Why the first use case matters
The first agent sets expectations for every project after it. A good first choice shows results within weeks and builds trust. A poor one — vague, rare or high-risk — can stall AI plans for a year.
The five-question checklist
- Frequency: does the task happen many times a week? Frequent work pays back faster.
- Clarity: could you write the rules down for a new employee? Clear rules make reliable agents.
- Systems: do the systems involved have an API or another dependable way in?
- Cost of a mistake: is an error easy to catch and cheap to fix? Start where mistakes are low-stakes.
- Measurement: can you count the result — tickets closed, invoices posted, leads answered?
Good first candidates
- Answering routine customer questions from your own documentation
- Entering supplier invoices or forms into accounting or ERP
- Replying to and qualifying new inbound leads
- Booking, moving and reminding appointments
- Producing a weekly report from data you already collect
Tasks to leave for later
- Rare tasks that happen a few times a month
- Decisions that depend on judgement nobody has written down
- Anything where a single error is costly and hard to reverse
Run a proof of concept, not a pilot programme
Build one agent for one workflow on your real data, with a person approving its actions at first. Measure how often it is right and how much time it saves. Then decide — with evidence — whether to roll it out, adjust it or try a different task.