Choose a task you can describe clearly
A good first candidate has a recognisable trigger, known inputs and an outcome someone can check. “Improve operations” is too broad. “Prepare a draft shipment update from a carrier email” gives you something concrete to evaluate.
Look for repeatable effort
Ask the team where they repeatedly search, copy, classify or prepare information. Request triage, routine record preparation and answers from approved documents can be useful candidates. Their suitability depends on the quality of the inputs and the consequences of a mistake.
Check whether a simple rule is enough
If the input and action are predictable, ordinary automation may be easier to test and maintain. AI becomes relevant when interpretation is needed. A workflow can use both: an AI step reads a message, then rules validate the record before any update.
Keep the first action reviewable
Drafting an answer is easier to reverse than sending it to a customer. Preparing a record for approval can be a sensible first scope. Define how a reviewer sees the source, checks the proposed action and rejects or corrects it.
Compare candidates on the same basis
For each task, record frequency, current effort, data quality, integration difficulty and review needs. Also note the consequences of an error and who owns the process. Do not prioritise a frequent task solely because it looks easy in a demonstration.
Use one illustrative comparison
Consider two possible projects: preparing a summary from approved internal documents, and changing financial records based on ambiguous emails. The first may offer a more contained pilot when sources and access are clear. The second needs stronger validation and review. The right choice still depends on the actual workflow.
Agree the baseline and the stopping point
Record how the task works today. Test the pilot with representative inputs and exceptions, then compare output quality, handling effort and operating cost. Agree what would justify expanding the scope—and what would mean changing direction.
Build for the people using it
Someone needs to understand the output, handle exceptions and keep the underlying information current. Include training and ownership in the first project. A narrow workflow that people use consistently is a better foundation than a broad demo with no operator.
Explore AI agents and assistants ↗
Further reading
The UK government’s guide to adopting AI tools explores the role of user needs, training and adoption in making tools useful.