How does an AI receptionist decide which calls are worth booking?

First real cold snap of the year and the phone's been going since six. You got to the ones you could get to. What you're wondering, if you're honest, is what a machine would've put on Tuesday if it'd been picking up instead of you.

Here's the short version. It doesn't size up your caller. It applies filters you set: the work you take, what counts as urgent at your shop, and how fast you want to hear about it. The judgment it makes on the call is classification, working out what the caller is describing and matching that against your rules. That part is ours, not yours, so when it reads a call wrong that is on us. What it never does is plan your day or decide a trip is worth the drive, so the calendar you wake up to is still built out of decisions you made.

Willison is one of the things you can put on that phone, so read this knowing who wrote it. Answering, asking and booking work broadly the same wherever you buy. The specific guards below are Willison's own design choices, not the category's, and two of the sections are about where this gets the decision wrong.

What are you actually afraid it'll book?

Not that a robot will rank your customers by revenue. Nobody's losing sleep over that. The worry is opening Tuesday and finding trips you can't bill.

Out of your area. A system you don't touch. A commercial building when you only do houses. A tenant who can't authorize a dime. Somebody who wanted a price over the phone and got a slot instead. A message for your tech that somehow turned into an appointment. That's the real question sitting behind "which calls are worth booking," and it's a fair one.

There's a second thing going on, worth pulling apart from the first. Invoca's Home Services Lead Conversion Benchmarks Report 2026 finds that 34% of calls to HVAC companies are answered by a person, against 74% for plumbing. Same kind of business from the outside, and a gap of more than two to one between them. The report doesn't say why, and there's no sense inventing a cause it never measured.

So there are two problems here, not one: calls that never got answered, and calls that got answered and shouldn't have been booked. This post is about the second. The straightforwardly junk calls, the listing scams and lead sellers, are a separate subject, covered in can an AI receptionist screen spam and still catch real customers.

So what does it judge, and what does it just look up?

Two different things happen on the call and they tend to get mixed together.

The judgment is classification. A caller says the unit is "making a noise," or that the house "won't get warm," and something has to work out which of your categories that actually is. That's a real call, and it's the only genuinely hard thing happening on the line.

Everything after it is a lookup. Whether that category is work you take, whether it's the kind of thing you'd drive out for tonight, whether you want a text or the call itself: none of those are opinions. They're rules, and they were written before the phone went live.

Which is why the flat version of this answer, that it doesn't filter anything, is wrong and worth not believing. Qualifying a job is a filter. It's just that the filter is yours.

Which of those rules can a setup actually hold?

Some of them. It's worth being plain about which.

What it asks is built around your own services and your own intake questions, so the categories it sorts into are yours rather than a generic script's. What that comes to on a no-heat call is your list, not a default one. The address gets read back on the call either way, so a wrong street gets caught while the caller is still there, and it asks one question per turn rather than stacking three into a sentence a rattled homeowner answers half of.

Two of the fears above are worth answering precisely, because the honest answer is not the same for both.

A price over the phone. It doesn't quote one. What it does is state your own posture on the service the caller described, and on HVAC those postures differ inside one shop: a repair on a system you already installed is a paid diagnostic visit, while a quote on a new system is a free visit. Which one gets said is resolved per service and spoken at the read-back, in your words, not priced by the software.

Out of your area. Here the honest answer is that it captures rather than screens. The address is asked for and read back, so an out-of-area job is visible to you before a truck moves, but treat the service-area line as yours to enforce today, not something the phone does for you.

And here's what no setup holds. It doesn't plan your day. It doesn't know that a Copley call at eight and a far-side-of-the-county call at nine puts your guy on the highway for forty-five minutes in between. It books against whatever your calendar makes available, and it stops there.

Willison's part ends at a booked job with the details attached, or a job handed to you because it wasn't bookable. Ordering your day around it is still yours.

What will it flatly refuse to book?

Work that isn't on your list. That sounds obvious until you see the shape it usually arrives in, which is rarely a caller asking outright for a trade you don't practise.

It arrives as the machine next door. Somebody rings an HVAC shop about a water heater. Somebody else wants radiators looked at, at a shop that has never touched radiant heat. Those sound like your work, they are adjacent to your work, and they are the calls most likely to put a truck in a driveway for a job the crew can't do.

So an edge-of-trade ask routes to a different beat instead of into booking, including when it turns up late, as an addition at the read-back, which is exactly where it slips past a script. The design fails toward the cheap mistake on purpose: asking a caller to confirm what they need costs a few seconds, and booking the wrong machine costs a truck roll and a customer.

One more refusal worth knowing about: it won't claim credentials on your behalf. Licensing and insurance language is set per shop, because a sentence there isn't marketing, it's a claim made in your name.

Who decides a call is an emergency, you or the AI?

You do, but not in the way that sounds, and this is where a rule written at setup time runs out.

The trouble is that urgency rarely lives in the symptom. Two callers say the same six words about the same broken thing, and one of them is a job for tonight for reasons that never come up on the call. You know which is which because you know the street, the building and usually the customer. A rule written at setup time knows none of that.

So the useful thing was never a rule that encodes your judgment. It's getting the call in front of you fast enough that you still get to make it yourself.

There are two shapes that handoff can take, and a shop picks one. The details can arrive written down, for you to act on when you pick the phone up, or the call itself can be put through to a number you name. The second one has a fallback underneath it: if nobody picks up, the caller is booked and the details are sent rather than left hanging.

