The Calls Your AI Voice Agent Should Never Handle Alone
David Selva · · 3 min read
AI Voice Agents, Veterinary Clinics, Automation Tradeoffs
Automation that answers a call it should have transferred does more damage than the voicemail it replaced.
Every conversation about AI on the phone starts in the wrong place. Everyone wants to know what the agent can handle. The question that decides whether this works is what it must never handle alone.
Here's what that looks like when it goes wrong. A clinic puts a voice agent on the main line to stop losing appointment calls during surgery blocks, and it works — bookings that used to hit voicemail at eleven now land on the schedule.
Then a client calls because her fourteen-year-old lab stopped eating three days ago and won't get up this morning. The agent, doing exactly what it was configured to do, offers her a wellness slot on Thursday.
That isn't an automation failure. It's a configuration failure, and it's entirely foreseeable.
The handoff list comes first
So I've stopped opening these conversations with what the software does. I ask the practice manager to tell me which calls would horrify her if a machine took them, and we build outward from that answer.
In a clinic the list is short, obvious, and non-negotiable:
- Anything the caller calls an emergency, in whatever words she uses
- Bloat, seizures, hit by car, labored breathing, ingested toxin, uncontrolled bleeding
- Euthanasia, quality-of-life conversations, and any call about a patient currently hospitalized
- Medication questions, dosages, and reactions
- A caller who's crying, or repeating herself, or plainly not getting through
Write that with your DVMs before anybody configures a prompt. Then set the rule: when any of it surfaces, the call rings a human immediately and the agent says so plainly.
What's left is still most of the volume
Carve out the hard list and what remains is exactly what's been eating your front desk. Routine booking, vaccine due dates, boarding availability, hours, whether you carry a diet, refill requests into the tech queue, and confirmations for the clients who never confirm.
In a general practice that's the majority of ring volume, and none of it needs a licensed person. Handing it over is what makes your two front-desk people available for the calls that actually need them.
Where automation quietly makes things worse
A system that can't recognize distress isn't neutral. It converts a client who needed help into a client who felt dismissed by a machine, and she will tell that story to everyone at the dog park.
Overreach is the same failure wearing a different hat. An agent guessing whether a symptom is urgent has taken on triage, and triage is a clinical judgment. It shouldn't be estimating urgency — it should be routing on keywords and tone and defaulting to a human whenever it's unsure.
An agent that transfers too often is a minor inefficiency. One that transfers too rarely is a liability.
Read every transcript for a month
Not a sample. All of them, weekly, looking for one thing: a call it handled that a person should have taken.
Every one you find is a missing rule, and the list stabilizes faster than you'd expect.
If you're weighing this for your clinic
Sit down with your lead DVM and write the always-a-human list before you evaluate a single vendor, ours included. If that list doesn't exist on paper, the technology decision is premature.
And if you'd rather have somebody build it around your list instead of the other way round, that's the conversation worth having — bring the list.