AI document assistant

In every organisation a good share of the day's questions already has a written answer somewhere: in a procedure, a closed file, a contract, a circular, the minutes from two years ago. The problem is not that the answer is missing — it is that finding it costs more than asking the colleague who "knows where to look". A document assistant reverses that relationship: the knowledge stays where it is, and becomes something you can question in plain language.

What the system does

• It indexes your documents where they live. Shared folders, the document management system, mail archives: documents do not move; they become searchable by meaning, not only by exact word.
• It answers with its sources cited. Every answer is built from passages retrieved from the archive and points to them: the reader can open the document and verify in one click. In this design, an answer without a source does not exist.
• It admits what it does not know. When the archive does not cover the question, the system says so — a designed and tested behaviour, because the alternative is an assistant that fills the gaps with imagination, and in a business that is expensive.
• It respects the compartments. Each person receives answers only from documents they already have access to. Existing permissions are the assistant's boundary, not a suggestion.

Typical cases

• Internal procedures and policies: "how do we handle a return from an overseas customer?" finds the right procedure, in the version in force.
• Case history: "have we dealt with a case like this before?" searches the closed files, which are the organisation's real memory.
• Contracts and suppliers: deadlines, clauses, terms — found in the signed text, not recalled from memory.
• Onboarding: the questions of the first months get answered without interrupting the most senior colleague every time.

A limit worth knowing in advance

The assistant returns the quality of the archive it questions. If documents contradict one another, if superseded versions sit beside current ones with nothing to tell them apart, the best the system can do is surface the conflict — resolving it is yours to do. That is why the project begins with an analysis of the archive, and in some cases the honest outcome is that before an assistant you need a pruning. Better to know at once than to discover it in production.

Frequently asked questions

Do our documents end up with an external provider?

It depends on the configuration, and it is decided beforehand: model providers are used in the business arrangements where submitted documents do not train their systems, and for the most confidential archives the whole apparatus — index and model — can run on your own machines. The choice is written into the project scope, with the costs and limits of each route.

Can the assistant make answers up?

A language model left to itself can; that is why the design constrains it. The assistant answers only from passages retrieved from your documents, every statement carries a link to its source, and when the archive does not contain the answer the system is instructed to say so. The risk does not vanish: it is reduced, made visible — a citation is checked in one click — and measured during the prototype.

Who may ask what? Not everyone here sees everything.

The assistant inherits permissions, it does not bypass them: it answers each person only from documents that person could already open. This is among the first things designed, because an assistant that fishes in the whole archive is, in effect, a breach of your internal compartments — convenient and dangerous.

Our archive is a mess — do we need to tidy it first?

Less than you fear, more than zero. The system copes well with imperfect archives — mixed formats, fanciful file names — but it cannot resolve substantive ambiguity: if three undated versions of the same policy coexist, no index knows which one is true. The analysis stage identifies the places where tidying genuinely matters, and they are usually few and well defined.