Your procedures, internal memos, standard contracts and reports contain the answers to most of the questions your teams ask. They still have to be found. An AI assistant connected to your internal documents answers a question in everyday language and indicates where its answer comes from. Here is how it works, what needs to be prepared, and what should not be expected of it.
An assistant that reads before answering
A consumer assistant answers from what it learned when it was built. It knows neither your internal rules nor your latest memo. The assistant we are talking about works differently: before answering, it fetches the useful passages from your own documents. Specialists call this method RAG, for “retrieval-augmented generation”.
How it works, in three stages
1. Preparing the document base
Your documents are split into short passages. Each passage is indexed so that it can be found by its meaning, and not only by its exact words. This work is done once, then each time a document is added or updated.
2. The search
When a member of staff asks a question, for example “what is the procedure for reimbursing travel expenses?”, the assistant searches the base for the passages that come closest to it. It keeps only those that this person is entitled to view.
3. The sourced answer
The assistant then writes its answer from these passages and cites the documents used. The member of staff can open the source and check. If it finds nothing relevant, a well-tuned assistant says so, instead of improvising.
What needs to be prepared
The quality of the answers depends first on the quality of the documents. Before the project, four points deserve your attention.
- Digital, legible documents. A collection that has stayed on paper must first be digitised and put through character recognition.
- Up-to-date versions. If three versions of the same procedure coexist, the assistant may cite the wrong one.
- A clear scope. It is better to start with a specific area, such as human resources or one department's procedures, than with the whole collection.
- Defined access rights. Who can query what? The answer must exist before going live.
- A person in charge of the base, responsible for adding, correcting and removing documents.
Its limits
An assistant of this kind is useful. It also has limits that are better known before committing.
- It can be wrong. It may misinterpret a passage, mix two documents or give an incomplete answer. Citing its sources reduces this risk without making it disappear.
- It knows only what your documents contain. A rule that is written nowhere escapes it.
- It does not decide. For a legal, financial or medical commitment, the answer is a starting point that the competent person checks.
- It requires upkeep. A base that is no longer updated gives outdated answers.
The assistant proposes and cites its sources. The competent person checks and decides.
Confidentiality: the questions to ask
Your internal documents contain sensitive information. Before any deployment, put these questions to your provider and require written answers: where are the documents and the exchanges hosted? Who can access them? Are users' questions logged, and who can read that log? Is your content used to train an external model? What happens to the data at the end of the contract?
Also set rules of use for your teams: what may be asked, what must not be entered, and when human verification is mandatory.
The MOKILIX offering
MOKILIX ASSISTANT is our product for this need: ingestion of your documents, chat interface on the web and on mobile, access control, logging and sovereign hosting. The entry price is stated as from 12,000,000 FCFA excl. VAT, for a timeframe of 8 to 14 weeks.
If your documents are still on paper, MOKILIX SCAN prepares the ground. In both cases, the project begins with two weeks of scoping, priced according to the analysis of your need, with no commitment to what follows. To discuss it, you can request a meeting.
