Should a Dog Daycare Use AI to Write Incident Reports?

Yes, but only for wording. Staff write down what they saw in fixed fields while it is fresh, AI turns those fields into plain, neutral sentences without adding or changing a fact, and the person who saw the incident checks the draft against their notes and signs it before the owner receives it.

The real danger is confident invention rather than clumsy grammar. An AI assistant fills gaps with plausible detail: it will write that the dogs were "separated immediately" when nobody noted how quickly, or turn "teeth touched skin" into "a minor scuffle". A report that reads beautifully but says something different from what staff saw is worse than a rough one, because the owner, your vet and possibly your insurer will all treat it as the record.

Follow me on Instagram@sagnikteaches

So the rule behind everything below is that the notes are the evidence and the report is a tidy copy of them. If a sentence in the report cannot be traced to a line in the notes, it does not belong in the report. Wording, order and tone are the only things AI is allowed to touch. The same logic runs through the daily report cards a busy room sends home, where the stakes are lower and the habit is identical.

Connect on LinkedInSagnik Bhattacharya

The rule also settles which incidents AI should never draft. A growl or a bump with nobody hurt suits this routine well. Anything involving an injury, a vet or a distressed owner is written by the manager from the raw notes, and a later section shows exactly where that line sits.

Subscribe on YouTube@sagnikteaches

Three ways AI quietly corrupts an incident report

Start with what goes wrong, because it shapes every rule that follows. These are rough notes of the kind a handler types on a phone straight after a scuffle: untidy, but every line is something the writer saw or did. The dog names are invented for the illustration.

14:10 yard 2 by water bowl. Rex (large) + Pip (small). Pip went to drink, Rex stiffened, growled. Rex pushed Pip onto his side, held him ~2 sec. I shouted, pulled Rex off by harness. Pip yelped, up straight away, shook, back to water. Looked at Pip's neck 14:12: wet patch, like saliva, parted fur, no puncture or blood. Didn't see how it started. [staff B] on gate, didn't see either. Rex to rest crate, Pip to quiet room w/ water. Rang Pip's owner 14:25 no answer, voicemail. Rang again 14:50, spoke to owner, told them all of the above. Collecting at 4.

Ask an assistant to "write this up as a professional incident report" with no other rules, and the result tends to look like the draft below. It is an illustration of the pattern rather than a quote from one product, and the exact wording changes from run to run, but every failure in it is a common one.

At approximately 2:10 pm, Pip was attacked by Rex, an aggressive dog, near the water bowl in Yard 2. Rex bit Pip's neck and pinned him down. Staff intervened immediately and separated the dogs. Pip suffered no injuries and was checked by a vet. Pip's owner was notified promptly and was understanding. The incident was caused by resource guarding, and staff supervision was adequate.

Set the draft against the notes and seven problems appear in six short sentences.

Those errors fall into three patterns, and each needs its own defence.

1. Invented detail. Speed words (immediately, promptly, quickly), people who were never involved (the vet) and emotional colour (understanding, relieved). The defence is a fact sheet with a visible "not recorded", so there is no blank left for the assistant to fill.
2. Softened or hardened facts. "Pushed Pip onto his side for about two seconds" became "attacked" (harder), while "no puncture seen" became "no injuries" (softer). Both change the meaning. The defence is a rule that every verb and every injury word must match the sheet.
3. Verdicts. Causes, fault and adequacy. These are the most dangerous because they sound like analysis. The defence is a flat ban: the report records and does not conclude. Conclusions come later, from a manager, in a separate document.

The general background on why assistants do this, and how to spot it elsewhere in your business, is in the tutorial on AI hallucinations for business owners. In an incident report one invented sentence costs more than it does almost anywhere else.

Fourteen fields staff complete before AI sees anything

A free-text box invites the gaps that AI then fills. A form with fixed fields turns every gap into a visible blank that staff must answer with a fact or with "not recorded". Fourteen is more than a rushed handler wants, which is why most fields take one line and why the whole sheet should take about five minutes. That timing is my estimate, so time your own staff on a real one before you promise anyone a number.

The most useful habit on the sheet is the marker at the start of each line of the sequence. SAW means the writer watched it happen. TOLD means someone else said so, and the line names who. ASSUMED means the writer is inferring, and a good report contains very few of those. The marker travels into the AI draft, so the owner reads "staff saw" and "staff were told" instead of a smooth narrative that blurs the two.

