Anna: the agent we built by taking things away
The customer we did not expect
Most of what Botzone builds is aimed at people who work at desks. Sage sits in your meetings. Cost wraps your API clients. Otto reads your calendar. All of it assumes a laptop.
Anna is for an electrician on a ladder.
She answers a trade business's WhatsApp, day and night. A customer messages at half nine at night, Anna replies in about four seconds, works out what is actually wrong, asks the two or three questions the owner would have asked, and then sends the owner the whole thing as a single job card. Urgency at the top, what is wrong, where it is, who it is, and the full conversation underneath.
She is live in beta now at anna.botzone.ai, and you can watch her answer a customer of your own business in about thirty seconds without signing up for anything.
The problem is not customer service
It would be easy to frame this as small businesses being bad at answering the phone. That is not it, and any tradesman will tell you so within about ten seconds.
You cannot answer the phone with your hands full. So it rings out. By the time you are back in the van and have washed your hands, they have rung the next fella.
There is a number attached to this, and it is worse than most people guess. 37% of people will ring somebody else if you do not answer inside a minute. Only 6 in 100 leave a voicemail. That is YouGov for Answer4u, 2,046 UK adults, June 2026. We put the source on the page because a tradesman will absolutely google it, and should.
The compounding part is that you never find out. A missed call leaves no trace. There is no report at the end of the month telling you that you lost four jobs because you were under a sink. The cost is entirely invisible, which is exactly why nobody fixes it.
Four options, and only one of them is new
When we talked to trades about this, we found they were already choosing between the same four things. So the site says so plainly rather than pretending the alternatives do not exist.
We think the honest competitor is not the other AI products. It is option one. Most of these businesses are doing nothing, and doing nothing is free.
The product is the list of things she cannot do
Here is the part that matters, and it is the reason this post exists.
The first question every single tradesman asks is some version of "what if it says something stupid and it costs me a job?" That is the correct question. Their name is on the van. A confident wrong answer from a bot they were talked into is a worse outcome than a missed call, and they know it.
So the design work went almost entirely into constraint rather than capability.
She never books anyone in. Anna has no way to put anything in a diary. She can say a slot looks free, but the owner confirms every job personally. This is not a setting that ships switched off. There is simply no such button.
She never invents a price. If she has not been told a rate, she says so and passes it to the owner rather than guessing.
She will not diagnose from a photo. Customers can send pictures, and half of them do, because most people do not know what the thing on their wall is called. Anna looks at the photo only far enough to tell whether it is the right shot. If it is the wrong end of the boiler, or too dark, she asks for another one there and then, which saves the owner finding that out on the doorstep. But she will not read a fault off it. No cause, no condition, no is-it-safe, no price. That is somebody's trade and somebody's name, and a guess off a photo is exactly the thing that would cost them a job.
Every Anna is attacked before she goes live. Each one is put through 20 adversarial probes, graded by a second and stronger model, that try to trick her into inventing a price, claiming work the business does not do, leaking the owner's details, or pretending to be the owner. A failure needs a majority verdict across independent gradings before it counts. Then a human reads the results and approves each one by hand, and regenerating a demo throws that approval away.
We are deliberately not claiming every Anna sails through all twenty. In one full sweep of 105 demos, not one hit a clean zero-failure pass. The claim is narrower and true: no Anna reaches a real customer without being attacked twenty times and then approved by a person.
What she does do
Constraints only sell if the thing still works. What she genuinely handles:
She is built from the business's own website, its reviews, and its trade registration where one exists. Every fact she holds carries the quote it came from. There is nothing for the owner to fill in before they can see her working.
Disclosure is about to stop being optional
One more reason this shape of product matters now.
Article 50 of the EU AI Act, the transparency obligation, becomes enforceable on 2 August 2026. It requires that a person interacting with an AI system is told so. Penalties run to 15 million euro or 3% of global turnover.
Anna already opens by saying who she is: "Hi Sinead, I am Anna, I look after Dave's messages." Not because of the regulation, but because pretending to be the owner was never going to survive first contact with a customer who then meets him. It is a useful illustration that the compliant design and the design that actually works are frequently the same design.
If you want that mapped properly across systems you already run, that is what Lex is for.
Where she is
Beta. €49 a month, first month free, nothing metered. Founding rate for the first ten businesses, locked in.
Trying a demo on your own business takes about thirty seconds. Getting a live one answering real customers takes a couple of days, because the WhatsApp side has to be approved by Meta and that part is not ours to promise.