AI CCTV Cameras for a Factory: Detection, Cost and Staff Consent

AI CCTV cameras can flag missing protective equipment, people in restricted zones, forklift and pedestrian near-misses and process steps. Software alone lists from $10 per camera a month on one cloud video-AI service, but the factory vendor I checked publishes no prices, so get itemised quotes. Staff need clear notice and usually an impact assessment; consent alone is weak.

That $10 is a build-it-yourself developer service, before cameras, set-up and alert wiring, and the rules on staff monitoring differ by place, so take a qualified adviser's view before switching on. Most of the trouble in these projects is not the camera. It is a mismatch between what a vendor demo shows and what a dim, dusty, crowded production line looks like, and a rollout that surprises the people being filmed. Both are cheap to fix before purchase and expensive afterwards.

Follow me on Instagram@sagnikteaches

What AI cameras can realistically detect

Video analytics software looks at a camera stream and raises an alert when something matches a rule. The capabilities below are the common ones on factory-focused vendor pages. One factory vendor page I reviewed in November 2026 lists forklift and overhead-crane tracking, restricted-area monitoring, parts counting, conveyor monitoring, SOP checks for assembly and packaging, operator performance, perimeter and gate monitoring, and warehouse area and rack monitoring. Protective equipment appears in its headline claims and a customer quote rather than as a listed module, it offers a 14-day pilot on existing cameras, and it states no prices. Expect that pattern elsewhere too: broad claims, little published detail. Ask for proof on your own footage.

Connect on LinkedInSagnik Bhattacharya

Protective equipment checks

The system looks for hard hats, high-visibility vests, gloves, glasses or masks on people in view. It works best where the camera sees the full body at reasonable size, lighting is steady and the equipment is a contrasting colour. It struggles with partial views, people carrying loads, reflective glare and equipment that blends with the background. Ask what the false alert rate is on footage from your own floor, not the vendor's.

Subscribe on YouTube@sagnikteaches

Zone and exclusion alerts

You draw a virtual boundary round a hazard, such as a press, robot cell or loading edge, and the software alerts when a person crosses it. This is generally more reliable than equipment detection because it needs less visual detail: it is "is a person here". It suits places where a light curtain or fence is impractical.

Vehicle and pedestrian proximity

Alerts when a forklift and a person come within a set distance or when a vehicle enters a walkway. Accuracy depends on camera angle and calibration. Treat this as an extra warning, not a safety system. Physical controls and traffic management come first.

Process and counting tasks

Counting parts, checking that an assembly step happened, and watching conveyor jams are quality and output uses rather than safety. They can be valuable and carry less staff concern if pointed at machines rather than people, which is a design choice worth making.

What it should not be asked to do

Do not use it as the only means of detecting a life-safety hazard. Do not use it to score individual workers' productivity without a separate, careful decision, even though at least one vendor lists an operator-performance module. Do not enable face recognition for a safety purpose. A camera that notices a missing hard hat does not need to know whose head it is.

Reusing the cameras you already have

This is where the largest saving sits, and where most disappointment arrives. Some analytics products run on existing IP cameras and advertise pilots on "existing cameras". Whether that works for you is a survey question. Check these five points:

1. Stream access. Can the cameras produce a video stream the software can read (commonly RTSP, a standard streaming protocol)? Older analogue systems often need an encoder.
2. Resolution and frame rate. Enough pixels on a person at the working distance. A camera 25 metres from a doorway may show a vest as a smudge.
3. Angle. Ceiling-mounted cameras looking straight down see the tops of heads, which makes equipment checks hard. A mounting angle that shows the torso works better.
4. Lighting. Dim corners, bright windows and welding flash confuse detection. Infrared night modes can reduce colour information.
5. Recorder limits. A recorder may not be able to serve extra streams while still recording.

A sensible approach is to choose two or three representative cameras, send a day's footage to the vendor, and ask for results on it. If their model works on your worst camera, reuse is plausible. If it only works on the new cameras in their demo, price new cameras in.

On-site or cloud processing

This choice changes the cost shape and, more importantly, the privacy position.

The factory vendor page I reviewed describes on-premises installation, yet its set-up steps also include a 48-hour footage feed for remote custom model training. So even with on-site processing, some footage can leave the building during set-up. Whichever you choose, ask: which frames or clips leave the building, who can view them, how long are they kept, and can the vendor use them to improve its models?

