Blog
All AI surveillance technology is not the same. Some of it is not even good. Everyone has a demo that works, and a demo is not the same thing as a system that works at two in the morning on a camera that has been on a wall since 2019.
You do not need to be an engineer to tell the difference. You need six questions, and you need to know what a good answer sounds like.
THE SHORT VERSION
Ask where video is stored and processed, whether a human supervises the machine learning, whether installing the system opens anything up on your network, where the AI's training data came from and who owns it, who actually responds when an alert fires, and who owns the footage. The sixth question is the one most buyers skip and the one most likely to matter later.
Video is big data and AI is expensive to run, so this question is really about who carries the cost, the maintenance and the risk.
There are three places it can happen and each has a legitimate use. Onsite puts the hardware in your building and makes you responsible for it — disk arrays, software updates, backups, and enough cooling to keep it alive. Edge means the analysis runs in or beside the camera itself, and that is genuinely the right choice where a decision has to be instant and simple: unlocking a door, triggering a gate, anything where waiting for a round trip to the cloud is too slow. Cloud sends the video out to be analysed and stored, which handles the heavy computation and keeps a copy somewhere the building cannot take with it.
The trouble starts when one of the other two is made to carry everything. Build the whole system on the edge and you are limited by what a small device can compute — the analytics stay simple, storage runs out, and you cannot search across sites or improve the system without replacing hardware. Build it all onsite and a fire takes your evidence along with the building.
What a good answer sounds like: the cloud. It is the safe answer for storage, search and the analytics that actually need computing power, and it is the only one that keeps a copy somewhere the building cannot take with it. If a provider also runs some processing at the edge for the time-critical pieces, that is not a red flag. What should give you pause is edge or onsite as the whole architecture, because that is where you end up owning the hardware, the limits and the risk.
Follow-up worth asking: how long is footage kept, and what happens on day 31? Retention is where quoted prices quietly differ.
This one sounds technical and is not. Supervised learning means people check the system’s work and correct it when it gets something wrong. Unsupervised means it teaches itself and nobody is looking.
In practice the difference shows up in your alerts. Every property has things that move and do not matter — headlights sweeping a wall, rain, a flag, a cat, a branch in the wind. A system nobody is correcting learns to treat all of it as worth flagging, and you start getting dozens of alerts a night that mean nothing.
What happens next is the real problem. People stop looking. Once the alerts are almost always nothing, the one that is something lands in a queue nobody reads any more. The usual fix is to turn the sensitivity down, which trades a noisy system for a quiet one that misses things.
Supervised means somebody is correcting it. The system gets told that the headlights do not matter and that the person walking the row trying door handles does, and over time it gets better at telling them apart instead of simply getting louder.
What a good answer sounds like: people are in the loop, and there is a way for your own team to teach the system about your property — what matters at your site and what does not — rather than leaving it frozen at whatever it knew the day it was installed.
There have been some genuinely catastrophic breaches of enterprise video surveillance. Ask about cybersecurity specifically, and listen for what the installation requires from your network.
An old-fashioned recorder keeps video inside the building, so watching it from anywhere else means propping open a door in the network and leaving it open. That works, and it is also a way in — and a way in for you is a way in for anyone who finds it.
What a good answer sounds like: no fixed IP address required, no holes in your firewall, no port forwarding in your router. A system that only ever reaches outward. Ask the follow-up too: is video encrypted on the way to the cloud and once it is there?
Two sub-questions do all the work here, and almost nobody asks them.
Where does the training data come from?
Garbage in, garbage out. Computer vision is built by showing a model millions of images until it learns to recognise something. But the images have to look like what the camera will actually see. A model trained on catalogue photographs of handguns — clean, centred, well lit, side-on — has learned to recognise a product listing, not a person holding a weapon in a stairwell at night, partially out of frame, at a bad angle.
Most commercially available models are not trained on real surveillance footage. It is the single biggest reason a system that dazzled in the demo disappoints on your cameras.
What a good answer sounds like: the provider trains on real video surveillance data, from real cameras, at real angles.
Who owns the computer vision — the provider or a third party?
Every large technology company now licenses computer vision. A surveillance provider without its own AI has to license someone else's, mark it up and resell it to you. That is expensive at scale, and it means critical features can be switched off by a company you have no relationship with. Facial recognition in particular has been turned on and off by large providers to suit the market rather than their customers.
What a good answer sounds like: the provider develops and owns its own AI.
Good AI filters out most of the noise. On a portfolio with dozens of cameras per building, most of the noise is still a lot of alerts.
