An assistant that answers from your documents. And cites its sources
Procedures, contracts, product sheets, internal regulations: I build assistants that answer your team or your clients from your own documents, with the source of every answer, and that say when they do not know.
How it works, and why most demos fail
The assistant does not answer from memory. It first finds the useful passages in your documents, then writes its answer from them and cites the source. This technique is called RAG.
Built well, it makes things up far less often. Built poorly, it gives plausible but wrong answers. The difference is the groundwork: document splitting, search, instructions, fallback answers, and monitoring.
- Answers drawn from your documents
- A source for every answer
- "I don't know" rather than making something up
The assistants I build
Where finding the right answer wastes time today.
Internal assistant
Procedures, HR rules, product documentation, sales methods: an answer in a few seconds, with the source.
Client support
Answers based on your documentation, with handoff to a person and a history of the exchange.
Sales support
Product sheets, competitive arguments, and references: your sales team finds the information during the call.
Compliance and internal rules
Regulations, internal policies, contract clauses, for legal, HR, or finance teams.
Several knowledge bases
The assistant chooses where to search, asks a question when the request is unclear, and can trigger a simple action.
How I work
The same pace as the Growth plan: an audit, then one delivery every month.
- Step 1
Prepare the documents
We decide together what goes into the knowledge base, what is out of date, and what needs rewriting.
- Step 2
Tune the search
Splitting and search are adapted to your documents, not left on the default settings.
- Step 3
Cite or hand off
Every answer cites its source. When the answer is not in your documents, the assistant says so or passes the question on.
- Step 4
Test with your questions
A list of real questions, with the expected sources, is the test for every change.
- Step 5
Monitor in production
Exchanges are monitored: answers with no source, questions with no answer, documents to complete.
What it looks like in practice
Three common situations, so you can picture it.
Tenant questions
The assistant answers from the building rules, the standard lease, and the FAQ, and passes the rest to the property manager.
Sales support
The sales team queries the product documentation and client references, with the source of every answer.
Internal HR assistant
Leave, expense reports, health insurance: every answer cites the rule, and ambiguous cases go to human resources.
Illustrative examples of common situations, not client references.
Two ways to start, at a fixed price
An assistant in production on a defined document base, with its tests, its monitoring, and training for your team.
Fixed price, with no overrun. You move to the subscription only if the result suits you.
Start with a projectThe opening audit, then a delivery every month on this subject or on all your other tools, maintenance, and training for your teams.
For the first 4 clients, the first year, first month free. Only 3 places left.
On top of that, paid directly to the providers: AI model and hosting fees, depending on volume, about 50 € to 500 € per month depending on usage. I price them precisely during the audit.
Frequently asked questions
How is this different from an AI agent?
The assistant answers questions from your documents. An agent acts in your tools. The two often work together.
Aren't ChatGPT or Copilot enough?
Sometimes, for public information. A dedicated assistant makes sense when your documents are private, every answer must be sourced, and you want to measure quality over time.
How many documents can it handle?
From a few hundred to several tens of thousands. The architecture adapts to the volume.
Does it learn from its mistakes?
Conversations are monitored and mistakes are caught. The fix is usually in the documents or the search, not in the model.
How do you measure whether it is useful?
Rate of correct answers on your test questions, correct sources, questions passed to a person, user satisfaction, and fewer repetitive requests.
What information does your team look up every day?
A free introductory call, to see if I can save you time. You leave with concrete next steps, whether you work with me or not.
Free, with no commitment. The button opens your email with a message ready to send. You can also write directly tocontact@votre-domaine.fr. Reply within [24 business hours].
Your message already contains three questions to prepare the call
- How many people work in your company?
- What sector are you in?
- Which task wastes the most time for your team?