The cost lines to ask for

The turnkey factory vendor I checked publishes no prices, and quotes vary with camera count and scope. One published anchor exists for the software layer alone: Google Cloud's video AI service (its pricing page is currently titled Agent Platform Vision) lists PPE detection at $0.10 a minute pay-as-you-go, or $10 per camera stream a month on its monthly model, plus a data charge of under a cent per gigabyte streamed. Its occupancy analytics suite is listed at $20 per stream a month. That is a developer service: someone still has to connect the cameras, build the alerts and maintain it.

The billing model matters more than the rate. Billed by the minute, one camera watching a 10-hour shift for 22 days is 13,200 minutes, or $1,320 a month; on the monthly model the same stream lists at $10. Ask any vendor or integrator which model their quote is built on. Beyond the software, here are the lines to make every vendor quote separately, so quotes compare. The amounts are for you to fill in.

1. Site survey. Often free for a pilot, sometimes paid. Ask.
2. Cameras and mounting. Zero if existing cameras pass the survey; otherwise per camera, plus cabling and power.
3. Processing hardware. A local computer or edge device if on site. Ask the specification and the replacement cycle.
4. Software licence. Per camera, per module (equipment, zones, counting), per site, or flat. Per-module pricing multiplies fast.
5. Set-up and calibration. Drawing zones, tuning alerts, installing. Ask how many tuning visits are included.
6. Alert delivery. Integration with a radio, lights, a messaging app or a dashboard.
7. Storage. Footage retention, which you should keep short.
8. Support and updates. After year one, what is charged?
9. Your time. Reviewing alerts for several weeks, staff briefings, the assessment paperwork.

A worked comparison for a 40-person plant

Say a 40-person metal fabrication plant wants two alerts: people entering the press-brake zone while the machine is live, and missing high-visibility vests at the loading dock. Illustrative scoping, with no prices invented:

• Cameras needed: 3 at the press area, 2 at the dock. All five existing cameras are IP, three pass the survey, two need replacing.
• Option A, on site: one edge computer, per-camera licence for two modules, set-up visit and a tuning visit.
• Option B, cloud: no edge computer, higher monthly licence, bandwidth upgrade, data agreement.
• Software floor, for scale: if an integrator built the dock vest check on the cloud service above, its two streams would list at about $20 a month ($720 over three years) before any integration labour, which you need quoted separately.

Ask both vendors to price the same scope in the same table, with a first-year and a three-year total. The cheaper first-year option often loses over three years if the licence is recurring and per camera. A pilot on two cameras for 14 to 30 days, with a clear success test, limits the risk before you commit to all five.

Staff notice, consent and data protection

This section is practical guidance, not legal advice. The rules differ by place and by the type of data, so take advice from a qualified adviser before you switch anything on. The steps below are the ones I would expect any adviser to want to see.

1. Define the purpose narrowly. "Reduce incidents at the press and the dock" is a purpose. "Monitor staff" is not. A narrow purpose also keeps you from drifting into uses staff never agreed to.
2. Check you need AI at all. Would a physical guard, interlock or light curtain do the job with no recording of people? Write down why not.
3. Carry out a documented assessment. If you have staff in the EU, the GDPR requires an impact assessment for processing likely to be high-risk, and the official EU guidance on video devices says it is reasonable to assume many video-surveillance uses will need one. It also warns that tools turning ordinary cameras into smart cameras raise the risk of footage being reused for things like performance monitoring. Record what you collect, why, the risks to staff and how you reduce them.
4. Choose your lawful basis carefully. The same EU guidance on consent gives workplace camera monitoring as an example where an employee is unlikely to be able to refuse freely, so consent is a weak basis here. Safety obligations or legitimate interests, properly assessed, are more common bases. Your adviser will confirm.
5. Tell staff in plain language before go-live. What is detected, where cameras are, what happens to alerts, who sees footage, how long it is kept, and what it will not be used for.
6. Consult representatives. If you have staff representatives or a union, involve them early. Where the law requires consultation, skipping it can invalidate the whole project.
7. Limit access and retention. A short retention period, named viewers and an access log.
8. Agree what alerts can be used for. Safety coaching, yes. Disciplinary action from one alert, no, unless written policy and a human review say otherwise.
9. Review after 90 days. Did it reduce incidents? Are staff comfortable? Is every alert type still needed?

Staff trust is not a soft extra. A system staff resent gets covered, blocked or ignored. Asking first costs little: how to survey your staff before an AI rollout gives questions you can adapt, and how to talk to staff who fear AI covers the conversation itself.

