What Invisible AI's own how-it-works, security and product pages say its Vision Execution System covers, and what sits outside the device, written for industrial engineering and manufacturing operations leads. Every constraint below carries the documentation behind it, the operational impact, and the workaround.
Every constraint below comes from Invisible AI's own current how-it-works, security and privacy, product and FAQ pages at www.invisible.ai, checked on August 17, 2026, which is where the older invisible-ai.com address now resolves. Where Invisible AI documents nothing, this page says so rather than guessing, and an absence on a public page is named as an absence rather than treated as a no. Two specifics that circulated in older write-ups are dropped here rather than repeated, because they no longer appear on those pages: an ISO 27001 certification and a per-device SSD capacity. Spot AI sells a competing platform, so nothing here rests on an anonymous source or an aggregated user rating.
Each device brings its own camera, its own compute and its own storage, and nothing crosses the factory firewall. That completeness is the product, and it is also where the coverage ends.
A constraint list is only useful next to an honest account of the product.
Each edge device is published as an NVIDIA-powered AI chipset with synchronized depth-sensing cameras, and the retention arithmetic sits beside it: all video processing and storage, up to 3 months and 2TB per device, done on the edge. A plant can work out how far back an investigation reaches and what a station's history costs before opening a negotiation.
All data is stored on edge without leaving the factory firewalls, in Invisible AI's own words, with an on-premise solution stated to maximize information security and minimize IT overhead and no data leaks, ever. On worker privacy it publishes no facial recognition, absolutely zero collection, use or storage of biometric data, and one easy click to blur operators without affecting AI performance.
Cycle-level data across every station feeds assembly work analysis, ergonomics, cycle time, quality deviation detection and throughput, with deployment stated as simple without labeling and on-site expert consultants supporting it. The install base is specific: a Toyota partnership across North America, deployments across 14 North American auto plants and more than 1,500 edge devices at a single automotive facility.
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 product is Invisible AI's own edge device, and the pages describe installing those devices rather than ingesting streams: the how-it-works page opens with install our AI edge devices and integrate with your existing manufacturing systems. Whether an installed camera estate can be read is not publicly specified, and no ONVIF or RTSP support appears on the pages this page read.
A plant that already has cameras over the aisles, the docks and the yard gets no value from them under this design, and the coverage it gains is exactly the number of devices it buys. On an estate where the cameras have been recording to nobody for years, that is the largest asset in the room going unread.
Ask directly whether an RTSP or ONVIF stream can be ingested, and get the answer in writing rather than inferred from the product pages. If the answer is no, split the estate on paper: stations for the device product, and everything else for a platform that reads cameras you already own.
Invisible AI publishes more than 1,500 NVIDIA-powered edge devices at a single automotive facility, and each device carries its own depth-sensing cameras, AI chipset and on-board storage. That is an honest shape for a device product and it makes the growth curve a hardware curve rather than a license curve.
Phase two of a rollout is a procurement cycle rather than an afternoon. Forty more stations means forty more devices, each specified, installed, cabled and eventually refreshed, and the second purchase order usually lands in a different budget year from the first.
Price per device and per line rather than per plant, and agree the phase-two device count at the same time as phase one so the second order is a release against an agreed number rather than a new negotiation. Ask what the on-site expert consultant support during implementation covers, since it is described as part of deployment rather than as a separate service.
What the pages describe is assembly work analysis, ergonomics, cycle time, quality deviation and throughput, with agents pointed at industrial engineering, quality assurance, production planning and new process introduction. Security, the yard, the dock, after-hours activity and any deterrence behavior do not appear on the pages this page read.
The same plant usually has two problems and one budget. If somebody also needs to know who was at the goods-in door at 02:00, or why a trailer sat on the dock for three hours, that is a different vendor on different cameras, with its own contract, its own review and its own installation window.
Walk the plant and count both populations before either vendor quotes: the stations that need cycle-level analysis, and the cameras already watching everything else. Most plants find the second number is the larger one, and that count decides which platform is the bigger line rather than a datasheet.
Invisible AI states all data is stored on edge without leaving the factory firewalls, describes an on-premise solution that maximizes information security and minimizes IT overhead, and publishes no additional bandwidth overhead because processing and storage are done on the edge. It also publishes integrating workforce movement data with your PLC and MES datasets, so the data does move sideways into plant systems even when it does not leave the site.
Cross-site comparison, remote support and central reporting all become questions rather than defaults. A group running the same line in four plants and wanting one view of takt time across them has to establish how that view is produced, and what has to cross a boundary for it to exist.
Ask three questions with your network team in the room: how a multi-plant view is produced, what the vendor can see when it supports you, and what leaves the site to make either possible. Ask the same of the published PLC and MES integration, since that is a second path out of the device, and put the answers in the security submission rather than raising them after the first plant is live.
