Strategy
Asking is table stakes. Trusting the answer is not.
Natural-language search over camera footage shipped in 2024 and is now expected. Verkada, Spot AI, Coram, Protex AI and Intenseye each let a user ask, and NVIDIA publishes a reference blueprint for visual Q&A over video. The interface problem is solved.
The thesis
Where the risk now sits
The unsolved problem is one step later. A vision-language model will describe a hazard it did not see as fluently as one it did. An EHS director who opens an investigation, stops a line or reports a near-miss on the strength of that answer owns the consequence, and today the only check is a person scrubbing the footage again.
FactoryLens is the trust layer between the model and the decision: a typed event ontology the model is held to, answers that must cite their evidence or abstain, a deterministic gate that holds anything contradicting what is known, and a question log a works council can audit.
Why now
Three shifts that make trust the buying question
Models can read the floor
General vision-language models now answer free-form questions about workplace stills well enough to be useful, and video foundation models such as TwelveLabs Marengo 3.0 reached Amazon Bedrock in December 2025. Capability is no longer the bottleneck; verification is.
Every vendor ships a search box
Verkada launched natural-language video search in May 2024. Spot AI, Coram, Protex and Intenseye followed with search or copilots. Buyers now expect to ask, and the next question they will face is which answers they can defend.
Regulation raises the evidence bar
The EU AI Act has banned emotion recognition at work since 2 February 2025. AI used to monitor or evaluate workers is listed as high-risk, and the Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026) moved those obligations from August 2026 to 2 December 2027. Plants deploying video AI now need an audit trail before that date, not after it.
Positioning
A layer, not another detector
FactoryLens does not ask a plant to replace its cameras, detectors or models. It adds the layer where an answer turns into an action.
- 4Trust: evidence gate, ontology, question logToday a person re-watching the footage. FactoryLens makes this layer software.
- 3Ask: search and copilotsNow offered by most of the above
- 2Perception and detectionVoxel, Intenseye, Protex AI, Spot AI; models from TwelveLabs; NVIDIA's blueprint for builders
- 1Cameras and video managementExisting RTSP/ONVIF estates; cloud platforms such as Verkada
Competitive landscape
Who already lets you ask, from their own published material
Each card states what the company publishes, with the source. The right-hand line is how FactoryLens would relate to it, not a claim about what it lacks.
Verkada
Cloud physical securityAI-Powered Search: natural-language search across camera footage (beta, May 2024), with query moderation and audit logs.
FactoryLens: Security-first and horizontal. An estate like this is where a trust layer would sit, adding safety semantics and an answer gate.
Spot AI
Video AI for operationsNatural-language video search and pre-built agents for forklift near-misses, missing PPE and unattended workstations.
FactoryLens: Closest in scope. The pilot question is whether each answer is checked against evidence before it reaches a supervisor.
Coram AI
Video AINatural-language video search; $13.8M Series A (January 2025) aimed at security, workplace safety and efficiency.
FactoryLens: Search-led. The same gate and ontology would apply to its answers.
Protex AI
Safety visionSafety detection with on-premises, anonymised processing, plus Copilot, a generative-AI tool for natural-language questions over safety data.
FactoryLens: Answers questions about the events it has detected. A neutral answer gate complements rather than replaces it.
Intenseye
Safety visionReal-time hazard detection plus Chief, a multimodal AI assistant that analyses safety data, surfaces trends and recommends actions.
FactoryLens: The benchmark for detection breadth. FactoryLens competes on how an answer is verified, not on the count of detectors.
Voxel
Safety visionReal-time hazard alerts on existing cameras; $44M Series B (June 2025). Reports a 91% reduction in recordable injuries.
FactoryLens: Sets the outcome bar any safety purchase is judged against.
NVIDIA VSS blueprint
Reference architectureA reference build for video search, summarisation, visual Q&A and verified alerts, using vision-language models and retrieval.
FactoryLens: The build path for a manufacturer with its own AI team. The ontology, gate and labelled evaluation set are the parts a blueprint leaves to the builder.
TwelveLabs
Video foundation modelsMarengo 3.0 and Pegasus video-understanding models, on Amazon Bedrock since December 2025.
