Reviewed by the AiGreenTools Editorial Team · Last Updated: October 2026
| Product | Protex AI — proactive safety video analytics platform |
| Vendor | Protex AI · founders Dan Hobbs (CEO) & Ciarán O’Mara (CTO) · 60+ staff |
| Founded | Concept 2020 · operations began in Limerick 2021 · now HQ Dublin, with a Boston office |
| Category | ⚠ Video analytics detection layer — NOT an EHS system of record |
| Best for | Vehicle-dense logistics and industrial sites with written safety rules |
| Architecture | ⭐ Existing fixed CCTV · on-site edge appliance required — raw video stays on site; short encrypted event clips and metadata go to the cloud |
| Rules | ⭐ Configurable rule builder — you set zones, thresholds and dwell times |
| Detections | PPE · vehicle–pedestrian · area breaches · housekeeping · behaviour · ergonomics |
| AI Classification | AI Native |
| Analyst | ⭐ Verdantix Smart Innovators: Video Analytics For Safety, 2026 — market-leading functionality (Intenseye and viAct are also named in the same report) |
| Funding | ~$54M · $36M Series B led by Hedosophia with Salesforce Ventures, Feb 2025 |
| User rating | G2: 4.8 / 5 (4 reviews — small sample, read with that in mind) |
| Pricing | Enterprise quote — not published, no free trial |
Until 2026, Nobody Could Tell You How to Compare These Products
Video analytics for workplace safety has been sold for several years without an independent framework for judging it. A buyer comparing two platforms had vendor demonstrations, vendor case studies and vendor claims, and no way to know which capabilities the market considered load-bearing.
That changed with the Verdantix Smart Innovators: Video Analytics For Safety report, published in 2026. It assessed ten video analytics providers — nine specialising in safety risk detection, one in security — against five core capabilities (PPE detection, behavioural safety, unsafe conditions, area controls and vehicle safety) and five technical differentiators: configurable rule-setting and adaptable models, detection accuracy and alert management, EHS-system integration, analytics and insight, and advanced AI features such as vision-language models, agentic AI and predictive safety intelligence.
Protex AI was recognised in that report for market-leading functionality. Intenseye and viAct are also named in the same report — this is a category-wide benchmark, not a single-vendor distinction, and the framework matters here more than any one recognition does, because two of the five technical differentiators — configurable rules and EHS-system integration — are precisely where this product makes its architectural bets.
Who Decides What “Unsafe” Means at Your Site
The two computer vision platforms in this comparison look almost identical until this question, and then diverge completely.
Intenseye ships a library of vendor-defined detections — 38 on its public library page, “50+” on its homepage — and asks you to tune sensitivity. Protex AI ships a configurable rules engine and asks you to write the rules — translating your own site policies into zones, thresholds and dwell times through visual interfaces.
That is a disagreement about who is qualified to define unsafe behaviour in a specific facility, and Protex AI’s position flatters the buyer: nobody at a software vendor knows your aisles, your blind corners and your tolerated exceptions better than your safety lead does.
⚠️ Configurability is a dependency dressed as flexibility, and the interface hides it. Every rule you write is a judgement about what matters — enforced continuously, without exception, for years. If your site safety policies are written, specific and agreed, this is a decisive advantage. If they live in the head of a supervisor who has been there nineteen years, you are about to encode one person’s model of risk and then measure your entire operation against it. Get the policies onto paper before the first configuration workshop, and put someone senior in the room. This is governance, not a settings screen.
Edge Processing — the Architecture That Decides Whether the Project Happens
An edge compute appliance on the local camera network analyses the live feed on site; the vendor says live streams are never stored. Only short encrypted event clips and metadata go to the cloud, and four privacy modes are available: face blurring, full-body blurring, silhouette ghosting and a videoless mode in which video is never recorded or transmitted. Protex AI’s own G2 reviews confirm the mechanism directly: deployment requires installing an on-site server, which means local IT involvement from day one rather than a purely cloud rollout.
Described without that detail it sounds like a technical preference. In practice it determines whether the deployment is possible at all, because a system that watches workers continuously must clear two gates that appear in no feature comparison. IT asks where the factory-floor video goes. Employee representatives ask who is able to watch it.
“Raw video stays in the building, and only short blurred clips leave it” answers both — and videoless mode, where nothing is recorded or transmitted, answers them more strongly still. For a site in Germany, France, the Netherlands or the Nordics, that distinction is worth more than a longer detection library.
⚠️ The on-site server is confirmed — its commercial terms are not. A G2 reviewer explicitly flags the requirement to install on-site server hardware, which needs local IT support to stand up. What is still not established in any source we read is who pays for that hardware: the vendor, you, or whether it runs on compute you already have on site. Cloud-based alternatives such as Intenseye carry no equivalent line item. Ask for this explicitly in the proposal, not after signature.
Protex AI vs Intenseye — the Only Comparison That Matters Here
| Dimension | Protex AI | Intenseye |
|---|---|---|
| Who writes the rules | ⭐ You do — rule builder (zones, thresholds, dwell times) | Vendor — ~50 predefined use cases |
| Where video is processed | ⭐ On-site edge appliance — raw video stays on site; short event clips to cloud | Cloud, private cloud or hybrid on-premise |
| Privacy mechanism | Edge processing + configurable redaction | Irreversible anonymisation at capture |
| EHS system integration | Stated: EHS and WMS integrations | ⚠ G2 reviewers ask for stronger integration |
| Analyst position | ⭐ Verdantix 2026 — market-leading functionality | Also named in the same 2026 report |
| User rating | G2 4.8/5 (4 reviews) | G2 4.6/5 (71 reviews) |
| Funding raised | ~$54M | More than $90M |
| Deciding question | “Are our safety rules written down?” | “Do we want someone to tell us what to look for?” |
| AiGreenTools Score | 68 | 65 |
Neither is a system of record. Both sit alongside Quentic, Ideagen EHS or EHS Insight rather than replacing them — and our best EHS software ranking covers the platforms of record separately. Worth noting for anyone comparing roadmaps: Verdantix has reported that Quentic plans to launch its own video analytics capability in 2026–27, which would put this function inside an EHS platform rather than beside one.