A staff notice you can adapt

NOTICE TO ALL STAFF: SAFETY CAMERAS WITH ALERTS (illustrative wording)
From [date] we are trialling software on five existing cameras at the
press area and the loading dock. It raises an alert when (1) a person
enters the marked press-brake zone while the machine is live, and (2)
a person at the dock is not wearing a high-visibility vest.
The software does not identify who anyone is.
Alerts go to the shift supervisor to prompt a safety conversation.
They will not be used on their own for disciplinary action.
Footage is kept for [number] days and viewed only by [named roles].
The trial lasts 30 days, then we will review it with staff
representatives. Questions or concerns: [name and contact].

Replace every bracketed item before use, and have your adviser read the final wording. Notice that it names the purpose, the limits and the review date. Those three do most of the work of building trust.

Reading a vendor demo with a sceptical eye

A demo shows the best footage the vendor has. Before you believe a number, ask the questions below and write the answers down, so you can compare vendors on the same terms later.

• Whose footage was this? If it is a customer's floor, ask how similar to yours it is: ceiling height, lighting, crowding, clothing colours.
• What was the false alert rate, and over how long? "Ninety-five per cent accurate" says nothing about how many alerts a supervisor must dismiss per shift. Ask for alerts per camera per day.
• What happens at night, in glare or in steam? Ask to see those conditions.
• Who tunes it, and how often? A system tuned once at install and never again drifts as layouts, uniforms and lighting change.
• What does it do when it is unsure? A good system logs low-confidence events for review instead of raising alarms.

A fair test is to give two vendors the same two hours of footage from your worst camera and ask for the alert log. You will learn more from that than from any slide.

A filled-in alert rule, so the pilot has something to test

Vague rules produce noisy alerts. Write each one in a fixed format. Here are two for the fabrication plant; the thresholds are illustrative and need tuning on your floor.

RULE 1: PRESS-BRAKE ZONE
Camera: PB-2 (angle: side-on, 3 m height)
Alert when: a person is inside the marked yellow zone for more than
2 seconds while the machine status is LIVE
Send to: shift supervisor's handset, no sound
Not an alert: a person outside the line, or inside it while the
machine is locked off
Review: every alert tagged true / false / unclear, daily for 14 days

RULE 2: DOCK HIGH-VISIBILITY VEST
Camera: DK-1 (angle: front, 4 m height)
Alert when: a person is in the dock area for more than 10 seconds and
no high-visibility colour is detected on the upper body
Send to: log only for the first week, then supervisor
Not an alert: visitors with a visitor vest on, drivers in the cab
Review: tune the 10-second threshold after 100 events

Notice that Rule 1 ties the alert to machine status. A person standing near a locked-off machine is not a hazard, and an alert for it teaches everyone to ignore alerts. If the software cannot read machine status, ask whether a simple signal from the machine controller can be wired to it.

When vest alerts swamped the supervisors

Illustrative: a plant switched on vest detection at the loading dock across all shifts on day one. The first week produced over 200 alerts a day. Most were drivers in their cabs, visitors in coloured jackets and one stack of orange pallets the model read as a person in a vest. Supervisors stopped looking after day three, and one asked for the cameras to be unplugged. The recovery was to go back to quiet mode, add exclusions for the cab area and the pallet store, set a minimum time in view, and relaunch with a staff briefing about what had gone wrong. The lesson is the one in the pilot plan: quiet mode first, tag every alert, and fix the noise before supervisors see any of it.

Who looks at the alerts, and what they do next

An alert nobody owns is a cost with no benefit. Decide before go-live who receives each type, and what they are expected to do within what time. For the press-brake zone, the supervisor walks over and has a short conversation, which is logged in one line. For the vest rule, a weekly summary goes to the safety lead to spot patterns by shift or area, not by person. For out-of-hours perimeter alerts, someone on call needs a clear instruction: check the live view, then call a named contact. Write these as one-line procedures and put them in the same file as the staff notice. When an incident does occur, you will want a record showing that alerts led to action and that the action was proportionate.

Finally, set a trigger to stop. If the pilot misses its success test twice running, or staff concern is serious and unresolved, pause the rollout. A system you switched off deliberately costs far less than one you quietly stopped trusting.

