What to Look for in an AI Front Office, Before You Buy One

The category has filled up quickly, and the marketing has converged. Nearly every vendor describes itself the same way: answers your calls, sounds natural, books appointments, saves your front desk time. Sit through three demos in a week and they blur together, which is how practices end up choosing on price or on whichever salesperson was easiest to talk to.

The differences are real, they are just not in the pitch. They show up in specific, slightly awkward questions that most demos are not structured to answer. What follows is the list worth asking, and more usefully, what a real answer sounds like compared with a rehearsed one.

Which calls is it actually built to answer

Vendors in this space are built around one of three shapes: after-hours only, after-hours plus business-hours overflow, or full inbound replacing the front desk entirely. A vendor who cannot say plainly which of the three theirs is has either not decided, or is hoping you will not notice.

The shape matters more than the feature list because it determines what the system is optimised for. Something built to cover evenings handles a narrow set of predictable requests very well. Something built to replace a front desk has to handle everything, which is a much harder problem and usually a much larger bill.

There is a practical reason the shape is chosen rather than emergent. A system covering evenings can be tuned to be cautious, handing off anything ambiguous, because a missed handoff at 9pm costs little. A system replacing a front desk cannot hand off constantly without defeating its purpose, so it has to be tuned to push through uncertainty. Those are opposite settings and no product runs both at once.

Be suspicious of any answer that claims to be all three at once. Coverage models involve genuine trade-offs in how the system is tuned, what it escalates, and how much configuration you are expected to do. A vendor claiming no trade-offs is describing a demo, not a product you will run for a year.

Ask which shape it is. If the answer is "all of them, basically", ask again.

The practical version of this question: what percentage of our calls do you expect to handle end to end, and what happens to the rest. A vendor who has deployed in practices like yours will have a real answer with a real number attached, and will be comfortable naming the calls their system is not for.

Where a booking request actually lands

A booked appointment that lives inside a vendor's own dashboard is not the same as one in the calendar your team already works from. The distinction sounds pedantic until you picture a Tuesday morning where nobody remembered to check the other system, and the request sat there for two days.

This is the single most useful thing to watch during a demo, and the one most likely to be skipped. Do not accept a description of the integration. Ask them to make a test booking on the call and then show you where it appeared, on screen, in the actual tool your practice uses.

Ask also what the system does when it cannot identify the caller. An established patient ringing from a number you have never seen, or a new caller who gives a name that matches two existing records, is routine rather than exotic. How a product handles ambiguous identity says a lot about whether it was built for real practices or for a pitch deck.

The follow-up question is what happens when the calendar is wrong. Double bookings, a provider who blocked out the afternoon, a slot that was filled thirty seconds earlier by someone at the front desk. Real scheduling is messier than a demo, and how a system behaves at those edges tells you far more than a clean happy-path walkthrough.

Worth asking explicitly whether the system books directly into your calendar or captures a request for your team to confirm. Both are legitimate designs and they suit different practices. What is not legitimate is a vendor blurring the two, because the difference determines how much work is left on your side after the call ends.

What happens when it cannot help

Every system hits calls it should not handle. In healthcare this is a safety question rather than a product-quality question, and it deserves more demo time than the pleasant parts. A caller with an urgent clinical problem must reach a human, and the system must never attempt medical advice.

Ask what specifically triggers a handoff, where the call goes at 10pm when no line is staffed, and what the caller hears while it happens. Those three answers, taken together, tell you whether escalation was designed or bolted on afterwards. Vagueness here is the most expensive kind of vagueness in the whole evaluation.

Patients are unambiguous about wanting this path to exist. In a 2024 Talkdesk survey of 1,000 US adults, 81% said they would rather consult a human for medical advice, even while a majority were comfortable with automation handling administrative tasks [1]. A system with no clean route to a person has solved the easy half of the problem.

The same survey is useful on objections. The top concern about AI in healthcare was inaccurate responses at 26%, followed by data privacy and losing the human touch, both at 24% [1]. Notice that these are answerable concerns rather than general hostility, and that a vendor willing to talk about how their system behaves when it is unsure is addressing the biggest one directly.