Two fields cause most of the trouble. Field 10, owner contact, is where reports go soft. Record every attempt, including the ones that reached voicemail, because a 25-minute gap before the first call may be perfectly reasonable, but only if the report shows it. Field 7, the injury check, needs the phrase "none seen" rather than "no injuries", since a quick look through fur is not a vet's examination.

If your daycare software already has an incident form, use its fields as the fact sheet and paste them in. If it has only a single notes box, keep the sheet as a shared form or a paper pad and treat the software as the place where the final report is filed.

A prompt that tidies the wording and touches nothing else

Here is the same incident as a completed sheet, using the markers from the previous section. This is the only input the assistant gets.

Date: [date]   Time first seen: 14:10
Place: Yard 2, beside the water bowl
Dogs: Rex (file 0412, large, male); Pip (file 0877, small, male)
Present: [staff A] saw it; [staff B] was at the gate and did not see the start; 11 dogs in the yard
Sequence:
1. SAW: Pip walked to the water bowl; Rex was beside it.
2. SAW: Rex stiffened and growled.
3. SAW: Rex pushed Pip onto his side and held him there for about 2 seconds.
4. SAW: [staff A] called out and pulled Rex off by the harness.
5. SAW: Pip yelped, stood up straight away, shook himself and walked back to the bowl.
6. NOT SEEN: how it started.
Injury check: [staff A] looked at Pip's neck at about 14:12. Small wet patch, looked like saliva. Fur parted: no puncture or blood seen.
Actions: Rex to a rest crate; Pip to the quiet room with water. Times not recorded.
Vet: not contacted.
Owner of Pip: called 14:25, no answer, voicemail left. Called 14:50, spoke to owner, described the sequence above. Owner said: "I'll collect at four."
Owner of Rex: not yet contacted.
Follow-up: [manager] decided at 15:00: Rex and Pip in separate groups for their next two visits; both owners updated at end of day.
Photos: none.
Written by [staff A] at 15:05.

Above it goes the instruction. Save it as a note or a project instruction in your assistant so that every shift runs identical rules.

You are tidying a dog daycare incident report. Use ONLY the facts in the FACT SHEET below.

Rules:
1. Do not add, infer, estimate or change any fact: times, counts, durations, body parts, names, actions, injuries, or who said what.
2. If a field is empty or says "not recorded", write "Not recorded" in the report. Never fill it in.
3. Keep the marker on each statement: write "Staff saw", "Staff were told by [who]" or "Staff assumed", as the sheet does.
4. Describe behaviour, not character. Never use "aggressive", "attacked", "vicious", "bad", "dangerous" or "fight" unless the sheet quotes someone using the word.
5. Do not state or hint at a cause, a reason, fault, or whether supervision was adequate.
6. Describe injuries only as the sheet does. "None seen" stays "none seen"; it never becomes "no injuries".
7. Use the dogs' names from the sheet. Do not name or describe any owner except the one the report is addressed to.
8. Plain English, past tense, short sentences, 150 to 250 words, under the headings Date and time, Place, Dogs, What staff saw, Injury check, Actions taken, Vet, Owner contact, Follow-up.
9. After the report, add a section headed CHECK listing every place where the sheet was silent on something a reader might expect.

FACT SHEET:
[paste the completed sheet here]

Three of those rules do most of the work. Rule 2 stops the assistant from filling blanks. Rule 6 protects the difference between "none seen" and "none". Rule 9 turns the assistant's habit of noticing gaps into something useful: instead of quietly bridging them, it lists them, and staff answer each one with a real fact or leave it as "not recorded".

Run on the sheet above, a draft of this kind comes back looking like the one below. It is an illustration written for this tutorial, and I have left two realistic slips in it.

