Assistants that answer questions from company documents are among the most practical uses of generative AI. Their limits should be understood before they are deployed.
A private document assistant lets staff ask a question in plain language and receive an answer drawn from the organisation's own documents. The technique behind it is retrieval: the system finds the passages most relevant to the question and asks a language model to answer using only those passages.
What it does well
It finds and summarises information that exists in the documents, across many files at once, in seconds. It can show the source passage for each answer, which lets the reader check it. It respects access rules when these are built in, so people see only what they are permitted to see.
What it does not do
It does not know anything that is not written down. If a policy is out of date, the answer will be out of date. If two documents conflict, the assistant may quote either one.
It is weak at questions that need calculation across many records, such as totals or trends. Those belong in a database and a report, not in a document assistant.
It can still produce a fluent answer that the sources do not support. Showing sources reduces this risk and does not remove it, so high-stakes answers need human review.
How to judge one
Prepare a test set of real questions with known correct answers before the build begins. Measure the assistant against it, and measure again after every change. An assistant whose accuracy has been measured can be trusted to a known degree. One that has only been demonstrated cannot.