FactoryLens: A likely supplier, not a competitor: perception is bought, and the model layer stays swappable.
Market
The budget already exists
Computer vision in manufacturing
~$7.9B (2026)to ~$16.2B by 2032, 12.7% CAGR
Third-party estimate, for sizing only.
Where the capital is going
- Voxel raised a $44M Series B in June 2025, $61M in total, for AI safety on existing cameras. Source
- Spot AI raised $31M in October 2024 for video AI across operations and safety. Source
- Coram AI raised a $13.8M Series A in January 2025, aimed at video security, workplace safety and efficiency. Source
- Voxel reports a 91% reduction in recordable injuries and one site saving over $2.2M in direct costs over two years. Vendor-reported, not independently verified, and not FactoryLens data. Source
The budget already exists. Plants pay for cameras and, increasingly, for detection. FactoryLens does not ask for a new camera budget; it makes the answers those systems produce defensible.
The buyer who signs is the EHS leader; the people who can veto are the CISO, the data-protection officer and the works council. A product that arrives with masking, a question log and a published accuracy method removes three objections before the first meeting.
Product principles
Make every answer a camera system gives about the plant floor something a safety leader can defend: cited, checked against what is known, logged, or honestly withheld.
North Star
Cleared answers acted on, per site per week
It rises only when people ask, the gate clears the answer, and someone acts on it. Its counter-metric is the false-hold rate: held answers a reviewer later finds correct. A gate that holds everything is as useless as one that holds nothing.
Grounded or silent
Every answer cites the stills it relies on. If the footage cannot support an answer, the product says so instead of guessing.
The gate decides, not the model
Deterministic checks grade every answer. One blocking failure holds it, whatever the model's confidence or the average score.
Privacy before perception
Faces are masked before a frame leaves the device. In production, raw video never leaves the site.
Safety lands, operations expands
The same event layer answers the EHS question that opens the account and the throughput question that renews it.
- Not a camera or hardware vendor: it rides existing RTSP/ONVIF estates.
- Not a safety-rated machine-control system.
- Not worker surveillance: it measures hazards and processes, never an individual's performance or emotions.
Defensibility
Why this can be defended
Ontology with guards
Typed events that say what must not count: a cap is not a hard-hat, a parked forklift is not a near-miss. Encoded domain judgement that can be tested line by line.
Measured trust
A labelled evaluation set per site that grows with every pilot. The gate's hold rate and false-hold rate are numbers a buyer can audit, not adjectives.
Neutral layer
Works over any camera, detector or model, so it can be sold alongside incumbents or embedded by them. The model underneath stays swappable.
Governance by default
Masking before perception, purpose-limited questions and a hash-chained question log: what a works council and a data-protection officer ask for first.
Go to market
Who buys, who can veto, and how it is priced
Multi-site discrete and process manufacturers and 3PL warehouses (automotive, electronics, food and beverage, logistics) with camera estates already installed, an EHS function, and often a recent incident, insurer or regulatory trigger.
EHS / Safety Director
Cares about: Incidents, liability, proving programmes work
“Act on answers you can defend.”
Plant Manager / Operations
Cares about: Throughput, downtime, OEE
“Ask why the line slowed, and see the evidence.”
CISO / OT security
Cares about: Data egress, access, integration risk
“Masked before it moves; raw video stays on site.”
DPO / Works council
Cares about: GDPR, purpose limitation, the AI Act
“Hazards, not people. Every question logged.”
CFO / Plant GM
Cares about: Payback and risk
“Fewer wrong actions, faster investigations, one layer across vendors.”
Pricing
A per-site platform fee for the gate, ontology and question log, plus a per-camera fee for footage indexed. The site fee prices the trust layer; the camera fee scales with the estate.
Anchor on avoided cost, not compute: the investigation hours saved and the actions not taken on wrong answers. Detection vendors already sell against injury outcomes (see Voxel's reported figures), so the trust layer sells on top of that, not instead of it.
Paid pilots priced to clear a plant manager's signing limit, converted to the site fee once the pilot's accuracy bar is met.
