Build it
Make it sell
Keep it running
Custom AI assistants
An assistant that knows what it must not say.
Most chat widgets have one of two problems. Either they answer things they should not, or they refuse so much that nobody uses them twice. We write ours, which is why they can do neither.
What building one actually found
Four of the shop’s own recommendations contradicted its own warnings.
While building the assistant for Apotheke zum heiligen Joseph, we checked every cross-selling suggestion against the upper limits for each nutrient. Four of them proposed combinations that went past a limit the product page itself warned about. Those texts had been in the shop for years and nobody had added them up, because nobody had ever needed to. The assistant did not cause that — building it simply made it visible, which is a fair description of most of the work.
How the line is drawn
Four levels, not two.
Almost every assistant is built with two categories: allowed and forbidden. That is why so many of them are useless. The interesting work happens in the two levels in between — and the example below is a pharmacy, because that is where the limits are tightest.
-
01
A question about the product
What is asked
Is this vegan. How many are in a pack. What does a day cost. How long does it last. What is actually in it.
What it does
Answers in full, from checked data. These are the questions that quietly stop a purchase, and they are the entire commercial case for having an assistant at all.
-
02
A question about what something does
What is asked
What is good for the immune system. Which of your products contain magnesium. What is the difference between these two.
What it does
Answers in full, with products and links. This is the level most assistants get wrong: they refuse because the question sounds medical. It is not — it is the wording the shop is allowed to use, and refusing it is the largest silent loss of the whole system.
-
03
A person describing themselves
What is asked
I am tired all the time. I sleep badly. Anything that is about the person rather than about the product.
What it does
Stops naming products and hands over to a person, with the phone number. It may still explain generally, because refusing to speak at all is unhelpful — but it does not turn a description of a person into a recommendation.
-
04
Something it must not answer
What is asked
What helps against a named condition. Medication, pregnancy, children, anything with a diagnosis in it.
What it does
Hands over, and — this is the part almost everyone gets wrong — does not sidestep. “I cannot discuss that, but for general wellbeing we do have…” is worse than saying nothing, because it makes exactly the connection that was not allowed. It has to be forbidden explicitly, or the model does it by itself.
Built rather than installed
Four things a subscription cannot give you.
It knows your actual range. Not by reading your marketing copy and guessing, but from fields your own team fills in and maintains: quantities, composition, what it costs per day, what it is not suitable for. That is the slow part — twenty to thirty minutes per product — and it is the reason the answers are right.
It costs a few euros a month. Around four, for a range of that size at a hundred conversations a month, because we chose a small model and cache the unchanging part of every request. The task is reading facts from checked material under strict rules, not free thinking. A larger model is one line if you ever need it.
The account is yours. Registered in your name, with us invited as a member. If we ever part company, it keeps running and you change nothing. That is the same principle as your domain and your hosting, applied to a thing most agencies quietly keep hold of.
Nothing is stored. The conversation lives in the visitor’s own browser and ends with the tab. No server log, no cookie, nothing to delete later. On a shop where people ask about themselves, that is not a detail.
Two days of work for a range of about twenty products, most of it on the data rather than the code. And the limit, plainly: if what you want is a system that acts on behalf of the visitor — placing orders, changing bookings — that is a different project and we will say so. This one answers, hands over, and knows the difference. What it answers about is up to your range; we have built it for a pharmacy, and the same structure works for a hotel, a service business or a B2B catalogue.
Try the one that is already running.
There is a working assistant on a client shop right now. Ask it something awkward and see where it stops. Then tell us what yours would need to answer — and, more importantly, what it must never answer.