Describe the task before choosing a product
Observe how the team does the work now. Record the repeated steps, exceptions and corrections. Decide what a useful output looks like and who will judge it. A product demonstration should use that example rather than an unrelated polished scenario.
An illustrative trial might prepare a draft answer to a routine question about opening arrangements. The approved source facts, human check and sending responsibility still need to be clear.
Set boundaries for the first trial
Agree which information can be used, what the tool may do and what requires a person's decision. Use fictional information when exploring an option unless an authorised arrangement has been established for real records.
A tool that drafts text should not silently become a system that sends it. Changes involving clinical decisions or sensitive information require their own appropriate assessment and responsible oversight.
Test awkward cases
Include missing information, contradictory instructions and an unavailable source. Ask what happens when the output is uncertain or wrong. The team needs a way to stop, correct and continue through the normal process.
Measure checking and correction time as well as time apparently saved. If staff spend longer repairing drafts than preparing the original answer, the trial has revealed something important.
Expand only after a useful result
Agree a small scope, a named owner and a review point. Keep the existing workable process available while evaluating the trial. Choose the next step from observed results, rather than treating an AI label as proof of improvement.
Sources and further reading
Choose a useful next step
If repetitive admin is the problem, start by describing one task and the result your team needs to check.
Healthcare AI and automation · Plan a first automation project