Date and time: [date], about 14:10.
Place: Yard 2, beside the water bowl.
Dogs: Rex (file 0412) and Pip (file 0877). Eleven dogs were in the yard. [staff A] saw the incident; [staff B] was at the gate and did not see how it started.
What staff saw: Pip walked to the water bowl, where Rex was standing. Rex stiffened and growled, then pushed Pip onto his side and held him there for about two seconds. [staff A] called out and pulled Rex away by the harness. Pip yelped, stood up straight away, shook himself and walked back to the bowl. Both dogs settled quickly in their separate areas.
Injury check: At about 14:12 [staff A] looked at Pip's neck and saw a small wet patch that looked like saliva. With the fur parted, no puncture or blood was seen. Pip appeared unharmed.
Actions taken: Rex was taken to a rest crate and Pip to the quiet room with water. Times were not recorded.
Vet: Not contacted.
Owner contact: Pip's owner was called at 14:25 (no answer; voicemail left) and again at 14:50, when staff described the sequence above. The owner said they would collect Pip at four o'clock. Rex's owner has not yet been contacted.
Follow-up: [manager] decided at 15:00 that Rex and Pip will be in separate groups for their next two visits. Both owners will be updated at the end of the day.

CHECK
- Time the dogs were separated: not recorded.
- Whether Pip's behaviour changed after the incident: not recorded.
- Who, if anyone, examined Rex: not recorded.

What I would change before anyone signs it:

1. "Both dogs settled quickly in their separate areas" is not on the sheet. Delete it, or have the handler add a real fact such as the time Rex lay down in the crate.
2. "Pip appeared unharmed" is broader than the sheet, which says only that no puncture or blood was seen on the neck. Put the sheet's words back.
3. The CHECK list is the most valuable part. Staff answer the three items from memory or by asking a colleague, and the sheet is updated before the final version is produced. Anything nobody can answer stays "not recorded", and that is a perfectly respectable thing for a report to say.

A second pass can catch slips like the two above. Paste both the sheet and the draft into a fresh chat with this instruction:

Below are a FACT SHEET and a REPORT. List every statement in the REPORT that the FACT SHEET does not state, quoting the sentence and saying what the sheet says instead. If you find none, reply "none found" and name the three sentences most worth checking by hand.

FACT SHEET:
[paste]

REPORT:
[paste]

Treat this as a helper rather than a guarantee. It is another AI reading text that an AI wrote, so "none found" means "now do the human read-back", not "safe".

A swap list for no-blame wording about dogs

No-blame wording does not mean vague wording. It means describing what happened precisely enough that nobody needs to add a verdict. A dog's owner reading "Rex attacked Pip" will argue about the word. An owner reading "Rex pushed Pip onto his side and held him there for about two seconds" has only the facts to argue with, and they are the same facts your staff will give if asked.

Notice how many of the replacements are numbers, quotations or lists. Adjectives are where opinions hide, and opinions are what an incident report must not carry. If a replacement still sounds harsh, that is usually because the event was harsh, and softening it would be the very change this routine exists to prevent.

Put the swap list into your prompt as an extra rule if your assistant keeps reintroducing a particular word. Two or three examples of "instead of / write" inside the prompt work better than a long list of banned words.

Telling the owner: call first, write second

The written report is the second conversation, never the first. For anything where a dog's skin may have been touched, a dog is limping or bleeding, or a vet is involved, a person phones before any document is sent. The order below keeps the call and the paper consistent.

1. Finish the fact sheet, or at least the sequence, injury check and actions fields, before you ring. You will want the facts in front of you.
2. On the call, give the facts as written: what was seen, what was checked, what has been done. Offer no cause and make no promise you cannot keep.
3. Record the call on the sheet straight away: time, who you spoke to, what you said, what they said in their own words. Do this before the AI draft is produced, so the draft includes it.
4. Send the written report afterwards. It must say nothing the call did not, and the call must have said nothing the sheet does not.
5. If nobody answers, log each attempt and leave a short, factual voicemail. Two attempts with times beat a vague "we tried".

For a minor incident where the owner has been told by phone, the follow-up message can be drafted by AI from the signed report alone, never from memory. A version for Pip's owner, with the other dog described by size rather than by name or owner, might read as below. The placeholders are deliberate; staff fill them in.

Hi [first name], following our call at 14:50, here is the written note on Pip's afternoon.

At about 14:10 in Yard 2, another dog, a larger male, stiffened and growled as Pip reached the water bowl, then pushed Pip onto his side for about two seconds. [staff A] pulled the other dog away by the harness. Pip yelped, stood up straight away and walked back to the water. At about 14:12 [staff A] looked at Pip's neck and saw no puncture or blood. Pip is resting in the quiet room until you collect him at four.

