An assistant that answers from your documents, not from its imagination.
It reads your documentation, your product pages, your procedures. It cites the source of every answer. And when the information does not exist on your side, it says it does not know and hands over — that last property is what decides everything.
- From
- 1 990TND
- or
- 1 490€ excl. VAT
- Live in
- 3weeks
- Quote within
- 24hours
The trouble with assistants is that they always answer.
A raw language model was trained to produce a plausible answer, not a true one. Ask it about your prices, your lead times or your returns procedure: it will answer, confidently, and invent. That is what turns an assistant into a commercial liability — a customer promised something that does not exist.
The technical answer is grounding: the assistant answers only from documents you give it, it says which ones, and it is explicitly constrained to say "I don't know" rather than fill a gap.
That constraint is the whole point. An assistant that admits ignorance and offers to pass the question on is useful. An assistant that always answers is a complaints generator.
Four uses that hold, and two that do not.
What works. Answering recurring questions from your documentation. Helping someone choose within a catalogue by asking the right questions. Guiding a visitor to the right page or the right form. Qualifying an incoming enquiry and routing it to the right person with a summary.
What does not. Replacing a human adviser on sensitive matters — health, law, an open complaint, a dispute. And negotiating or committing the business: an assistant must never state a price or a lead time that is not written in black and white in the documents you gave it.
We set that scope at the first meeting and we write it into the quote. Anything outside it goes to a human, with the context of the conversation.
An assistant in service, and the means to steer it.
The entry package covers an assistant fed by your documents, up to two hundred pages of content.
Included in the price
- Scope set: what it answers, what it refuses
- Your documents collected and prepared
- Assistant grounded in your content, with sources cited
- Explicit refusal when the information does not exist
- Handover to a human, with a summary of the conversation
- Interface built into your site, in your own identity
- History of the questions asked, readable by you
- Documents updatable without any involvement from us
- Guardrails: forbidden topics, tone, answer length
- Two hours of handover
Billed separately, and quoted up front
- Writing the documentation that does not exist yet
- Connecting to a CRM, an ERP or live stock
- A voice or telephone assistant
- A version in a language beyond the one delivered
- The cost of model calls, passed through at cost
- Human moderation of the conversations
A setup fee, then a consumption cost.
- Tunisia
- from1 990TND
- France, Europe, international
- from1 490€ excl. VAT
Setup is a one-off figure. On top sits the model's consumption, passed through at cost with no margin from us: on a site taking a hundred or so conversations a month, that stays in the order of a few tens of euros monthly. We give you a costed estimate based on your actual traffic before starting.
We prefer models that can be self-hosted whenever the nature of the data calls for it — internal documents, personal data, a public contract with a sovereignty clause. It takes longer to put in place and we say so, but it avoids sending your content to a third party.
Three weeks, and you approve the answers before it goes live.
Scoping
The twenty questions your customers actually ask, taken from your emails and your conversations. The scope: what the assistant answers, what it refuses, what it passes on. Then an inventory of the documents available.
Days 1 to 4Preparing the content
Your documents collected, split and indexed. This is the step that decides answer quality: a badly structured document gives approximate answers, whatever the model.
Days 5 to 10Tuning and acceptance
Guardrails tuned, then the twenty reference questions run through. You read the answers and correct them. Nothing goes live before you have signed off that list.
Days 11 to 17Launch and follow-up
Built into the site, put into service, and the questions asked in the first month recorded. That record is what shows which documents are missing — the assistant then becomes a diagnostic tool on your own documentation.
Days 18 to 21
What people ask before they sign.
Can it still invent an answer?
The risk is never zero, which is why we handle it by construction rather than by promise: the assistant answers only from your documents, it cites its sources, and it is constrained to refuse when the information is not there. Acceptance on your twenty reference questions exists precisely to check that behaviour before launch.
Where does our data go?
That depends on the option chosen, and it is a decision you make with the facts. With a model hosted by a provider, your questions pass through them — acceptable for public documentation, arguable for internal documents. With a self-hosted model, nothing leaves your infrastructure. We put the cost of both in front of you before you choose.
Do we need a lot of documentation?
You need good documentation rather than a lot of it. Twenty precise, current pages beat three hundred contradictory ones. If your documentation is thin, the assistant's first value will be showing you what is missing, through the questions it cannot answer.
How long before we see an effect?
On the volume of recurring questions arriving by email or phone, the effect is measurable in the first month — provided the assistant is visible and not hidden behind a bubble at the bottom of the page. On sales conversion it is slower, and it depends mostly on the quality of your content.
Does it replace someone?
No, and we do not sell it that way. It takes repetitive questions away from people who have better to do, and it answers at three in the morning. An assistant asked to replace an adviser on a sensitive matter always ends up creating a problem more expensive than the time it saved.
Two ways to start.
You know what you want
Describe what you need in a few lines. You get a priced quote, with the scope and the dates.
Request a quote Answer within 24 business hoursYou would rather talk it through
Message us on WhatsApp or call. Fifteen minutes are usually enough to know whether your need fits the package — and we will tell you plainly either way.
Message us on WhatsApp TN +216 58 044 066 · FR +33 7 69 78 69 98