AI Assistants That Answer From Your Documents, Not the Internet

An AI assistant for business documents should answer from approved policies, price lists, and guides, cite the source behind the answer, and hand the question to a person when the documents cannot support it. Its value depends on the quality of those documents and the checks around the answer, not how confidently it writes.

The useful assistant is the boring one. A teammate asks where a request belongs and gets the current instruction, with a link they can open. Nobody has to search an old chat or guess which copy of the policy is right.

Which questions belong in a document assistant?

Start with questions whose answers already exist and change through an agreed process. Booking rules, onboarding steps, service boundaries, and internal instructions are reasonable candidates. Negotiating a contract, approving an exception, or interpreting a complaint still needs a person who can take responsibility for the decision.

Write down the question, its source, and the person who owns that source. If the team disagrees about the answer, settle the disagreement before teaching an assistant to repeat it.

The same check applies to public-facing answers. A customer should receive approved information, not an employee’s private notes. Separate internal and public document sets, and decide which signed-in roles can read each set before connecting them to the assistant.

Make the documents answerable before uploading them

A useful source has a clear title, an owner, an effective date, and enough context to answer without another document open. Remove superseded copies from the approved set. A folder containing conflicting price lists gives an assistant a conflict to manage, not a reliable price.

Review tables, screenshots, and scanned PDFs. A price shown only inside an image may need a checked text version. Keep conditions beside the statement they qualify: a cancellation rule without its exceptions can produce a technically quoted but misleading answer.

Use a small source register:

Source	Owner	Check before approval
Service guide	Delivery lead	Scope and exclusions agree
Price list	Business owner	Current version is clearly marked
Booking policy	Operations lead	Exceptions and contact route are included

For material that changes throughout the day, consider a controlled lookup in the underlying system rather than another static document export.

How should the assistant find and cite an answer?

The assistant should search the approved set, use relevant passages, and attach a source reference the reader can inspect. A citation is useful only when the cited passage actually supports the answer. Showing a document title beside an invented explanation does not make the explanation trustworthy.

For example, OpenAI’s file search documentation describes searching uploaded files and returning file citations. The application still has to display those citations clearly and check that its document permissions and answer rules are correct.

Ask for the smallest useful answer. If someone asks where to send an onboarding form, give the destination and the source. A long summary of the whole onboarding process can hide the instruction they needed.

Keep source links usable after the conversation ends. An answer copied into a ticket should still identify which approved document and section it came from.

Treat missing evidence as a handover

When the approved documents do not cover a question, the assistant should say that directly and provide a working route to a person. A good fallback identifies what is missing without pretending to have resolved it. “The current guide does not cover this exception” is more useful than a confident guess.

An escalation should carry the question, the sources checked, and any information the requester has already supplied. Do not make the person repeat the conversation. Also avoid copying private document extracts into a public ticket or email.

Document text needs a boundary too. Instructions found inside an uploaded file should be treated as source material, not permission to change the assistant’s rules or send data elsewhere. Keep actions such as refunds or account changes outside the assistant until their permissions and approvals are defined separately.

A hypothetical booking-policy example

Suppose a studio’s guide says a booking can move before work begins, but custom rush bookings need the owner’s approval. A customer asks to move a rush booking. The right answer cites the exception and offers to pass the request to the owner. It does not treat the general rule as approval.

Now suppose the customer asks for a discount that the guide never mentions. The assistant should explain that it cannot confirm a discount from the approved documents and route the request. Silence in a policy is not permission.

After the owner changes the rush policy, the document owner updates the source, refreshes the assistant’s searchable document set, and repeats both questions. The test checks the answer, the citation, and the escalation. This is a hypothetical acceptance test, not a client result.

Test the awkward questions before launch

Use questions the assistant ought to answer and questions it ought to refuse. Include an outdated policy, conflicting passages, a request for private material, and wording that does not match the document heading. A fluent answer is not the pass condition; a supported answer or a correct handover is.

Record the expected source and acceptable response beside each test question. When a source changes, run the relevant checks again. Keep examples of failed answers so later changes do not reintroduce the same mistake.

Assign someone to review unanswered questions. They may reveal a missing guide, an unclear policy, or a request that should stay human. Our handover checklist explains why the document register, tests, and update instructions belong to the team that owns the assistant.

Frequently asked questions

Does a document assistant still need human review?

Yes. A person should own the sources, check important answers, and handle exceptions. Searching approved documents reduces the scope of the answer; it does not guarantee that every generated sentence is correct.

Can internal and customer documents share one assistant?

They can share an interface if access is enforced for each request, but the document sets should follow the reader’s permissions. Do not assume a prompt alone will protect internal material.

What happens when a policy changes?

Update the approved source, refresh the searchable document set, and repeat the affected questions. Check that the answer cites the new version and that the old version is no longer available as an approved source.

When is an assistant the wrong choice?

When answers depend on undocumented judgment, sources conflict, or nobody can maintain them. Clarify the process first. A searchable guide or a simple form may be enough until the knowledge is ready.

Choose a small set of repeat questions and test whether the approved documents can answer them. If the gaps are in the process rather than the documents, start with a systems audit before adding an assistant.