How much knowledge management do you need before using AI?
You do not need to organise every company document before starting useful AI work. You do need a manageable area where the relevant information is available, its authority is clear and someone owns corrections. Begin with a recurring business question and improve the knowledge needed to answer it safely.
Ger Perdisatt, Acuity AI Advisory
4 min read
- Write itOne current version, with an owner
- Save itWhere the team expects to look
- Name itA title a colleague would search for
- Find itCheck that the source is accessible and searchable
- Review itUpdate or retire it when it changes
Access permits retrieval; it does not guarantee that the right source is found or interpreted correctly.
A whole-company tidy-up is a difficult starting point
It is easy to turn an AI conversation into a proposal to reorganise the entire intranet. Years of duplicate documents and inconsistent folders provide plenty of reasons to begin. They provide much less agreement about when the organisation would be finished enough to try anything useful.
A better-sized starting point is a recurring question that costs people time: which procedure applies to this job, what has been agreed with this supplier, or what does a new colleague need to understand before handling a request? Choose an area whose owner can judge the answer.
This is a practical recommendation, not a claim that information quality stops mattering. It limits the work to a scope where the organisation can establish whether better information and an AI assistant produce a worthwhile improvement. Wider problems discovered there can be recorded and addressed in proportion to their consequences.
Make authority visible before adding more documents
A large collection can contain plenty of relevant information while leaving the governing answer unclear. A draft procedure, a meeting note proposing a change and an approved instruction may all discuss the same subject. Search relevance alone cannot settle which one should guide an employee.
For the chosen area, make the current source and its owner easy to identify. Explain the period or circumstances it covers and how a reader can tell it has been replaced. Where multiple sources legitimately apply, describe their relationship. A local exception may qualify a general policy without replacing it.
Retention needs are separate from everyday guidance. An older document may have to remain available as a record. Decide how staff should distinguish historical evidence from current instructions, following the organisation's existing retention arrangements. Deleting old material simply to improve an AI answer can create another problem.
Some of the important knowledge is unwritten
The folder may contain the procedure but omit the judgement that makes it workable. An experienced colleague knows which request is an exception, why a familiar shortcut causes trouble, or when to ask someone else. An assistant cannot reliably recover that local knowledge from a document that never recorded it.
Capture the explanations people repeatedly give when someone joins or takes over a task. Have the responsible colleague review the result. The aim is an accurate account of how the work should be handled, including its limits. An automatically generated transcript or summary needs editing before it becomes guidance.
Distinguish an approved practice from one person's habit. Both may be useful to understand, but publishing the habit as official procedure would give it authority it has not earned. This is a management decision that an AI-generated knowledge base cannot make on the organisation's behalf.
Keep the information where the team can maintain it
Use an existing approved home for the material where it can support the job. A separate copy created solely for an assistant has a maintenance cost: somebody must keep it consistent with the operational source. That may be manageable for a small, controlled reference set, but it should be a deliberate choice.
Microsoft's preparation guidance addresses permissions and content governance, including ownership and inactive sites. Those administrative capabilities can help at scale. Their availability does not mean every small team needs to begin by purchasing an additional system or running a tenant-wide transformation.
Agree who can change the material and what happens when a colleague identifies a mistake. If a correction exists only in a private chat, the next person can encounter the same problem. A useful improvement leaves the underlying shared source more reliable.
Check the information with the people who depend on it
Use ordinary questions from the selected work to assess the knowledge available. Include a question with a clear answer, one with an exception, and one the material cannot yet answer. The final case matters: a useful assistant should leave an unresolved question visible instead of filling it with plausible advice.
Have the relevant staff member inspect the cited material as well as the prose. Record why an answer failed. Missing knowledge, contradictory instructions and poor retrieval call for different changes. Counting all failures as "bad prompting" hides the organisational work that needs attention.
Consider the cost of keeping this reliable. If every answer requires a senior colleague to reconstruct the source from scratch, the proposed arrangement may not yet save time. A smaller scope or clearer reference set may be a more sensible next move than introducing another tool.
What a proportionate first engagement should settle
A useful initial piece of work should leave the organisation clear about the job being improved, the information it depends on and the responsibilities that remain with people. It should also identify whether the existing approved tools are sufficient and what evidence would justify extending the approach.
The result may be a workable reference area and a better way for staff to use it. It may also reveal that an important decision has never been documented consistently. Finding that early is useful, provided the responsible person can resolve it and the change reaches the colleagues who need it.
Acuity helps teams examine those questions using the systems they already have, then develops training around the agreed work. Full training packs, exercises and editable materials are supplied privately under an agreed commercial engagement. The public guidance here explains the reasoning so you can judge whether that kind of help fits your organisation.
Sources and product references
Checked 3 October 2026. Features and access can change, so check what your own licence includes.
Editorial guidance informed by Acuity's work with teams. Examples are fictional and client materials stay private. To record who owns an AI use and when it is reviewed, see AI Register.