The Minutes Before Anyone Knows

On Friday afternoon, a fire started in plastic pallets stored outside a plant in South Haven, Michigan. It spread into the building. Propane tanks reportedly exploded early in the sequence. Six agencies responded to what was classified as a three-alarm fire, the plant was evacuated, and no injuries were reported. The city then shut off power to every customer it serves, restoring it roughly six hours later. The local school district cancelled its homecoming events. The cause has not been established.

I have spent the last couple of years working in and around AI for physical security, and the thing I notice about an account like this one is not the fire. Industrial fire is a well-understood risk. Sites that store polymer in volume know what polymer does. They have suppression, fire lanes, hydrant placement, evacuation procedures, mutual-aid agreements. By every available indication South Haven executed the response well.

What I notice is that almost every one of those controls begins operating after somebody knows. They govern the response. The interval before the response, between a thing starting to burn and a person recognizing it, is largely ungoverned at most sites, and it is the interval in which the scale of the eventual outcome is decided.

That yard was very likely already under camera coverage. Outdoor storage usually is, for theft, for liability, for vehicle movement, for insurance. If video of the first minutes exists, and it probably does, the gap was not coverage.

A facility like this typically has cameras pretty much everywhere: on perimeter fencing, the main gate, the truck court, the loading docks, the outdoor material yard, the waste and recycling area, the compressed-gas storage, the rooftop mechanical space, the parking areas, the employee entrances. That is a great deal of video, and considerably more than any operator can hold in attention at once. The material yard is rarely the feed a person chooses to watch at four o'clock on a Friday, and that, rather than any gap in coverage, is what the account comes down to.

What an agent is actually for

This is where I want to be precise, because the public discourse about AI agents has drifted somewhere unhelpful.

Much of the coverage right now frames agents and robotics as a story about replacing people and invading privacy. I understand why that frame is attractive to write, and I think it is both wrong and a waste of the opportunity in front of us.

An agent watching an outdoor material yard at four on a Friday is not taking work that a person wanted, and it is not taking work that any organization was realistically staffing. No site hires someone to watch a pallet stack on the chance that it ignites. That watch simply did not exist, at that site or at most others, and there is no scandal in that. There are more cameras than there are eyes, and the ratio has been getting worse for twenty years.

What an agent can do is hold that attention continuously and, crucially, bring context to it. A thin column of smoke at the edge of a yard is ambiguous on its own. It could be exhaust, steam on a cool afternoon, dust off a passing truck. A small thermal anomaly near stacked material is likewise weak evidence. So is an absence of personnel in a zone that is usually busy. Any one of those signals is not worth waking anyone for, and a system that alerts on each of them separately produces so much noise that operators learn to ignore it, which is worse than having nothing.

Evaluated together, against what that specific yard normally looks like at that specific hour, the same signals stop being ambiguous. They become one legible situation, and they become legible while the fire is still outdoors and still small.

The privacy question deserves a real answer, not a deflection

The other half of the public objection is privacy, and I do not think our industry has earned the right to wave it away. When people hear that an AI system is watching a site continuously, nobody is worried about the fire getting caught early. They are worried that continuous attention turns into continuous identification, and that a tool bought to protect a facility ends up profiling the people who work in it.

That worry is answerable, but only architecturally. It is not answered by a policy document, and it is certainly not answered by telling people their concern is misplaced.

The distinction that matters is between recognizing a situation and recognizing a person. A system can be built to read behavior, context and the normal rhythm of a space without ever identifying who is in it. Detection can run on what is happening rather than on who it is happening to. No facial recognition, no personally identifiable characteristics used to detect, no personal data retained. Video can stay under the operator's control, in their own environment, with retention and deletion decided by them rather than by a vendor. Systems can be built to align with GDPR and CCPA rather than to work around them.

Notice that this is a choice made in the architecture, not a setting toggled afterward. Privacy by design means the system cannot do the thing people are afraid of, because the capability was never built into it. That is a materially different promise from a system that can identify people and has been configured not to, and buyers have become good at telling those two apart.

As my esteemed colleague James Connor, head of customer engagement at Ambient.ai with decades of experience, puts it: trust has to move at parity with value, especially in physical security. A capability that outruns the trust placed in it does not get deployed, or it gets deployed and then withdrawn after the first uncomfortable headline. Privacy by design is not a concession the industry makes to critics. It is the thing that lets the useful version of this technology exist at all.

The part that stays human

The agent does not fight the fire, and it does not decide. It notices, it reasons about what it is seeing, and it puts a single meaningful alert in front of a person with the frames attached. That person looks, judges, and makes the call.

I want to push on this, because there is a weaker version of the argument that I hear constantly and that I think the industry should stop making. The weak version is that AI makes the existing security workflow faster, running the same monitoring and the same alarm queue and the same escalation tree at higher speed. That framing is comfortable because it asks nobody to change anything, and it is why so much of this technology gets sold as an efficiency upgrade and then quietly underused.

The stronger and more accurate version is that much of the existing workflow only exists because security never had real-time reasoning available to it. The alarm queue, the after-the-fact footage review, the practice of assigning a person to stare at a wall of feeds: these are all workarounds for a capability that did not exist. Now that it does, the honest question is which of those workarounds should still exist at all, rather than how fast we can run them. Do not automate the mess. Remove the work, and give the human back the part that was always theirs.

The part that was always theirs is judgment. Watching was a coverage tax paid by people capable of much more, and it is work human attention is measurably bad at, which is a limit of the species rather than a failing of any individual operator. Deciding what a situation means and what the organization should do about it is skilled work, and no agent I would trust is anywhere near taking it. The useful way to describe the shift is from tools you operate to a system that operates alongside you, with the person moving up rather than out.

That is also where the economics change, because the cost of detecting something late in an industrial setting is not linear. An outdoor pallet fire and an involved structure with propane in it are two different emergencies, with different costs to property and different risks to the workers and the responders who have to go in.

None of this replaces a fire system. Sprinklers, smoke detectors and suppression do a job that video cannot, and the argument here is for a second layer over the places those systems do not reach, which in most industrial sites means everything outside the walls. Nothing watching a camera would have stopped that plastic from igniting either. The claim I would make is narrower and, I think, more useful: the interval between ignition and recognition is the piece of this that is addressable, compressing it changes what the rest of the sequence can become, and agents are now genuinely capable of compressing it.

The cameras in that yard were almost certainly running the whole time. What a site gets to decide now is whether anything is reading them while a fire is still the size of a pallet.