Who this is for
Enterprise architects and AI leads who've been asked to scale AI across core systems, and who need to be able to say where any AI answer came from.
Step 1: List the questions people already ask
Collect a month of real questions from service desks, change boards, programme teams and copilot logs. Group them by system and process. These are your test set.
Step 2: Pick one system and compile it
Start with the system that gets the most questions. Compile its code and documents into one model, so every fact can be traced to source.
Step 3: Label every answer
Require every AI answer to say whether each claim is grounded in source, inferred, or a gap. Anything that can't be labelled isn't ready for production.
Step 4: Connect your existing tools
Point your copilots and agents at the same model, so they stop building their own partial pictures. Compare their answers before and after on your test set.
Step 5: Govern from evidence
- Track the share of answers that are grounded versus inferred.
- Keep a register of declared gaps and assign owners.
- Only allow agents to act where the model's coverage is complete.
A checklist for your next AI review
- Can every answer be traced to a source?
- Do tools say when they don't know?
- Do all tools use the same model of the estate?
- Who owns each declared gap?
- Where are agents allowed to act, and why?