Can an AI receptionist screen spam and still catch real customers?

Your phone rings. Unknown number, an area code you half recognize. You're on a roof, or under a sink, or three minutes into an estimate. It's either the homeowner from Tuesday finally calling you back, or the fourth call this week about your Google listing. You let it ring. That reflex is completely rational, and it is one of the most expensive habits in the trades, because the junk call and the paying customer arrive on your phone looking exactly the same.

So the question underneath every demo is a fair one: can something sort out the garbage without sorting out the job?

Can an AI receptionist screen spam and still catch real customers?

Yes, but only if it screens on what the caller says, not on the number they called from. That distinction is most of the answer. Number-based screening comes in two flavors and both leak. A carrier spam label judges the calling number by its own outbound history, so it catches a lot of dialers and misses the human with a fresh line. A block list, or the setting that silences unknown callers, judges the number by whether you have seen it before, which flags your next customer by definition.

An AI receptionist works the other way around. It answers everything and sorts on what the caller is actually there for. A recorded pitch never becomes a job on your calendar. A homeowner with no hot water gets qualified and booked. Nothing is turned away at the door, so nothing real is lost at the door.

What actually counts as a spam call for a service business?

Ask an owner what the junk is and you'll hear robocalls, though the one most of them can quote from memory is the outfit claiming your Google listing is about to be suspended. Your phone is carrying five different things, and only some of them are what anybody means by spam.

• Automated junk. Recorded pitches, autodialers, and the dead-air ones that sit silent for a few seconds after you say hello. No human on the line at all.
• Human cold sales. Somebody selling search rankings, listing fixes, lead lists, merchant cash advance, extended warranties, commercial insurance, staffing. A real person, a real script, and increasingly a spoofed local number so your screen shows your own town.
• Wrong numbers and misdials. Not malicious, just noise, and there is enough of it that call-data researchers report answer rates with and without it.
• Existing-job traffic. Where's my tech, can we move Thursday, the invoice is wrong, he left and it's still dripping. Not new work, not remotely junk, and a real share of the day.
• The gray zone. The supply house saying the coil didn't ship, a sub who isn't showing tomorrow, an adjuster, somebody answering your help-wanted ad.

What most owners end up doing is running one rule across all five on their personal cell: don't answer numbers you don't recognize. That's a single setting aimed at five different problems, and your next customer is standing in the blast radius with the rest of them.

It costs the most on the days the phone is busiest. In the first hot week of a cooling season a listing pitch and a no-cool call land on the same number minutes apart, which is why sorting a heating and cooling line on what the caller wants beats sorting it on who is calling.

Why does screening by the number cost you jobs?

Because a phone number carries no information about what the caller wants. Screening on it is backward-looking by design: the most it can know is how that number has behaved before. It cannot know that this particular unknown number is a homeowner who moved into your service area last month and has a sump pump that quit.

Then there's the part that settles the argument. The two mistakes do not cost the same. Screen out a robocall and you lose no work at all. Screen out a customer and you lose the whole job, plus whoever they would have told about you, and you never find out it happened. There's no bounce-back and no notification.

Per Invoca's platform data, fewer than 3% of callers pushed to voicemail leave a message, a behavior we dig into in do customers leave a voicemail when you don't answer. The ones you filter by accident don't complain. They call the next name.

That asymmetry is why "block harder" is the wrong instinct, even though the junk is real: the FTC logged more than 2.6 million Do Not Call complaints in fiscal year 2025. The volume is not in your head. It's the response to it that costs you.

How much of your phone is actually a job?

Worth putting numbers on, because it explains why "just answer everything yourself" collapses by Wednesday. In Invoca's 2026 home services benchmarks, which cover nine home services sub-industries including plumbing, HVAC, construction and restoration, 52% of callers to home services businesses speak with a person.

The interesting part is what happens when you count only the calls that lasted. Per that same report, the answer rate rises to 65% past 15 seconds, a cut they describe as filtering out misdials and quick hangups, and to 73% past 30 seconds. The spread between those three numbers is the shape of the junk, sitting right there in the data.

