Most AI pilots stall because they start with the technology and look for somewhere to put it. The ones that stick start from a task that is done the same way every week and takes longer than anyone would like to admit.
The tasks that suit an agent
Three properties make a task a good candidate. It repeats on a predictable schedule. It follows rules that can be written down. And its output can be checked quickly by a person who knows what right looks like.
Reconciling a bank feed, triaging a support queue, drafting a first pass of a document from a template: all three qualify. Deciding whether to take on a client does not.
Keep a human at the decision point
The agents worth deploying draft, sort, and prepare. A person approves anything that leaves the building or changes a record of consequence. That boundary is what makes the time saving safe to bank.
It also makes failure cheap. A bad draft costs a minute to reject; an unreviewed action can cost considerably more.
Start with one role
Pick the single task your team complains about most, measure how long it currently takes, and deploy against that. One measured result is worth more than a broad rollout nobody can evaluate.
