Existing cameras or new AI cameras? A decision guide for safety analytics

Most sites that look at AI safety analytics already have CCTV. The first budget question is whether those cameras can do the job, or whether the project needs new ones. The honest answer is usually "some of each", and the split depends less on the cameras than on where your hazards are.

This article sets out the decision: what to check first, the four ways analytics reaches a camera, when to keep and when to replace, and how to compare cost and supplier risk over five years. For a camera-by-camera test of what you already own, use our existing CCTV readiness checklist alongside it.

Start from the hazard map

Security cameras are installed to watch doors, perimeters and stock. Safety analytics needs to see forklift aisles, dock edges, line-side zones, crossings and exclusion areas. A good camera pointed at the wrong place is no help.

Map the hazards first: where people and vehicles meet, where PPE (personal protective equipment) is required, and where people enter restricted zones. Then lay the existing camera fields of view over that map. Three results are possible for each hazard:

• Covered well. The camera sees the area at a useful angle, with enough resolution and frame rate, in all lighting.
• Covered badly. The camera sees the area, but from too high, too far, at too low a frame rate, or into glare.
• Not covered. No camera sees it.

Only the first group is a true brownfield (keep the cameras) case. The second group can sometimes be fixed by moving or re-aiming a camera. The third needs new hardware.

Four ways analytics reaches a camera

How the analytics is delivered decides how much of your existing estate you can use. The guide chapter on how AI video safety works explains the models in depth. In short:

On-camera AI. The camera runs detection on its own processor. This is now common on mid-range and premium cameras: Hanwha Vision says AI cameras were 49% of its network camera sales in the fourth quarter of 2025 ^[1]. Axis says its ARTPEC-9 chip has enhanced deep learning capabilities and triples video analytics performance ^[2]. Camera makers also offer app platforms where specialist vendors run their own models, such as Axis ACAP ^[3] and the i-PRO Camera Application Platform, which runs partner apps on Ambarella-based cameras ^[4]. Cisco Meraki lets customers deploy their own TensorFlow Lite models of up to 40 MB to its MV cameras and receive detections over MQTT ^[5]. On-camera AI only helps if your installed cameras already have a suitable processor and firmware, which older cameras usually do not.

Edge box or on-site server. A box on site pulls streams from existing cameras and runs the models. This is the main way analytics reaches existing cameras. Coram, for example, says it works with any IP camera through an on-site device ^[6]. Rhombus sells a Relay Core appliance that supports up to 10 third-party cameras for US$1,999 at list price, with no streaming licence fee ^[7].

Cloud platform with a connector. A connector appliance brings existing cameras into a vendor's cloud. Verkada's Command Connector and Avigilon Alta's Cloud Connectors are examples ^[8][9].

VMS plugin. Analytics runs as an app inside the video management system (VMS) you already use, and alerts appear in the operator's existing screens. The installed bases are large: Milestone reported EUR 298 million of 2025 revenue ^[10], and Genetec cites analyst firm Omdia ranking it first worldwide in video surveillance software ^[11].

What third-party cameras lose on cloud platforms

Vendors that sell their own cameras usually support other makers' cameras with a reduced feature set. Verkada's Command Connector page, for example, lists basic analytics such as people, vehicle and motion search for other makers' cameras, "limited support" for AI search with higher latency, trajectory analysis that may not be available on all channels, no license plate alerts and no firmware updates ^[8]. Other platforms have their own limits.

Get the feature list for your camera models in writing before a pilot. A pilot run on the vendor's own cameras tells you little about how the system will perform on yours.

When to keep, when to replace

A hybrid is the most common outcome: keep existing cameras where they are adequate, add a small number of AI cameras at the critical points, and run everything through one edge box or VMS plugin. This keeps capital cost down while fixing the coverage gaps that matter.

Five-year cost per camera

Each delivery model moves cost into a different line, so compare the full cost per camera over five years with these lines kept separate:

• Hardware: new cameras, edge boxes or connectors, and their replacement.
• VMS licence: often charged per device. In Milestone's 2021 US price list, for example, device licences cost US$169 to US$329 each depending on the product tier, on top of a base licence of US$499 to US$3,183 ^[12]. These are old list prices, and real quotes vary.
• Analytics licence: per camera, per site or per detection type.
• Cloud fees: storage and platform fees per camera.
• Installation, network and power: cabling, switch capacity and mounting.
• Support and patching: who patches cameras, boxes and models, and for how long.

Timing matters too. Novaira Insights warned in July 2026 that memory and hard disk shortages were stretching lead times to between 12 and 52 weeks and could raise prices ^[13]. Treat that as a forecast, but build lead time into any plan that depends on new hardware.

Supplier and security risk

Restricted brands. Hikvision and Dahua video equipment is on the US FCC Covered List for public safety, government facility, critical infrastructure and national security purposes ^[14]. Section 889 bars federal agencies from buying it and from contracting with companies that use it ^[15]. From October 13, 2026, the FCC will not authorize new devices that contain logic-bearing components from Covered List entities, but the rule does not affect equipment already authorized ^[16]. For what this means for an installed estate, see the FCC component rule and Hikvision and Dahua restrictions by country.

Old streaming standards. ONVIF is retiring Profile S because its username-token authentication is now too weak, and manufacturers cannot submit Profile S products after March 31, 2027 ^[17]. ONVIF points instead to digest authentication in Profile T or to TLS ^[17]. Profile M covers analytics metadata and events over MQTT ^[18]. For new cameras, specify Profile T, and Profile M where analytics metadata is needed.

Chip suppliers. Edge boxes and AI cameras depend on a small number of chip makers, and these change hands. Microchip completed its purchase of Hailo in September 2026 and says it will keep supporting existing products ^[19]. Ask what support commitment covers the chip in your box, and for how long.

Certificates. Cybersecurity claims vary in scope. Axis publishes an ETSI EN 303 645 certificate covering more than 200 products ^[20]. IQSIGHT claims IEC 62443-4-1 and 4-2 and UL 2900-2-3 ^[21]. Ask any vendor for the certificate number, the models covered and the firmware versions.

What AI video cannot do

AI video is an alerting system. It informs people; it is not a certified protective device that stops a machine. Functions such as area guarding or stopping an automated vehicle need sensors rated to standards such as IEC 61496 and ISO 13849. The SICK nanoScan3 safety laser scanner, for example, is rated Type 3 under IEC 61496 and PL d under EN ISO 13849 ^[22]. NVIDIA's Halos Outside-In blueprint uses fixed cameras to decide when it is safe for a forklift to enter a docked trailer, yet NVIDIA states it is not for production use in safety-related systems without your own certified safety layer ^[23]. Plan cameras and rated devices as separate layers.

What to do next

• Map hazards first, then overlay existing camera fields of view and mark each hazard as covered well, covered badly or not covered.
• Run the readiness checklist on the cameras that cover hazards well.
• Ask each vendor which delivery model it uses and which features are missing on your camera models, in writing.
• Build a five-year cost per camera with hardware, VMS, analytics, cloud, installation and support as separate lines.
• Screen every camera, recorder and edge box for restricted brands and Profile S-only support before deciding what to keep.
• Use the buying and piloting chapter to set pilot acceptance criteria on your own cameras, not the vendor's.