How to prepare company documents for an AI agent

Prepare company documents for AI by making their authority, scope and current version explicit. Start with the documents needed for one workflow, not a full archive import. A smaller set of well-owned sources is easier to validate than a large collection of conflicting files.

Choose a task-sized collection

Imagine building an assistant for new-client onboarding. The useful starting set might include the approved onboarding checklist, current service scope and contact ownership rules. Old sales decks and abandoned process drafts can remain outside the first collection.

For each document, ask what decision it supports. If nobody can answer, defer it. This keeps the initial evaluation focused and avoids granting unnecessary access.

Add a simple source header

• Owner: the person or team responsible for accuracy.
• Status: approved, draft or retired.
• Effective date: when the content applies.
• Scope: business unit, service or customer group covered.
• Replacement: the current source when this document is superseded.

Use descriptive headings and separate unrelated topics. A retrieval system has a better chance of returning a useful passage when the document itself has clear boundaries. Preserve tables and qualifiers that change the meaning of a rule.

Resolve conflicts before ingestion

If two documents disagree on a cancellation policy, ask the process owner to identify the authoritative version. Do not delegate a policy decision to the agent because both files happen to be searchable. Keep retired material clearly labeled or outside the active retrieval collection.

Also inspect attachments and embedded screenshots. Important text hidden in an image may need a verified transcription, especially when numbers or conditions matter.

Run a source acceptance check

Create questions that require each important document. Confirm that the assistant retrieves the right version and can point back to it. Include a question the collection cannot answer. A useful knowledge base supports both correct answers and honest limits; uploading files is only the beginning of that work.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.