What Veesion's own documentation says its gesture detection does not do, written for loss prevention and store operations teams running due diligence. Every constraint carries its documentation, its operational impact, and its workaround.
Every constraint below comes from Veesion's own current documentation, checked on August 17, 2026, and an unusual amount of it comes from a section Veesion publishes itself about what its system does not detect. Where Veesion 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 this page corrects something Spot AI has itself published about Veesion. We wrote that it analyzes the cameras you already have with nothing to replace, no downtime and no extra hardware. The camera half is right and the rest is not: Veesion's own installation documentation lists a server, a power supply, two ethernet cables and an adapter going into the shop, and it states that a store open around the clock has to accept one hour of server downtime a day for a maintenance restart.
Almost every box here comes from Veesion's own pages, including a section headed by the question of what the system does not detect. A vendor that publishes its boundary makes the rest of its claims easy to check.
A constraint list is only useful next to an honest account of the product.
Eleven gesture types across three categories, with concealment described as the most important and always active, personal bag and backpack always active, and the rest switchable per store and per camera. Alongside it sits a published section on what the system does not detect: an item carried openly with no concealment, a theft committed entirely with the person's back to the camera, and suspicious looks, eye movements or expressions.
A compact server, which Veesion describes as about as simple as an internet box, is installed on the existing recorder and reads the streams continuously. Compatibility is stated as any video system that supports an RTSP protocol, which Veesion says is the case with all recent installations, and the install documentation is engineering grade, down to the stream settings and menu paths for common recorders.
Alert history is kept for 30 days, after which Veesion states videos are permanently deleted from the system. It performs no biometric processing, states its system cannot identify or re-recognize a person, and publishes the three components behind that claim: human detection, pose estimation and object recognition. Every employee signs an ethics charter, and 6,000+ stores across 55+ countries sit behind it.
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 AI is trained to detect 11 specific gesture types organized into 3 main categories, and Veesion's own wording is that behaviors outside these categories will not be detected. The published route for anything missing is to report it through the app so the team can retrain the algorithm, with no timeline attached, and there is no path for a customer to define a detection of their own.
A chain whose pattern is concealment into a bag or clothing is buying a precise instrument. A chain losing stock to open walk-outs, to staff at the till or to delivery discrepancies is buying an instrument aimed elsewhere, and the published non-detections say so directly rather than leaving it for month three.
Map your last hundred incidents against the eleven gestures before booking a demo, and count what share falls outside them. Use the Essential, Moderate and Maximum presets or expert mode to switch gestures per camera, and treat anything genuinely missing as a roadmap conversation with a number attached.
The published sequence ends with a push notification carrying a short accelerated clip, delivered typically in under 30 seconds with a median around 18 seconds in Veesion's own figures, with a full-length clip available on request. No speaker, strobe, talk down or any other on-site response appears anywhere in the documentation.
In a staffed shop during trading hours this is the design and it works, because a colleague walking over is what changes the outcome. On a thin shift, or in a large store where the alert lands with somebody already at a till, the value of a detection is capped by who is free rather than by how fast the model is.
Measure how long it actually takes a colleague to reach the far aisle on your thinnest shift and use that, not the alert latency, to size the benefit. Use the suspicion-level filters, the notification schedules and the confirmed-thefts-only subscription so the alerts that do arrive are ones somebody will act on, and price a response layer separately for closed hours.
The published capability is gesture detection linked to concealment on a retail sales floor. No safety detection appears, no operations detection, no queue or dwell analysis and no non-retail deployment. Even the third gesture category, holding consumption, deblistering, shelf sweeping and a slightly suspicious class, is still shop-floor theft.
For a chain that also wants back-of-house safety, loading bay activity, queue length at the tills or equipment compliance on the same cameras, that is a second platform and a second bill. The architecture is built around pose and gesture rather than around a general detection engine, so this is a boundary rather than a feature gap waiting to close.
Decide early whether the video budget is a loss prevention budget or a store budget. If it is the former, Veesion is a clean fit. If operations and safety are contributing to the funding, price a platform whose published set spans them on the same cameras, and compare totals rather than feature lists.
Veesion publishes the placement rules rather than leaving them to an installer: one aisle per camera, the subject 1.5 to 12 meters from the camera and the camera 1.8 to 3.8 meters high. It processes the sub stream rather than the main stream, and recommends D1 resolution or higher, a bitrate above 1000 kbps and 10 frames per second.
The one-aisle-per-camera rule turns coverage into arithmetic. A 30-aisle store with 18 cameras is not a store that is covered, and cameras mounted above 3.8 meters in a high-ceilinged unit fall outside the published range. Re-aiming and adding cameras across an estate is real cost, and none of it appears in a license discussion.
Audit the fleet against the published geometry and stream settings first: aisle coverage, mounting height, distance to the subject, sub stream resolution, bitrate and frame rate. Veesion publishes the menu paths for common recorder brands, so the stream side is usually a configuration afternoon. Put any camera outside the range into the budget as a move or an addition.
Veesion states that data is secured on its own servers, that connections are secured per device and that alerts stay confidential to the store, alongside the 30-day deletion policy. It does not publish a hosting region, an architecture or end-of-contract behavior beyond that deletion, and no SOC 2 or ISO 27001 statement appears. It also states that a store open around the clock has to accept one hour of server downtime a day for a maintenance restart, at a window the customer chooses.
