When people come to me asking “which neural net should we install in production”, I usually ask back: which process do you want to change? In about half the cases the conversation ends there, because there is no answer.
The working sequence looks like this.
First the process. Find the bottleneck where people do routine work with their eyes and hands: reading certificates, reconciling documents, listening to recordings, spotting defects in photos. The more boring the task, the better.
Then the data. An honest answer to whether it exists at all, and in what shape. More than half of all ideas die at this stage, and that is fine. Better now than after signing a contract.
Next, a pilot on one section with a measurable result. Not “improve efficiency”, but specifically: an operation took 40 minutes, now it takes 12. A pilot without a before-and-after number is a slide deck, not a pilot.
Only then, model selection and scaling.
A separate topic is people. If operators see the system as a supervisor, they will find a way around it. If they see it as an assistant that takes the dirty work away, things move by themselves. That is decided not by technology, but by who sat at the table during the design.