I is widely seen as the future of security – but most projects never make it past the pilot stage. In this article, Reece Downs examines why, arguing that the real challenge is not the technology itself, but how it is delivered. From integration failures to unclear ownership, he outlines where organisations go wrong and what it takes to turn AI ambition into operational reality.


A


The conversation around artificial intelligence in the security sector has been dominated, for some time now, by the possibility. The potential to automate threat detection, reduce response times, analyse vast volumes of surveillance data, and free up skilled operatives for higher- value work is genuinely compelling. Organisations across the industry have launched pilots, invested in platforms, and issued press releases about their AI journeys. And yet, when you look at what has actually been delivered, what is working in practice, at scale, across real operational environments, the picture is considerably less impressive.


The gap between AI ambition and AI delivery is not a technology problem. The tools available today are sophisticated and, in many cases, genuinely ready for enterprise use. The failure points, time and again, lie in how projects are scoped, integrated, scaled, and handed over. Understanding those failure points and how to address them systematically is what separates organisations that extract lasting value from AI from those that spend significant budget to end up exactly where they started.


The pilot trap


Most AI security projects do not fail catastrophically. They simply never leave the pilot stage. A proof of concept is commissioned, it produces encouraging results in a controlled environment, leadership nods approvingly, and then the project quietly stalls. Six months later, it is either still ‘in evaluation’ or has been quietly shelved.


The root cause is almost always the same: the pilot was never designed with delivery in mind. It was designed to demonstrate capability, not to answer the questions that matter for real-world deployment. Questions such as: How does this integrate with our existing systems? Who owns it operationally once it goes live? What happens when it produces a false positive


Why most AI security projects


fail in delivery and how we fix it


at 2 am on a Sunday? What does retraining look like when the environment changes?


A well-run pilot should be a miniaturised version of the full delivery, not a standalone experiment. That means defining exit criteria before it begins, identifying integration dependencies early, and involving the operational teams who will actually use the system from day one, not as an afterthought once the technology has been ‘proven’.


Integration: where ambition meets reality


Security environments are rarely clean. Legacy access control systems, analogue CCTV infrastructure, fragmented data sources, manual incident logs, and a patchwork of third-party platforms are the norm, not the exception. AI tools are generally built and demonstrated on well- structured, clean datasets. The real world is neither.


Integration failure is the single most common cause of AI delivery delay in my experience. Projects underestimate the time and complexity involved in connecting an AI platform to live operational data, and they underestimate the quality work required before that data is fit for purpose.


13 © CITY SECURITY MAGAZINE – SUMMER 2026


A video analytics system that performs brilliantly on the vendor’s demo footage may produce unreliable results when pointed at cameras with inconsistent lighting, varying frame rates, or coverage blind spots.


The fix here is to invest properly in a data and integration assessment before any AI platform is selected, not after. Map your data sources, assess their quality and consistency, identify the gaps, and build remediation time into the project plan. Choose vendors willing to work with your actual infrastructure rather than ones who require you to reshape your environment around their product.


Workflows and the human factor


Technology does not operate in a vacuum. Every AI tool ultimately sits inside a human workflow, and if the workflow design does not change to accommodate the tool or worse, if the people expected to use it do not understand it or trust it, adoption will be minimal regardless of how technically sophisticated the system is.


Security personnel are, understandably, cautious about systems that make automated decisions or flag incidents


www.citysecuritymagazine.com


Page 1  |  Page 2  |  Page 3  |  Page 4  |  Page 5  |  Page 6  |  Page 7  |  Page 8  |  Page 9  |  Page 10  |  Page 11  |  Page 12  |  Page 13  |  Page 14  |  Page 15  |  Page 16  |  Page 17  |  Page 18  |  Page 19  |  Page 20  |  Page 21  |  Page 22  |  Page 23  |  Page 24  |  Page 25  |  Page 26  |  Page 27  |  Page 28  |  Page 29  |  Page 30  |  Page 31  |  Page 32  |  Page 33  |  Page 34  |  Page 35  |  Page 36