Industrial safety

Deploying AI video analytics on an industrial site

Camera analytics projects usually stall for one of four reasons, and none of them is the model. The cameras cannot carry the workload, the detection scope was never agreed, nobody owns the alerts, or there was no definition of success. This checklist covers those four before anything is installed.

Deployment checklist Published September 7, 2026 6 min read

Arrange a site review Industrial safety monitoring

01Why these projects stall

The model is rarely the problem

By the time a site is considering camera analytics, the detection question is largely settled: recognising a person without a hard hat is not the hard part any more.

What stalls these projects is more mundane and more expensive. Cameras positioned for recording after an incident rather than for detection during one. A detection scope that grew during the project until nothing was finished. Alerts arriving somewhere nobody watches. And no agreed definition of what a successful deployment would look like, which means it can always be argued to have failed.

02Before anything is installed

Can the cameras you have carry this?

Most sites can reuse a substantial part of their existing estate, and most sites also have some cameras that will never work for detection. Finding out which is which is the first task, not a surprise for month three.

A camera inventory with, per camera: resolution, frame rate, and whether the stream is reachable over the network the analytics will run on.
A view assessment: is the area of interest large enough in frame, and is it in frame all the time or only when the camera happens to be pointed there?
A lighting note per location, including night and, outdoors, direct low sun — the two conditions that most often separate a camera that works from one that does not.
A network path: where the analytics runs relative to the cameras, and whether the site link can carry what that implies.
A list of the cameras that are not suitable, with the reason, so the conversation about adding cameras is a costed decision rather than a mid-project discovery.

03The scope decision

What is being detected, and what is deliberately not

Scope creep in a detection project is not adding features, it is adding conditions the system must never miss. Each one is a promise.

A written list of the detections in scope, in priority order, with the first one named as the one that must work.
An explicit not-in-scope list. This is the more useful half and the one usually missing.
For each detection: what a true positive looks like, and — harder and more important — what a false positive looks like and how annoying it is.
A decision on whether the system detects a state (a person is not wearing a hard hat) or an event (a person entered a restricted zone). They are different problems with different failure modes.
A statement of what happens on the first day the system is wrong in public, which it will be, and who explains it.

04The alerting decision

Who receives an alert, and what are they expected to do

Recipient	What they can actually do	What that requires
Control room operator	Act within minutes, on a screen they are already watching	The alert has to fit into what they already do, or it becomes a second screen nobody looks at
Site supervisor on a phone	Act at the location, but not instantly	Alerts filtered hard enough that the phone is worth carrying — one nuisance alert a shift is enough to end this
HSE, as a daily or weekly summary	See patterns, not incidents: which area, which shift, which recurring condition	Aggregation and trend, not a feed. This is often the highest-value output and the one designed last
Nobody, initially	Nothing — and that is a legitimate first phase	An agreed period where the system records and is reviewed, so its behaviour is known before anyone is asked to trust it

Before deployment

People and privacy

A safety system that workers experience as surveillance stops being a safety system

The technical question — what is stored, for how long, and who can see it — has to be answered before deployment, and it is the easier half.

The harder half is that the same system can be deployed as "we are watching for hazards" or as "we are watching you", and the difference is almost entirely in how it is introduced and what happens with the output. A deployment where detections are used to coach and to fix recurring conditions is one workers will point out gaps in. One where the first visible use is a disciplinary case is one that gets worked around within a fortnight, and a worked-around safety system is worse than none, because it reports that everything is fine.

05The last decision, made first

How the deployment will be judged

Agreed before it starts, in writing. A pilot with no pass mark is a pilot that can always be argued either way.

The one detection that must work, and what "working" means for it in numbers the site can check itself.
The review period, and who reviews. A fortnight is usually too short to see a pattern and a quarter is usually too long to keep attention.
The nuisance-alert threshold at which the deployment is judged to be failing, agreed in advance so it is not renegotiated afterwards.
What is measured before deployment, so there is a before. This is the step most often skipped and the one that makes the after meaningless.
What happens at the end: expand, adjust, or stop. Naming "stop" as an acceptable outcome is what makes the other two credible.

06Sequencing

A rollout that can be stopped at each step

Each phase produces a decision rather than a milestone, and each one is a place the project can legitimately end without having failed.

undefined
undefined
undefined
undefined
undefined

Phases, not a schedule. The review period in phase three is the one that cannot be compressed.

07Buyer questions

What buyers ask us

Direct answers to the questions that come up in real evaluations. Anything missing, ask us at the bottom of the page.

Do we need new cameras?

Usually not all of them, and usually some. Existing CCTV was generally positioned to record what happened after an incident, which is a different requirement from detecting a condition during one — so a camera that gives a fine overview of a yard may show a person too small in frame to assess. The camera inventory above is what turns this from an assumption into a costed list.

How accurate is AI video detection?

Accuracy is a property of a specific detection on a specific camera in specific conditions, not of a product, and any figure quoted without those three is not telling you anything you can use. The useful question during an evaluation is narrower: on these cameras, for this detection, how often does it miss and how often does it cry wolf — measured on the site, over an agreed period.

How long does a deployment take?

The install is short. What sets the timeline is the camera assessment, the scope agreement, and the review period — and the review period cannot be compressed, because it is measuring behaviour over time. A deployment planned without a review period has not been planned, it has been installed.

What happens to the video?

That is a decision to make explicitly rather than inherit from the analytics product: what is retained, for how long, who can view it, and whether detections are stored as clips, as stills, or as records with no image at all. Each is defensible; the failure is not choosing, and discovering the answer during an incident.

From the insights library

Work the checklist against your site

Describe the site, the cameras you have and the one detection that matters most. The camera assessment is where this conversation usefully starts.

Or directly

+966-11-2180999 info@interkey.com.sa

Tawuniya Towers, North Tower, 7th Floor, King Fahad Highway, Olaya, P.O. Box 56835, Riyadh 11564, Saudi Arabia

or Industrial safety monitoring

Published by Interkey. Last updated 2026-09-07. Interkey is registered in Riyadh, Saudi Arabia under commercial registration 1010156897.