So on an ordinary Tuesday, most of what you personally answer is not new work. A big part of why owners stop picking up is sitting right there, and it isn't laziness, it's arithmetic. It is also why the miss rate on a trades phone runs higher than most owners will admit out loud, broken down in what percentage of calls home service businesses actually miss.

What that does not mean is that everything which isn't a lead is garbage. A price shopper might be next month's job, and "where's my tech" is your existing customer. The point isn't that most of your phone is worthless. It's that most of it doesn't need to be you.

How does an AI receptionist tell a robocall from a customer?

By having the conversation a filter can't. The sorting happens after the answer instead of before it, and it comes down to four moves.

• It answers every call. No number is judged before it says anything, which removes the one error that actually costs you money.
• It asks what they need. This is the sort. An autodialer can't answer it and a recorded pitch doesn't react to it, so the automated half of your junk falls out here on its own.
• It qualifies against your real business. Your services, your service area, your hours, the work you actually take. A caller two counties out isn't spam, but they aren't a booking either, and those two deserve different endings.
• It escalates only the real ones. The job goes on the calendar. The pitch stays a line in the call log instead of a buzz in your pocket while you're halfway under a sink.

That last one is what owners underrate. You were never going to make junk calls stop existing. What you're buying back is not being the one who has to sort them. The plain-English version of how that runs on a live call is in what an AI receptionist actually is.

What about the cold caller who just lies?

This is the honest limit of the approach and it deserves saying out loud. Sorting on intent works cleanly on anything automated. It works far less cleanly on a trained human opening with "I'm calling about the property on Maple," often from a spoofed local number, because that person is coached to get past whoever asks what they need.

What saves you is that the pitch has to arrive eventually, and when it does it still isn't a job, so it doesn't become one. That's a smaller win than blocking the call outright, and anyone claiming otherwise is selling. It's still the right kind of win: the cold caller burned two minutes on software instead of pulling you off a roof. And weigh it the way you weigh everything else here. A screen that lets a salesman through costs you an eye-roll. A screen that stops a homeowner costs you the job.

What if it screens out a real customer anyway?

Fair question, and the honest answer is that no screen is perfect. But the two failure modes are not equivalent.

A number filter fails silently and permanently. The caller goes straight to voicemail, or gets nothing at all, and you never learn their name or what they needed.

A conversation fails out loud, in front of somebody who can push back. The ugly version is real and worth naming: mishearing a street name twice, talking over a caller, looping on a question already answered, or a caller who just wants a human.

None of that is pleasant, but all of it is visible, in a call you can pull up and listen to. That's the difference between suspecting you're losing work and knowing where it went, and it's why "hands the caller to a person when they ask" belongs on the checklist below, not in the fine print. What it does with a question it cannot answer is the same test run from the other direction.

The risk worth worrying about isn't over-filtering. It's a setup that puts a gauntlet in front of the caller: press one for service, press two for billing, state the reason for your call. That isn't screening, it's a phone tree with a nicer voice, and it's a hard thing to sit through with water running across a floor. Any system that makes a customer prove they're worth talking to has traded your annoyance for your booking rate.

What happens to the callers who are neither?

This is where a lot of screening setups quietly fall down, because they know only two endings: book it, or bin it. The supply house saying the coil didn't ship fits neither. Neither does the sub who isn't showing tomorrow, the adjuster, or somebody answering your help-wanted ad, which in this labor market can be worth more than the job you were hoping for.

The rule for all of them is route, not reject: a message with a name, a number, and what they wanted, on the timetable that caller actually deserves. A "job or junk" filter that drops your supplier into a log you read on Sunday hasn't saved you anything, it has moved the problem somewhere quieter.

Any system worth paying for lets you say which of these matters and how fast. That split runs trade by trade, and which calls in your trade are the emergency and which are fine to book is a line you draw, not one a vendor draws for you.

Does this look different for a restoration company?

It gets sharper, because so much restoration work arrives from numbers nobody in the office has ever seen. The 11pm call is an unknown mobile. The adjuster is an unfamiliar corporate line. The property manager with four units under water is calling from something that looks like a call center, because it is one.

