The useful question is never "where can we add AI?" It is "which specific step in this workflow costs a person time every week, and is that step describable?" If the answer to the second half is no, the automation will fail, model or no model.
A step is a candidate when it is repetitive and bounded
- It happens on a predictable trigger, not on intuition.
- The inputs are already digital, or can be without a new data-capture project.
- A competent person could write down how to do it in under a page.
- An imperfect result is recoverable — a draft, not an irreversible action.
- Someone can tell, quickly, whether the output was good.
That last point is the one most projects skip. If nobody can evaluate the output cheaply, you cannot improve it and you cannot trust it, so you end up with a system that produces work rather than removing it.
Where AI genuinely earns its place
Drafting, classification, extraction and summarisation of material that a human then approves. Routing an incoming request to the right queue with a confidence score and an escalation path. Turning structured data into readable prose. In each case the model handles the tedious middle and a person keeps the decision.
Where it usually does not
Anything that must be exactly right every time and can be expressed as rules — use rules. Anything irreversible without review. Anything where the real problem is that the underlying process is undefined; a model will simply produce confident output on top of an unresolved disagreement.
Automating a broken process makes it break faster and more consistently.
How we scope this in an Automation Sprint
We map the workflow as it actually runs, not as it is documented. We time the steps. We rank them by hours recovered against effort to automate, and we build the top of that list — sometimes with a model, often with plain integration work, because a reliable webhook beats a clever prompt whenever both would work.
Then we measure the same workflow again. If the time did not move, the automation was not worth keeping, and we say so.