Measure the manual process first
The first risk in any automation project is automating a process that should have been removed instead. So we start by watching the work end to end: how often it runs, how long it takes and where it goes wrong. Without those numbers there is no way to know whether an automation paid back.
Rank every candidate by return and feasibility
Most operations have more candidates than budget. Rank each one on two questions: what it would save, using the real hourly cost of the roles involved, and whether the data it needs is available and clean enough to act on.
Candidates that score high on both go first. Those with a big return and poor data usually point to work on the core system, which is why clean data in one system so often comes before any agent.
Decide where a person must approve
Every automation needs confidence thresholds, and a clear rule for what happens below them. Agree in writing what the agent may do on its own, what it must route to a person, and what it must never do. The failure to design out is an agent taking an action in front of a customer that nobody approved.
Shadow mode before write access
Before an automation acts, it should run alongside your team on real work without taking any external action, so its decisions can be compared with theirs. The cases where they disagree are the most useful test data you will get.
Only then does it go live, with a person approving low-confidence cases and monitoring switched on, and with thresholds tuned as volume grows.
Where it usually does not pay back
- Processes that run rarely, however annoying they are.
- Work whose inputs live in spreadsheets nobody owns.
- Anything where a wrong action cannot be reversed and a person would have to check every case anyway.