Insight

AI automation: what should and should not be automated

Automation is most useful when it makes a stable, understood process easier to run and inspect. It is a poor fit for work that has not yet been agreed, owned or simplified.

Map the work before choosing a tool

Describe who starts the process, what information enters it, which decisions are made, where people wait and how exceptions are resolved. A simple map often shows that part of the problem is duplicated input, unclear ownership or an unnecessary approval rather than a missing system.

Choose a bounded first automation

A good first candidate has a clear trigger, repeatable rules, a defined owner and a safe way to recover when something fails. Avoid beginning with the most politically complex process simply because it is visible. A bounded workflow gives the team a chance to learn about data quality, adoption and operational responsibility.

Design for exceptions and human judgement

Not every case should follow a rule automatically. Identify the signals that need review, who can override an outcome, and how an automated action is explained. Automation should make human judgement more deliberate where it is needed, not remove accountability from ambiguous work.

Assess a process before automating it

Bring one real process, its owners, inputs and recurring exceptions to a planning conversation. That is enough to decide whether the next step is simplification, integration or a focused automation.