Every AI surveillance company fits one of six shapes. Pick the shape before the vendor.

An AI surveillance company runs automated inference on video feeds in real time, rather than recording for later review. The automation has to decide what a camera is seeing (a person, a vehicle, a package, a zone entry), score whether it matters (time of day, zone rules, dwell), and deliver a useful alert to an operator inside seconds, not minutes. The 'AI' label is generous in the market right now. A lot of vendors badge themselves as AI surveillance companies when the underlying product is motion detection with a dashboard. The practical test is latency: if the time from 'actor enters a restricted zone' to 'a human receives an actionable alert' is measured in seconds, it is a real AI surveillance company. If it is measured in minutes or requires someone watching a monitor, it is an old product with new packaging.

Because they are solving different problems even when the marketing copy looks identical. A 16-camera Class C multifamily property can get a quote of $450 one-time plus $200 a month from an edge-AI adapter vendor, a quote of $35,000 one-time plus a cloud subscription from a rip-and-replace smart camera company, and a quote of $4,000 a month from a remote-guarding AI company. All three label themselves AI surveillance companies. They are pricing completely different shapes of deployment. The adapter shape taps the DVR you already own. The rip-and-replace shape assumes you will throw out every camera. The guarding shape includes live humans. Buyers who do not understand the shape before they shop spend weeks comparing apples to helicopters.

Rip-and-replace shapes (Verkada, Rhombus, Lumana, Coram AI, Spot AI, Turing AI, Eagle Eye Networks) require replacing cameras with their first-party IP cameras or a certified subset. Cloud-AI overlay shapes (Scylla, Cloudastructure) work with some existing IP cameras as long as they expose RTSP or ONVIF with usable credentials. Remote-guarding shapes (Deep Sentinel, Silver Star, Stealth Monitoring) typically bring their own cameras or use IP cameras. Retail-specific shapes (Veesion) work with existing IP cameras in stores. Edge-AI adapter shapes (Cyrano) require none of the above: they read the DVR's HDMI multiview output and work on analog and IP cameras regardless of brand, firmware, or credentials. If a property has working analog cameras with a 2015-era DVR, the adapter shape is the only one that does not require a truck roll.

Edge means the inference runs on a device physically at the property, not in a cloud region. For a surveillance context, that implies three things. One, the video does not leave the building, which satisfies tenant privacy and keeps bandwidth low. Two, the alert latency is bounded by the property's LAN and the notification channel, not by upload to a data center. Three, the cost model is a box plus a small recurring fee, not a per-camera cloud subscription that scales with feed count. Cyrano is edge-AI in this specific sense: the physical unit sits next to the DVR, ingests the HDMI multiview at 25 tiles, runs person detection and zone rules locally, and only pushes event thumbnails and metadata off the property. Nothing full-frame ever leaves.

The question is whether the value is AI acting on every camera every second, or a human acting on a filtered alert. Remote guarding companies (Deep Sentinel, Silver Star) offer live humans who can talk down an intruder through a speaker, but they cost between $1,500 and $4,000 a month per site because you are paying for attention-time. Pure AI surveillance companies sell automation: the AI does the watching and pages a human at the property only when the rule fires. For most Class B and C multifamily, the math works out in favor of AI surveillance with an on-property on-call, not a remote guarding subscription. The exception is high-value construction jobsites and remote facilities where a live voice matters more than the per-month cost.

Ask: 'What is the capture point?' That is, where physically does the inference process get its pixels from. The answers sort every vendor on the market. If the answer is 'your IP cameras via RTSP,' you are committing to cameras that expose RTSP with valid credentials. If the answer is 'our cameras,' you are committing to a camera replacement. If the answer is 'a cloud ingest per stream,' you are committing to uploading every camera's feed 24/7. If the answer is 'the DVR's HDMI multiview output,' you can keep whatever cameras and DVR you already have. This question alone decides whether you are buying a product or a project. The rest of the feature list is downstream of this one architectural choice.

For the detections a 200-unit Class B or C multifamily operator actually has to act on (trespassing, package theft, tailgating, loitering, parking-lot activity, zone entries outside hours), yes. Cyrano catches the same categories as Verkada, Coram, and Lumana using person detection plus zone and dwell rules, on the cameras already installed. What Cyrano does not offer is license plate OCR at 4K per camera, biometric ID, or integration with the enterprise access-control stack. If the property needs those, an enterprise AI surveillance company is the right shape. If the property needs fast incident detection on existing cameras at a fraction of the cost, the adapter shape is the right shape. That is the whole selector.

Approximate list prices from public vendor sites and publicly reported multifamily deployments: Verkada cameras run around $1,200 to $2,500 per camera with a $300 to $600 per camera per year license, which for 16 cameras is $24,000 one-time plus $6,000 a year. Coram AI and Lumana sit in the same band. Cloud-AI overlays like Cloudastructure and Scylla typically charge $20 to $80 per camera per month. Deep Sentinel starts around $500 per month for a small residential kit and scales to $2,000 to $4,000 per month for commercial with live guarding. Cyrano is $450 one-time plus $200 per month for the whole property, regardless of whether there are 8 cameras or 25. The math only makes sense because the capture point is a single HDMI port, not 25 separate cameras.