What Intenseye's own platform, Sentinel, Hard Stop and privacy pages say its workplace safety AI does not cover, written for EHS, operations and plant leaders running due diligence. Every constraint below carries the documentation behind it, the operational impact, and the workaround.
Every constraint below comes from Intenseye's own current platform, Core AI, Sentinel, Hard Stop, Connect, solutions and privacy pages, checked on August 17, 2026. Where Intenseye documents nothing, this page says so rather than guessing. Spot AI sells a competing platform, so nothing here rests on an anonymous source or an aggregated user rating, and two claims that circulated about Intenseye are corrected here because its own pages have moved past them. It is no longer safety only: Operations and Quality Assurance are now published as solutions of their own. And its response no longer stops at an alert or a speaker, because Hard Stop is published as connecting directly to PLC systems to stop machines.
The hazard catalog and the response path are documented further than anything else in this category. Scope, camera qualification and retention are the three places the pages stop.
A constraint list is only useful next to an honest account of the product.
Core AI is stated as 50+ categories of unsafe acts and conditions, listed individually rather than summarized as a count. Personal protective equipment alone names apron, cal suit, glasses, gloves, hard hat, hearing muff, mask, reflective vest and sleeve, and area controls name crane area, line of fire, machine area controls, worker-count limits and time limited area. An EHS lead can map that against a hazard register line by line before a single call.
The Sentinel Speaker delivers audio warnings on the floor, and Hard Stop is published as commanding machines to stop when Sentinel detects danger, connecting directly with your PLC systems for sub-second intervention. Very little of this category will put machine actuation in writing with the integration named, and Intenseye does.
Video is processed on site with face blurring and masking applied before any footage leaves the facility, each model is trained only on that customer's own data, and the platform holds SOC 2 Type II with GDPR stated alongside. The Sentinel Hub runs the models locally on an NVIDIA Jetson Orin NX, which an operations technology team can place on a network diagram and take to a security review.
Each one is a consequence of how the platform is designed rather than a defect. What matters is whether it collides with your starting point.
The published solutions are Safety, Operations and Quality Assurance: hazard detection across six families, then pathway analysis, heatmaps, cycle time and cross-site benchmarks on the operations side, and hygiene, QA stations, cleanrooms and cold-chain on the quality side. No security or loss prevention set appears on the pages this page read, so intrusion, after-hours access and register incidents are not part of what is published.
The same cameras over the same floor are usually asked more than one question, and the plant-side questions are now well covered. What is not covered is the night: a plant that also wants to know who was at the goods-in door at 02:00, or who took a pallet off the dock at the weekend, is looking at a second platform on the same feeds, with its own contract and its own review.
List the questions by department before the demo and mark which are safety, operations or quality, since all three now have published answers, and which are security. Price the security half separately rather than assuming the platform will grow into it, and put both totals in front of whoever is funding the project.
Intenseye's Connect page states you can use the camera systems a site already has for AI-powered insights, enabling real-time detections without replacing infrastructure, and the Sentinel Hub is published as joining Sentinel devices and existing site video into one network. What the pages do not state is whether all 50+ categories are available on any connected stream, or what a camera has to meet in resolution, frame rate, angle or lens to qualify.
That single answer decides how much of the estate is really covered. A plant with 90 cameras at 2MP on ten-year-old mounts may find the hazard list it bought applies to twelve of them, and the rest of the coverage arrives as hardware.
Send the camera list with make, model, resolution, frame rate and mounting height, and ask for a written mapping of which of the 50+ detections run on each. Do it before the pilot rather than after, and make the mapping an appendix to the quote so the device count is a consequence of it rather than a later conversation.
Sentinel Core is published as standard IP cameras preloaded with 50+ safety detections from day one, with Sentinel Traffic for vehicle and pedestrian zones, Sentinel Thermal for heat hazards, Sentinel Solar for remote or outdoor coverage without new wiring and Sentinel Depth for 3D distance in robotic zones, and a Sentinel Hub running the models locally on site.
Those four device types cover the corners where an EHS program normally stalls, which is a genuine strength and also where an estimate built from a software subscription quietly becomes a capital request. A site with a yard, a furnace line, a robot cell and an outbuilding is four device families plus a Hub before the software is counted.
Ask which of your questions need a Sentinel device and which run on an existing stream, then get the device count per site into the same quote as the software rather than a later phase. Stage the device purchase behind a pilot on cameras you already own, so the hardware is bought against a measured gap rather than a plan.
Deployment is published as flexible cloud, private cloud or hybrid on-premise, alongside a clear processing position: video is processed on site with face blurring and masking before any footage leaves the facility. No retention period, storage location or export behavior appears next to it. The pages state what happens to a frame, not how long the record of it lives.
An incident investigation that opens six weeks after the event, an insurer asking for the clip, or a works council agreement that fixes a deletion date all need a number rather than an architecture. Without one, the answer arrives during the first serious incident.
Get three answers in writing during the evaluation: the retention period per deployment model, where the footage and the derived metadata are stored, and the export path for an investigation. Name the documents you need alongside them, the data processing agreement and the retention schedule, and attach both to the works council submission.
Hard Stop is published as a product: when Sentinel detects danger, Hard Stop commands machines to stop instantly, and Intenseye states it connects directly with your PLC systems for sub-second intervention. What does not appear beside it is a named protocol, a controller family or a stated safety integrity level, and the wider actions list covers smart plugs, network speakers, industrial sensors and other operational technology equipment.
