What Your AI Receptionist Should Never Say

The Short Version

Almost every serious failure of an AI receptionist in a dental practice is the same failure: it answered a question it should have refused.

Not a misheard name. Not a wrong transfer. A confident, fluent, wrong answer about money, coverage, clinical risk, or somebody’s private information. Fluency is the problem, because a system that sounds certain gets believed.

The fix is boring and it works. Before go-live you write a short refusal list: the specific things the agent must never say, and the exact sentence it says instead. This page is that list, and the reasoning behind each line, so you can hand it to a vendor and check what comes back.

We are a vendor, not your legal or clinical adviser. The list below is the operational starting point most practices land on, not advice about your obligations.

Why Refusals Matter More Than Capabilities

Vendor demos are built around what the agent can do. Your risk lives entirely in what it will do when it should not.

The asymmetry is brutal. An AI that fails to book an appointment costs you one appointment. An AI that tells a caller their crown is covered at 80 percent, when their plan pays 50 percent and they have not met their deductible, costs you a furious patient at the front desk, a write-off argument, and a review. An AI that offers reassurance about facial swelling costs you something far worse.

So the question to put to a vendor is not “can it handle insurance questions?” It is “show me what it says when it does not know.”

The Four Things It Should Never Say

1. A coverage amount, a benefit level, or what the patient will pay

This is the big one, and it is the one most likely to be quietly wrong.

An agent can reasonably confirm whether you are in network with a carrier, because that is a fact about your practice that you control and can keep current. What it must never do is state a benefit level, a percentage, a remaining deductible, an annual maximum, or a dollar figure the patient will owe.

The reason is that none of that is knowable from the phone call. Real benefits depend on plan design, the specific procedure codes, the patient’s history against their annual maximum, waiting periods, frequency limitations, and downgrade clauses. A number that sounds authoritative and turns out to be wrong is worse than no number, because the patient made a decision on it.

The replacement sentence. Something close to: “I am not able to quote your benefits, because that depends on the details of your plan. Our team will verify your coverage and call you back with exact numbers. I can book you in now and hold the appointment while we check.”

That sentence does three jobs. It refuses, it explains why the refusal is in the patient’s interest, and it keeps the booking. A good vendor will recognise all three.

Collection is a different matter from quoting, and it is safe. Capturing the carrier, the member ID, the group number and the subscriber is data entry, not interpretation. Can AI collect or verify dental insurance details covers where that line sits, and what insurance questions AI can and cannot answer goes through the spoken version in detail.

2. Anything that sounds like clinical judgement

The agent must not assess, diagnose, rank severity, or reassure about symptoms. Not “that sounds like it can wait,” not “that is probably just sensitivity,” and above all not “that does not sound serious.”

What it does instead is recognise and route. It hears the words, matches them against the protocol you wrote, and either reads your pre-approved instruction or escalates to a human. It does not evaluate how bad the bleeding is. The distinction between recognition and assessment is the entire safety model, and where that line sits after hours is worth reading before you write your protocol.

The replacement sentence. Your own approved wording, read verbatim. The point of pre-approving it is that the agent never improvises under clinical pressure.

Note the failure mode runs both ways. An agent that treats every reported problem as an emergency will bury your on-call rota and train your team to ignore it. The protocol has to say what is routine, too.

3. Patient information to someone who has not been verified

A caller saying they are Sarah’s husband is not verification. The agent should not confirm appointment times, treatment details, balances, or even that someone is a patient, until whatever identity check you have specified has passed.

This gets sharpest in two places. Family calls, where a parent, spouse or adult child expects access they may not have. And orthodontics or pediatrics, where a patient turning eighteen changes who the personal representative is, and no system will flag that date unless you build the rule.

The replacement sentence. Something like: “I am not able to share details on the account without confirming a few things with you first.” Then your verification steps, or a transfer.

How the underlying data is protected is a separate question from who may hear it, and HIPAA compliance for an AI receptionist covers the storage, encryption and BAA side.

4. A promise your practice has not agreed to

Fee quotes for treatment. Guarantees about outcomes or timelines. Commitments about what a dentist will do at a visit. Anything financial beyond what you have explicitly published.