What the monthly report should show

"Analytics" can mean a genuine monthly summary of what happened, or it can mean a login you are expected to remember to check. Ask to see a real sample report before signing anything, not a description of what one could contain, and not a screenshot of a dashboard with placeholder data in it.

Ask when it arrives, too, and whether it is pushed to you or waits to be fetched. Reporting that requires someone to remember to log in is reporting that stops happening in month two, which is exactly when you most want to know whether the thing is working.

A report worth paying for answers three questions without you doing arithmetic. How many calls came in, how many turned into a booking request, and how many follow-ups went out to people who went quiet. Anything beyond that is nice. Anything short of it is a call log with better styling.

The right-hand column is the pattern. Real answers are specific, contain nouns, and can be demonstrated. Weak answers are reassurances. This is not about catching anyone out, it is that the specific answer is genuinely harder to give, which is exactly what makes it informative.

What it costs, and what moves that number

Pricing in this category splits into per-minute usage and flat monthly, and the two behave very differently in a busy month. Usage pricing means a good month for your practice is an expensive month. Flat pricing means the number is known before the month starts and does not move.

Velaire is built on the flat model: Recovery covers after-hours calls at $599 a month, Growth adds business-hours overflow at $1,199, and Pro covers full inbound from $1,999. Whichever model you are quoted, ask what is excluded, because that is where the surprise usually lives.

Ask what the bill looks like in your busiest month, not your average one.

Specific things to ask about: setup or onboarding fees, whether call minutes are capped, what happens when you exceed a cap, whether SMS follow-up is included or metered separately, and what the term commitment is. None of these are unreasonable to charge for. All of them are unreasonable to discover in month three.

The comparison only becomes real when you put your own volume into it, which means working out what a missed call is worth to your practice before you shop. Without that number you are comparing prices with no idea what you are buying against.

What to ask about patient data

Any system answering a healthcare practice's phone will touch protected health information, because callers volunteer it whether or not the system asked. In the United States that means a signed business associate agreement, and it means one with every vendor in the chain, not just the one whose logo is on the invoice.

The phrase to listen for is "HIPAA compliant", which sounds like a certification and is not one. There is no HIPAA certificate and no body that issues one. Compliance is a configuration plus a contract, so a vendor describing themselves as certified has told you something useful about how carefully they have read the rules.

Ask who is in the chain. An AI front office usually sits on top of a telephony carrier, a speech provider, and a model provider, and each of those touches call audio. A vendor should be able to name them and confirm the agreements are in place without treating the question as unusual, because it is not unusual.

Ask what happens to the data if you leave. A practice ending a contract should be able to take its call records with it and have the vendor's copies deleted, and the mechanism for that should exist before you need it rather than being invented during an unhappy exit.

Disclosure is the part practices forget and patients do not. Pew Research Center, polling 3,488 US adults in June 2026, found 72% said it was extremely or very important that a provider tell them when AI was used in their care, and 63% wanted more say in whether it was used at all [2]. KFF, polling 1,343 adults over 7 days a few months earlier, put concern about the privacy of medical information given to AI tools at 77% [3]. Neither figure obliges you to announce the system in your greeting. Both suggest a practice that can answer "is this a robot" plainly will spend less time managing that question than one improvising an answer.

Retention is the other half. How long is call audio kept, how long are transcripts kept, who can access them, and can you have them deleted on request. A practice signing this contract is taking on responsibility for those answers, so getting them in writing beforehand is simply doing the job properly.

Putting the shortlist together

None of this is exotic and none of it requires technical knowledge to evaluate. It is a short list of specific questions, asked in a demo, where you watch what the vendor does rather than listen to what they say. The specificity of the answers is the signal, more than the answers themselves.

A reasonable process is three demos, the same questions in the same order, and notes written the same afternoon while the differences are still clear. Vendors blur together within about a week, and a scoring sheet written before the first call is worth more than a strong impression remembered after the third.

The strongest single move is to pull a week of your own call logs first. Knowing what actually arrives at your phone converts most of these questions from abstract to concrete, and it stops a demo from setting the agenda. If the after-hours share turns out to dominate, the coverage question answers itself and the rest of the evaluation gets much shorter.