How to evaluate AI CCTV monitoring: the questions worth asking

Every CCTV system sold today claims some form of AI. The label covers an enormous range - from motion filtering that has existed for a decade to systems that read behaviour and context in real time - and the marketing rarely distinguishes between them. If you are responsible for security at a site, or for the monitoring service your clients depend on, the difficulty is not finding vendors. It is working out what you are actually being offered.

These are the questions we think are worth asking, whoever you are talking to. Several of them have no single right answer; what matters is that the vendor can answer them clearly, and that the answer fits your sites.

1. Does it detect objects, or understand behaviour?

This is the first fork in the road, and most of the market sits on the near side of it.

Object detection identifies what is in the frame: a person, a vehicle, a bag. It is mature, reliable and useful, and it is what most "AI CCTV" means in practice. Behavioural understanding is a different thing - reading what is happening rather than what is present. The practical test is a simple one. Object detection can tell a person from a fox. It cannot tell a contractor doing legitimate work from someone who should not be there, because both are people, in the same place, at the same time of night.

Ask what the system does with a person who is present but not a threat. If the answer is that a human reviews the clip and decides, you are buying a filter. That may be exactly what you need - filters materially reduce alert volume and are cheaper to run - but you should know which you are buying.

2. Where does the processing happen, and what happens when the connection drops?

Systems process footage in the cloud, on the site, or in some combination. Each has real trade-offs and no option is simply better.

Cloud processing is easier to update and can bring more computing power to bear. On-site (edge) processing keeps working when connectivity fails and does not depend on uploading video continuously, which matters on rural compounds, construction sites and anywhere with poor or metered connections. Most systems are hybrid to some degree, and you should be sceptical of anyone claiming a pure architecture without explaining what runs where.

The question that gets a useful answer: what happens when the site loses internet for an hour? If the answer is that monitoring simply stops, decide whether that is acceptable for the sites you are protecting.

3. Where does the footage go, and is any of it biometric?

For UK sites this is a contractual question, not a technical curiosity. Some platforms - particularly those built for the US market - process footage in overseas data centres, which means your site's video crosses borders. If your organisation has data-processing obligations, or your clients do, you will need to be able to answer for that in writing.

Separately, establish whether the system uses facial recognition or any other biometric identification. Behavioural analysis and facial recognition are different technologies with very different legal and reputational profiles in the UK, and the distinction matters enormously in residential, retail and public-facing settings. A vendor should be able to tell you plainly which it does, and where the data is held and for how long.

4. Does it work with the cameras already on site?

Some systems require their own cameras, sold with mandatory licences. Others work with the CCTV you already have. This is the difference between a capital project - new hardware, new cabling, downtime, a procurement cycle - and a software change on an existing estate.

Neither is automatically wrong. If your cameras are genuinely at end of life, replacing them may make sense anyway. But be clear which you are being sold, and be wary of the assumption that new cameras solve the underlying problem. Modern cameras with embedded analytics still, for the most part, perform object detection. Replacing the estate can reproduce the same judgement gap on newer equipment.

5. What actually arrives when something happens, and how quickly?

Ask to see a real alert, not a slide. There is a considerable difference between a notification that says motion was detected in zone three, and one that describes what is happening in a sentence a person can act on without opening the video.

Ask about timing too, and be specific: how long between the event occurring and a human being told? Systems that batch, queue or route through a review step can take a minute or more, which for an intrusion is the difference between a warn-off and an insurance claim.

And ask how the alert reaches the person who needs it. An alert that lands in a dashboard nobody is watching at two in the morning has not solved anything.

6. Can the system stand down an alert on its own?

This one is asked less often than it should be. Any system that reduces noise is, by definition, deciding not to tell you about some things. Ask what it is permitted to suppress without a human confirming, and what evidence exists that the suppression is safe.

The safer design principle is asymmetric: the system can raise attention on its own, but cannot lower it without a person confirming that decision. Ask how the vendor handles this, and ask specifically about false negatives - the events the system missed. Vendors will happily discuss false-positive reduction, because it is the flattering number. Missed events are the ones that cost you a contract.

7. Who stays accountable, and how does it fit what you already run?

If you are a security firm or installer-monitor, this is the commercial question underneath all the technical ones. Some vendors are effectively competing monitoring services with their own operators, which puts them between you and your client. Others provide the technology and leave the relationship, the billing and the decisions with you.

Both models exist and both can work, but they have very different consequences for your business. Establish which you are dealing with before the pilot, not after.

8. How will you know if it works?

Finally, agree what proof looks like before you start. A pilot without a baseline proves nothing: you need to know what the site's alert volume, response time and incident history looked like beforehand, so that "it feels better" can be replaced with a number.

Sensible pilots run long enough to cover the site's real patterns - normal working days, quiet weekends, bad weather - and end with a decision that both parties agreed in advance how to make.

Volpex builds real-time monitoring intelligence on the cameras sites already have, for UK security firms, installer-monitors and commercial site operators. We are happy to be asked all of the above. If you would like to talk it through, get in touch.