Hundreds of cameras already installed, none of them smart: how to add AI without replacing the infrastructure

1. The installed base problem

A portfolio with 30 properties and an average of 20 cameras per property has 600 cameras in service. That installed base is a sunk cost: cameras, cabling, DVRs, switches, monitors. Replacing it costs $1.5M to $3M and takes 12 to 24 months. Most portfolio owners do not have that capital appetite.

The cameras themselves are usually fine. They were installed in the last 5 to 10 years, they record continuously to a DVR, and the footage quality is workable for incident review. The thing that does not exist is real time intelligence: nobody is monitoring the feeds 24/7, and the DVR has no concept of 'what is in this frame.' Incidents are caught after the fact, if at all.

This is the installed base problem. The infrastructure works for documentation. It does not work for prevention.

2. What AI actually changes when added on top

AI on top of the existing camera feed adds three concrete capabilities. First, real time object detection: the system knows when a person, vehicle, or environmental motion event happens, in which camera tile, at what time. Second, rule-based alerting: zone + time + object class composable rules that fire alerts only when the conditions match.

Third, persistent metadata on every event: the AI overlay stores a clip, a thumbnail, a class label, a zone tag, and a timestamp for every alert. This makes search useful in a way it never was on raw DVR footage. 'Show me every clip last week where a vehicle entered the loading bay between midnight and 6 AM' is one query against the AI metadata, not a 4 hour scrub through the DVR.

3. The architecture that preserves the installed base

The architecture is an edge AI device that taps the existing DVR's HDMI multiview output. The DVR already produces a composite feed (4x4 or 5x5 grid of all cameras) for the office monitor. The AI device reads that same composite, runs object detection on the whole composite at native frame rate, and de-tiles each detection back to a per camera ID using a fixed grid template.

Nothing about the existing camera + DVR setup changes. The cameras keep recording to the DVR. The office monitor keeps showing the same multiview. The new AI device is a separate physical box that taps the multiview a second time via an HDMI splitter or the DVR's secondary output.

Cyrano is one product built in this architecture. There are a small handful of others. The category exists because the installed base problem is universal and replacement is rarely viable.

4. Rollout pattern across a 20 to 50 site portfolio

The pattern that works is not 'order 30 devices and install everywhere in a quarter.' The pattern that works is start with one site, ideally the one with the highest historical incident count, install on a low traffic day, run for 30 days, and review the alert history.

The first 30 days exposes the rule grammar: which zones matter, which time windows are real, what false positives look like at this site. After tuning, the rules transfer to the next 3 to 5 similar sites with site-specific overrides for unique zones.

From there, batch rollouts of 5 to 10 sites at a time work well. The total time to portfolio-wide deployment is typically 2 to 4 quarters depending on site count and travel logistics.

5. Cost vs full camera replacement

On a 20 camera site, edge AI overlay is roughly $450 in hardware and $200 a month in software, $2,850 first year and $2,400 a year thereafter. Total cost over 5 years per site: $12,000 to $13,000.

Full camera replacement on the same 20 camera site is $30,000 to $80,000 capital plus $1,000 to $5,000 a year in cloud and management fees. Total cost over 5 years per site: $35,000 to $105,000.

Across a 30 site portfolio, that is the difference between $390,000 and $3,000,000+ over 5 years. The math on retrofit is straightforward.

6. What retrofit cannot do

Retrofit overlay is bounded by the quality of the underlying camera. If the camera is producing a grainy night image, the overlay AI sees the same grainy image. If the camera is misaimed, the overlay AI works on the wrong frame.

Some advanced features (license plate recognition at acute angles, face recognition at distance, 4K-resolution forensic search) require a higher-quality input than typical 5 to 10 year old cameras provide. For those features, the camera itself needs replacement.

For ordinary trespass / loitering / vehicle / package alerting at typical 1080p resolution, retrofit overlay is at parity with much more expensive systems.

7. Evaluation checklist before rollout

Pull a 1 AM clip from each camera. Can you see a person at 15 feet? If yes, retrofit candidate. If no, the camera needs replacement first.

Check the DVR. Does it record continuously for 30 days, no gaps? If yes, retrofit candidate. If no, the DVR is the problem.

Identify the rule shape per site: zone polygons, time windows, object classes. The richer the rule grammar, the better the alert quality.

Define the on call workflow: WhatsApp group, SMS shortcode, dispatch contact. AI alerts that land in a dashboard nobody opens are wasted.

Pilot one site, tune for 30 days, then roll forward.