Use AI to tidy what you saw, never to decide what it means. Jot the facts (what the child did and said, with whom, for how long), let AI shape them into your observation format and suggest possible links to your framework, then the key person confirms the links and writes the next step. Photos stay out of general AI tools.
The quality risk is interpretation creeping in. Models love praise, so a note that says "threaded four large beads, tried a fifth for about a minute" comes back as "demonstrated excellent fine motor control and great perseverance". That's less useful to the next practitioner and less honest to parents. A good observation records evidence; the professional judgement about it belongs to the person who watched.
What AI can and can't do in observation work
- Can: correct spelling and grammar; turn a rushed note into full sentences in your house format; suggest which areas of learning an observation might relate to, if you give it your framework's wording; shorten long entries; turn several observations into a parent-friendly summary; translate a summary for a family whose first language differs from the setting's.
- Can't: assess development or decide whether a child is on track; choose next steps without the practitioner's knowledge of the child; see or interpret photos safely; handle safeguarding information at all.
Where the AI should live
The safest place is inside the learning journal platform you already use, because the data doesn't leave a system your setting has already assessed. Several platforms now include writing help. Famly, for example, includes a writing assistant called Sidekick on all its packages at no extra charge; it checks spelling and grammar, changes tone (professional, polite, observant or conversational) and has a "make it shorter" option. Ask your own supplier what they offer, where it processes data, and whether it's covered by your data processing agreement.
| Option | Suitable for | Rules |
|---|---|---|
| Writing assistant built into your learning journal platform | Everyday observations and posts to parents | Check the supplier's data terms once; then use freely for text |
| Business AI plan (ChatGPT Business, Claude Team, Gemini in Workspace, Microsoft 365 Copilot) | Summaries, templates, planning, parent letters | Text only; first names or initials; no photos; no safeguarding or health details |
| Personal or free AI apps on staff phones | Nothing involving children | Not for observations, even with names removed |
Business plans don't train on your content by default, which is why they're on the list and personal apps aren't. For the wider question of what's safe, whether it's safe to use AI with children's data in a nursery goes further, and your setting's data protection lead should sign off the approach before anyone starts.
From a jotted note to a finished observation
In a pre-school room of 24 children (an illustrative one), each key person aims for two meaningful observations per key child each week. Written from scratch at the end of a session, each took about 10 minutes; staff often caught up at home. Here's one from start to finish.
The key person's jotted note, typed on the room tablet during tidy-up time:
J (3y 5m). Block area w/ A. Built tower, fell 3x. 3rd time said
"too wobbly, need big ones bottom". Swapped big blocks to bottom,
tower 9 high. Counted to 7 then "lots". Asked A to hold it. ~6 min.The prompt used, saved as a template (text only, initials only):
Rewrite this note as an observation for a learning journal.
Rules:
- Record only what was seen and heard. Keep direct quotes exactly.
- No praise words or judgements (e.g. "excellent", "confident",
"resilient") unless they are in my note.
- Past tense, 60-100 words, warm and factual.
- Then list POSSIBLE links to these areas of learning, quoting the
evidence from the note for each: [paste your framework's areas]
- Do not suggest a developmental stage or age band.
- Do not write a next step; I will.The output (illustrative):
J built a tower in the block area alongside A. The tower fell three times. After the third fall, J said, "Too wobbly, need big ones bottom," and moved the larger blocks to the base. The tower then reached nine blocks high. J counted the blocks aloud to seven and then said "lots". J asked A to hold the tower steady while adding blocks. J stayed with the activity for about six minutes.
Possible links: problem-solving ("need big ones bottom", then changed the base); early number (counted to seven); communication ("too wobbly"); working with others (asked A to hold it).
The key person's edits took about two minutes. She kept three of the four links, dropped "communication" as the weakest, and wrote the next step herself: "Offer J blocks of mixed sizes and ask 'which will go at the bottom?' before building, to see if J predicts rather than fixes." Total time about four minutes instead of ten. Across 12 key children at two observations each, that's roughly two and a half hours a week back for one practitioner, most of it time that had been spent at home.
Linking to areas of learning without letting AI assess
The prompt above asks for "possible links" with quoted evidence, and that wording is deliberate. Paste your framework's own statements into the prompt (or a Project on a business plan) so the model uses your language rather than something it half-remembers. Then treat its suggestions as a checklist to accept or reject, not a verdict.
