Conversational bots and applied AI
A model is superb at conversation
and terrible at guarantees.
So whatever cannot fail is not trusted to it. The model listens, understands and replies; decisions with consequences are taken by code that can be audited, or by a person. It sounds less modern and it survives the first week in production better.
How responsibility is split
What the model does and what the code does
This division is the whole proposal. It is not a technical limitation to be overcome with a better model: it is the decision that lets you put a bot where mistakes cost something.
What we do trust the model with
- ·Understanding someone who writes badly, in a hurry or in another language
- ·Drafting a reply that sounds like a person and not a form
- ·Summarising a long conversation for whoever picks it up later
- ·Deciding which topic a query most resembles
What we never trust it with
- ·Working out a dose, a price or an amount
- ·Deciding whether someone meets a risk criterion
- ·Confirming availability, stock or a booking
- ·Anything whose error shows up on an invoice or on a person
A case where the difference can be measured
In a health project, the risk questionnaires are not answered by the model: they run on buttons and rules. The conversation around them is natural, but the part that decides is code.
100 %
accuracy in its test battery
17
red-team attacks, zero leaks
With a language model answering those questions you cannot claim the same, however good it is. The hallucination surface is not small: it is zero, and that is a property of the design, not a promise.
What comes as standard
Official WhatsApp API with verified signatures
Language detection
Handover to a person with web inbox
Sending technical datasheets
Frequently asked questions
Where to put the model and where not to
Why do you say the language model should not be trusted with everything?
Because a model is excellent at conversation and terrible at guarantees. There is no way to prove it will never get a dose or a risk criterion wrong: it can be right a thousand times and wrong on the thousand and first, and you will not know which. So anything with consequences goes into deterministic rules that can be tested one by one, and the model keeps what it does well: understanding and drafting.
Does that not make the bot dumber?
It makes it more boring and more reliable, which in production is what you want. The person writing notices no difference: they still converse in natural language. What changes is what sits behind the answers that matter. A risk questionnaire on buttons and rules has zero hallucination surface, and that is not an opinion: it is a property of the design.
Do you have any measurement of that?
In a health case, risk questionnaires run on buttons and rules rather than on a model. In their test battery they scored 100 % accuracy, and withstood 17 red-team attacks without a single leak. With a language model answering those questions you cannot claim the same, however good the model.
Do you work with the official WhatsApp API?
Yes, with the official WhatsApp Business API, with request signatures verified. Telegram too. It matters because bots built on unofficial WhatsApp work until they stop working, usually with the account blocked and no warning.
What happens when the bot cannot continue?
It hands over to a person, with the whole conversation and a summary, into a web inbox. A bot with no route to a human does not save work: it shifts it to the angry customer who phones instead. Language detection and sending technical datasheets come as standard.
Can you do anything beyond chat?
Yes. We have built physical interfaces with a 3D robot, camera gesture recognition and voice with a wake word, configurable per client. The conversation does not have to happen on a phone screen.
What happens if your bot gets it wrong?
If the answer is «nothing serious», almost any tool will do. If the answer involves an invoice, a dose or a person, let's talk.
Tell us your case