Surveillance buying has gotten harder, not easier. Every spec sheet claims 4K-per-camera 24/7 retention with AI analytics on a PoE switch that draws 120W. The reality is harder to pin down. Here's how to buy cameras that actually capture the incident, without burning your budget on features you'll never use.
01Resolution: pixels-per-foot is what matters.
A 4K camera on a 60-foot field of view delivers the same effective resolution as a 2MP camera on a 15-foot field of view. Marketing talks megapixels; investigators need pixels-per-foot at the object of interest.
Forensic-grade identification requires 50+ pixels per foot at the subject. License-plate capture needs 60+ at the plate location. Faces, 20 pixels per inch. These are the numbers to design to, not camera resolution in isolation.
02Retention: the number most specs get wrong.
"30 days of retention" is the industry standard ask. It's also routinely over-quoted because the math changes dramatically with resolution, bitrate, and motion profile. A 32-camera 4MP system with 24/7 recording and normal motion activity wants 40–60TB usable, at modern compression.
H.265 (HEVC) and AI-driven smart codecs (Axis Zipstream, Hanwha WiseStream) can cut storage requirements by 40–60% vs H.264 at equivalent quality. Specify these; don't just accept what's defaulted.
03On-prem vs cloud surveillance.
Traditional pattern: NVR in a rack, cameras over PoE, monitor from a workstation or mobile app. Low recurring cost, high upfront, proven architecture.
Cloud pattern: cameras stream to cloud (Verkada, Rhombus, Eagle Eye), no local NVR, rich analytics, higher recurring cost. Dramatically better user experience, meaningful ongoing subscription.
Hybrid: local NVR with cloud failover, or cloud with local recording at the edge. Best of both, more complex. The right answer depends on retention requirements, bandwidth, and how much the operations team values the cloud UX.
04AI analytics: which ones actually matter.
License plate recognition (LPR) is genuinely useful, auto-populated entry/exit logs, stolen-vehicle alerts, visitor correlation. Person detection with filtering reduces false-alert noise from 200/night to 5/night. Intrusion/loitering detection in specific zones does real work.
Ignore: "facial recognition" in most commercial SMB contexts (accuracy is poor, privacy risk is high, legal questions are unsettled). "Smart search" features that are table stakes, not differentiators. Any AI feature that requires a separate server and a five-figure subscription.
05The integrator matters more than the product.
A great camera platform installed poorly is a bad system. An adequate platform installed correctly is a good system. The specification, placement, aiming, and commissioning all have more impact on outcome than the camera brand.
When evaluating proposals, ask for the integrator's last five installs and a reference visit to one of them. The gap between integrators is much wider than the gap between name-brand camera platforms.
Recent camera projects: 32-camera IK-10 install at a Cape Coral self-storage, 18-camera resort-entry system on Fort Myers Beach, 12-camera estate perimeter in Port Royal with long-range PTZs. Axis and Hanwha primary; Milestone VMS for anything over 16 cameras. Pixel-per-foot specified per camera, not a number we handwave.