The Business Case, Honestly

Camera analytics proposals are built on labour replacement and prevented incidents. Both are weaker than they look.

The claims that do not hold

"Replaces a person watching monitors." Nobody was watching the monitors. The realistic comparison is against reviewing footage after an event, not against continuous human attention.

"Prevents incidents." Detection is not prevention unless someone intervenes in time, which requires an operator, a response procedure and an alert rate low enough that alerts are read.

"Reduces shrinkage by X percent." Figures from other deployments, other sectors, other populations, published by the vendor.

"Pays for itself in N months." Built on the two claims above.

One challenged figure discredits the rest, which is a recurring theme in every measurement discipline.

What does hold up

Faster retrospective search. Finding the relevant two minutes in a week of footage is a real, measurable saving and does not require the system to be preventive.

Consistent detection of a specific, well-defined condition. A spill, a blocked exit, a vehicle in a pedestrian aisle, a missing component on a line. Narrow and checkable.

Quality inspection throughput, where the comparison is against manual inspection of every unit and the error rates of both can be measured.

Counting, where the alternative is a person with a clicker or an estimate.

Regulatory evidence, where a record of a specific condition is required.

Sizing it from your own operation

Count the events. How often does the condition actually occur? If the answer is unknown, that is the first finding and it can be established by review.

Cost one occurrence end to end, honestly.

Estimate detection realistically, using pilot numbers rather than quoted ones.

Cost the alerts. Operator time per alert, multiplied by the alert rate you measured, is a real recurring cost that appears in no proposal.

Subtract. The result is frequently smaller than the proposal and occasionally negative, which is worth knowing.

The costs that get missed

Camera work. Repositioning, adding, upgrading. Usually the largest line and rarely in the quote.

Network and storage.

Operator time on alerts, permanently.

Model maintenance. Performance drifts; retraining is not optional.

Evaluation. Someone must periodically check that it still works, with ground truth.

Legal and assessment work, which for anything involving identifiable people is substantial.

Decommissioning, which nobody plans and which has its own note.

What to commit to

Not a percentage reduction in incidents.

A specific, checkable claim: "detects blocked fire exits with recall above X and no more than Y false alarms per camera per day, measured against ground truth over a month."

Report it on the date, met or not.

A missed target reported honestly costs less than a met one nobody believes.

Costing an alert

The recurring cost that appears in no proposal and dominates the operating budget.

Time to triage one alert, measured rather than estimated.

Multiplied by the measured alert rate, not the quoted one.

Multiplied by shifts and by operator cost.

Annualised.

For many deployments this exceeds the licence, and it continues forever. It is also the number that makes the case for threshold work and zone restriction, which reduce it directly.

The retrospective-search case

The application that holds up when the preventive claims do not.

Finding the relevant two minutes in a week of footage is a real, measurable saving.

It requires no real-time alerting, so alert fatigue does not apply.

It tolerates a low threshold, because volume costs nothing when nobody is being interrupted.

It is checkable: time an investigation with and without it.

Lead with this where the preventive case is weak, because it is defensible and it delivers immediately.