CCTV vs AI Cameras on the Factory Floor: What Changes

CCTV records your floor so you can look back after something goes wrong. AI cameras watch the same feed live and interrupt you while it's still going wrong — a worker in the press bay with no helmet, right now, not on tomorrow's playback. The cameras are usually identical. The difference is software that reads the feed and either files it or calls you.

But framing it as "CCTV or AI" is the wrong decision. For a mid-size Indian plant, most of the payoff from analytics comes from two or three zones. The real question is not which system to buy — it's which cameras to make smart, and where. That is a triage decision, and almost nobody selling you cameras will frame it that way.

Key takeaways

• CCTV and AI usually run on the same cameras and feeds — CCTV is the recorder, AI is a software layer on top.
• CCTV gives you evidence after an incident; AI gives you an alert before it becomes one.
• Don't rip-and-replace, and don't make every camera smart. Start with two zones where a missed event costs real money — then expand.
• In India, STQC/BIS Essential Requirements and the DPDP Act 2023 apply to both — and post-1-April-2026 they change which cameras you can even buy.

Key facts (India, 2026)

• From 1 April 2026, CCTV cameras that don't meet MeitY's Essential Requirements may not be sold in India (MeitY Office Memorandum, 16 January 2026).
• The Essential Requirements regime was notified via Gazette S.O. 1652(E), dated 9 April 2024, amending the Compulsory Registration Order (effective 9 October 2024).
• India's DPDP Act, 2023 caps penalties for failure to take reasonable security safeguards at ₹250 crore.

What passive CCTV actually does

Traditional CCTV is a recorder. Cameras feed a DVR or NVR, footage sits on a disk for a few weeks, and a human looks when there's a reason to. That is genuinely useful: it settles disputes, documents accidents and contractor work, deters casual theft, and gives management a factual record after the fact.

What it doesn't do is notice. Nobody watches 16 or 32 live tiles all shift. An oft-quoted trade-press figure (attributed to Ainsworth, 2002) puts operator misses near 45% of activity after 12 minutes and 95% after 22 — treat it as folklore, not audited data, but the direction is right: attention collapses fast, and faster the more feeds you add.

We felt the passive gap on our own pilot floor — a ~40 m² machine shop we wired with three 2 MP PoE cameras and watched for weeks. On 13 July, power to the switch and cameras dropped at 11:08 and came back at 12:58 — nearly two hours dark — and nobody noticed in real time. We found it afterward in the power logs. That is what "reactive" means: the value shows up on rewind, if you go looking.

What AI cameras add

AI analytics interpret the video in real time and pull a human in only when a defined event happens. On a factory floor the useful signals are concrete: PPE and hazard-zone checks near presses and forklifts, blocked or open fire exits, people and vehicle counts at the gate, and machine idle-vs-running time.

The gap that convinced us wasn't a stat. On the same pilot floor a gas cylinder sat in a walkway, right next to the electrical panel, for an entire shift. On plain CCTV that's invisible until someone rewinds the day. That is the whole point: a static hazard nobody is watching, live, is exactly what a rule ("cylinder in walkway → alert") catches in the first minute.

Rank the use cases by rupee, not by brochure. The one people demo first — PPE detection — is often the hardest to turn into money. What pays first for a mid-size Indian plant is usually the pair where the loss is measurable: after-hours intrusion (someone in a restricted bay at 2 a.m.) and machine idle / unplanned downtime (a line stopped for 40 minutes nobody flagged). Prove those two, then expand.

Side-by-side: CCTV vs AI cameras

The honest limits no reseller prints

"Runs on your existing cameras" is often false economy. Retrofit vendors pitch this hard, and sometimes it's true. But analytics need resolution, frame rate, a clean line of sight to the hazard, and decent light. A sub-2 MP analog feed at a bad angle in a dim bay will not reliably read a helmet. And after 1 April 2026, many older cameras are non-ER-compliant anyway — so "keep everything you have" quietly runs into the certification wall.

The first month is noisy. Out-of-the-box analytics frequently increase nuisance alerts for the first two to four weeks, until the rules are tuned to your floor — your angles, your traffic lanes, your lighting. The "80–90% fewer false alarms" you'll see quoted is entirely a function of that tuning, not something that ships in the box. Anyone who promises a quiet system on day one hasn't run one.