The site quotes a customer describing savings per minute of downtime and per workstation removed, with a stated three to five times return. No sample size, measurement window or method sits beside it. It is presented as a customer's own arithmetic rather than as a study, and the pages do not claim otherwise.
That is fine as a customer's words and thin as a business case. A finance review that finds a return figure with no baseline behind it will discount the whole submission rather than the sentence, which is a poor outcome for a platform whose actual evidence is unusually strong elsewhere.
Build your own model from the cycle data during a pilot on two stations, and agree the baseline before the pilot starts: current cycle time, current variance, current rework rate. Ask Invisible AI for the customer's own measurement window rather than the headline multiple, and keep the reference figure out of the board paper.
The same five constraints in one view, sized to paste into an evaluation document.
Swipe the table sideways to see every column.
Invisible AI data comes from Invisible AI's own how-it-works, security and privacy, product and newsroom pages at www.invisible.ai, checked on August 17, 2026. Gaps are marked as not publicly specified.
If the number the plant is managed by is cycle time at a station, and the network team will not approve an outbound video stream on any terms, most of the constraints above stop applying, because the device answering both questions is the product you are buying. The on-edge retention figure, the statement that all data stays inside the factory firewalls and the no-bandwidth-overhead design are the three things that get a project past an automotive CISO, and very little in this market puts all three in public.
Spot AI starts from the cameras a plant already paid for, which is why its constraint list looks different. 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 installing a device at each new point. The aisles, docks, yards and doors that have been recording to nobody become the same estate the AI runs on.
What runs on them is broad rather than station shaped. 15+ pre-trained Video AI Agents cover vehicle break-in, fire, cash register theft, after-hours intrusion, personal protective equipment, forklift near-miss, falls and crowding in hazard zones, and Iris builds anything else in natural conversation in about eight minutes. Full-resolution video stays on the IVR in the building and only event metadata crosses the network, and the response runs through talk down, strobes and horns on standard speakers already on site.
Quote the station work per device and the estate work per camera, and never blend them. One line hides which half of the plant is actually being bought.
None of that makes Spot AI the right answer for every buyer. Invisible AI publishes an on-premise deployment where all data stays inside the factory firewalls with no additional bandwidth overhead, and Spot AI does not: its design is hybrid and event metadata crosses the network, with SOC 2 Type II, NDAA compliance and HIPAA alignment as its published pack. Invisible AI also publishes retention arithmetic per device, up to 3 months and 2TB, where Spot AI publishes no equivalent figure, and its cycle-level industrial engineering at a station is deeper than anything in Spot AI's shipped agent list. A constraint list earns its keep by holding each platform's shape against your starting point, and on a plant floor that starting point is usually two questions with one budget.
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%.
Liberty-Perry School District resolves an incident in about five minutes, after evaluating ten systems before choosing Spot AI.
"Even at 90% accuracy, Spot AI's vision beats someone standing there making notes."
Five, all documented: whether it can read cameras a plant already owns is not publicly specified, because the product is its own edge device; coverage scales as a device purchase, with more than 1,500 devices published at one facility; the published scope stops at the station, so security, the yard and after-hours activity do not appear; the on-premise design makes multi-plant views and remote support questions rather than defaults; and the published return figure carries no baseline, sample or window.
Not publicly specified, and the product shape points the other way. Invisible AI ships its own edge device carrying an NVIDIA-powered AI chipset and synchronized depth-sensing cameras, and the pages this page read describe installing those devices rather than ingesting a plant's installed streams. Treat it as an absence to confirm with the vendor rather than as a documented no, and price coverage as a device count either way.
On the device, inside the plant. Invisible AI states that all data is stored on edge without leaving the factory firewalls and publishes video processing and storage of up to 3 months and 2TB per device done on the edge, with an on-premise solution stated to maximize information security. On worker privacy it states no facial recognition, absolutely zero collection, use or storage of biometric data, and one easy click to blur operators without affecting AI performance.
That is the question to ask, because the pages do not answer it. With all data stored on edge and no additional bandwidth overhead by design, a group view across 14 plants has to be produced somehow, and whatever produces it defines what crosses a network boundary. Ask how the view is built, what the vendor can see during support, and what leaves the site, then put those answers into the security submission alongside the published PLC and MES integration.
It depends which half you are solving. If cycle-level industrial engineering at the station is the brief and nothing may leave the network, Invisible AI is strong and unusually specific about it. If the cameras already over the aisles, docks and doors have to earn their keep across safety, security and operations, 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 and 15+ pre-trained Video AI Agents run across the whole fleet.