Physical security still runs on a primitive invented in the 1970s: alarm, or no alarm.
A sensor trips. A signal fires. And at that instant the system's entire intelligence is exhausted — everything after the trip is human labor sorting signal from noise, usually by phone, usually with no information beyond which zone opened.
The sorting failure is measurable at national scale, and the numbers are worse than most people in this industry say out loud.
The numbers
94–98% of police alarm calls are false. That's the standing figure from U.S. Department of Justice data, and while the exact rate varies by jurisdiction, no serious source puts it below 90%.
Each of those calls consumes roughly 20 minutes of officer time. In aggregate that's about $1.5 billion a year of U.S. police resources spent responding to things that did not happen. In many cities, alarm calls run 10–25% of all police service calls — a quarter of the dispatch queue, in the worst cases, generated by a system that is wrong nineteen times out of twenty.
The canonical datapoint is DeKalb County, Georgia, which logged more than 144,000 alarm calls in a single year against 39 actual or attempted burglaries. That is roughly 3,700 dispatches per real event.
Set against that: only about 13% of burglaries end in an arrest. The system is simultaneously generating enormous dispatch volume and failing to convert it into outcomes. Those two facts are the same fact.
It is not a sensor problem
The instinct is to blame the hardware, and it's the wrong instinct. Consider what a contact sensor actually knows: a circuit opened. That is the entire content of the signal.
Everything a human would want to know before deciding whether to send someone — was anyone there, was it the wind, was the homeowner home, is this door always doing this — is absent from the signal by construction. So the only tuning available is a sensitivity dial with two bad ends:
- More sensitive → more trips → more false dispatches, more fines, more station-operator time burned on verification calls that reach nobody.
- Less sensitive → fewer trips → the real event doesn't alarm, which is the failure nobody wants to be on record for.
There is no setting on that dial that produces a good system, because the problem isn't where the threshold sits. It's that a threshold on a context-free signal is the whole architecture.
This is also why "better AI on the camera" is only a partial answer. A model that reliably tells you a person is in frame has improved the signal enormously and still hasn't told you whether that person belongs there.
The strongest evidence that the alarm itself is the problem
Salt Lake City adopted verified response: police respond to alarms only when the alarm has been verified by some second, independent means. The outcome:
- Alarm-response calls fell by about 87%.
- Burglaries fell 26%.
Read that second line again, because it's the one that matters. Removing unverified alarms from the dispatch queue did not make the city less safe. It made it measurably safer. The officer hours that had been going to 3,700-to-1 noise went somewhere more useful.
The conclusion we draw is uncomfortable for the industry that sells alarms: verification is where the value was all along. The trip was never the product. Everyone has been selling the cheap half of the transaction and leaving the expensive half — deciding what the trip means — to whoever picks up the phone.
What verification has to mean
If verification is the product, it's worth being precise about what qualifies. A second signal from the same sensor isn't verification. Neither is a phone call to a keyholder who doesn't answer, which is most of them at 3 a.m.
Working verification needs three properties:
It has to be independent. The second signal must be able to fail differently than the first. A door contact plus a camera on that door is independent. A door contact plus a second door contact is not.
It has to arrive fast enough to matter. Verification that takes four minutes has converted a dispatch decision into a post-incident report. The useful window is seconds, which means the evidence has to already be queryable when the trip happens — not retrieved and reviewed afterward.
It has to produce evidence, not just a verdict. "Verified" as a bare flag asks a dispatcher to trust a black box. "Verified — person in frame at the east door, 02:14:07, no badge event within ten seconds, clip attached" gives them something to act on and something to defend the decision with later.
Most of what makes this hard is that the second signal usually lives in a different system than the first. The door contact is on the panel. The person is on the camera. The badge is in access control. The homeowner's phone is somewhere else entirely. An event that is obviously real when you look at all four can pass every one of them individually.
What this changes commercially
Municipal false-alarm fines already run $25–100 and up per incident, and the schedule escalates with repeats. That's the visible cost. The invisible ones are larger: operator hours spent on verification calls, customer churn after a string of 2 a.m. false calls, and the slow erosion of police responsiveness to a given address.
Every one of those is a line item that verification moves in the right direction at the same time. Which is why we think verified response spreads from a municipal-policy curiosity to the default way monitoring is sold, and why we've built our operation around deciding what an event means rather than around how fast we can relay that one occurred.
Where this argument is weakest
Salt Lake City is one city. It's the strongest published result we know of, but municipal politics have reversed verified-response policies elsewhere, and a single natural experiment is not a law. We'd want to see the pattern replicated in a half-dozen jurisdictions before treating the 26% as a transferable number rather than an existence proof.
The 94–98% band is compiled, not measured uniformly. It comes from DOJ data across jurisdictions that count differently. The direction is not in doubt; the second digit is.
Verification adds a step, and steps can fail. A verification layer that is slow, or that is down, is worse than no verification if it sits between a real event and a dispatch. Anyone selling this — us included — should be measured on verification latency and availability, not just on false-alarm reduction.
Fewer dispatches is not automatically better. The Salt Lake City result is persuasive precisely because burglaries fell. A verification layer that suppressed real events would produce the same reduction in calls and a very different outcome. That is the number to watch, and it's the one that should be in every contract.