Camera-based analytics can answer real operational questions occupancy, dwell time, usage patterns but the moment a video pipeline stores more than it needs, or exposes raw footage instead of structured data, it becomes a liability the business didn't sign up for.
The easy way to build video analytics is to record everything and figure out privacy later. That approach creates exactly the kind of exposure a business doesn't want: raw footage sitting in storage, unclear retention policies, and no clean answer to "who can see this" when someone asks. The harder, correct way is to design the pipeline so the video itself is never the product the structured data extracted from it is.
Frames are processed for detection and tracking in near real time; the system persists structured events presence, dwell time, counts not a searchable archive of footage.
Downstream systems (dashboards, billing, reporting) consume aggregated occupancy and usage data, never raw camera frames, so access to insight doesn't mean access to surveillance.
Every detection carries a confidence score, and low-confidence results are treated as uncertain rather than silently folded into reported numbers.
What's retained, for how long, and who can access it are architectural decisions made before the first camera is connected, not a policy written after the fact.
Computer vision systems are built to answer a specific operational question well, not to become an unmanaged video surveillance system that happens to also report metrics the two are architecturally different problems, and we treat them that way.