AI agents and RAG assistants for business: answers that cite your documents

An AI agent answers from your own documents, procedures, contracts or reference standards, and can take action in your tools. Before answering, it looks for the relevant passages, builds its answer from them and cites the source. This is known as RAG, and Bonne Idée applies it to all its agents so that every answer can be checked.

A language model on its own answers with confidence, even when it is wrong. Grounded in your texts, the agent instead says it cannot find an answer when nothing in your documents matches.

It can also act: answer your teams from your procedures, check a contract or a file against a reference standard, gather the documents for a file, prepare follow-ups. Human approval remains built in wherever a mistake is costly.

Our own agent, CoBIos, co-built with the Orléans Commercial Court Registry, works this way. Our AI agents are built on your documents; CoBIos is our ready-made agent for accounting and law firms; BIA is our sovereign AI workspace, run locally.

A first agent can be queried from the first two-week sprint, after the two weeks of scoping. Three things change the timeline:

• The state of the corpus. Documents that are up to date, organized and free of duplicates shorten everything. Conflicting versions need sorting out, which is part of the project.
• Access rights. An agent that has to respect who is allowed to see what takes more work than an agent on documents open to everyone.
• Actions. An agent that answers is quicker to deliver than an agent that acts in your tools, because every action has to be tested and kept within limits.

A narrow first agent that is live and measured is better than an agent that covers everything and that nobody dares to use.

The risk of a made-up answer never falls to zero. It can be measured, and that is the condition for putting an agent in your teams’ hands.

The test set of questions

We write a series of real questions with you, each with the correct answer and the document that supports it. Every version of the agent is run on this set, and every answer is classified: correct and sourced, correct but poorly sourced, wrong, or « I don’t know. » An agent that admits it cannot find an answer is better than one that makes things up, so the last category counts as good behavior.

What we do with the result

Errors almost always point to a specific cause: a missing document, an out-of-date version, an ambiguous question. We fix the corpus, measure again, and go-live only happens once the score agreed during scoping is reached. The same set is then used every time the model or the documents change.

An agent can read, search, summarize, prepare a document, fill in a field or suggest a follow-up. The line between what it does on its own and what waits for human approval is drawn during scoping, action by action, according to what a mistake would cost.

The rule we apply: anything that goes out to a customer, commits money or changes reference data goes through a person. Every action is recorded in a log, so you know at any time what the agent did, based on which source, and who approved it.

When the agent needs to trigger sequences of steps in your tools, we back it with an automation hosted in your own environment. If your teams first need to learn how to use it, the Consulting practice prepares the ground with adoption and change management.

Take a common situation: an SME whose quality procedures, technical data sheets and internal rules are spread between a file server and a few people’s heads. The same questions come back every week, put to the same people.

• Scoping lists the most frequent questions, identifies the authoritative documents and sets aside out-of-date versions.
• The prototype answers these questions on a first batch of documents, each answer with a reference to its source.
• Measurement runs the agent on the test set of questions written with you, before any go-live.
• Go-live opens the agent to a team, in the tool it already uses, with a process for handling unanswered questions.

Unanswered questions show which documents are missing, and that is often the first benefit of the project.

An agent reads your documents: where they are stored and who has access to them matters as much as the quality of the answers. The agent respects the access rights already defined in your tools. A data processing agreement (DPA) is signed with the contract, everything is hosted in France by default, or on your own infrastructure when nothing must leave, and no model is trained on your documents.

We integrate AI for commercial court registries, public and ministerial officers who are demanding on this point. Human review and what GDPR and the EU AI Act require are set out on our AI guarantees and data sovereignty page.