbelt is involved.
• Roles: a named lead, a named first-aider, who carries the emergency contact sheet and medicines, and who is doing head counts at each move.
• Children's needs: a separate, paper-only sheet kept by the key persons. Allergies, mobility and behaviour plans do not go in the chatbot.
• Stop rules: weather, illness numbers or a staff absence that cancels or changes the trip.

For a wider template and a worked trip example see the school trip risk assessment template, and for sessions in woods or fields the forest school site checks tutorial. Those pages cover venue-specific checks that this one deliberately leaves out.

Keeping children's information out of the chat

The cleanest rule is also the simplest: no child's name, photo, date of birth, medical information, family detail or safeguarding concern goes into a prompt, ever. Not abbreviated, not initials, not "a child in our room with a condition." A room assessment does not need them, and an assessment that depends on them belongs on paper in the key person's file.

When a draft needs to reflect an individual need, leave a placeholder row: "Children with individual plans: see key person file, reviewed with parents." That sentence is generic and safe. The detail lives where it was always kept. If you want a fuller briefing for staff, the one-hour AI briefing on children's data sets out the same rule for the whole team, and is it safe to use AI with children's data in a nursery goes into the reasons.

Two practical points on tools. Check the privacy settings so that chats are not used to train the model. On consumer plans this is a switch you turn off (in ChatGPT it is called "Improve the model for everyone" and is on by default for Free and Plus), while on business plans such as ChatGPT Business and Claude Team training on your content is off by default. And use a work account on a managed device, not a personal login on a staff member's phone, so that history stays under the nursery's control.

A before-and-after: one hazard line, rewritten

This is the kind of edit that makes the difference between a draft and a document.

Draft says: "Staff will ensure children are safe at mealtimes."

Manager's version: "Choking: all food for under-threes is cut lengthways, grapes quartered. Children sit down to eat with a member of staff at each table. Staff trained in paediatric first aid are on every shift (rota check, deputy, weekly). Allergy information is on the kitchen wall and the key person sheet."

The first sentence cannot fail an inspection, because it says nothing. The second can be tested. Each clause names something a person does, or a check someone performs, which is what turns a hazard list into evidence of how the nursery runs.

Three more rooms, three different drafts

The same prompt gives different results because the descriptions differ. Here is what changes in practice, using illustrative rooms.

Baby room. The sleeping set-up dominates. Your description should say where cots sit, how often checks happen and how they are recorded, and the draft will usually surface safe-sleep points. The manager's job is to compare it with the nursery's own sleep policy and the checks staff really make. The September 2026 framework updated the safer sleep rules for under-twos (paragraph 3.86 in the group version): babies aged 12 months and under must only be placed to sleep in a cot, babies under six months must have an adult in the same room for every sleep, and all children must be frequently checked and kept within sight and hearing of staff. A model trained on older text will not know those lines, so check them against the framework, not the draft. A draft that says "sleeping children are checked regularly" fails the test from the previous section: regularly is not a number. Replace it with the interval your policy sets and the place the check is logged.

Garden at lunchtime. Describe the surface, the fence height, the gate latches, the shed, any water butt or planters, and the plants. Plants are a good example of where AI is a weak authority. It may list a plant as harmless or harmful with equal confidence. Walk the garden with a plant guide or ask a qualified gardener, and remove anything you cannot identify. The draft only reminds you the question exists.

Pre-school room with messy play. Trays, sand, water, glue and paint raise slip, hygiene, allergy and eye-contact hazards. The draft is useful for a menu of "what could go wrong" ideas. It is not useful for allergy decisions, which depend on individual children and stay on paper with the key person. Note the contrast: the same page has parts where the AI adds value and parts where it must stay out.

After three or four rooms you will see a pattern. The model repeats the same ten or twelve hazards, and the value in your own review is the two or three that are unique to the room. Keep a short "what the walk-round found that the draft missed" note on each file. After a term, those notes become a better prompt: you add the recurring misses to the instruction ("always check radiator covers, door closers, and blind cords") and the next drafts start closer to the truth.

Time, cost and a one-room pilot

For a manager doing this for the first time, a sensible pilot is one room. Write the description (20 to 30 minutes), generate the draft (a few minutes), walk the room and edit (30 to 45 minutes), and sign. That is roughly an hour and a half for a document that might otherwise take two or three. Later rooms go faster because the structure repeats.

The tool cost can be nil. A free plan is enough for a draft with no personal data. Individual paid plans such as ChatGPT Plus and Claude Pro list at $20 a month. Team plans (ChatGPT Business, Claude Team) list at $25 a seat a month, or $20 a seat billed annually, with a minimum of two seats. Only pay if several managers will share instructions and want them controlled in one place. A saved prompt and a saved setting description in a shared document do most of the same work for free.

To check it worked, ask three questions after a month. Did the walk-round find things the draft missed? If not, people are probably signing without walking. Did any signed assessment contain a control that did not exist? If yes, tighten step 2. And did the review dates actually happen? The AI only drafts; the system around it keeps the paperwork alive. For other routine nursery writing that benefits from the same pattern, writing up your self-evaluation is a natural next job.

If you would rather map which of your nursery's repeated documents suit this pattern with someone, that is what my AI implementation consultation covers.

Questions nursery managers ask before trying this

Can I paste last year's risk assessment into the AI?

Yes, if it holds no names, photos or details about individual children or staff. A room assessment is about the space, so it rarely needs any. Strip anything personal first, and describe a child's need in the draft only as a placeholder line that you complete by hand.

Does the AI need to know the exact legal wording of the framework?

No, and you should not trust it to quote it. Tell it to leave a reference column blank, then fill the column yourself from the current statutory framework and your local authority or inspector guidance. A chatbot's memory of requirement numbers is the least reliable part of its output.

How often should an AI-drafted assessment be reviewed?

On the same schedule as any other assessment you hold: when the room, equipment, group or activity changes, after an accident or near miss, and at your set review date. The date you walked the room and signed is what matters, not the date the draft was generated.

Is a free chatbot good enough?

For a draft built from a setting description with no personal data, a free plan can do the job. Check the privacy settings so your chats are not used for training. If several staff will use one shared set of instructions, a paid team plan with shared projects is easier to control.

Further reads

• Can AI Write a Nursery Self-Evaluation Form From Your Evidence? — Turn your assessments and visit notes into inspection evidence.
• AI Chatbots for Nursery Enquiries and Visit Bookings — Handle parent enquiries without adding to staff workload.
• Choosing Nursery Software With AI: What to Compare — Compare the nursery software you may already hold.
• How Nurseries Use AI to Write Daily Reports for Parents — Another routine writing job that suits a human-checked draft.
• Tree Surgeon Risk Assessment Template: AI Drafts, You Sign Off — See the same draft-then-sign pattern in a high-risk trade.
• How to Draft Risk Assessments With AI and What a Human Must Check — The general method for any workplace, with the human checks.
• Can AI Write a Beauty Salon Risk Assessment? What It Must Not Skip — How to draft a salon risk assessment with AI by treatment, and the on-site checks, records and sign-off that a competent person has to do themselves.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the statutory framework for the early years foundation stage for group and school-based providers, in force from 1 September 2026 (paragraphs 3.86, 3.91 and 3.93) and the government's publications page; the ChatGPT and Claude pricing pages and OpenAI's data controls help article.