Putting a call through also has a clock on it. The hours when that is allowed to happen are yours to set, and the decision is worked out per call, so an after-hours caller reaches a shop that takes after-hours handoffs and nobody else. Get that window wrong and the call simply doesn't transfer, which is the safer way for it to break. What counts as urgent enough to interrupt you is the same question we worked through for after-hours and weekend calls.

What happens when the call is a hazard instead of a job?

It stops being a booking conversation. A caller describing something hazardous goes down a separate path, and that path is built to fail closed: the only thing that routes them back to ordinary booking is saying that a professional has already dealt with it. There's no configuration in which that path hands the live call off instead.

That costs you something, and it's fair to say so out loud. A caller who mentions a gas smell and then walks it back isn't getting booked on that call. That's a bookable job you lose. It's a trade made on purpose, and the right way round: a phone call isn't the tool for a gas smell. If there's a real question about gas or electricity, the utility and the fire department are the call, not a slot on your board.

Does it book the same job twice?

One conversation can't produce two appointments. That's enforced rather than encouraged, so a caller who confirms a no-heat visit and then brings up the thermostat in the back bedroom doesn't walk away holding two slots.

Be clear about the scope of that, though, because it's per call and not per job. The homeowner rings, then his wife rings, then the property manager rings, all about one furnace. That's three calls, and one-booking-per-call has nothing to say about them.

What does carry across calls is narrower. A caller ringing from a number it already knows is recognised and greeted by name, their details ride along, and there's a check against the booking they already have before another one is offered. That catches the homeowner ringing back. It does not catch the wife on her own phone, and no setup claim should pretend otherwise.

What does hold on every one of those calls is that nobody's told they're booked until the booking has actually been written. The confirmation is a fixed line that can't be rewritten and can't be talked over, so the read-back happens even when the caller is mid-sentence. The mechanics of how an appointment lands on your calendar are their own subject.

Where does it get the classification wrong?

Here's where the original worry is right, just aimed one step off target. The judgment is real, which means it can be wrong.

It can mishear an address. It can file a job in the wrong category. It can get interrupted and restate something badly. It can ask two things at once. It can take a message and fail to get it to you.

The answer to each of those is structural rather than a promise that it doesn't happen. The address gets read back while the caller's still on the line. Nothing's called booked until it exists. One question per turn. One booking per call. Where you've asked to be told, that alert is wired to fire before the caller can leave the call, which is the guard against a message going nowhere. And every call is recorded and transcribed, with the caller told so at the start, so what happened on it exists as a recording and a transcript rather than only as a calendar entry.

What that buys isn't a clean sheet. It's that mistakes land somewhere you can still catch them.

What should you settle before it answers a single call?

Here's the part most write-ups skip: you probably don't have any of this written down. You've been carrying it around in your head for years, which works fine right up until something other than you has to answer the phone.

Worth settling anyway, roughly in this order.

• What work you take, and what you don't. Houses or commercial. Systems you'll touch and systems you won't. This is the rule that decides what a call is allowed to book at all.
• Who can actually authorize a repair. An owner and a tenant are two different calls, and a tenant holding a booked paid visit is a wasted trip.
• What counts as urgent at your shop. Not in general. At yours, in January, with the trucks you've actually got.
• How you want to hear about it. A text with the details, or the live call in your ear, and what hours the second one is allowed to happen.

With Willison the account is reviewed personally by the founder before it goes live, which exists for exactly this reason. It isn't a form you fill in on your own.

One more, which isn't about setup at all. Whoever ends up answering your phone, you or somebody you hire or software, make sure asking for the booking is part of the job. It is the step most easily left out, and nothing downstream recovers a call that ended without one.

The calendar you wake up to

"How does it decide what's worth booking" turns out to have a plainer answer than it looks. It decides what the caller is describing, then it does what you told it to do about that. The deciding you were worried about happened earlier, at a desk, by you.

So the thing worth getting right was never the software's taste. It's those four answers, and whether what you put on your line tells you quickly when it's out of its depth. To hear how that goes on an actual call, press play on a real call on willisonhq.com. It answers for a sample HVAC shop, so the questions are that shop's rather than yours, and the engine underneath is the same one.

Frequently asked questions

That's what intake is for. What it asks is built around your own services and intake questions, so "we don't touch that" becomes a rule the call runs on rather than something a tech finds out in a driveway. The sharper you make that list up front, the less there is to argue about later, and the honest catch is that a category you never mention is a category nobody wrote down.

Only if that's how you set it up. The owner text is optional, which matters because an owner may not want a 2am text about a job that's already on the calendar and can't be worked before seven anyway. The other way round is the live transfer, and that one runs inside hours you pick rather than around the clock.

It still answers it and it still writes it down. What protects you there isn't the software knowing it's off the map, it's the guards underneath: nothing gets confirmed as booked until it's been written, the address gets read back, and the call is recorded and transcribed either way. An odd call reaches you the same way the rest do, and you get to be the one who decides what it was.

Every call is recorded and transcribed, and the caller is told so at the start, on every call, new or returning. That's the part worth knowing before the morning you need it: "why is this on Tuesday" is a question with a record behind it rather than only a calendar entry, so how you get at that record is a fair thing to pin down before you hand over the line.