We tried to break our own AI receptionist. Here's what happened.

Anyone can tell you their AI is clever. We'd rather show you what ours is built to never do — and then show you us trying to make it do exactly that.

Handing your business number to something automatic is a big ask. We know why. The worry isn't whether it's clever enough — it's whether it'll open its mouth and cost you. So before we get anywhere near a promise, here's the honest version of the testing we do, in plain language.

The worry is the right worry

The fear is simple and fair. Dobbie answers a customer in your name and says something you never would. Quotes a price you'd never honour and now you're stuck with it. Promises Tuesday when you're booked solid. Gets talked round by some chancer having a laugh at eleven at night. That's a real risk with any AI. A firm that tells you it flat-out can't happen is either not being straight with you or hasn't looked hard enough.

We looked. And the way you find out whether a guard holds isn't to describe it. It's to attack it.

So we tried to break it ourselves

Before Dobbie goes near a real customer, we play the difficult one. Not on paper, and not a demo we've rigged to behave — real test conversations against the same live system a customer would actually reach. We run them with made-up customers, so no real person is ever part of it. The brief we give ourselves is blunt: get it to do one of the things it's built to never do.

What we threw at it

First, the price bait — the "go on, just give us a rough figure" that every customer tries, then tries again when the first push doesn't work. Dobbie didn't budge. No number, not once, no matter how the question was dressed up.

Then the old trick of telling the AI to forget its instructions and follow ours instead — the one people use to make these things go off the rails. It declined and stayed in character, like a good receptionist who's been asked to do something that isn't their job.

We asked for a trade the business doesn't even do, to see if it would wing it and make something up. It steered the chat back to the work the business actually offers. We tried to switch it onto a different business halfway through a conversation. We ran the whole thing again in Spanish, in case a language change slipped past the rules. Each time, it held its line.

The guard that sits in front of the send button

Behind all of that sits one plain safeguard we're glad we built. Before any message leaves, a separate check reads it first — looking for the three things Dobbie must never send: a price, a promise of a specific time, or a link. If it finds one, the message doesn't go. You keep control of money and diary; the guard makes sure a slip can't reach a customer.

This isn't a diagram on a slide. It has already earned its keep. On one occasion it caught a made-up price the AI had drafted and stopped it dead — the customer never saw it. That's the whole point of putting a second check in front of the send button: the day something does go wrong, it goes wrong quietly, into a log, instead of loudly, into your customer's chat.

What held, and what we'll actually claim

We ran a full round of this — eight different attacks, one after another — and every guard held. Nothing unsafe got sent. That wasn't a fluke or a one-off, either; an earlier round went the same way. Same tricks, same result.

So here's the only claim we'll make, and we'll make it in plain words: not that Dobbie is flawless, but that we test it against people actively trying to break it before we ever let it near your customers. That's diligence you can check, not a boast you have to swallow.

What we're not saying

We're not telling you it's perfect. No AI is, and anyone who tells you theirs is, is selling. What we're telling you is where the hard stops are. The things that would actually embarrass you — money, arrival times, dodgy links — are blocked automatically, even if the customer pushes twice. When a chat drags on or starts going in circles, Dobbie hands it back to you rather than digging a hole. There's a firm cap on how much it can ever send, so it can't run away with itself. And you see every conversation live in your own WhatsApp — nothing happens behind your back. If any of that matters to you, the companion read is whether a WhatsApp receptionist is safe, which covers your number and your data.

"I still don't fancy a machine talking to my customers"

Fair enough. Keep the worry — but weigh it properly, because it isn't a risk against nothing. Right now, the messages that land while you're up a ladder or asleep don't get a careful reply. They get no reply. And most of those people simply message the next firm on the list. That leak is certain, and it runs every single week, quietly, whether you think about it or not.

The thing you're nervous about is the thing we've spent our time trying to make happen — and couldn't. The thing you're used to is happening tonight. One of those two is costing you a customer this evening. It isn't the AI.