What changed: cameras stopped being passive.

A traditional CCTV system records and waits. Someone reviews footage after the loss happens, usually hours or days later. By then the truck has left, the inventory's gone, the worker who skipped PPE is at home, and the only useful output is a police report.

Modern AI cameras read the feed in real time. The detection runs on the camera (edge AI on chips like Ambarella, Qualcomm RB5, or vendor-specific silicon) or in the cloud. Either way, the system sees an event and pages a human inside seconds. It also ignores the noise: animals, leaves in wind, a delivery driver walking the same path every Tuesday. Your team stops reviewing footage and starts reviewing alerts. The footage becomes evidence, not the source of detection.

The detections worth turning on (and the ones to ignore).

AI vendors ship dozens of detection types. Most sites only need a handful to cover 80% of the actual loss. Enabling everything generates alert fatigue and trains the security team to ignore the dashboard. Turn on the high-signal detections first; add more after a documented incident drives it.

  • After-hours intrusion in unstaffed zones. The highest-yield detection across retail, manufacturing, and SMB. The system knows when you're closed. Trip an alert when anything human-shaped enters.
  • Loitering at sensitive points. Front door, loading dock, ATM vestibule, pharmacy, server room. Someone standing outside a door for 90 seconds is rarely innocent.
  • License-plate awareness. Allow-list employee plates, flag unknowns at the dock, match against a watch list of plates from past incidents.
  • PPE compliance. Hard hat, vest, safety glasses on the manufacturing floor. Cuts OSHA exposure and gives supervisors a real-time coaching tool.
  • Tailgating at access points. Two people walk in on one badge. The log says one entry, the camera sees two. The discrepancy is the alert.

Leave facial recognition off by default unless there's a documented use case (high-shrinkage retail with known repeat offenders, or controlled access at a regulated lab). State biometric privacy laws (Illinois BIPA, Texas, Washington) carry real liability and the detection rarely earns its keep.

The cameras and platforms we actually install.

Tec-Tel is camera-agnostic. We install and integrate Verkada, Avigilon (Motorola Solutions), Genetec, Axis Communications, Hanwha Vision, Eagle Eye Networks, Coram AI, Pelco, Rhombus on the video side. Each ships its own analytics layer, and a few (Genetec, Milestone, Avigilon) accept third-party AI plugins on top. The right pick depends on whether you want cloud-managed (Verkada, Avigilon Alta, Eagle Eye) or on-prem with a VMS (Genetec, Milestone, Axis-direct). For sites that already own cameras, most analytics platforms run on top of existing IP cameras as long as they meet a minimum resolution and the network can carry the streams.

For NDAA Section 889 buyers (federal agencies, contractors with FAR 52.204-25 in their contracts): every vendor in our stack publishes an NDAA compliance statement. Don't take a sales rep's word for it. Pull the public statement and attach it to the purchase order.

Where AI video pays for itself.

Not everywhere, and not on day one. AI surveillance pays back fastest at sites with high after-hours risk and no security staff at those hours: single-store retail with a stockroom door, mid-size manufacturing without a 24/7 SOC, multi-site SMB rollouts where central monitoring replaces a guard at every site.

Payback is slower at sites that already have a guard force, sites where the loss is internal theft during staffed hours (the camera helps, but the answer is process and pre-employment screening), and sites with persistent network problems where alerts won't reliably leave the building. The free consultation puts a real number on the payback: we walk the building, list the loss patterns, name the detections that match, and write a phased rollout if the full system is too much for year one.