What to reject: any link where the quoted evidence is thin ("shows understanding of number" from a child saying "lots"), and any link that sounds like a developmental judgement ("working within the expected range"). If the model starts producing age bands or milestone language, remove it and strengthen the "do not" line in the template.
A second illustrative observation shows the sorting in practice. The note: "M at water tray, poured from big jug to small, overflowed, said 'that one's full up'. Tried again slower. ~4 min." The model returned three possible links: "Mathematics, capacity: 'that one's full up', then poured more slowly"; "Physical development: shows good hand-eye coordination"; and "Personal and social development: shows age-appropriate independence". Keep the first, because the evidence is quoted and it fits. Reject the second, which has no quote behind it, since the note never says how steady the pouring was. Reject the third outright: "age-appropriate" is an assessment, and the key person hadn't made one. Two of three rejected is normal in the first weeks and a good sign that someone is actually reading.
Next steps that come from the child, not a template
AI-suggested next steps tend to be generic enough to fit any child: "continue to provide opportunities for J to explore building." A next step should come from what you know about this child's interests and what you've just seen. Compare:
- Generic: "Support J's counting skills through play."
- Specific: "J counted to seven then said 'lots'. At snack time, count the cups together to ten, and pause at eight to see if J carries on."
If you do ask AI for ideas, ask for three options based on the child's interests you describe, choose one, and rewrite it in your own words. The key person is still the author.
Turning a term of observations into a journal summary for parents
Summaries for parents are where AI saves the most time. Paste a term's observations for one child (initials only, text only) and ask for a parent-friendly summary under 200 words: what the child has enjoyed, two or three things they've started doing, and how parents might build on it at home. An illustrative line from one: "J has spent a lot of time in the block area this term and has started changing his plan when things go wrong, like moving the big blocks to the bottom so his tower stays up. At home, you could build together and ask him what he'll put at the bottom first."
Check every sentence against the observations; summaries are where the model is most tempted to generalise ("J loves maths"). For daily notes rather than termly summaries, how nurseries write daily reports for parents with AI covers a lighter routine.
A before and after from the same term. The first draft said: "J is a natural leader who loves organising his friends and is showing real talent for maths." The observations behind it were one tower where J asked A to hold it steady, and one count to seven. The corrected sentence reads: "J has started asking friends to help with his building, like when he asked A to hold his tower steady, and he likes counting the blocks as he goes." A simple rule stops the first version: ask the model to put the date of the supporting observation in brackets after each sentence, check them, then delete the brackets before the summary goes to parents.
How the routine changes from baby room to childminder
The baby room. Babies don't give you quotes, so notes are about movement, gaze and sound: "reached for the ball with right hand, rolled onto side, babbled 'ba-ba' when it rolled away". AI is useful here mainly for structure. Watch for it interpreting feelings ("was frustrated") that you didn't note; describe what you saw instead.
A settling-in note shows how easily this happens. The jotted version: "R cried when mum left, about 4 min. Stopped when given soft bunny. Looked at door several times over next 10 min." The draft came back as: "R was very upset and anxious about the separation but settled well once comforted." Every judgement in that sentence ("very upset", "anxious", "settled well") is the model's, not the practitioner's. The fixed version keeps to what was seen: "R cried for about four minutes after her mother left and stopped when given the soft bunny. Over the next ten minutes she looked towards the door several times." A parent reading the second version learns more, and the next practitioner has something to compare against tomorrow.
A toddler room with high staff turnover. New staff often write very short or very long observations. A shared template in the platform's writing assistant, plus the "no praise words" rule, brings entries to a consistent standard quickly, and the room leader can see at a glance who needs coaching.
A childminder working alone. With no colleague to read your entries, AI is a useful second reader for spelling and clarity. Keep it inside your journal app if possible. If you use a business AI plan instead, keep a note of the tool in your privacy notice for parents.
How errors slip into journals, and how they show up
Most mistakes in AI-assisted observations are small and easy to catch if you know where to look.
- Two children merged. A practitioner pastes notes for two children into one prompt to save time, and the output credits one child with the other's quote. It shows up when a parent says "that doesn't sound like her". Rule: one child per prompt, always.
- Roles swapped. In a paired activity, the model decides who led. The note said A held the tower; the draft says A suggested the big blocks. Check every sentence involving another child against your note.