We did not see how it started. Pip and the other dog will be in separate groups for their next two visits, and we will message you at the end of the day to say how he got on.

Two things stay out of the AI's hands altogether. The first is any apology that admits fault, or any offer to cover a vet bill. Those are decisions for the person who owns the business, ideally after a word with the insurer, and a drafting tool will happily write a generous one without asking. The second is advice about the other dog. Staff can describe what was seen; whether the other dog is suitable for daycare is a separate decision, made by the manager, with the dog's own file and assessment in front of them.

Your insurance policy may set a time limit for reporting an incident that could lead to a claim, so ask your insurer what counts and write the answer into your procedure.

Four tiers: which incidents AI may draft

Tiers stop the debate on a bad afternoon. The version below is a suggested policy to adapt, not an industry standard. What matters is that the boundaries are agreed in advance and staff do not have to decide in the moment whether to trust a draft.

Many trainers and behaviour professionals grade bites on Dr Ian Dunbar's six-level dog bite scale, which assesses severity from the wound itself: level 1 is aggressive behaviour with no skin contact by teeth, level 2 is teeth touching skin without a puncture, levels 3 and 4 are one to four punctures from a single bite graded by depth, and levels 5 and 6 cover multiple-bite attacks and a death. Because it grades evidence rather than character, it suits a fact-first report. If a vet or behaviour professional has graded an incident, copy their grade and their name onto the sheet. Do not let staff, and certainly not an AI, assign one.

When a staff member is the one who is hurt, the same fact-first logic applies to the staff injury record, which belongs in your workplace incident report rather than in the dog's file. Keep the two separate, and cross-refer by incident number.

Photographs follow the same rule as words. Take them, attach them, and record who took them and when, but do not ask an AI to describe an injury from a photo. An AI's reading of an image is not a dependable judge of a wound, and a confident description in the file is worse than no description.

The read-back and a trap test before you trust the prompt

The read-back is the step that makes the whole routine safe, and it takes about two minutes. The person who saw the incident does it, sitting down, with the fact sheet next to the draft. The tutorial on a five-minute fact-check routine for AI output covers the general habit; for incident reports the specific checks are these.

1. Every number in the draft (times, counts, seconds) appears on the sheet.
2. Every name, file number and initial matches.
3. Every "not recorded" survived, and none has been replaced with a plausible fact.
4. Every SAW, TOLD and ASSUMED marker is still attached to the right statement.
5. No word describes a dog's character, and none describes a cause, fault or adequacy.
6. The injury wording matches the injury field exactly.
7. The writer can read the draft aloud and say "that is what I saw", then sign and add the time.

Keep three versions together in the incident file: the raw notes, the AI draft and the signed final. If anyone later asks what the tool changed, you can show them, and you can show that a person approved it.

Before staff start using the prompt, give it a trap test. Take the completed sheet above and delete three facts: the headcount, the time of the injury check, and who looked at Pip's neck. Run the prompt three times. If any of the three drafts supplies a headcount, a time or a name, the prompt needs a stronger rule, not more hope. A second trap is a sheet in which a handler has written "Rex went mental" in the sequence. A good result keeps the meaning, rewrites it as behaviour that the sheet supports (or flags the gap in CHECK), and does not invent details to make the sentence respectable.

A quick sum helps keep expectations honest. Say a daycare has 20 incidents a month, and writing each from scratch takes a handler 12 minutes. With a fact sheet (5 minutes), an AI draft (1 minute) and a read-back (2 minutes) it takes 8, so the monthly saving is 80 minutes. Those times are assumptions; replace them with your own. The point of the exercise is that the time saved is modest, and the real gain is that every report reads the same way and says only what was seen.

What to paste into the AI and what to keep out

The assistant needs the incident, not the identities around it. Paste dog names, file numbers and staff initials. Leave out owners' names, phone numbers and addresses, payment details, and the other dog's owner altogether. Where a vet's comment matters, paste only the words the vet used about this incident, not the dog's wider medical history.