Running a 30-day pilot that tells you the truth

1. Week 0. Choose the two alerts. Write the success test: for example, at least 8 in 10 zone alerts are genuine, and no more than three false alerts per camera per shift.
2. Week 1. Install on two or three cameras. Run in quiet mode: alerts go to a log, not to the floor. This is the same idea as piloting AI in shadow mode.
3. Week 2. Review the log with a supervisor each day. Tag each alert true, false or unclear. Adjust zones and sensitivity.
4. Week 3. Switch alerts on for supervisors only. Keep tagging.
5. Week 4. Compare against the success test, count incidents and near misses against the same period before, and ask staff and supervisors what annoyed them.

At the end, decide: expand, adjust or stop. Stopping on evidence is a good result, because it costs one pilot rather than a five-year licence.

Four more situations, briefly

A print shop with a guillotine: use a physical guard instead

One machine, one hazard. A fixed guard or two-hand control is cheaper and more reliable than analytics. Verdict: no AI cameras.

A warehouse with forklifts and pickers: zone and proximity alerts

Mixed traffic is the classic case. Start with a single aisle junction, calibrate carefully and treat alerts as coaching prompts. Add physical measures such as marked walkways and mirrors in parallel.

A food packing line: counting and SOP checks

Camera views of the packing step can check that a lid or label is present. Point cameras at product and hands rather than faces, which also eases staff concerns.

A site gate: perimeter alerts out of hours

Alerts when someone enters a fenced area at night. Staff concern is low because nobody is working there, and false alerts from animals and weather are the main tuning problem.

Measuring whether the cameras changed anything

Four measures after the pilot: genuine alerts as a share of all alerts, incidents and near misses against the baseline, supervisor hours spent reviewing alerts, and a short staff pulse check. If more than a third of alerts are false after tuning, people will stop reading them. If nobody can say what changed on the floor because of an alert, the system is a recording device with extra steps. Choose vendors who let you export the alert log so you can measure for yourself.

For the wider choice of vendor and contract, questions to ask an AI vendor before you sign applies directly, and what to check in an AI tool's privacy policy and terms covers the data terms. If you would like to map your own plant against these steps, my AI implementation consultation covers finding the jobs where AI pays off and where it does not.

Questions factory owners ask about AI cameras

Can AI run on the cameras I already have?

Often yes, if the cameras give a clear, steady image at enough resolution and you can pull a video stream from them. Some vendors advertise a pilot on existing cameras. A site survey decides it, because lighting, angle and distance matter as much as the camera model.

Is on-site processing better than cloud?

On-site keeps footage inside your building and works if the internet drops, but needs a local computer. Cloud is easier to set up and update, but sends video or frames off site. For footage showing identifiable staff, many owners prefer on-site or ask for frames only to leave.

Do I need staff consent to monitor them with AI?

It depends on where your staff work and on the lawful basis you rely on. Notice is almost always required, consent alone is often a weak basis at work, and a documented impact assessment is commonly expected. Take advice from a qualified adviser before switching on.

Should the system identify individuals by face?

For safety use cases, no. Detecting that someone is not wearing a hard hat does not require knowing who they are. Face recognition raises the legal bar sharply and staff trust problems. Ask vendors to switch identification off and show you it is off.

Further reads

• How to Write a Health and Safety Policy With AI and What to Check — Write the safety policy the cameras are meant to support.
• AI Camera Software for CCTV Installers: Analytics You Can Sell Clients — How installers compare camera analytics products.
• Can AI Plan CCTV Camera Placement From a Floor Plan? — Plan camera positions before you buy anything.
• Will AI Replace My Employees? An Honest Answer for Small Firms — An honest answer for staff worried about monitoring.
• How to Get Your Staff to Actually Use AI Tools — Bring the floor team along with a new system.
• What If Your AI Vendor Shuts Down? Checks Before You Commit — Checks to make before you commit to a camera vendor.
• AI Footfall Cameras for a Retail Store: What They Count and Privacy — What a footfall camera really counts, how to turn the numbers into a better rota, and the privacy checks and signage to settle before you install one.
• CCTV Installation Checklist: Commissioning, Signage and Handover With AI — The commissioning tests, settings, retention sums, signs and handover pages for a CCTV job, plus prompts that draft the customer guide from your sheet.
• CCTV GDPR: Can AI Draft Privacy Notices and Signage for Clients? — How a CCTV installer can use AI to draft sign and notice wording for a client, with an adviser check list and recorder settings that must match.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: SeeWise product page (modules, on-premises installation, 14-day pilot on existing cameras, no published prices); Google Cloud Agent Platform Vision pricing page; EU data-protection guidelines 3/2019 on video devices and 05/2020 on consent. All checked November 2026.