What happens when a robot cold-calls your AI receptionist

Picture what actually happens. An outbound agent opens with a warm line about your Google listing. Your front desk, doing exactly what it was designed to do, listens, acknowledges, and tries to help. The caller has a script and infinite patience. Your receptionist has good manners and no reason to hang up. Neither one gets bored, neither one has somewhere else to be, and neither one is the party paying for the minutes.

The call does not end because someone loses interest. It ends when it hits a limit — which, on most AI phone products, is a maximum call length measured in tens of minutes. That is a long time to pay for a conversation that was never going to become a booking, and it can happen several times a week without you noticing, because from the dashboard it looks like a normal answered call.

The direct cost is minutes. Voice AI is metered, and a handful of long calls with diallers can eat a meaningful slice of a small plan before a single customer has rung.

The indirect costs are worse, because they corrupt the things you use to make decisions. A cold caller shows up in your call volume, so a quiet week looks busy. It shows up on your leads list with a phone number, so someone eventually rings a dialler back. If you have automated follow-up texts to callers who did not book, it gets one of those too — you have now paid to send a marketing robot an SMS.

None of this is catastrophic. It is just money and attention leaking somewhere that produces nothing, which is exactly the kind of leak that runs for a year because no single instance is big enough to investigate.

The behaviour you want from a human receptionist is well understood. They hear the opener, recognise it inside a sentence, say some version of "we don't take sales calls, thanks," and put the phone down. They do not take a message, they do not offer a callback, and they certainly do not book the salesperson in for a chat with the owner.

An AI front desk should do the same, and it needs three things to manage it. It has to recognise the shape of a pitch as distinct from a customer question — "do you do Google reviews for my salon?" is a customer, "I'm calling about your Google listing" is not. It has to be willing to end a call, which cuts against everything else it is trained to do. And it needs a ceiling on how long it will keep talking to someone who is plainly not trying to buy anything.

The last one matters most, because it is the one that does not depend on catching the opener correctly. A call where nobody has asked about a price, a time, or an availability — where no booking, no lead, no quote has been attempted — has a natural length, and it is not thirty minutes.

Estric AI screens for this on every voice call. The assistant is instructed to treat sales, marketing, SEO, recruitment and survey calls as non-customers: one polite line, then the call ends. It will not take their details or promise a callback.

Underneath that, the line watches for the tells of an automated caller — replies that begin faster than a person decides to speak, reply timing that barely varies from turn to turn, a whole conversation without a single hesitation or interruption, a pitch opener, a caller that announces itself as an AI. No single one of those decides anything, because each has an innocent explanation. Two together end the call in well under a minute.

When that happens, three things follow. The call is logged as spam rather than as a customer call, so it does not inflate your numbers or land on your leads list. The number is flagged on your Risk page, where you can block it permanently or dismiss it — the system never bans anyone by itself, because the occasional misjudged call is a real person who must still be able to reach you. And the minutes are not charged to your plan.

Whatever you run your front desk on, three questions are worth asking your provider. What is the maximum length of a single call, and who pays for it? What happens to a caller who never asks about anything bookable? And can you see, after the fact, which calls were junk — or do they all look identical in the dashboard?

Then look at your own call log for last month and read the longest five calls. If one of them is a sales pitch your receptionist sat through to the end, you have found the leak.