Choose a task with a clear finish
“Use AI in operations” is too broad to estimate or test. “Prepare a draft of the weekly support report from these approved sources” gives a team something concrete to evaluate.
Look for repeated work where people spend time reading, sorting or preparing information. If a calculation or a fixed rule solves the problem, ordinary automation may be sufficient. Our starting recommendation is to test the smallest workflow that could be useful.
Anthropic discusses the tradeoffs between fixed workflows and agents.
Write a brief before choosing tools
- Input: which documents, messages or records begin the task?
- Output: what should the person receive, and in what format?
- Owner: who can say whether the result is usable?
- Limits: which systems can the workflow read or change?
- Review: what must a person approve before the next step?
An example: preparing a weekly report
Consider a hypothetical support team that combines a ticket export and a spreadsheet every Friday. A pilot could group the approved records, draft a short summary and flag gaps for the team lead. Sending the report would remain a separate approval step.
Set the rules before testing: every reported number must match the source, unresolved cases must stay visible, and missing data must be identified. These are observable checks. “The report sounds professional” is useful feedback, but it cannot replace them.
Compare the pilot with the current process
Use representative examples, including incomplete or conflicting records. Keep a few examples aside so you can check changes against work the system has not already been tuned to.
Record the time spent preparing and reviewing the result. Also note corrections, failed runs and tool costs. A draft produced quickly can still create more work if someone has to reconstruct its sources.
- Could the reviewer find the source of an important statement?
- Did the workflow stop or ask for help when required information was missing?
- Was the total effort acceptable after review and corrections?
Agree on what happens after the pilot
Before connecting the workflow to daily operations, decide who maintains the instructions, handles access and responds to failures. Keep an operating note that explains how to run, review, pause and update it.
A quote should separate the implementation work from recurring model, tool and maintenance costs. Expand the workflow once the first task meets the agreed criteria. The next useful step may be a better source document or a clearer instruction, rather than a larger system.