Counting things is not understanding them
The most common computer vision deployment I see is a counter. Units through a line, vehicles through a gate, people in a zone. It works, the numbers are accurate, and six months later nobody looks at the dashboard.
The reason is that the count was never the missing information. Somebody was already counting, roughly, on a clipboard, and roughly was good enough for the decisions being made.
Where the value actually sits
What a camera gives you that a clipboard doesn’t is continuity. Not the total, but the shape of the total over time — the ten minutes each afternoon when throughput halves, the difference between shifts that nobody had quantified, the fact that line three degrades gradually after a changeover rather than all at once.
None of that is visible in a daily total. All of it is visible in a minute-resolution series, and all of it points at something actionable.
So the design question isn’t “what should we count?” It’s “what variance are we blind to, and would seeing it change what anyone does?”
If the honest answer is that the process is stable and well understood, a camera will confirm it accurately and change nothing. That’s a real outcome and worth predicting before you spend the money.