Preparing documents for an AI assistant starts with choosing current sources for one task, checking readability and recording who may access them. Give each source a version and a content owner. Then prepare questions with expected answers, including questions the documents cannot answer. This gives you a basis for evaluating a pilot before connecting a wider company archive.
Suppose employees regularly ask which form they need to request equipment and who receives it. The first collection needs the approved procedure and the form instructions it refers to. Our AI solutions for business service covers the wider development, integration and support scope.
Which documents should you include first?
Start with recurring questions in one team. For each question, identify the source employees currently use: a procedure, instructions, an approved price list or an internal knowledge page. Someone who knows the job should confirm which information applies and which exceptions the answer must retain.
Create an inventory recording the document identifier, title, original location, version and effective date. Add its content owner, permitted audience and status: current, draft or archived. Record permission to use the content too. Being able to open a file does not automatically authorise sending it to your chosen AI provider.
Include only material needed for the agreed questions in the pilot collection. Remove personal data and business secrets from demonstration copies. Before using real content, agree processing location, data retention and access with the people responsible for those decisions in your company.
Is uploading PDF and Word files enough?
Check what is actually extracted from each document. For a scanned PDF, review the recognised text, especially identifiers, amounts and tables. If a table heading becomes separated from its values, the assistant may receive a number without its meaning. Keep conditions, exceptions and units with the information they qualify.
Use clear section headings and consistent form names. Remove repeated headers and signatures that disrupt reading, while retaining the original. For longer materials, the technical team can split content into smaller units for search. Microsoft's document chunking guidance describes approaches based on structure and meaning. The size of a useful unit needs checking against your documents and questions.
Each extracted unit should retain a link to its document and location within it. When an answer cites a source, the employee should be able to open the relevant page or section and check the whole claim. A document link alone does not establish that the source supports the answer.
How should you handle versions and access permissions?
If two procedures give different rules, the process owner needs to decide which applies. Keep the older version in the business archive where required, but exclude it from the collection used for current-process answers. If the assistant must also answer historical questions, define that scope separately, including the date the question concerns.
The application and search layer must check permissions before a document reaches the model. Telling the model not to disclose other people's information is not access control. Microsoft's document-level access overview describes filtering sources against the user's identity. Actual support depends on the system; check how the chosen solution handles the source folder's permissions and labels.
Track permissions for extracted sections and displayed citations as well as whole documents. Check employee role changes, revoked access and copies in caches or conversation history. Agree what is removed and how long the change takes. Access to a restricted document through an older answer also belongs in this review.
What does a small pilot collection look like?
The following example uses fictional internal equipment-request rules. It is material for planning and proposed checks, with no AI-system test reported. For a question about the current process, a team member should receive the current procedure and the associated instructions.
Scroll the table sideways to see all columns.
| Source | Status and audience | Preparation decision |
|---|---|---|
| PROC-01, version 2 | Current procedure for team members. | Include the rule that requests go to the team lead through form ZA-01. |
| PROC-01, version 1 | Archived procedure naming the former recipient. | Exclude it from the collection for current-process answers. |
| UP-01, version 1 | Current instructions for completing the form. | Retain the required fields and the additional information for replacing a device. |
| RAC-01, version 1 | A document available only to accounting. | Exclude it from search for team members without access. |
Download our text example of a collection and expected answers. It contains fictional extracts, statuses and six proposed checks. It is not connected to an AI service. Replace its rules with content approved by the responsible person for your own copy.
How do you check whether the assistant uses sources correctly?
Record the user's role, permitted sources and what a good answer must contain beside each question. For “Who receives my request for a new monitor?”, the expectation is the team lead and form ZA-01, citing PROC-01, version 2. “How many hours will approval take?” has no answer in this example. The assistant should state that no timeframe is specified and refer the user to the responsible person.
- Ask about the current procedure and check the answer and cited source version.
- Ask about replacing a device. Check that the additional requirement from the form instructions is retained.
- Ask for information that is absent, then ask a question without enough detail. Check for invented facts and whether clarification is useful.
- Sign in with a role that cannot access accounting. Check that the restricted document appears in neither retrieved sources, answers nor citations.
- Change the current version or withdraw a document. After the agreed refresh, check new answers, search results and the display of previously saved results.
- In a separate check, add content that attempts to steer the assistant outside its task. Check access boundaries and approval of actions.
Documents can contain text that attempts to influence model behaviour. OWASP describes these indirect prompt injection attacks and recommends several controls, including limited privileges and human approval for sensitive actions. Editing documents alone does not remove this risk. For an initial search project, decide separately whether the assistant may send or change anything at all.
Who maintains the documents after rollout?
For each source, assign someone to approve changes and define how they reach the search collection. An updated original file does not establish that the assistant already has the new version. You need a record of a successful refresh, error handling and checks for questions affected by the change. Document withdrawal and permission changes need agreed completion times.
When answers use data from business applications, coordinate this process with their exchange method. Our guide to connecting software without an API covers checks for file transfers and other supported routes. Show the user how current the source is when that affects the answer.
What should you prepare for an AI project estimate?
Prepare a description of one task, anonymised source examples and questions with expected answers. Identify user groups, the content owner and how documents change. Include source checks, the no-answer procedure and pilot acceptance criteria in your software specification.
For an AI assistant project estimate, send an outline of the process and the types of documents involved. An initial estimate can start with that overview; we agree how to exchange confidential materials before you send them. Preparation work, access and collection maintenance form part of the project scope.