Four weeks between grooms, and the lump behind Moose's elbow had not been there last time.
It was small. He pulled the leg away when Steph went near it, then let her carry on. Nothing else about the day was unusual, so she finished the groom, and while he was drying she typed what she had found into a chat window, because it was 3pm on a Tuesday, the vet's line was engaged, and the answer she was hoping for was "that sounds normal".
She got better than that: a calm, well-organised answer, three bullet points long, and the shape of it was common enough, often nothing, keep an eye on it, mention it next visit. So she mentioned it at pickup the way you mention a broken nail. Moose saw a vet on the Saturday instead of the Tuesday.
He came out of it fine. It cost his owner more than it needed to, cost him four days he did not have spare, and nothing anyone did that afternoon was reckless.
The AI risks in pet care are not the ones people write about. The risk is not that a chat window leaks your client list, real as that is. The risk is that it will answer a question about an animal it cannot see, in complete sentences with no hesitation in them, plausible enough to change what you do next.
There are five places it must never go: anything clinical, medication and diet, behaviour, incident reports, and any decision an animal's welfare turns on. In all five, the substitute is the same one sentence. You describe what you saw. You never name what it is.
I run Hound Health Bondi. Steph and Moose are worked examples we follow through this series rather than a real client and a real dog, but the afternoon is the ordinary one, and I have had versions of it.
The rule that covers every AI risk in pet care
Describing what you found is your job. Naming what it is is not.
You are the person with hands on that animal every four weeks, in good light, over every inch of it. Nobody else has that baseline, not the vet who sees him once a year and possibly on a bad day. What you noticed is valuable, and you should say it out loud.
The moment you turn it into a name, you have stepped from the thing you are qualified to do into the thing you are not. That step used to be hard, because you had to reach for the word yourself and you could feel yourself doing it. AI makes it frictionless, in your own tone, and it will never tell you a line has been crossed. It does not know the line is there.
The words to use when you have found something you must not name
This is the part worth keeping. Say it roughly like this, at handover, with the dog in the room:
"I'm not a vet, so I can't tell you what this is and I don't want to guess. What I can tell you is what I found: a small lump behind his left elbow, and he pulls the leg away when I go near it. It wasn't like that four weeks ago. I'd rather you had your vet look at it than have me tell you it's probably nothing. I've written down what I saw with today's date, and I'll send it to you before you go."
Five moves, and each is doing work. You disqualify yourself out loud, so nothing after it can be mistaken for a verdict. You give the observation, specific and located.
You give the change since last time, which is the piece only you have. You refuse to reassure, which is the hardest one and the whole point. Then you make a record, so the conversation exists somewhere other than two people's memories.
The written version, for the message you send afterwards:
There is a second reason to hold the line, and it is not the main one. The ACCC's position on the claims a business makes about its own services is that they have to be accurate, truthful, and based on reasonable grounds, and that it makes no difference whether you meant to mislead. That is about what you sell rather than what you say about somebody's dog, but it is one habit either way: a sentence you cannot stand behind does not become safer because a machine composed it. The law behind the safety spine is post four's job rather than mine, and the difficult conversation itself is a skill of its own.
Where the AI risks in pet care actually come from
A chat window cannot see the animal. It cannot put a hand on him, smell him, watch him walk across the room, or feel him tense a half-second before you reach the spot. That is where your judgement actually comes from, and none of it survives being typed.
It also has no idea what happened last visit unless you tell it. Everything it can see is the conversation in front of it, and what you typed is already an interpretation rather than the dog. Then it answers in the same steady register whether it is right or wrong, which is why these tools get things wrong.
No tool in this series is safer for this. Not the paid tier, not the newest model, not the one your cousin swears by. And by this level you are working with something that reads your files and does jobs rather than sitting in a chat box, so you re-read less than you used to. That is why the five lines below want writing down before you get comfortable, alongside the settings that decide what it may do without asking.
One: anything clinical
Lumps, limps, skin, ears, eyes, wounds, weight, appetite, and the dog who is just not himself today. Add the car-park version: an owner asking a direct question while you hold the lead, and you can feel yourself wanting to be useful.
