How to choose an AI pilot that is worth taking forward.

Choose a problem, not a technology

Look for a task that repeats often, uses accessible information and has a recognisable definition of success. Internal knowledge search, support triage and document extraction are possible candidates; the right choice depends on your operations.

Know the baseline

Measure the current handling time, error rate, escalation rate or cost before introducing AI. Include the time people spend reviewing outputs. A faster demonstration is not the same as a better end-to-end workflow.

Put boundaries around data and actions

Use only approved sources, preserve access permissions and define where human review is required. Decide what happens when the system is uncertain, a connection fails or an input falls outside the intended task.

Evaluate before expanding

Use representative examples, including difficult cases. Compare quality, operating cost and user adoption against the baseline. Expand only when results and control requirements support the next step.