Bystander reaction vs CCTV after the fact: the difference is response time, not video.

Because the bystander triggers a response inside the only window where a response can do anything. A neighbor who hears a window break and dials 911 puts a patrol car en route within roughly 60 to 90 seconds of the event. By the time the suspect leaves, police are already in the area or pulling onto the block. CCTV-only review usually starts after the next business day, sometimes after a tenant complaint days later. Police arriving 24 hours late do not catch anyone. The camera saw it; nobody acted on what it saw.

Across multifamily and small commercial properties the median time-to-review for a routine after-hours incident is somewhere between 6 and 72 hours. Faster if a tenant complains the same morning. Slower if the only signal is missing inventory or vandalism noticed on a walk-through. For an 8 to 16 camera site, even when staff know which night to look at, scrubbing every camera at 8x playback takes 90 minutes to several hours of one person's day. Many incidents go unreviewed because the labor cost exceeds the perceived loss.

Bystander: roughly 30 to 120 seconds from event to a 911 call placed and a dispatcher pulling up the address. Live-monitored CCTV with AI: roughly 5 to 30 seconds from event to a phone or radio alert to the on-call operator. After-the-fact CCTV: 6 to 72 hours typical, sometimes longer. Two orders of magnitude difference between the live channels and the after-the-fact channel. Everything else in the security stack (locks, fences, lighting) lives or dies on whether someone is paying attention inside that window.

Two reasons. First, insurance carriers and city ordinances often require cameras at multifamily and commercial properties; the cameras get installed to satisfy a checkbox, not to drive an outcome. Second, operators conflate documentation with deterrence. A camera that records footage feels like a deterrent because a reasonable person would assume someone is watching. Professional offenders learned in roughly the 2010s that almost nobody is watching, and adjusted accordingly. The fix is not more cameras; it is making the cameras you already have act on what they see.

Yes. Every modern DVR or NVR draws a multiview composite to its HDMI output (a 2x2 grid for 4 cameras, 3x3 for 9, all the way up to 5x5 for 25 cameras). An edge AI device plugged inline on that HDMI cable reads the same picture a wall monitor would, splits the frame back into per-tile crops with a one-time calibration, and runs detection on each tile on-device. The cameras do not change. The recorder does not change. The wiring does not change. The added behavior is a real-time alert pipeline that previously did not exist.

A guard at $3,000 a month is a single human covering one walking patrol or one monitor wall. Attention degrades over a shift; a 2 AM event during the back half of a 12-hour shift is the worst-case time and the most common time for incidents. Live monitoring services are roughly $200 to $1,200 a month and rely on operators watching dozens of properties at once, so each property gets a fraction of one human's attention. An edge AI box monitors every tile every second, never blinks, and only escalates to a human (or directly to dispatch) when an event fires. The handoff to a human still matters; the difference is what the human gets paged about and when.

It works on basically any DVR or NVR with a working HDMI output from roughly 2014 forward. That covers most Hikvision, Dahua, Lorex, Reolink, Amcrest, Swann, Annke, Uniview, Q-See, and Night Owl recorders sold in the past decade. The retrofit is electrically isolated from the camera network; the box reads pixels off the HDMI port and sends alerts out a separate uplink. There is nothing on the recorder to upgrade and nothing to expose to the internet.

Two layers. First, the detection model classifies objects (person, vehicle, package) rather than firing on raw motion the way 2018-era DVR alerts did, so leaves, shadows, and house cats do not trigger anything. Second, an intent layer scores the activity against context: a person walking through a leasing path at 4 PM is not an event; the same person climbing the dumpster enclosure at 2 AM is. Operators get a triaged stream that is closer to bystander judgment than to motion-detector noise. The Fort Worth deployment surfaced 20 actionable incidents in month one, not 2000.

It helps. Live monitoring centers are quality-limited by how many tiles a single operator can scan, which on a busy night is well below the number of cameras any sizeable property has. The edge AI layer pre-filters: instead of an operator scanning 16 tiles every 90 seconds, the operator gets a queue of pre-triaged events with a tile, a timestamp, and an intent score. The same operator handles more properties more accurately. Several monitoring providers now sell the combined offering for exactly this reason.

Two things. Alert volume becomes meaningful: instead of 200 motion notifications a day that staff stopped opening, you get a handful of actual incidents per week with a tile and an intent label. Second, the time-to-action collapses from 'next business day' to under a minute. Whether you escalate to police, to a courtesy patrol, or to a tenant text, the action lands inside the window where it can still do something. The Fort Worth comparison: 20 incidents caught and acted on in month one, including a break-in attempt, on cameras that had been in place for years.