Most industrial sites already own the hardest part of a video-AI deployment: hundreds of correctly positioned cameras, cabled, powered, and pointed at the places where incidents happen. The pitch to replace them with “AI cameras” asks you to write off that asset and run a quarters-long capital project to buy back a capability you can add in days.
This is the practical playbook for the retrofit path — what actually happens between “we have CCTV” and “our policies are enforced on it,” based on how pilots really run.
Step 1: Inventory what you have
The only camera requirement that matters is a stream the inference layer can consume — in practice, ONVIF or RTSP, which covers the large majority of fixed IP cameras installed in the last decade. Pull the camera list from your VMS (make, model, resolution, location) and check each model. The Compatibility Checker answers per model in seconds; a mixed estate with a few analog stragglers is normal and rarely a blocker — encoders exist, and coverage can start with the compatible 90%.
Step 2: Choose use cases by consequence, not by demo appeal
A pilot earns its rollout by preventing things your organization already pays for. Rank candidate use cases by severity and frequency of what they prevent — forklift–pedestrian conflicts, missing PPE, blocked exits, zone incursions — and pick the two or three where cameras already cover the risk area. The Use-Case Library maps roughly ninety of them by industry with a sample rule for each.
Step 3: Size the edge, don't guess it
Retrofit intelligence runs on an on-site GPU appliance, and its size is a function of camera count, resolution, and use-case load — not a mystery. As a rule of thumb, plan around three to four standard use cases per camera at standard processing weight, and budget roughly half a gigabyte per camera per day for sealed event clips (not continuous recording — your VMS keeps doing that). The CamEdge Sizer turns your real numbers into a hardware spec per site.
Step 4: Write the policy as rules — before anyone installs anything
This is the step that surprises teams used to model-training projects. The policies you already have on paper become the deployment: “flag anyone on the floor without required PPE — notify the shift supervisor after two events” is a complete, enforceable rule. Zero-training detection means there is no annotation phase; writing and revising rules takes minutes in plain English. Draft them with the people who own the SOPs, and you have done the integration work most projects spend months on.
Step 5: Pilot on 10–50 cameras with success criteria in writing
The proven pattern is deliberately small: one or two sites, the cameras covering your chosen use cases, and detection targets plus alert-quality thresholds agreed before anything is signed. Go-live is days, not quarters, because the physical work is connecting streams — not mounting hardware at height across a live operation. Everything the pilot produces — rules, zones, integrations, evidence — carries directly into rollout.
The retrofit question is ultimately financial: a working camera estate is a sunk cost that can start earning, or a write-off waiting for a capital budget. Days versus quarters is the difference.
If you want the two paths side by side with your own camera count, the TCO Comparison models retrofit against rip-and-replace over a multi-year horizon — and the Pilot Planner turns the result into a scoped, fixed-price pilot.
