Is there a repeatable problem?

Good early opportunities occur frequently enough to matter and have an outcome people can recognise. “Use AI somewhere” is not a use case. Helping staff find approved policy answers or triage a known document type is much easier to assess.

Are the information sources usable?

An assistant grounded in business knowledge depends on current, relevant and appropriately accessible source material. Conflicting documents, unclear ownership and missing policies will affect the system as surely as they affect staff.

Prepare your source material

  • Identify the approved sources
  • Remove or label outdated information
  • Decide who can access each source
  • Assign responsibility for future updates

What happens when the output is wrong?

AI output can be incomplete or incorrect. The design should reflect the consequence. Low-risk drafting may need a quick human review. Advice involving safety, rights, finance or sensitive decisions can require specialist controls or may not be suitable for automation.

Can success be tested?

Define representative questions, documents or workflow cases before launch. Useful evaluation considers correctness, completeness, safe refusal, escalation and whether the result actually saves effort.

Is there an owner after launch?

Models, source information, integrations and costs change. Someone in the business needs responsibility for feedback, access, source updates and operating decisions. An AI system is an ongoing capability, not a static brochure.