The shared drive contains three onboarding guides, an old price list and a presentation describing a retired service. Uploading the folder does not settle which information the team should use.
Photo by Wesley Tingey on Unsplash.
Prepare an AI knowledge base by selecting authoritative documents, resolving contradictions and defining access and ownership. Begin with a narrow set of questions. A useful collection is easier to maintain than an indiscriminate archive.
Choose questions before files
Write real staff questions, such as which documents a client must supply or who approves a scope change. Identify the source that should govern each answer.
Exclude questions the business has not resolved. An AI tool should not invent a policy because nobody has agreed one. Put unresolved questions on a separate decision list with an owner.
Accept sources deliberately
For each candidate, record:
Purpose and intended audience.
Business owner and authoritative location.
Current version and effective date where applicable.
Scope, exclusions and limitations.
Access restrictions and approval to use it.
Replacement or retirement rules.
Remove duplicates from the proposed collection and resolve conflicting instructions with the owner. Keep historical material separate where another purpose requires it. Document why a questionable source was excluded.
Make documents understandable
Use descriptive headings and explicit names. Replace 'the usual process' with the actual procedure or a reliable reference. Keep important conditions beside the relevant instruction.
For an illustrative onboarding guide, distinguish mandatory information from information needed only for one service. Neither a colleague nor a retrieval system should have to infer that boundary from an unrelated email.
Check the platform's supported formats and test your actual documents. Do not assume scanned pages, tables and images behave like clean text.
Verify against the approved set
Try the original staff questions and inspect the sources supporting the answers. Include something the collection cannot answer. The workflow should expose the gap rather than create a plausible company rule.
Keep those questions for later checks when sources change. Preparation is the beginning of maintenance, not a one-off upload exercise.
Related guidance: guide to mapping the work before changing it.
