Start with the work. Then decide where AI belongs.

An organisation can experiment with several AI tools without improving how its work gets done. A stronger starting point is a specific workflow: what happens today, where it becomes difficult and what a better outcome would look like.

Describe the current process

Map the steps, inputs, systems and decisions involved. Look for repeated effort, slow hand-offs or inconsistent information. Distinguish a problem caused by manual work from one caused by unclear ownership or poor source data.

Choose a bounded opportunity

Drafting a first version of a status update from approved notes is more contained than allowing a system to make commitments to customers. Start where people can check the output and where an error is detectable before it causes harm.

Define what success means

Choose measures before the pilot begins. These might include time spent, correction effort, completeness or turnaround. Compare the new process with a baseline. Include the effort required to review the output and maintain the workflow.

Keep judgement explicit

Decide which outputs require human review, which information is permitted and who owns the final decision. Consider access controls and the organisation's policies before introducing confidential material. The pilot should have a clear stop condition as well as a success condition.

Make adoption part of the design

A technically capable tool can still fail if it adds friction. Test the workflow with the people who will use it. Keep what improves their work and change what does not. The objective is a useful operational outcome, not the mere presence of AI.