After several years of running adoptions, I have a simple map of where AI pays off in business processes and where it does not. Sharing it as is, no marketing.
Pays off almost every time:
- Document flow. Sorting incoming mail, extracting fields, reconciliation, routing. Boring, measurable, visible effect in the first month.
- First-line support. Typical questions get closed automatically, the hard ones go to people together with the context already collected. People stop burning time on "how do I reset my password".
- Monitoring and anomalies. The system watches event streams and calls a human when something goes off. It does not replace the person on duty, but the person sleeps better, not the system.
Pays off poorly or not at all:
- Full autonomy in processes where errors are expensive. If the price of a mistake is client money or legal risk, a human in the loop is mandatory. Removing them eats all the savings in the first incident.
- Automating a mess. If a process does not work with people, with AI it will not work faster, just more expensively.
- Anything where the effect cannot be counted. If you cannot name a number before and a number after, that is hype, not a project.
My main criterion is simple: AI takes the routine, the human keeps the judgment. Everything that breaks this rule gets reworked sooner or later.