Choose the account carefully. Business plans from the big assistant providers (ChatGPT Business, Claude Team, Gemini in Workspace, Microsoft 365 Copilot) do not train on your content by default, while consumer plans give you a switch in the privacy settings that you must check yourself. Deleting a chat also does not erase the provider's copy straight away, which is one more reason to keep identities out of it.

If your daycare software offers its own AI-written incident reports, ask three questions before switching it on. Does it work only from the fields I fill in? Can I see its draft beside those fields? Does it keep the original notes? A single notes box with a "make it professional" button has none of the safeguards above, and it is the easiest way to end up with a polished report nobody can trace.

Spotting patterns across finished reports

Once reports are accurate, AI earns its keep in a different place: reading the pile. A month or a quarter of signed reports is data, and the fields you already fill in (place, time, tier) are exactly the columns an assistant can count. The assistant is analysing verified records here, not producing evidence.

Take 30 reports from a quarter, entered in a sheet with columns for date, time, place, tier and the trigger as recorded. If you score dogs at assessment, add that score too, which lets you check whether incidents bunch in one band; the temperament test score sheet tutorial shows how to build the scoring.

Here is a table of incident reports (date, time, place, tier, trigger as recorded). Count incidents by place, by hour band and by tier. Then list the three place-and-time combinations with the most incidents. Do not suggest causes. Do not guess missing values: list rows with blanks separately.

A reply might look like this (illustrative counts that add up to the 30 rows):

• By place: water station 11, arrival gate area 7, indoor play room 6, other yard areas 6.
• By hour band: 14:00 to 15:30 had 13; 07:30 to 09:00 had 8; the other 9 were spread across the day.
• By tier: tier 1 had 17, tier 2 had 11, tier 3 had 2.
• Top combinations: water station between 14:00 and 15:30 (7), arrival gate area between 07:30 and 09:00 (5), indoor play room after 16:00 (3).

What you would do with that is practical and cheap: a second water station, a staggered arrival time, or a smaller afternoon group. Two cautions apply. Check the counts yourself with a pivot table, because assistants can miscount rows. And treat the pattern as a lead rather than a cause, since the sheet records where and when, not why.

Questions daycare owners ask before trying this

Do I have to tell owners that AI tidied the wording?

A signed report is still your staff's account, so the main duty is accuracy rather than disclosure. A short line at the foot, such as 'Staff notes, wording tidied by software, checked and signed by [name]', costs nothing and answers the question before it is asked. If you are unsure whether any disclosure rule applies to your records, ask your adviser.

Can staff dictate the notes instead of typing them?

Yes, dictation is a quick way to fill the fact sheet on a phone. Speech tools often mangle dog names, times and numbers, so read the transcript and correct it before it goes near the AI. Treat the corrected transcript as the raw notes and keep it alongside the draft and the signed final.

What if two staff members saw it differently?

Record each person's account separately under their initials and let the report say so. Do not let AI merge two accounts into one smooth version, because the merge hides the disagreement and may quietly pick a side. A manager can then deal with the difference in writing, separately from the incident report.

How long should incident reports be kept?

That depends on your insurer's conditions and the rules where you operate, so ask both and write the answer into your procedure. Whatever the period, keep the raw notes, the AI draft and the signed final together, so you can show exactly what the tool changed and that a person approved it.

Further reads

• Dog Boarding Intake Form: Vaccines, Feeding and Medication Fields — The intake fields that should already sit in each dog's record.
• Writing Pet Owner Emails and Treatment Explanations With AI — Writing to owners about health matters without overstating anything.
• Can AI Draft a Dog Grooming Consent Form? What to Cover — What a consent form should cover before problems start.
• AI for Dog Trainers: Session Notes and Homework Plans for Clients — Another notes-to-writing routine where the facts must stay put.
• How to Catch Made-Up Figures in AI-Drafted Proposals — A method for finding numbers the AI made up.
• Dog Trainer Intake Form: Behaviour History Questions With AI — The behaviour history questions a dog trainer's intake form needs, from bite history to medical flags, with escalation rules and AI build and summary steps.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: report fields adapted from published pet-care incident form templates and a kennel-software knowledge base; business-plan data-handling defaults from the OpenAI, Anthropic, Google and Microsoft help pages; Dr Ian Dunbar's Dog Bite Scale in the official version published by the Association of Professional Dog Trainers. Prompts and sample outputs are illustrations written for this tutorial.