What Voxel's own documentation says its industrial safety platform does not cover, written for EHS and operations teams running due diligence. Every constraint carries its documentation, its operational impact, and its workaround.
Every constraint below comes from Voxel's own current documentation, checked on August 17, 2026. Where Voxel documents nothing, this page says so rather than guessing, because an absence in a document is a question to ask rather than a finding to claim. 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 Voxel. We wrote that Voxel publishes no certifications on its public pages. It does: SOC 2 Type II audited controls with annual penetration testing, encryption in transit using TLS 1.2 and at rest using AES-256, and infrastructure on ISO 27001-certified AWS, alongside faces and bodies blurred by default and no biometric identification.
The cameras are the ones already on the wall and the evidence carries customer names. The pages go quiet at two points: the moment of the incident, and the terms around the video afterwards.
A constraint list is only useful next to an honest account of the product.
Voxel states it works with over 95% of existing IP cameras, that a site can be deployed within 48 hours using its existing camera infrastructure, and that the system goes live within 48 hours of installation. For a distribution network with a decade of mixed hardware, that turns a video AI program from a capital project into a scheduling exercise, and 48 hours is short enough to prove on one building.
Piston Automotive is published at an 86% drop in vehicle safety incidents, Port of Virginia at a 50% decrease in truck speeding on docks and an 85% increase in safety team efficiency, Carlex Glass at an 86% increase in safety vest compliance and Verst Logistics at an 82% drop in vehicle safety incidents. Customer-reported figures attached to named companies and named metrics are the most checkable form this evidence takes.
SOC 2 Type II audited controls with annual penetration testing, encryption in transit using TLS 1.2 and at rest using AES-256, and infrastructure on ISO 27001-certified AWS, alongside faces and bodies blurred by default, role-based permissions, single sign-on, configurable access by location or camera and no biometric identification of employees. That combination removes most of the objections that stall a floor-monitoring project.
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 documented output is an alert feeding workflows of task assignments, follow-ups and coaching opportunities, with executive-level reporting above them. No on-site audio, strobe or horn response appears on any published page. The documentation describes what the safety program receives, not what the person near the forklift hears.
For a coaching-led safety program this is a coherent design and the published outcomes suggest it works. It is still a real gap on a night shift, at a gate, or anywhere the useful intervention has to reach somebody in the next four seconds rather than the supervisor next morning, and it means the same cameras cannot cover a deterrence requirement.
Split the requirement by urgency. Keep Voxel for the trend, the coaching and the leading-indicator reporting it is built for, and where an event needs a response on the floor within seconds, ask which platform on the shortlist drives the speakers and strobes already installed, then price that alongside rather than after.
The named coverage is ergonomics, material handling, vehicle interactions, personal protective equipment compliance, pedestrian zone violations and near-miss detection, with examples running to improper bends, safety vest compliance, forklift incidents, truck speeding on docks and no-stop at end-of-aisle. Intrusion after hours, loss, tailgating at a door and evidence packs for an investigation do not appear.
A distribution center usually funds cameras from two directions, and this platform answers one of them. If the security manager expects the same feeds to cover the yard fence at 02:00 and produce an evidence pack for a claim, that expectation has to be corrected before a rollout rather than after, or the estate ends up running two vendors on one camera network.
Settle the scope question with whoever is funding the project before the shortlist closes. If EHS is the only sponsor, this focus is a strength. If security is contributing, price either a second platform on the same feeds or one whose published set spans both, and compare the totals rather than the feature lists.
Voxel publishes the workflow as part of the product: detections feed task assignments, follow-ups and coaching opportunities, with executive-level reporting above them, and it publishes Port of Virginia at an 85% increase in safety team efficiency as evidence that the workflow itself is part of the result.
That is the honest design of a coaching-led program, and it means the license is not the whole cost. Someone has to triage the queue, assign the task, hold the coaching conversation and close the loop, every week, at every site. A three-person EHS team covering nine buildings can absorb that or drown in it, and the published outcome figures come from organizations that resourced it.
