A filter, not a backlog of everything
The first automation should be boring. High volume, a stable set of rules, inputs you have seen before, a clear place to put exceptions, and a number you can recount after the pilot. If a process fails any of those, it can wait.
The five
These are categories, not a ranking of your company. Pick the one you can already describe on a whiteboard.
- Moving structured data between two systems you trust, such as a form into a CRM, with a stop on unmatched fields.
- Assembling a report people already know how to check, from a source you name, on a schedule.
- Routing work by a rule you already operate, such as region, product line, or priority.
- Reading one repeating document type into fields a person confirms before anything posts.
- Sending a status update that reports a fact and does not decide an exception.
Leave these until later
Wait if the decision is regulated, if the input is different every time, if a wrong action is expensive, or if you have no system of record. A model will not supply the missing policy.
Also wait if the only reason to automate is that the work is unpleasant but rare. The build cost will not come back from a handful of cases a month.
One owner and one queue
Each pilot needs a person who can say whether an exception was handled correctly. The queue of exceptions is part of the product. If that person has no time to review, the automation is not ready to run, however finished the software looks.
Related service: Agentic AI and automation