So: who watches them — you, the provider, or a third party nobody mentioned? Detection without a human response is an alarm, not a deterrent. Using AI surveillance the way it is meant to work, to stop something while it is happening rather than document it afterwards, needs remote guarding software and remote guarding people.
What a good answer sounds like: fully integrated remote guarding, software and service together. Ask who employs the people watching — the provider, or an unaffiliated subcontractor — and ask where they sit. Both answers tell you something.
Read the fine print. It is your business and your cameras, but a number of providers write into the contract that they control the data. You want footage from last Tuesday; they run the investigation for you and charge you for the privilege of learning what it found.
There is a second use worth asking about. Your footage is a marketing asset to a vendor — a dramatic clip of a break-in at your property is exactly what a security company wants for its next case study or trade show loop. If the contract permits it, your property’s worst night becomes somebody else’s advertisement, with the address identifiable and your residents in frame. That is a reputational problem before it is a legal one, and the resident who recognises themselves brings the complaint to your office, not the vendor’s.
The one to close on: does it share with law enforcement automatically?
Some platforms are built to push footage into a law enforcement network continuously, or to let agencies search it on demand. These arrangements are sold as a public safety benefit, and they are often described to property owners in terms of response: join the network and the police will get there faster.
That claim is worth examining, because it runs two different things together. Nobody on the law enforcement side is watching your cameras. These networks are overwhelmingly investigative — a detective searching footage after something has already happened. That is genuinely useful for solving a crime. It does nothing to shorten the time it takes somebody to arrive while one is in progress. Faster response needs a person watching in real time, which is Question 5, not this one.
What has actually been traded is standing consent. You agreed once, at setup, to an unlimited number of future disclosures — and because each one is technically voluntary, the protection a warrant requirement would otherwise provide never engages.
Our own position, stated plainly: this is not a risk worth taking, and the record is now long enough to say why.
Once footage or plate data enters a shared network, the property owner no longer controls who looks at it or for what. That is not hypothetical. More than 50 US law enforcement officers have been charged with or accused of misusing automated plate reader systems, and in 26 of those cases the officer was tracking a wife, a girlfriend, an ex, or a woman he wanted to meet. One woman found that an officer had queried her and her daughter’s location 600 times. In Georgia alone, roughly 20 officers faced discipline or criminal allegations tied to a single network in one year.
The second failure is where the data ends up. Plate and location data collected at shopping centres, office parks and residential parking has been feeding private databases for years. One holds roughly 1.8 billion vehicle location records and adds around 70 million scans a month, and that data is sold on to insurers, banks and repossession firms, with searches offered for as little as $20. Unlike a police database, a commercial one can be queried by nearly anyone, for nearly any reason.
For a multifamily operator that lands uncomfortably close to home. Insurers use this data to find what the industry calls garage fraud — drivers who registered a vehicle at a cheaper address than the one they actually park at. Which means a camera at your own property can become the evidence that raises your resident’s premium, and neither of you will ever know it happened.
State regulators have started auditing these arrangements and finding violations, including data leaving a state in ways its own law prohibited. The property owners whose cameras answered those queries did not know, were not asked, and could not have stopped it.
This is being audited and litigated now and the liability picture is unsettled. What is already settled is the shape of the problem. “We share with law enforcement” is not one decision with a knowable scope. It is a standing permission whose uses get defined later, by other people, in other states, under laws you do not control and cannot predict.
You can weigh that and decide the trade is worth making. We think the asymmetry is too steep. The benefit was oversold, the exposure is real, and the consequences land on your residents and your reputation rather than on the company that sold you the network.
Ask four things specifically. Can you access your own footage directly, without asking anyone. What may the provider do with it — and is that written down, or left vague. May they use it in their own marketing, and if so, does that need your permission every time. And what happens when law enforcement asks: does the system hand footage over automatically, or is that your decision to make?
What a good answer sounds like: you own your footage outright, it says so in the service agreement, access is always yours, and the provider does not sell it or share it with third parties. Any use of your data should be disclosed in the contract rather than buried in it, and marketing use should need your written permission each time. And no footage should leave for law enforcement on an automatic basis — if a system can do that, you should be able to switch it off, and the decision to share anything should land with you, each time, while you still have the facts in front of you.
Getting answers to six technical questions sounds daunting, and there are providers who will meet all of it without flinching.
In any case, congratulations on asking. Moving to AI surveillance can be a moon landing for a security programme: you go from having footage to hand the police after you have been robbed, to the same cameras working as a real-time tool that stops the robbery happening.
Two of these questions have enough underneath them to deserve their own answer:
.svg)
.svg)
.webp)