On the rating gap above: Protex AI’s 4.8/5 comes from only four G2 reviews, against 71 for Intenseye. A higher average on a sample that small is not evidence of a better product — it is a reason to read the four reviews yourself rather than cite the average.
The 64% Claim, Examined Rather Than Repeated
Protex AI states an average 64% reduction in risk within three months of deployment, a figure repeated across press coverage of its Series B and attributed to Fortune 500 partners.
Separately, on G2, at least one reviewer representing a single enterprise deployment reports an 82% reduction in at-risk events. That is a materially different number from a materially different kind of source: one customer’s own stated result on their own site, not a vendor-wide average, and it should not be confused with the “up to 80% reduction in incidents” figure that circulates on unrelated aggregator pages with no attribution at all and no stated methodology — that aggregator figure is not published here.
🔍 Two real numbers, two different things, and neither tells you what will happen at your site. The vendor’s 64% is a cross-customer average with no published baseline or measurement window disclosed beyond “three months.” The reviewer’s 82% is a single site’s result, reported by that customer rather than the vendor. Both are more credible than an unattributed aggregator claim, and neither is a prediction. There is also a trap specific to this category worth naming: a fall in detected events after threshold tuning is not the same as a fall in risk, and the two are easy to conflate in a short window. Ask what the denominator is, and ask to speak to a reference customer who can describe the measurement rather than the outcome.
What Implementation Actually Requires
The technical deployment centres on an on-site server for edge inference, connected to your existing fixed cameras — no new capture hardware, but a real local-IT dependency that cloud-only competitors do not carry. G2 reviewers also describe two concrete friction points worth planning around: ergonomics detection producing a high volume of false alerts in at least one deployment, and camera selection at pilot sites with hundreds of camera views becoming time-consuming enough to complicate a broader rollout.
The organisational work sits almost entirely in rule definition, which is the inverse of Intenseye’s documented pattern, where the burden is threshold tuning and initial alert volume.
💡 The precondition is a document, not a system. Before the first configuration session, write down what your site policies actually say — segregation distances, PPE zones, permitted exceptions, who is allowed where. Most sites discover at this point that the rules exist in three incompatible versions: the written procedure, what supervisors enforce, and what actually happens. Reconciling those three is the real project. The software will faithfully encode whichever one you hand it, and it will do so consistently, forever.
Who Should Not Buy Protex AI
Organizations whose safety rules are not written down. This is first because it is the specific failure mode of a configurable platform. You will encode assumptions rather than policy, and then measure against them for years with the authority of a system.
Anyone looking for an EHS system of record. No incident workflow, no audit scheduling, no recordkeeping, no CAPA. Buy Quentic, Ideagen EHS or EHS Insight first, and add detection afterwards.
Sites without dense vehicle–pedestrian traffic. The detection library reflects warehouses, plant floors and ports. A laboratory or a process-industry site whose dominant risks are chemical or thermal will see less of its risk register represented than the marketing suggests.
Organizations without a usable camera estate, or without appetite for an on-site server. The architecture needs cameras that are fixed-position and cover the zones where risk actually occurs, plus local infrastructure and IT support for the edge compute Protex AI’s own reviewers confirm is required.
The Verdict on Protex AI
Protex AI makes two architectural bets that a European industrial buyer will recognise immediately as the right ones. Edge processing keeps raw video on site and sends only short encrypted clips, with a videoless mode for the strictest settings, which answers the IT objection and the works-council objection with one sentence instead of a negotiation. And the configurable rules engine puts the definition of unsafe where the knowledge actually sits, rather than in a vendor’s use-case library. The company’s announcement that Verdantix recognised it for market-leading functionality in the first independent benchmark of this category gives a procurement committee something to check; Intenseye is also mentioned in the same report.
The reservations follow from the same choices. Configurability transfers the hardest work to you, and a visual interface makes that work look easier than it is — an organization without written safety policies will encode one person’s judgement and measure against it indefinitely. The on-site server that edge processing requires is now confirmed by Protex AI’s own users, though who pays for it still is not. The outcome claims are real but should be read as what they are: a vendor average and a single reviewer’s result, not a guarantee. And the G2 rating, while high, rests on only four reviews — too few to treat as a market verdict.
Between the two computer vision platforms here, the choice is unusually clean and does not depend on feature counts. If your site rules are documented and your raw video must stay in the building, this is the one. If you would rather someone told you what to look for, Intenseye is. Neither replaces the EHS platform you already have, and any vendor suggesting otherwise is selling you a measurement system as a filing cabinet.
Assessment date: 2 October 2026. This assessment would change if the commercial treatment of the on-site server is published, removing the one cost variable this profile cannot currently bound; if the G2 review base grows enough to make the 4.8/5 rating statistically meaningful; or if the 64% and 82% figures are published with a stated methodology and baseline, converting two separate claims into comparable evidence.
