Two years into the AI wave, the pattern is clear enough to state plainly: the wins are narrow and boring, and the failures are broad and ambitious. The clients getting real value are not replacing departments. They are removing specific, well-defined drudgery.
The use cases that earn their keep share a shape. The input is messy human language, the output is structured data or a draft a person reviews, and a mistake is cheap to catch. Summarizing intake forms, drafting first-pass responses, extracting fields from documents, routing support requests: these work because a human stays in the loop and the cost of an error is an edit, not an incident.
The use cases that burn money share a shape too: anything where the AI acts on its own authority and a mistake reaches a customer or a ledger. The technology is not ready to be unsupervised in those seats, and pretending otherwise is how AI projects end up quietly shelved.
Our advice is unglamorous. Pick one workflow your staff visibly dreads, measure how long it takes today, pilot the smallest possible AI assist, and measure again. If the numbers move, expand. If they do not, you have spent little and learned a lot. That is the whole strategy.
