Where AI automation pays back first

“We want to use AI” is not a starting point. It is a budget looking for a problem, and it usually finds an impressive demonstration rather than a saved hour.

A better starting question: which task does a specific person do repeatedly, where the outcome can be checked?

Four signatures worth looking for

• The same information is typed twice. Somebody is a human API between two systems. This is the highest-return work in most businesses and it is usually integration rather than AI.
• The input is unstructured but the decision is not. Invoices, forms, emails, scanned documents — messy in, but a rule-bound outcome. This is where current models genuinely earn their place.
• A queue forms behind one person. Not because the decision is hard, but because only they know how. Much of that can be prepared automatically so they approve rather than assemble.
• You only find out afterwards. The data existed, nothing was watching it. That is monitoring and alerting before it is intelligence.

The one that looks promising and is not

Anything where the outcome cannot be checked. If nobody can tell whether the output was right, you cannot tell whether the automation is working, and you will find out it was wrong when a customer tells you.

Automate where you can verify. Assist where you cannot.

Design for being wrong

Any system built on a model will be wrong sometimes. That is not a reason to avoid it; it is a design constraint. Before anything is built, agree what may run without a person, what needs review, who approves, what happens to low-confidence cases, what is logged, and how an action gets corrected afterwards.

A process that cannot answer those has not been designed. It has been hoped for.

Start with one

Pick the single task that consumes the most time for the least judgement. Do that one properly, measure it against what it replaced, and only then look at the next. A programme of eight automations started at once produces eight half-finished ones.

Common questions