Can AI answer questions about our private documents?

Choose a small set of questions staff genuinely need answered: where a procedure is recorded, which version applies, or what a policy says about a particular situation. Include questions that the documents cannot answer. This gives the project a meaningful check beyond producing fluent text.

The proposed output should let a person inspect the source passage or document. Agree what the system should do when information is missing, contradictory or out of date, and who will decide whether an answer is good enough for the intended use.

Specify where original files, extracted text, search indexes, model processing and conversation logs may be stored or processed. A local chat model does not by itself settle the behaviour of every other component.

Separate teams may need different access. Agree whether the first scope is a single-user tool or a shared service, which collections each user can access, and how documents are removed or replaced. Use sample material for the first discussion rather than sending a complete confidential archive.

A clean folder of text documents is a different project from scanned PDFs, mixed file types, several user groups or a live connection to another application. Document preparation, access controls and evaluation determine the work.

The quote should identify the collection, allowed formats, update process, expected users and acceptance examples. Extra integrations, optical character recognition and ongoing operation should be named explicitly if they are needed.