The seven AI security camera features being marketed in May 2026, scored by whether you have to replace your cameras to get them.

Because for the multifamily, small-commercial, and HOA segments that already have 16 to 25 cameras wired to a DVR, the difference between 'this feature exists in the marketing' and 'this feature is reachable from where I am today' is the entire conversation. A feature that requires ripping out 25 cameras, rerunning cabling, replacing the recorder, and reconfiguring every motion zone is in a different cost universe (roughly $50,000 to $100,000 per building) than a feature that drops in over the existing HDMI port in under 30 minutes. Marketing pages collapse those two into one bullet list. The decision a property operator actually has to make does not.

The DVR or NVR already paints a 1920x1080 composite mosaic of all its cameras to the wall monitor in the back office, that is what the live monitoring screen has always shown. An overlay device taps that HDMI signal, runs one inference pass over the full composite, and maps detected bounding boxes back to per-tile coordinates (480x270 per tile at a 4x4 grid, 384x216 at a 5x5 grid). One inference pass covers all 16 to 25 cameras at once, no per-camera RTSP credentials, no network closet changes, no firmware fleet to manage. The cameras keep doing what they already do; the overlay does the AI on top.

Natural-language search runs against a per-event description index. The model sees the composite tile (480x270 or 384x216 pixels), writes a one-line sentence per surviving event, and the search hits that sentence in plain text. 384x216 is well inside the working range of nano-class detectors and vision-language models. Cross-camera person re-ID is different: it tries to match the same individual across multiple cameras using subtle appearance embeddings (clothing texture, gait stride, facial geometry). Those features collapse below roughly 256 pixels of head-and-shoulders height, and a tile that gives you 80 pixels of person is not enough. To match identity across cameras with reasonable precision you need each camera's native 1080p or 4K stream, which means either pulling per-camera RTSP after the alert (slow, manual, often impossible because credentials were lost) or replacing the cameras (expensive). This is the one capability where the new-camera path genuinely pulls ahead.

Motion detection on a stock DVR fires whenever pixels change inside a region. A leaf, a shadow, a passing car, a maintenance worker, a delivery driver, a crow, and a person climbing a fence all generate the same event. Intent-aware alerts add a classifier that scores the event on a small set of axes: is the subject a person or vehicle, are they inside an armed zone, how long have they been stationary, are they approaching a threshold (door, window, gate), do they match a vehicle profile, and is this within or outside the operating hours of the property. The output is a small label like LOW THREAT (delivery driver, daytime, no dwell) versus HIGH THREAT (person, after hours, dwell over 30 seconds at the rear gate). The classifier runs on the same composite tile the description model uses, so it adds no per-camera infrastructure.

Not if the deterrence devices have their own API. A camera that triggers a smart light or a siren is just sending a webhook on a detection event. If the lights live on Hue, Shelly, Sonoff, Lutron, or any HomeKit / Matter bridge, the webhook goes to that bridge, not to the camera vendor's app. The same is true for door locks (August, Yale, Schlage, Salto), sirens (DSC, Honeywell, Ring), and PA systems (Algo, AtlasIED). The detection layer can sit anywhere as long as it can fire an HTTP call. The bundled vendor solutions (Ring's automated lights, ADT's command-and-control) are convenient for residential, but they tie deterrence to the camera vendor; an HDMI overlay path lets the property mix and match.

The map is a top-down rendering of the property with detected objects placed on it. The map itself is a one-time configuration step (drag camera positions onto a floor plan or site plan). The objects placed on the map come from per-event detections plus the per-camera position. Once detections are running, mapping them to a floor plan is a coordinate transform and an overlay, not a new sensor capability. Vendors who advertise this as a new 2026 feature are mostly shipping the map UI, not the underlying detection pipeline. If you already have intent-aware alerts running off your DVR composite, adding a top-down map view is a frontend project, not a hardware project.

Edge AI keeps the raw video on the device. Only event metadata (timestamp, camera ID, classification, one-line description, low-resolution thumbnail) leaves the property. Cloud AI requires either streaming all 16 to 25 camera feeds out continuously (high bandwidth, regulatory exposure) or sending clips to a hosted analyzer per event (lower bandwidth, but the clips themselves still leave). Multifamily and HOA operators in 2026 are increasingly being asked by tenants and boards to confirm that camera footage does not leave the property in raw form. An on-device or edge-AI architecture answers that question cleanly. The DVR composite path is naturally edge: the inference happens on the same physical box that ingests the HDMI signal, and only the resulting alert metadata is sent out.

Two clusters worth being skeptical about for a property-operator buying decision in May 2026. First, vendor-locked AI Box products (Reolink AI Box, Ring's AI features, Nest's Familiar Faces) are real, but they only work with that vendor's cameras; if your property has a mixed Hikvision / Dahua / Lorex / Swann fleet, none of these apply and the marketing tour is a distraction. Second, 24 megapixel triple-lens flagship cameras and other resolution flexes are technically impressive but solve a problem most properties do not have. Recall and identification accuracy plateau well before 24 MP; the binding constraint at most multifamily sites is night-time low-light detection, not daytime resolution. A camera that captures usable evidence at f/1.0 with infrared assist outperforms a 24 MP daytime sensor when the actual incident happens at 02:14.

Three weeks, three steps. Week one: take honest inventory. Count cameras, identify the recorder make and model, confirm there is a working HDMI output to a wall monitor, list the kinds of incidents you most want to catch (after-hours intrusion, package theft, parking-lot loitering, vendor no-show). Week two: do one pilot site with an HDMI overlay. The install is under 30 minutes, the cost is the device plus a monthly subscription, and you will know inside 14 days whether the alerts are useful, the description search returns what you remember, and the false-positive rate is liveable. Week three: compare cost per useful alert against the projected cost of replacing the camera fleet at one building. If the overlay covers 80 percent of your real use cases at 1 percent of the replacement cost, you have your answer for the rest of the portfolio. If it does not, the pilot tells you which two or three feature gaps would push you toward replacement.

Yes for natural-language search of footage, intent classification with LOW / HIGH threat labels, real-time alert dispatch (SMS, phone call, webhook into Slack / Teams / WhatsApp), edge-AI inference on the property device with no cloud round trip, and a top-down map view per property. We do not do cross-camera facial re-identification at the high-precision threshold described above (that is the architectural limit of composite tile resolution, not a roadmap gap), and environmental sensor fusion (smoke, glass-break, CO) is out of scope because we are a camera-AI overlay, not a multi-modal sensor product. For everything in the 'works with overlay' column we either ship today or have it in active development. The honest version is in the table above.