AI triage risks on the phone: what an AI should never decide

What an AI receptionist can safely do

An AI receptionist can safely handle any job where the answer is the same for every caller and has nothing to do with the pet's condition. Once the answer depends on how the pet is doing, it's not a job for software.

AI triage risks on the phone: five failure modes

The AI triage risks on the phone tend to show up in five ways, and the first is false reassurance. A caller describes a dog retching with nothing coming up. The AI hears "vomiting" and books a routine slot for Thursday. But signs of bloat, poisoning or breathing trouble need care now.

Second, missed keywords. People who are frightened don't use clinical words. "She's breathing funny" and "he got into the garage" can both be emergencies, and neither sounds like one to a system listening for textbook terms. Third, over-confidence: language models produce smooth, fluent answers even when they should say "I don't know", so something plausible but wrong comes out sounding like it knows.

Fourth, slow escalation. A system that insists on collecting a full history before transferring burns minutes that matter. And fifth, dead ends after hours. If the clinic's closed and the AI only takes a message, an emergency sits there until morning.

Design rule: when in doubt, escalate. A false alarm costs a minute of staff time; a missed emergency can cost a life.

What Ontario's rules imply for AI on the phone

No Ontario rule is written specifically for AI receptionists, but the existing standards all point the same way.

Advice needs a vet and a relationship. The CVO's telemedicine standard (revised June 2022) says practising veterinary medicine remotely is only permitted within a veterinarian-client-patient relationship, and only to the extent possible without a physical exam. General, non-specific information is a different thing from specific advice. An AI that tells a caller what to do about their pet's symptoms lands on the wrong side of that line.

Clients also have to know how to get after-hours care. Regulation 1093 under the Veterinarians Act requires vets to tell clients how they can access services outside regular hours, so an after-hours AI should give that information every time. Remote communications belong in the record, too: the CVO says information collected via telemedicine becomes part of the medical record, and its records guidance expects client communications to be documented.

And be upfront that it's AI. The Office of the Privacy Commissioner of Canada's generative AI principles say public-facing tools should make clear that people are dealing with AI, and that organizations stay accountable for the outcomes.

How urgent calls should reach a person

Urgent calls should reach a person immediately, and escalation is the single most important feature of any AI phone system. You should be able to explain it in one sentence. Here's a practical design:

1. The AI says it's the clinic's AI assistant at the start of the call.
2. Any urgent word or phrase, or any mention of a sick or injured pet that the AI can't clearly place as routine, sends the call to a person right away, before collecting more details.
3. If no one picks up, or it's after hours, the caller hears the emergency hospital's name and number straight away.
4. The caller can ask for a person at any point and get transferred.
5. Every call leaves a record the team can review, so misses get caught and the setup gets better.

Then test it like a fire drill. Have staff call in with realistic, messy descriptions of emergencies, the way a panicked owner actually talks, and check that every one reaches a person or the emergency number within seconds.

A checklist for evaluating an AI phone system

Put these to any vendor before you forward a single call:

• Does it ever give medical advice or say a call can wait? It shouldn't.
• What exactly triggers a transfer, and can we add our own words?
• What happens after hours, and can we set the emergency hospital?
• Does it say it's an AI at the start of every call?
• Can callers reach a person whenever they ask?
• Where do call records go, who can see them, and how long are they kept?
• Can we switch it off instantly and fall back to our normal phone setup?

How PetGrid's AI Front Desk handles this

PetGrid's AI Front Desk answers the calls you forward to it, using what your team has taught it: hours, services and policies. It can book, move and cancel visits if you let it, and it takes messages. It says at the start of every call that it's the clinic's AI assistant, and it never gives medical advice. An urgent call goes straight through to your team, and after hours the caller gets the emergency hospital number you've chosen. It's a paid add-on that's rolling out as clinics' phone lines are connected; see pricing, and the AI questions for more.

Preguntas frecuentes

Lea las preguntas y respuestas para clínicas

En resumen

AI triage risks on the phone are real, but you can design them out with one firm rule. The AI handles logistics. People handle anything about a pet's condition. Build the escalation first, test it with realistic calls, and keep records so the team can look at every miss.

Want to see an AI receptionist that never gives medical advice and puts urgent calls straight through? We'd be glad to show you.