Choose a problem before a model
Ask teams which recurring tasks require reading, sorting or drafting information. Shortlist tasks with enough examples to test, a clear outcome and someone able to assess quality. Avoid beginning with an organisation-wide assistant without a defined job.
Use four selection questions
Is the task frequent? Can we access appropriate data? Can a person check the output? Would a mistake be recoverable? A task that passes all four is easier to pilot than one with rare inputs, hidden rules or irreversible actions.
Agree a baseline and a stopping rule
Measure current handling time and quality. Decide what improvement would justify continued work, including review effort and running costs. Include difficult examples in the evaluation and stop if the system cannot meet the agreed quality threshold.
Plan the working process
Decide where the result appears, who approves it and how feedback is collected. Review data access and retention with the business before testing. A useful pilot proves a way of working, not just a convincing demonstration.
Turn the question into a practical next step.
Bring the process you want to improve. We can help you assess the options and decide what is worth doing first.
Explore the relevant service →