First four quarters of selling
- Q1
Two or three design partners on retrospective safety footage. Build each site's evaluation set; publish the gate's numbers to the partner.
- Q2
Convert to paid; switch on operations events; produce two case studies with measured investigation time.
- Q3
One EHS-platform partner and one systems integrator; begin EU outbound built around the AI Act timeline.
- Q4
First multi-plant customer; offer the gate to detection vendors as an embedded layer.
Roadmap
Each release is gated by evidence, not a date
Answers you can defend
Prove that an EHS team will act on gated answers, and that the gate's holds are right. One site, retrospective footage.
- Edge ingestion on one node
- Core safety events: PPE, forklift proximity, restricted zone, spill
- Event record, gate and question log
- Site evaluation set: labelled clips per event type
Exit gate: Per-event precision and false-hold rate agreed before the pilot and met on the site's own clips; one documented case where a held answer prevented a wrong action.
Prove it pays
Add the operations events that justify the renewal, and route answers into existing EHS workflows.
- Idle station, micro-stoppage and cycle events
- Weekly digests with cited evidence
- Read integration with the EHS system (Enablon, Cority, Intelex or similar)
- Human review queue for held answers
Exit gate: Three quantified operational loss sources per site, and a partner commitment to expand from safety to operations.
Platform of record
Deployable across a plant network and through enterprise security review by default.
- Multi-site rollup and cross-plant questions
- SSO/SAML, role-based access, residency and retention
- Chain head anchored in write-once storage
- Exportable AI Act and DPIA pack
Exit gate: One multi-plant customer live on three or more sites with no raw-video egress.
Risks
What could break the plan
Detection vendors build their own answer checks
HighBe the neutral, measurable layer: publish per-event numbers, work across vendors, and offer the gate for embedding.
The gate only knows what is labelled
HighEvery site gets a labelled evaluation set before go-live, and unlabelled answers are marked 'unverified', never 'cleared'.
Works council or DPO blocks deployment
HighMasking, purpose lock and question log from day one; lead with the DPIA pack.
A wrong answer is acted on
CriticalCite or abstain; one blocking failure holds the answer; held answers route to a person.
Vision-model supplier price or quality shifts
MediumModel layer kept swappable; the ontology, gate and evaluation set are model-independent.
Long enterprise sales cycles
MediumPaid pilots under a plant manager's limit; sell through EHS platforms and integrators.
Regulation
Constraints that decide enterprise and EU deals
EU AI Act
Emotion recognition in the workplace has been prohibited since 2 February 2025. AI used to monitor or evaluate workers is high-risk under Annex III; the Digital Omnibus (Regulation (EU) 2026/1744) moved those obligations from 2 August 2026 to 2 December 2027. FactoryLens is scoped to hazards and processes, never an individual's performance, with human review of every held answer.
GDPR and works councils
Video of workers is personal data: it needs a lawful basis, minimisation and purpose limitation. In Germany, monitoring technology is subject to works-council co-determination (Works Constitution Act, §87); in Italy, remote monitoring needs a union agreement or labour-inspectorate authorisation (Workers' Statute, Art. 4).
What that means in the product
Masking by default, no audio, a purpose lock that keeps questions to safety and operations, a question log, configurable retention and residency, and an exportable data-protection impact assessment pack.
Sources
Every external claim on this page
- Verkada, “Introducing Verkada’s AI-Powered Search” (May 2024)
- Spot AI, “Forklift & pedestrian safety with video AI”
- SiliconANGLE, “Spot AI reels in $31M” (Oct 2024)
- Business Wire, “Coram AI raises $13.8 million in Series A funding” (Jan 2025)
- Protex AI, “Protex Copilot”
- Intenseye, “Chief”
- PR Newswire, “Voxel raises $44M in Series B funding” (3 Jun 2025)
- NVIDIA AI Blueprint for video search and summarization (GitHub)
- Amazon, “TwelveLabs launches Marengo 3.0 on Amazon Bedrock” (Dec 2025)
- Research and Markets, “Computer vision in manufacturing” market report
- White & Case, “EU AI Omnibus enters into force, amending the AI Act” (Jul 2026)
- European Commission, “AI Act: regulatory framework” (application timeline)