5 Questions to Ask Your AI Surveillance Vendor

The AI surveillance market is crowded. Every vendor promises real-time detection, fewer false alarms, and seamless integration with your existing systems. Some of those claims are true. Many are not.

The problem isn't that vendors are dishonest. It's that the questions most buyers ask during procurement don't expose the gaps. You ask about features. You see a demo in ideal conditions. You sign a contract. And six months later you're managing a system that needs constant tuning, runs on hardware that costs twice what you budgeted, or flags so many events your operators have stopped paying attention.

The right questions to ask aren't about what the system can do. They're about what happens when conditions aren't ideal, what the real total cost of ownership looks like, and whether the vendor's claims hold up in the real world.

Here are five questions that will tell you more about an AI surveillance vendor than any product demonstration.

1. Is your AI 100% accurate, and should you trust a vendor who says it is?

If a vendor tells you their AI is 100% accurate, that is your cue to end the conversation.

No AI surveillance system is perfect. Camera angles change. Lighting shifts between day, night, and adverse weather. Any system operating in the real world will produce some false positives and some false negatives. That is not a design flaw. It is the nature of machine learning applied to dynamic environments.

The claim of 100% accuracy is not a sign of a superior product. It is a sign that a vendor is more interested in closing a deal than in preparing you for what you will actually experience.

What you should be asking instead is more specific: What is the false positive rate in a real deployment, in an environment similar to mine? How does the system handle edge cases such as unusual lighting, occlusion, and crowded scenes? What happens when the AI is uncertain: does it flag or suppress? And critically: how does the system's accuracy change over time, and what does improvement look like?

A vendor who can show you how their system performs across hundreds of cameras in a live environment is worth far more than one who promises perfection in a controlled demo.

2. How many rules do I need to configure, and who maintains them?

This question separates self-learning AI from dressed-up rule-based analytics, and the answer has significant long-term cost implications.

Traditional video analytics systems work from a set of predefined rules. You draw a virtual line, and the system alerts when something crosses it. You define a zone, and the system alerts when someone enters it at the wrong time. These rules have to be created, tested, and maintained by your team, and updated every time your environment changes.

The problem is that environments change constantly. Staff patterns shift. Seasonal foot traffic alters what "normal" looks like. A new construction zone changes the routes people take. In a rule-based system, every change potentially means broken alerts, missed incidents, or a flood of new false positives until someone logs in and reconfigures the logic.

Ask your vendor: does the AI learn the normal patterns of my environment automatically, or do I need to define what normal looks like? Does it adapt as my environment changes, or does it require manual reconfiguration? Who on my team will be responsible for maintaining the rules, and what training do they need?

A system that learns continuously without requiring manual input is not just more convenient. It is more reliable, because it reflects what is actually happening in your environment rather than what you anticipated when the system was first installed.

3. Will it work with my existing cameras and VMS, or do I need to replace hardware?

This question is often overlooked during procurement and frequently becomes one of the largest hidden costs of an AI surveillance deployment.

Most organisations already have significant infrastructure in place: cameras installed across multiple sites, a Video Management System managing footage and access, and an operations team that knows how to use those tools. A new AI system should extend the value of that infrastructure, not replace it.

Some AI surveillance vendors require proprietary cameras or their own VMS platform to operate. Others integrate with the leading platforms but only support certain camera models or firmware versions. The sales team may gloss over these requirements, but they become very real once your technical team starts the implementation process.

Ask directly: which Video Management Systems does the AI integrate with natively? Which camera manufacturers and models are supported? Does integration require additional middleware, custom development, or ongoing licence fees? What does the implementation process look like, and who is responsible for it?

Vendors with genuine integration depth (such as native support for platforms like Genetec Security Center and Milestone XProtect) can typically deploy across your existing infrastructure without disruption. Those with limited integration options will often ask you to replace hardware you don't need to replace, at a cost that was not in your original budget.

4. What happens to my false alarm rate, and how does the AI improve over time?

False alarm fatigue is one of the most significant and underreported problems in physical security. When operators receive too many alerts that turn out to be nothing, they start to discount them. Response times slow. Real incidents get missed. The very problem the AI system was supposed to solve becomes worse than it was before.

Studies consistently put the false positive rate of traditional CCTV analytics between 70 and 90 percent. An AI system that does not address this problem at its core is not a solution. It is a more expensive version of the same problem.

Ask your vendor what their approach is to reducing false positives, and be specific about what you want to understand. Is the AI reducing false alarms by learning the normal behaviour of your environment, or by simply raising the threshold for what triggers an alert? Those are fundamentally different approaches with very different outcomes.

A learning-based approach, where the AI builds a model of what is normal for each camera, at each time of day, across different conditions, produces far fewer false positives because the system understands context. An alert is generated not because motion was detected, but because the motion is unusual given everything the AI knows about that location at that time.

Also ask how the system improves after deployment. Does accuracy increase as the AI processes more footage, or does it stay static? Can operators provide feedback that refines the model? A system that improves continuously is a fundamentally different investment to one that performs the same on day one as it does in year three.

5. Does your system work on-premise, or only in the cloud?

For many organisations, particularly those in government, corrections, defence, healthcare, or regulated industries, this is not a preference. It is a requirement.

Cloud-based surveillance AI routes video data through external servers. That creates questions about data sovereignty (where is your footage stored, and under which jurisdiction's laws?), latency (how quickly can events be detected and escalated if footage is being processed remotely?), and connectivity (what happens to your security coverage if your internet connection goes down?).

Ask your vendor clearly: can the AI run entirely within my own environment, on my own hardware, without any data leaving my network? Is that a supported deployment option, or an afterthought? Are there feature limitations in the on-premise version compared to the cloud version?

A vendor with genuine on-premise capability (not just a box that phones home to a cloud service) will be able to give you clear answers about where processing happens, what data leaves your environment, and how the system behaves during a network outage.

Ask about camera-to-server ratio. This is a cost most buyers miss. On-premise deployments typically involve physical hardware, and that hardware has limits on how many camera feeds it can process simultaneously. Vendors do not always make this clear upfront. If your facility operates 200 cameras and each server handles 50, you need four servers. If that was not in your original quote, you are looking at a significant unplanned cost before the system even goes live. Ask for the camera-to-server ratio in writing, across your specific camera resolution and frame rate requirements, before you finalise any commercial agreement.

What Good Answers Look Like

If you ask these five questions during your next procurement process, you will quickly separate vendors who have genuine answers from those who are repeating marketing copy.

Good answers are specific. They reference real deployments,and real integration requirements. They acknowledge limitations honestly rather than deflecting. And they give you enough information to make a decision based on what will actually happen in your environment, not what looks best in a controlled demonstration.

At icetana AI, our system is self-learning, requires no rules to configure, deploys in under 24 hours across your existing cameras and VMS, and is available as both a cloud and fully on-premise solution through Antara Core, our private AI for organisations where data sovereignty is non-negotiable.

If you want to see how it performs in an environment like yours, book a free demo with our team.