Why Leo won't diagnose your dog, and why that's the point

Leo, the AI companion inside Maggie, will not diagnose your dog. He will not name a condition, prescribe a medicine, change a dose, or tell you something is “safe” or “dangerous”. This is not a disclaimer added under a confident answer. It is enforced in the system: a rule-based check handles emergencies before any AI runs, and every reply passes checks that block diagnosis-shaped text. Here is how that works, and why we built it this way.

The problem with confident AI medical answers

General AI assistants will happily produce a diagnosis for a limping dog. The answer will be fluent, plausible and possibly wrong, and the assistant has no idea which. The World Health Organization warns that AI answers “can appear authoritative and plausible” and still be “completely incorrect”, especially on health. A wrong but confident answer is worse than no answer, because it can delay the vet visit that matters.

There is also a simple professional line. In the UK, the Veterinary Surgeons Act 1966 defines veterinary surgery to include “the diagnosis of diseases in, and injuries to, animals” and advice based on that diagnosis, as the RCVS guidance sets out. A diagnosis needs someone who can examine the dog. An app on your phone cannot.

We still think an AI can be useful for pet health. It just has to know its job: help you understand what is going on, keep your records straight, and get you to a professional at the right moment. Careful by design, not by disclaimer.

The guardrails, concretely

Emergencies never reach the AI. A rule-based check runs first on every message. If it spots an emergency, such as breathing trouble, collapse, seizures, bleeding, or poisons like chocolate and xylitol, the reply is instant and fixed: first-aid steps read from reviewed data, your own vet’s number, and the right poison line for your region. No model writes it, it does not count against your daily answers, and it is free for everyone. The check is tuned to err towards flagging, because missing a real emergency is the worst failure. A hypothetical such as “can dogs eat chocolate?” does not set off a false alarm.

Output checks block diagnosis-shaped replies. Every answer Leo writes passes through a validator before you see it. Replies shaped like a diagnosis or a medical verdict are blocked. This is enforcement, not a prompt asking the model nicely.

Your records are read back exactly. Ask “what vaccines has my dog had?” and the answer comes straight from the database with no AI generation. No model sits between you and your own data, so nothing can be misremembered.

Clinical content goes through vet review. First-aid steps, emergency contacts and the regional risk calendars go through a vet review workflow. The calendars for the US, the UK, India, Canada, Australia and Singapore are vet-approved today. Content still in review, such as the wellness-plan templates and part of the toxin guidance, carries a draft guidance label in the app, never passed off as signed off.

Numbers are calculated, not generated. Feeding and weight answers use fixed WSAVA calorie and body-condition tables applied to your dog’s logged data. The model does not do the arithmetic.

Confirmation before action. On Pro, Leo can create reminders, build vaccination plans and log events, and every change waits for you to confirm before anything saves.

What a good answer looks like instead

Say you write: “Bruno has been limping on his back leg since yesterday. What is wrong with him?”

A diagnosis-shaped reply would name a condition. Leo’s reply is shaped differently. He can:

1. Say plainly that this is one for the vet, and that he cannot tell what is wrong from a message.
2. Help you describe it well. Which leg, since when, whether it is worse after rest or after walks, whether Bruno will put weight on it, and anything that happened before it started.
3. Check the record. A past injury, a joint supplement, or a vet visit last month that is worth telling the vet about.
4. Help you act. Your vet’s number from your saved vets, and on Pro a vet-visit briefing you can share.

If Bruno cannot stand, is in obvious pain, or the leg looks badly hurt, the emergency path applies and you should call your vet now.

What Leo will do

Within those rails, quite a lot. He answers health, behaviour and nutrition questions grounded in your dog’s records, breed knowledge and WSAVA-derived guidance. He flags what is due and what has quietly slipped. He remembers facts you tell him, such as a fear of thunderstorms, and you can correct him or ask him to forget. On Pro he prepares a shareable briefing before vet visits: conditions, medicines, allergies, logged concerns, what changed since the last visit, and questions to ask. And he ends the right conversations the right way: “this one is for your vet, and here is what to tell them.”

To see how this compares with asking a general assistant, read why a general AI chatbot is the wrong place to ask about your dog.

The full detail, including what happens to your data, is on our safety page, and every capability is described on the Leo page.

Leo is an AI companion, not a veterinarian. For medical concerns, always consult your vet. That sentence is not small print for us. It is the spec.