The privacy argument is strong enough to satisfy a works council and it is not the document a security review asks for. A 24-hour convenience format also has to choose an hour a day when the detection is not running, and that should be a deliberate choice of the quietest hour rather than a default accepted during installation.
Request the hosting region, the certification pack and the end-of-contract terms in one email at the start of the evaluation. Choose the daily maintenance window against your own footfall data, record it in the runbook, and write the 30-day export rule into the loss prevention procedure so evidence leaves the platform before it is deleted.
The same five constraints in one view, sized to paste into an evaluation document.
Swipe the table sideways to see every column.
Veesion data comes from Veesion's own product, FAQ, installation and commitments documentation, checked on August 17, 2026. Gaps are marked as not publicly specified.
If the loss is people hiding product in clothing, bags, backpacks or a stroller, and there is somebody on the floor who can walk over inside a minute, most of the constraints above stop applying. The eleven gestures can be switched per camera so a shop whose customers carry their own bags does not drown in alerts, and 6,000+ stores across 55+ countries have already been through the installation. Very few vendors in this market document their own boundary as plainly, which is what makes the rest of their claims easy to trust.
Spot AI reads the same store cameras with a wider remit, which is why the constraint lists differ rather than compete. 15+ pre-trained Video AI Agents cover cash register theft, after-hours intrusion, vehicle break-in, fire, personal protective equipment, forklift near-miss, falls and crowding in hazard zones, so the stockroom, the loading bay and the car park sit inside the same deployment as the sales floor, and Iris builds anything else in natural conversation in about eight minutes.
The response and the storage answers are the other difference. An event at 02:00 is answered on site through talk down, strobes and horns on standard speakers instead of waiting for somebody to read a notification. Any ONVIF or RTSP IP camera works at full functionality and legacy analog comes in through the Intelligent Video Recorder, and full-resolution video stays on the IVR with only event metadata crossing the network, so how far back an investigation can reach is set by the recorder rather than by a fixed window.
The useful question is not which platform has fewer constraints, but whether your loss is concealment on a staffed floor or something happening where nobody is standing.
None of that makes Spot AI the better instrument for concealment, and Veesion beats it outright on disclosure. Veesion publishes its detection set in full and then publishes what it will not detect, which Spot AI does not do for its own agents, so a buyer can test Veesion's claims against their own incident history in a way Spot AI does not make possible. Veesion also publishes a flat 30-day retention policy with permanent deletion, which is the unambiguous commitment a data protection officer wants, where Spot AI's retention depends on the recorder in the building. Its installation documentation, down to stream settings and camera height, is more specific than anything Spot AI publishes. A constraint list earns its keep by lining each platform's shape up against your starting point.
A live pilot on your store cameras answers in a week what a spec sheet cannot.
Customer-reported outcomes from named Spot AI customers.
Don Franklin Family of Dealerships had incident footage on responding officers' phones within four minutes of an alarm, and reported five recoveries within the hour.
Storage Asset Management runs about 50 virtually managed facilities, where a 1 a.m. notification brought officers on site while an incident was still in progress.
Blackmon Oil Co. runs a vehicle loitering filter across its round-the-clock stores so the lot stays watched with a smaller overnight crew.
"The ability to formalize our incident reporting with Spot AI, keep every case in one database, and attach video directly to those cases has been a game changer."
Five, and Veesion documents most of them itself. The gesture set is closed at 11 types in 3 categories, with behaviors outside these categories stated as not detected and no customer-defined path. The published sequence ends with a notification, so nothing happens in the shop without a person. Everything published is shop-floor theft. Coverage follows published camera geometry, one aisle per camera at 1.5 to 12 meters. And the hosting region, the end-of-contract terms and any certification are all request-only.
Yes. Veesion states it is compatible with any video system that supports an RTSP protocol, which it says is the case with all recent installations, and it names HIK, DAHUA, TVT and Uniview as brands its customers generally use. A compact server sits on the existing recorder and reads the sub stream. The real conditions are the published ones: D1 or higher, above 1000 kbps and 10 frames per second, the camera 1.8 to 3.8 meters high and the subject 1.5 to 12 meters away.
It publishes the answer, which is rare enough to be the most credible thing on its site. An item picked up and kept in plain view will not trigger an alert, because the system focuses on concealment. A theft committed entirely with the person facing away from the camera cannot be analyzed. Suspicious looks, eye movements, head rotations and expressions are outside what it reads. Veesion also describes itself as complementary to people rather than a replacement for them.
Thirty days. Veesion states that qualified alert history is kept for 30 days and that after that period videos are permanently deleted from the system, and it advises downloading anything important inside that window. It is short and unambiguous, which is what a works council or a data protection officer wants in writing, so the practical step is an export rule for anything that acquires a case number.
It depends on where the loss is. If it is concealment on a staffed floor, Veesion is a precise instrument for it and documents itself better than most. If the same cameras also have to cover the stockroom, the loading bay, the car park and the hours when the shop is closed, a platform such as Spot AI fits: 15+ pre-trained Video AI Agents run on any ONVIF or RTSP camera plus legacy analog through the Intelligent Video Recorder, footage stays in the store, and an event is answered with talk down, strobes and horns.