Plugging a model into a folder of documents is easy. Getting correct answers is much harder, and the work happens almost entirely before the model.
The problem is not volume
We often see companies postpone a project while they wait to have “enough documents”. It is the opposite: a few hundred clean pages give better results than ten thousand contradictory ones. The real obstacle is not quantity, it is contradiction.
What to exclude
Drafts, outdated versions, internal exchanges and documents nobody can vouch for. A model does not distinguish a 2019 price list from a 2026 one: it will answer with the most similar, not the most recent. Exclusion is the highest-return action in the whole project.
Structure without rewriting everything
This is not about reformatting your archives. It is about adding three pieces of information to each document: what it covers, from what date it is authoritative, and who validated it. Those three pieces of metadata turn a pile of files into a searchable base.
An assistant does not correct your documents. It amplifies them, mistakes included.
Control the answers
We always require the source to be cited. An answer with no document cited is treated as no answer at all. That single rule removes most invented answers and makes review possible by a non-technical person.
Where to start
Pick one domain — HR procedures, technical sheets, commercial terms — and do it properly. Extending to the rest of the archive is then mechanical.