The substitute is the observation note: what you found, where it was, when you first noticed it, and what has changed since last time. Four facts, all of them yours, none of them a diagnosis. Then the script above, then the record.
AI is genuinely useful on the note itself. Talking a handover note into your phone and having it come back in fixed fields is one of the best uses in this series, and that post draws the boundary in the same place: it may write down what you said, and it may not add a condition, a cause, or a treatment. Read the output for exactly that.
Two: medication, diet, and "is this safe for dogs"
Never. Not a dose, not an interaction, not a supplement, not a food or a topical, and not the one your best client has already decided about and just wants confirming.
This category is the most tempting because it does not feel like judgement. It feels like looking something up, and the answer arrives with units in it, which reads as authority. Nothing in that confidence tells you whether the number is right, and you cannot tell either.
The substitute is one sentence you never have to compose again. "I can't advise on anything you'd give him or put on him, and I wouldn't want to guess. Your vet can, and they'll have his history in front of them." Say it warmly, say it every time, and let it be boring. Boring is the objective.
Three: behaviour, which people underestimate
This is the one I would put first if the order were mine. Getting the clinical line wrong costs an animal time, which is what the top of this post cost Moose. Getting the behaviour line wrong can hurt a person the same afternoon, and a dog that bites someone in a salon has a much shorter list of futures than the one it woke up with.
It also slips past you, because a behaviour plan does not read like a medical question. It reads like process advice, the same genre as rostering or pricing, arriving structured, reasonable, and staged over four weeks. Apply that to a frightened dog and you may well get somebody bitten, and the dog surrendered by an owner who thought they were following something sensible.
Your handling notes are worth a great deal. What the dog did, what you did, what worked, what made it worse, how long it took to settle. Write those, hand them to the owner and a qualified behaviour professional, and let the plan come from whoever can watch the dog. Low-stress handling is a practice you build in the room, not a document you generate.
Four: incident reports, in your words, the same day
You write this one yourself. All of it.
An incident report is the document that gets read months later by somebody looking for the distance between what you wrote and what you remember: an insurer, a council officer, a lawyer, or an owner who has stopped being reasonable. A rough, honest account in your own voice survives that. An AI-smoothed one does not, because the smoothing is what gets noticed: tidy paragraphs, no gaps, no uncertainty, and a register that is not yours.
Never paste into an AI chat:
And never let it:
Write it the same day, before your memory reorganises itself into a version where you come out better. There is one legitimate use for AI here: questions rather than answers.
That last instruction is the whole value of it. The thing you leave out of an incident report is never the thing you were thinking about. For the structure of a proper report, incident reporting for a pet business has it.
Five: anything the animal's welfare turns on
Including the ordinary one nobody calls a welfare decision: whether to keep going with this groom.
Steph's policy is that a distressed or aggressive dog means she stops, and the owner gets a call. Her matting policy is that if the coat has to come off short she says so before she starts, not after. Both are judgements made by a person standing next to an animal, watching it, and both are exactly the sort of question a chat window will answer if you ask.
It will not refuse. It will give you a measured paragraph about a short break and reassessing, and you are now holding permission you did not earn from something that cannot see the dog shaking. Ask the question and you have already given away the decision.
Make the call, then use AI for what comes after: the message explaining what you stopped and why, the note on the file, the refund line. That is writing, and writing is what it is for.
What this actually reclaims for you
Nothing in this post hands you back an hour. What it gives back is narrower and worth more.
Steph's two groomers see around 40 dogs in a week. Say two of those, in an average week, have something on them worth raising with the owner. That is roughly 100 conversations a year, and what these five lines buy her is that not one contains a sentence she could not stand behind, in front of an owner, an insurer, or a vet who asks where it came from.
The number that matters most is four days. That is the gap between the Tuesday and the Saturday, and it is the only cost here that nobody can refund.
The right place for those five lines is not this page; you will not come back to it at 3pm with a dog on the table. It is the file your tool reads at the start of every session, which is the next post on the ladder.