These creep in because they feel helpful and because a generative system will produce a plausible answer if nothing stops it. A caller asking “roughly what does an implant run?” is asking a reasonable question, and the wrong answer is a confident range the agent invented from general knowledge of dentistry.

The replacement sentence. “That depends on the exam and what the dentist recommends, so I would not want to guess. I can book a consultation and the treatment coordinator will go through exact costs with you.”

The Rule Behind All Four

Every line above is the same rule wearing different clothes.

The agent may state facts you control and keep current. It must not interpret, estimate, or reassure.

Your network participation, your hours, your locations, your appointment types, your parking, your policies: all facts you own. A patient’s benefits, their clinical situation, their identity, and your dentist’s future recommendations are none of those things.

If you can only remember one sentence going into a vendor conversation, make it that one. It generalises to questions this page has not thought of.

How to Test It Before You Trust It

Write the refusals down, then call your own line and try to break them. Fifteen minutes, before go-live and again a month later.

1. Ask what your insurance covers. Ask twice, the second time pushing for “just a rough idea.” Pressure is where soft refusals fail.
2. Describe a symptom. Use your test script. Confirm it routes on your protocol without characterising the symptom.
3. Ask for someone else’s appointment. Claim a plausible relationship. Confirm it does not confirm.
4. Ask what a crown costs. Listen for whether it invents a range.
5. Ask it to guess. “I know you can’t be sure, but what would you say?” A well-configured agent refuses the invitation to speculate. A poorly configured one takes it.

Item five is the one that separates vendors. Refusing a direct question is easy. Refusing a friendly, reasonable, insistent caller is the actual engineering problem.

What a Good Vendor Answer Sounds Like

You are listening for specifics, not reassurance.

A vendor who has done this will talk about configured refusal rules, the exact fallback wording, what happens when confidence is low, and how escalation is triggered. They will be comfortable showing you a call where the agent declined to answer.

A vendor who says the model is very accurate has misunderstood the question. Accuracy is not the control. The refusal is the control, because the failure you are guarding against is a confident answer, and confidence and accuracy are not the same property.

If the agent will hand off when it hits a boundary, ask what the human receives. How AI hands difficult calls to your team covers what a clean handoff should carry, and a handoff that makes the caller start over is only half a handoff.

Key Takeaways

• The risk in an AI answering your phone is concentrated in confident wrong answers, not in misheard words
• Four refusals cover most of it: no benefit or coverage amounts, no clinical judgement, no patient information without verification, no fee quotes or promises
• Every refusal needs a pre-written replacement sentence that also keeps the booking, or your team will disable the rule within a fortnight
• The general rule is that the agent may state facts you control and must not interpret, estimate or reassure
• Test by pressing for a guess, because refusing an insistent caller is the hard case that separates vendors

Frequently Asked Questions

Does refusing these questions make the AI less useful? No, and this is the common worry. Almost none of the value is in answering coverage or cost questions. It is in answering the call at all, booking the appointment, and routing what it should not handle. A refusal that ends in a held appointment loses nothing.

Who writes these rules, us or the vendor? You own the content, because they are decisions about your practice. A vendor should bring the structure, the standard wording and the questions you have not thought of. If a vendor hands you a finished ruleset without asking anything about how you run, that is a flag.

What if a patient gets annoyed at being refused? Far less often than practices expect, provided the refusal explains itself and offers the next step. “I will not guess at your benefits, our team will confirm and call you back, and I will hold the appointment meanwhile” reads as careful, not obstructive. Whether patients are annoyed by AI answering at all covers the wider evidence.

Should the AI say it is an AI? Separate question, and in some states it is not purely a matter of preference. Should the AI tell callers it is AI has the detail, including where a proactive duty applies.

Can the AI still verify insurance in the background? Collecting the details and passing them to your team for verification is routine and safe. What it must not do is speak the result of a verification it has not actually performed against the payer.

Where to Start

Take the four refusals above, put them in a document, and add your own replacement wording for each. That document is the single most useful thing you can bring to a vendor conversation, and you can write it in twenty minutes without buying anything.

Then book a demo and ask to hear the agent refuse something.

---

Related reading: the boundaries above need an owner — who’s accountable when the AI books it wrong.