And on plenty of losses the first call isn't the homeowner at all, it's a program dispatcher or a third-party administrator assigning a claim with a clock on it, or a plumber handing you a referral. Those are among the highest-value calls the division gets, and every one shows up dressed as the thing a screen is built to reject.

The caller is also in less of a mood to wait. Somebody standing in water is unlikely to try you twice, and with fewer than 3% of voicemail-bound callers leaving a message, you should assume they work down the search results until a voice answers. Screening on caller ID in restoration points straight at the emergency work, which is why the after-hours call carries the weight it does, covered in how to handle 2am emergency calls.

What should you hold a spam screen to?

Whatever you end up using, hold it to these. It should:

• Answer every call and sort on intent. Nothing rejected on the number alone, and no menu in front of a customer.
• Hand the caller to a person when they ask for one. The first thing to test, and the thing most demos skip.
• Take more than one call at once. Eight o'clock Monday is where a single-line setup quietly costs you.
• Have somewhere to put the gray zone. Suppliers, subs, adjusters, applicants and existing customers get routed, not binned.
• Reach a real person after hours. A text into the void at 11pm is a notification, not coverage.
• Give you the number, the transcript and the callback fast, and keep them, so you can look back at a week instead of guessing at it.
• Cost you nothing for junk you never wanted. Ask any vendor how automated and dead-air calls are counted and billed. The answer is telling.

Where Willison fits

Willison answers every inbound call, 24/7, in seconds. It greets the caller, answers their questions, works out what the job is, and books qualified work straight to your calendar. It can text you the details when something needs you. It never decides whether to pick up based on the number, because it picks up every time, and the sorting happens inside the conversation, where the information actually is. The booking half of that is its own subject, covered in how an AI receptionist books jobs to your calendar.

On a restoration loss the ending is different, as it should be. Willison triages the emergency, what is leaking, where it came from, whether the source is shut off, how urgent it is, and the address, then hands those details to your on-call person, even if that is you. It does not send a truck and it does not promise an arrival time, because that call is yours to make.

So what you get isn't a blocklist. It's a front desk. The recorded pitch takes a few seconds of a machine's time instead of yours, and the homeowner with a dead water heater gets qualified and booked. The same sorting on a plumbing line runs deeper than two lanes.

The only real way to judge whether it can tell the difference is to be the caller. Talk to the live Willison demo right in your browser on willisonhq.com and try it from both sides: pitch it something first, then call it like a customer with an actual job. You'll know inside of thirty seconds which side of the line you're on.

Frequently asked questions

Yes, but differently than a call blocker does. It does not refuse calls based on the number, it answers every call and sorts on what the caller is there for. Anything automated falls out immediately, since an autodialer cannot answer a question about what it needs. A trained human reading a sales script gets further, but the pitch still never turns into a job. The trade-off is deliberate: because nothing is rejected before it speaks, a first-time customer calling from a number you have never seen still gets through.

It should never block anyone, because blocking is not how it works. The failure mode is misreading a caller, not refusing one, and misreading fails out loud, in front of somebody who can say what they actually need. A number filter fails the other way, silently and for good, and you never hear about it: per Invoca's platform data, fewer than 3% of callers pushed to voicemail leave a message.

Because a phone number tells you nothing about intent. Number-based screening only knows how a number has behaved in the past, so a homeowner who just moved into your area, an adjuster on a corporate line, and a robocall all look identical to it. The two errors also cost wildly different amounts: filtering a robocall costs nothing, and filtering a customer costs the whole job without ever telling you it happened.

A phone menu makes the caller do the sorting before anyone listens to them: press one for service, press two for billing. Screening on intent does the sorting after the call is answered, so the caller just says what they need in plain words. The difference matters most on urgent calls, where sitting through a menu is the last thing a homeowner watching water spread wants to do.

No, it means you stop being the first filter. Across home services, only 38% of calls answered by a person are leads, per Invoca's 2026 benchmarks, so most of what an owner personally picks up is not new work. An AI receptionist takes that whole layer, and the calls that reach you are the ones worth putting the tools down for.