AI receptionists for clinics: an honest buyer's checklist

What the technology is genuinely good at

Across practices that have deployed these tools, one pattern holds: AI handles 70-80% of routine calls well. That bucket is bigger than most owners expect - scheduling and rescheduling, office hours, directions and parking, whether you accept a given insurance, and routing refill requests to the right queue. These calls are scripted repetition, and a system that never sighs and answers on the first ring at 2 AM is legitimately better at them than a tired human juggling line two.

The overflow case is where the money is. One analysis of dental practice call data found 38% of calls go unanswered when the front desk is busy - not after hours, during them, while your team is checking in patients and taking payments. If even a fraction of those callers are new patients who dial the next practice instead of leaving a voicemail, catching overflow alone usually pays for the tool by itself.

The compliance bar comes first

Clear the compliance bar first, because failing it makes everything else irrelevant. Three items, all in writing: a signed Business Associate Agreement, encryption for call audio and transcripts in transit and at rest, and a written retention policy that says exactly how long recordings live and how they are destroyed. A vendor that hesitates on the BAA is telling you they have not built for healthcare - thank them and end the demo.

This is operational guidance, not legal advice - run your specific setup past your attorney or compliance officer, especially if you record calls in a two-party consent state.

Ten questions to ask in any demo

Vendors control demos. These questions take control back. Ask all ten and write down what they say - the pattern of hedges tells you as much as the answers do.

1. Will you sign a BAA before the pilot starts? Not after go-live. Before any real patient call touches their system.
2. Where does call audio live, and who at your company can listen to it? You want a region and an access policy, not "the cloud."
3. What is your retention and deletion policy, in writing? If it only exists verbally, it does not exist.
4. Does it book into my actual calendar? A shadow calendar your staff re-keys is not automation, it is a second inbox.
5. What happens when a caller interrupts mid-sentence? Then interrupt it yourself, live, and watch.
6. How does a call reach a human, and how fast? Seconds matter. Ask for the transfer path, not a slide about it.
7. What does it say when it cannot understand someone? Twice in a row. Does it loop the same prompt, or does it hand off?
8. Who writes the escalation rules for emergencies, and can I edit them? You should own the words a scared patient hears.
9. Can I read every transcript? If transcripts are sampled or summarized, you cannot audit the system - and you will need to.
10. What is the all-in monthly cost at my call volume? Per-minute overages turn a $299 tool into an $800 one. Get the number for your volume, not the brochure tier.

The calls AI should never take alone

Anything that smells like clinical triage does not belong in a conversation with a language model. Chest pain, post-op bleeding, a reaction to a new medication, a parent describing a child's fever - these need a deterministic rule that fires every time: interrupt, state the emergency instructions, transfer to a human or direct to 911. Not a friendly chat that eventually gets there.

The same goes for emotionally distressed callers. Someone crying about a diagnosis, an anxious patient calling a third time about biopsy results, anyone who sounds frightened - the correct move is a fast, warm handoff to a person, and the correct vendor is the one who agrees immediately. Be suspicious of any rep who claims the AI handles these calls fine; the honest ones say 'we detect distress and get out of the way,' because getting out of the way is the feature.

Break it before you buy it

All of these systems perform beautifully on a calm, clearly spoken, single-request call. Those are not the calls that will hurt you. In the demo, and again during the pilot, test the four known failure modes yourself:

• Heavy accents. Have the most accented speaker in your office call in. Note whether the system asks to repeat once and adapts, or spirals.
• Interruptions. Talk over it mid-sentence and change your request. Real callers do this constantly; brittle systems restart their script.
• Multi-part requests. Ask for three things in one breath: "Reschedule my Tuesday cleaning, add my daughter to the same slot, and do you take Guardian?" Count how many of the three actually get handled.
• Distress. Roleplay an upset caller. The only acceptable response is a fast escalation, not an attempt to soothe.

Fifteen minutes of adversarial calling tells you more than any feature sheet.

Pricing, without the fog

AI receptionist tools typically run $199-$399 per month. Live answering services with real coverage - after-hours, overflow, weekends - run $500-$1,500 per month, and the humans on those lines read a script about your practice without seeing your schedule. The gap is real, but it should not drive the decision.

The number that should: what one recovered appointment is worth to you. A new dental patient is commonly worth $900 or more in first-year treatment; one converted med spa consult can cover several months of the fee on its own. If the tool catches the overflow calls your front desk is already dropping and books two or three of them a month, the pricing debate is over. And if it cannot book reliably, it is not worth $99 either - which is why you pilot.

The two-week pilot protocol

1. Scope it to after-hours and overflow only. Your main line stays human. The AI gets the calls you are currently missing outright - the lowest-risk, highest-value slice.
2. Read every transcript. Not a sample, not a dashboard summary. Two weeks of full transcripts is a few hours of reading, and it is the only way to catch quiet failures.
3. Verify every booking landed. Cross-check each appointment the AI claims it booked against the real calendar: right patient, right provider, right slot. Ghost bookings are the failure mode that loses patients silently.
4. Audit the escalations. Pull every call flagged as an emergency or distress and confirm a human got it, fast.
5. Then, and only then, expand. Daytime overflow next, and a bigger share of the line only after the second review looks as clean as the first.

A confident vendor will welcome this protocol. One who pushes you to go live on your main line in week one is telling you how their product performs under scrutiny. In a category drowning in hype, the sellers willing to admit what the machine cannot do are the ones worth shortlisting - and the buyers who verify keep their patients.