Certification is a vendor-selection risk, not a checkbox. ER certification is granted per camera model, and it isn't cheap or fast. A small AI-camera startup may never put its models through it — which means an AI camera you buy today could quietly become unsellable and unsupported after the deadline. Ask which specific models are certified before you sign, not whether the company "is compliant."

The cost and effort, in plain terms

On hardware the two are close, because the cameras are often the same. India price signals (indicative, mid-2026, aggregated retail — verify against a live quote):

Note the UPS line. On our pilot the laptop survived that two-hour outage on its own battery — but the PoE switch, router and cameras had none, so they went dark and recording stopped cold. If cameras matter, they need a small UPS; that's true whether or not analytics run on them.

Where AI adds cost is the analytics layer: smarter cameras or edge boxes, or a server plus a licence, often billed per camera per month. In our own pilot, running the reasoning that produces one plain-language summary costs roughly $0.05 per message — cheap enough that the cost question is labour and loss, not compute. Where AI saves is fewer missed safety events, less manual counting, less time scrubbing footage. Date and pressure-test any vendor ROI claim before you sign — most are modelled, not measured.

India: the rules apply to both — and they're tightening

Two India-specific points a plant head should not skip:

1. BIS registration and STQC testing are two things, not one. Under MeitY's Essential Requirements — notified via Gazette S.O. 1652(E), 9 April 2024, amending the Compulsory Registration Order — cameras need BIS CRS registration (an R-number, per model) and the surveillance requirement is backed by STQC security testing of the software/VMS layer. The government has withdrawn the relaxation that let older non-ER stock keep selling: per a MeitY Office Memorandum dated 16 January 2026, from 1 April 2026 non-ER-compliant CCTV cameras may not be sold in India (industry coverage of the MeitY memorandum). Because STQC's scope is the software layer, the analytics in an "AI camera" sit squarely in the tighter part of the net — not a lighter one. Ask for the BIS R-number and STQC status by model.

2. DPDP Act, 2023. Footage of identifiable people is personal data. That means posted notice, a documented purpose, a retention limit, and reasonable security — with penalties for failing at reasonable safeguards running up to ₹250 crore (Act No. 22 of 2023; text on MeitY). The Act is enacted; the operational specifics (notice, retention, consent) firm up under the DPDP Rules, and some employee-monitoring may fall under the Act's "legitimate uses" rather than requiring consent. AI doesn't loosen any of this — storing detections tied to individuals raises the bar. Set notice, retention and access policy before you scale either system.

Where Mama fits

The hard part of moving to AI isn't the software — it's knowing which cameras, watching which zones, would catch the events you care about. That's the triage gap Mama is built for. You record a short phone walkthrough of the floor; it returns a plain-language read of the space plus a placement plan — how many cameras, where, ceiling or wall — and the safety and efficiency signals worth an alert. You add analytics where it pays first, not everywhere at once, and you get the plan in a day without waiting on a site survey.

FAQ

Do AI cameras replace my existing CCTV? Usually no — analytics run on feeds from cameras you already have. But "keep everything" isn't automatic: sub-2 MP, poorly-lit, or badly-angled feeds may not support reliable analytics, and older cameras may be non-ER-compliant after 1 April 2026. Keep what qualifies; add analytics where live alerts matter.

Is AI video analytics worth it for a mid-size factory? It's worth it where a missed event is expensive — a safety incident, theft, or unplanned downtime. Start with two zones where the loss is measurable (after-hours intrusion and machine idle usually pay before PPE), prove it, then expand.

Do AI cameras need the same STQC/BIS certification in India? Yes, and the software layer faces the sharper end. Essential Requirements apply to CCTV cameras regardless of analytics, and STQC testing covers the software/VMS scope. With certified-only sales expected from 1 April 2026, ask each supplier for the BIS R-number and STQC status by model.

Does the DPDP Act change how I use factory cameras? It applies to both. Footage of identifiable people is personal data: you need notice, a stated purpose, a retention limit and reasonable security, with safeguard-failure penalties up to ₹250 crore. Operational duties firm up under the DPDP Rules; some employee monitoring may fall under "legitimate uses." Set policy before scaling.

Can I run AI analytics without a fast internet connection? Often yes — analytics can run on the camera, an edge box, or an on-site server, sending only alerts out. That suits Indian plants with patchy connectivity; confirm with your vendor whether processing is on-prem or cloud, as it affects both bandwidth and DPDP data-handling.