The new AI receptionists will run your schedule for you. The real question is who's checking their work.

01 The judgment nobody's supporting

It's a Monday morning — across North American dental data, Mondays run 20–40% busier than an average weekday as weekend pain and Friday's unbooked issues land at once (nobody has measured Winnipeg practices specifically, but the pattern has no reason to stop at the border). Your front desk has a line ringing, a patient checking in, and a voice on the phone saying a tooth has ached for three days — can you fit them in today? She has no triage protocol and no chart in front of her, and in that gap she decides: emergency squeeze-in, or next week. Whatever she jots in the notes field, there's no shared standard for what "urgent" means and no way to look back later at why the call was judged the way it was — so when a judgment goes wrong, it doesn't get corrected. It just repeats.

And that's only the calls that get answered. US industry trackers (Weave, Dental Intelligence, and similar — no Winnipeg-specific study exists, so treat the exact figures as directional) put missed inbound calls at 20–40% for a typical practice, worse during peaks and lunch. Roughly a quarter or more of patient calls come in after hours, when voicemail is all that answers — and most callers who hit voicemail don't leave a message. They dial the next clinic.

02 Your software is smart — in the wrong room

Whichever system your clinic runs — Dentrix, Curve, ABELDent, ClearDent, Tracker — it has probably grown real AI in the past year. Dentrix ships FDA-cleared Detect AI for reading X-rays and is rolling out voice-dictated clinical notes and a live claims workspace. Curve launched FLO AI: instant insurance verification, AI call summaries, X-ray diagnostics.

Look at where all of that lives: the operatory and the back office. It reads images, writes notes, chases claims — data already sitting on a screen. The moment that actually needs judgment — is this patient's pain urgent or not — happens on a phone call, and none of that software is in the room for it.

03 What the AI receptionists actually do — read this part carefully

The AI phone agents on the market are real products, and some are well built — most have handled far more calls than any single clinic's system starts out with. Credit where due: most also have an escalation rule for emergencies.

Here's the part to look at closely: the AI itself decides which calls count as urgent enough to hand to a human. A call it judges as routine gets booked straight into your schedule — and nobody reviews that judgment, because deciding whether a human needs to see it was the judgment. If it sorts a call into the wrong bucket, nothing stands between that mistake and your appointment book. The vendors aren't hiding this; it's the design. Full automation is what makes a subscription product scale.

Two more things that come with the subscription. The recording of a Winnipeg patient describing a health symptom typically lives on a US vendor's servers, under their retention policy — worth asking any vendor exactly where call audio is stored and for how long. And the pricing: homepage rates run $300–500 a month, but once setup fees, PMS integration surcharges, and per-minute overages are added, industry cost breakdowns put the realistic all-in at $700–1,400 a month, with year-two renewals commonly rising 20–40% once promo pricing expires.

One more thing worth knowing: your PMS is not going to close this gap for you. Curve has published its AI roadmap — it's training on claims, denials, and payment friction, where its data advantage lives. Dentrix leaves phone AI to its third-party marketplace entirely. That's structural, not accidental: platforms build AI where one model serves ten thousand identical clinics. Per-clinic emergency rules, with real liability attached, is the opposite of that. It isn't on anyone's roadmap.

04 The inverted design: every judgment goes through your desk

What I build flips the escalation logic. It's not the AI deciding which calls need a human — every judgment, urgent or routine, waits for your front desk before it acts.

A patient calls after hours. The agent answers, listens, and asks what your best front-desk person would ask. Then it drafts — never books — a judgment: likely emergency, suggest 8am squeeze-in, based on what the caller said. The draft sits in a review queue. In the morning, your front desk sees the reasoning next to the caller's own words, and approves — or overrides.

Fair question: if a human approves everything anyway, why involve AI at all? Because the hard part was never the moment of deciding. It's everything around it: being on the line at 9pm when nobody's staffed, drawing the relevant facts out of a rambling description of pain, and doing it the same way on every call regardless of who's on shift. The agent does that legwork and hands your desk a clean judgment to confirm or correct — not a blank decision to make from scratch. And this design doesn't need the smartest AI on the market to be safe: when a draft is wrong, it gets caught at the desk, corrected, and the rule gets fixed — with a trail showing exactly what the caller said and why the call was judged that way. The person who clicks approve owns the decision, same as today — except now she's backed by evidence instead of gut feel.

To be clear about what this is and isn't: I'm not claiming my system judges better than the venture-funded products — they've seen more calls than mine will for a long time. The difference is who holds the final say, and that's a design choice, not a horsepower contest.

Three things the subscription can't offer, whatever its training data:

You own the finished system. It's built for your clinic and it belongs to your clinic — not rented month to month from a vendor that raises the price at renewal. That's not a claim about being cheaper long-term; it's a different relationship to the thing entirely. Your data stays where you decide — hosted in the AWS Canadian region, recordings kept or deleted on your clinic's terms. And I'm in Winnipeg. When your emergency rules need adjusting, you're talking to the person who built the system, across a table, not a ticket queue.

And because most small-business AI fails at adoption, not installation: going live isn't the finish line — it's the start of a trial run. After launch, there's a trial period where your front desk uses the queue day to day, and if a judgment's off or a rule needs adjusting, I fix it remotely. When the trial ends, if it's working, it stays — and it's yours. If it isn't a fit, we pull it out, and you're out nothing. "Done" means your team finds it genuinely useful, not that I shipped something and left.

05 A second judgment gap: the bill nobody explained

The urgent-call decision isn't the only judgment that leaks. There's a quieter one, and it shows up after the visit — on the patient's statement.

Here's the shape of it: a patient comes in for a routine cleaning after a long gap. Because it had been a while, the visit ends up running longer and covering more than a basic cleaning. The clinical call is right. The work is right. But nobody said, before the chair, that a visit like this can push past what an insurance plan covers in a year, and that the difference lands on the patient. So the patient leaves, gets a bill they didn't expect, and reads it as a bait-and-switch. The clinic, fairly, points out that tracking every patient's coverage isn't its job — the onus is on the patient. Both are right. And in the gap between "we treated you correctly" and "nobody warned you," the trust goes.

Notice what this isn't. It isn't a billing error, and it isn't a clinic being dishonest — the treatment was sound and the coverage rule is real. It's a missing sentence, said at the wrong time or not at all: "Given how long it's been, today's visit may run past what your plan covers, and there could be an out-of-pocket portion — the front desk will confirm before we proceed." That sentence is cheap. The reason it doesn't get said is that it depends on someone, on a busy day, remembering to say it — and it's nobody's defined job to remember.

06 Where the approved-draft design fits — and where it doesn't

This is the kind of gap the design in section 04 is built for, with one honest boundary. The phone agent can't sit in the operatory at the moment the visit runs past what the plan covers — that moment is out of its reach, and I won't pretend otherwise. What it can do is earlier, at booking: when a patient books a cleaning after a long gap, the agent drafts a heads-up — "there may be an out-of-pocket portion today; front desk to confirm" — and routes it to your desk to approve, edit, or drop before it ever reaches the patient. The judgment about what to say, and whether to say it, stays with your team. The system's only job is to make sure the question gets raised every time, instead of on the days someone happens to think of it.

That doesn't make the out-of-pocket disappear. The patient may still owe the difference; the coverage rule doesn't change. What changes is that they heard it coming, from you, before the chair — so the bill is a number they were warned about, not a surprise they feel tricked by. Same as the urgent-call design: whoever approves the message owns it, and there's a record of what was said and when.