- Quotes "corrected". "Need big ones bottom" becomes "I need the big ones at the bottom". The child's actual words are evidence of their language, so the template insists quotes stay exact, and the key person checks.
- Signs and gestures reworded. For a child who communicates with signs or gestures, the note "J signed 'more' twice, then pointed at the jug" can come back as "J asked for more water". The sign is evidence of how J communicates, just like a spoken quote. Write it in the note in a fixed form, such as [signed "more"], and add a line to the template: "Keep anything in square brackets exactly as written."
- Translations that are too formal. A summary translated for a family whose home language differs from the setting's can come out stiff or oddly formal. Ask a bilingual colleague or the family themselves to look over the first one, and adjust the prompt from their comments.
Removing names doesn't make a note anonymous if the details identify the child; anonymising data before it goes into AI explains the difference.
Rules that don't bend, whatever the tool
- No photos or videos of children into general AI tools. Ever.
- First names or initials only; no dates of birth, surnames or addresses.
- Nothing about safeguarding concerns, injuries, health conditions or family circumstances goes into AI. Those follow your safeguarding and records procedures only.
- No additional needs or diagnoses in prompts. Describe the observed behaviour if needed, nothing more.
- The key person reads and approves every entry before it's published to parents.
- Staff use only tools the setting has approved, on setting accounts.
Put these rules in the setting's AI policy and go through them at a staff meeting with real examples; rules that live only in a document get forgotten by the second week.
Checking quality after the first month
The manager or room leader samples ten observations from different staff and asks four questions of each: Is it factual, with no invented praise? Is any quoted speech exact? Is the next step specific to this child? Did the practitioner edit the AI's links, or accept all of them? If links are always accepted unchanged, that's a sign people have stopped reading them. Ask staff too: time saved per observation, and whether writing at home has stopped. If you're reviewing your software at the same time, choosing nursery software with AI features lists what to compare.
An illustrative first-month result: of ten sampled observations, eight were purely factual and two had kept a praise word ("amazing", "super"); nine had exact quotes; six had next steps specific to the child; and in seven, the AI's links had been accepted unchanged. The last figure is the one to act on. The room leader sat with the two practitioners whose links were never edited and went through three of their entries together, which took twenty minutes and changed the next month's sample more than any rule in the policy.
It's also worth running the sum for the whole room, so the checking has a budget. At two observations a week for 24 children, the room writes 48 observations; saving about six minutes on each is nearly five hours a week across the team. The monthly sample takes the room leader about half an hour. Keep that ratio in view: if checking ever starts to take as long as writing did, the template needs tightening, not more checking.
What practitioners ask about AI and observations
Can AI decide whether a child is on track for their age?
No. Judging where a child is in their development is a professional judgement made by the practitioner who knows the child, drawing on many observations, conversations with parents and your framework. AI has no way to observe the child and will produce a confident-sounding judgement from a few sentences. Use it only to tidy the wording of a judgement you have already made.
Is it acceptable to dictate observations into a phone app?
Dictation inside your learning journal platform, or into a business AI tool approved by your setting, is usually fine for text. Avoid consumer voice assistants and personal apps, keep children to first names or initials, and never record audio of the children themselves for AI processing. Check your setting's data policy first.
Will parents mind that AI helped write their child's journal?
Most parents care that observations are accurate, specific and written by someone who knows their child. A short line in your parent handbook or privacy notice saying staff use AI tools to check spelling and structure, and that key persons write and approve every entry, is honest. If a parent objects, offer to write their child's entries without AI.
Further reads
- AI Chatbots for Nursery Enquiries and Visit Bookings — Using AI at the front door for enquiries and visit bookings.
- How to Classify Business Data Before Using AI Tools — Sort which setting documents can go into which tools.
- How to Roll Out an AI Policy So Staff Actually Follow It — Introduce the observation rules so staff actually follow them.
- AI Hallucinations Explained for Business Owners: Causes and Fixes — Why AI adds details and praise that were never observed.
- Tone of Voice Examples: AI Copy Before and After Editing — Before-and-after edits that sharpen your journal voice.
- How to Write an AI Usage Policy for Your Small Business — The one-page staff policy your observation rules belong in.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Famly's page on its Sidekick writing assistant (checked September 2026); OpenAI, Anthropic and Google help pages on business plan training defaults. Observations, children and timings are illustrative.