Stopping a machine from a camera is an engineering and safety-case conversation with your own controls team, your machine builder and often your insurer, whatever the integration is called. The capability being published as a product moves the question from whether it can be done to who signs off that it may be, and on which line.
Ask which controller families and protocols Hard Stop has been commissioned against, and for one reference running the same combination on a comparable machine. Keep phase one to audio warnings through the Sentinel Speaker, and scope Hard Stop as its own pilot on a single line with the controls engineer in the room from the start.
The same five constraints in one view, sized to paste into an evaluation document.
Swipe the table sideways to see every column.
Intenseye data comes from Intenseye's own platform, Core AI, Sentinel, privacy and engineering pages, checked on August 17, 2026. Gaps are marked as not publicly specified.
If the buyer is EHS, the business case is leading indicators against a recordable-injury target, and the shortlist will be judged on whether a platform covers the specific hazards on your register, most of the constraints above stop applying, because you are buying depth in one lane on purpose. Intenseye publishes more of that answer in public than anyone else in this category, its scope now reaches operations and quality assurance on the same feeds, and the floor response and the anonymization pack are the two things that usually get a safety project past operations and past a works council.
Spot AI is built wide rather than deep, which is why its constraint list looks different. 15+ pre-trained Video AI Agents span vehicle break-in, fire, cash register theft, after-hours intrusion, personal protective equipment, forklift near-miss, falls and crowding in hazard zones, so the safety questions and the security and operations questions run on one platform, one contract and one review. Iris builds anything not on that list in natural conversation in about eight minutes.
The camera side is settled before the pilot rather than during it. Any ONVIF or RTSP IP camera works at full functionality and legacy analog comes in through the Intelligent Video Recorder, so coverage grows by connecting a stream rather than by buying a device per zone. Full-resolution video stays on the IVR in the building while only event metadata crosses the network, and the response runs through talk down, strobes and horns on standard speakers already on site.
Quote both against the identical camera list, and state in writing how many detections run on the streams you already have rather than on the vendor's own devices.
None of that makes Spot AI the right answer for every buyer. Intenseye publishes 50+ hazard categories detection by detection where Spot AI publishes 15+ agents plus Iris, so an EHS lead mapping a hazard register line by line will find more of it already written down on Intenseye's side. Intenseye also publishes Hard Stop, which commands machines to stop through a direct PLC connection, and Spot AI publishes nothing equivalent: its documented response is talk down, strobes and horns through standard speakers, not an intervention on a machine. Spot AI's design is hybrid rather than fully on-premise, and event metadata does cross the network. A constraint list earns its keep by holding each platform's shape against your starting point.
A live pilot on your cameras answers in a week what a spec sheet cannot.
Customer-reported outcomes from named Spot AI customers.
Unique Industries covers more than a million square feet with a three-person safety team, catching near misses and falls on the cameras already installed.
Silver Bay Seafoods replaced fragmented legacy camera systems across 22 locations, including remote Alaska facilities, and lifted operational efficiency 15%.
Cambridge City cut footage search from two hours to 30 seconds after consolidating seven municipal locations on one platform.
"Having cameras that talk back to you changes everything. Spot AI has become a real part of our safety team."
Five, all documented: security and loss prevention sit outside the published scope, which now covers safety, operations and quality assurance; which of the 50+ detections run on a camera you already own is not publicly specified; the Sentinel device line is a hardware purchase per zone and per site; retention, storage location and export behavior are not publicly specified; and machine actuation through Hard Stop, while published down to the PLC connection, still needs a safety case rather than a configuration change.
Both are published. Intenseye states it integrates with existing cameras in minutes, and the Sentinel Hub is stated to connect Sentinel devices together with existing site video into one network without major infrastructure changes. Its own line adds Sentinel Core, Traffic, Thermal, Solar, Depth and Speaker for corners a standard indoor camera does not cover. What is not publicly specified is which of the 50+ detections run on a third-party stream.
Both are published. Intenseye's Connect page states you can use the camera systems a site already has for AI-powered insights without replacing infrastructure, and the Sentinel Hub is stated to join Sentinel devices and existing site video into one network. Its own line adds Sentinel Core, Traffic, Thermal, Solar, Depth and Speaker for corners a standard indoor camera does not cover. What is not publicly specified is which of the 50+ detections run on a third-party stream.
Privacy is one of the more documented answers in this category: video is processed on site with anonymization such as face blurring and masking applied before any footage leaves the facility, each model is trained only on that customer's own data, and the platform holds SOC 2 Type II with GDPR stated alongside. Retention is the gap. No retention period, storage location or export behavior appears next to the cloud, private cloud and hybrid on-premise options, so request all three in writing.
It depends how wide the brief is. If the program is safety, operations or quality assurance and the decision rests on hazard-by-hazard coverage, Intenseye is hard to beat on published detail, and Hard Stop puts machine intervention in writing. If the same cameras also have to cover intrusion, loss and after-hours activity, a camera-agnostic platform such as Spot AI fits: any ONVIF or RTSP camera connects as it is, legacy analog comes in through the Intelligent Video Recorder, 15+ pre-trained Video AI Agents run across the whole fleet and Iris builds the rest in about eight minutes.