Teams often worry that their knowledge isn't "AI-ready": documents are scattered, procedures are out of date and the best know-how lives in people's heads. The good news is that you don't need a perfect knowledge base to start. You need a clear view of what an agent will rely on, and a plan to improve it as you go.
Start from the decision
Don't try to prepare everything. Pick the decision your first agent will support and list the knowledge it needs: which documents, which systems, which past decisions. That list is your scope.
Check four things for each source
1. Is it current? Superseded procedures and old price lists are the most common cause of wrong answers. Make sure the agent reads only approved, in-date versions, and that retired versions are clearly marked.
2. Is it owned? Every source should have an owner who can say whether it's right and fix it when it isn't.
3. Is it accessible? Can the agent reach it through a system connection, or does it live in personal drives and inboxes? Moving key documents into a governed library pays off quickly.
4. Who may see it? Record who is allowed to see each source, so the agent only retrieves what the person asking is permitted to see.
Capture what people know
Some of the most valuable knowledge isn't written down: why a supplier is avoided, which clause tends to be misread, how a fault was fixed last year. Three practical ways to capture it:
- Record reviewer decisions and reasons. Every time a person changes or rejects an agent's output, ask for a short reason.
- Turn confirmed fixes into knowledge. When an engineer confirms a solution, save it where the agent can find it.
- Interview your experts. Short, structured sessions with experienced staff before they move on.
Mind the languages
Organisations in the Gulf work in Arabic and English, and often in Hindi, Urdu or Tagalog too, sometimes within one document. Make sure the agent handles the languages your teams actually use, and test it on real mixed-language material.
Keep personal data in check
Identify sources that contain personal data and decide whether the agent needs it. Where it doesn't, mask it. Apply your retention rules to everything the agent stores, including logs and memory.
Improve as you go
Use the agent itself to find gaps. Questions it can't answer, sources it cites that turn out to be wrong, and reviewer disagreements all point to knowledge that needs fixing. Review them monthly with source owners.
A simple checklist
- The first decision and its knowledge sources are listed.
- Each source has an owner and a current, approved version.
- Access rules are recorded for each source.
- Personal data is identified, and masked where not needed.
- Reviewer reasons are captured from day one.
- A monthly review of gaps is scheduled.


