There Are No AI Wins Without Understanding the Business

Every week another agency posts about AI. Very few of them ship anything with a return attached, and the reason is almost never the model.

Most companies aren’t ready for what’s being sold to them — not because they’re behind, but because the two things an AI win actually requires are usually missing. The process isn’t automatable: it lives in people’s heads, in email threads, in the judgment of the one person who knows how the exception gets handled. And the data isn’t reachable: split across systems that don’t talk, entered inconsistently, named differently in every department. Point a model at that and you get a demo, not a dividend — and a demo works because it only has to work once.

You cannot shortcut this with tooling, and that’s the part the market would rather you didn’t hear. Model access is cheap and getting cheaper; it was never the scarce resource. The scarce resource is understanding how a specific business makes money — where the margin hides, which steps are load-bearing, what quietly breaks when you change them. You cannot learn an industry like mortgage in a few months. No stack of API keys substitutes for having sat in the room while the money moved and seen why the process is shaped the way it is. The people who did that work carry it as instinct; most of the people selling quarter-long transformations don’t.

So the order matters, and it isn’t negotiable. Understand the business completely first — the workflows, the constraints, the data, the incentives, the exceptions everyone quietly works around — and only then build the thing that fits that business and no other. In that order the technology is almost anticlimactic: once the process is truly understood, the software to match it is the easy part. In the other order you get theater — a slick demo, a stalled pilot, and a quiet write-off two quarters later.

Which is why the honest first deliverable is sometimes a “no,” or a “not yet.” Some processes should be fixed before they’re automated; some data has to be made reachable before it’s worth reaching for. Saying so costs us the fast yes and saves the client the slow, expensive failure. The agencies racing to point models at unready businesses are selling motion. The return comes from the unglamorous work that happens before the model is ever switched on.