Name the owner of the queue at each site before the pilot, and size the weekly hours honestly rather than assuming existing capacity absorbs it. Measure closure rate rather than detection count in the pilot, because a rising detection count with a flat closure rate is the shape of a program that will not reproduce the published outcomes.
Over 95% of existing IP cameras is published as a coverage rate. The conditions behind it are not: no resolution floor, frame rate, codec or mounting requirement appears, so there is nothing published that says which cameras fall in the remainder. The figure is also stated about IP cameras, and no route for legacy analog channels appears anywhere.
A rate is reassuring at network level and unhelpful at site level. On a plant where a third of the coverage is analog on a hybrid recorder, or where the oldest aisle cameras run at a low frame rate, the fleet cannot be checked against anything before a survey, and the cameras that fail are usually the ones covering the areas the safety case was built on.
Ask which conditions put a camera in the excluded remainder, in terms of resolution, frame rate, codec and mounting, and ask specifically what happens to analog runs. Survey the oldest and lowest-specified cameras first rather than the newest, and put any camera outside the answer into the budget as a replacement or an addition.
96%+ detection accuracy is published with no baseline, sample, observation window or statement of which detections it covers. The customer outcome percentages are published without baselines or windows beside them, and one is explicitly a first-30-days figure. On the data side, cloud processing on ISO 27001-certified AWS with TLS 1.2 and AES-256 is published, and no retention period, export path or legal-hold behavior appears.
None of that makes the figures wrong, and it does mean none of them can serve as an acceptance threshold or be compared with another vendor's number. Meanwhile a site with an active claims process or a regulator to answer to cannot say from public pages how long the footage behind an incident survives, or how it is produced when a lawyer asks.
Agree in writing what accuracy means and what threshold the pilot has to clear, on your own cameras and your own lighting, before the pilot starts. Request the retention schedule, the export path and the legal-hold behavior in the same email, and ask for the baseline and observation window behind the two customer figures closest to your own operation.
The same five constraints in one view, sized to paste into an evaluation document.
Swipe the table sideways to see every column.
Voxel data comes from Voxel's own product, customer, security and platform documentation, checked on August 17, 2026. Gaps are marked as not publicly specified.
If the estate is a warehouse network or a plant with cameras already covering the aisles, docks and yards, and the business case is a measurable drop in the incidents you report every year, most of the constraints above stop applying. Voxel publishes more named customer evidence than almost anything else in this category, the security and privacy posture is documented to the level a works council asks for, and over 95% of existing IP cameras with a 48-hour go-live makes it one of the few platforms that can be proven on one building before a network decision.
Spot AI reads the same industrial cameras with a wider remit and a different ending, which is why the two constraint lists sit next to each other rather than on top of each other. 15+ pre-trained Video AI Agents span personal protective equipment, forklift near-miss, falls and crowding in hazard zones alongside after-hours intrusion, vehicle break-in, fire and cash register theft, so the yard fence at 02:00 and the aisle at 14:00 are covered by one deployment, and Iris builds anything else in natural conversation in about eight minutes.
The response and the storage answers close the other two gaps. An event can be answered on site through talk down, strobes and horns on standard speakers rather than waiting for a task to be assigned. Any ONVIF or RTSP IP camera works at full functionality and legacy analog cameras come in through the Intelligent Video Recorder, so an analog-heavy plant is not a coverage question, and full-resolution video stays on the IVR in the building with only event metadata crossing the network.
The useful question is not which platform has fewer constraints, but whether your incident needs a coaching conversation tomorrow or an interruption right now.
None of that makes Spot AI the better answer for an EHS program, and Voxel beats it outright in two places. Voxel publishes a detection accuracy figure and a 48-hour go-live, and Spot AI publishes neither, so on those numbers there is more to hold Voxel to than there is to hold Spot AI to. Its coaching workflow, with task assignments, follow-ups and executive-level reporting, is a deeper safety-management layer than Spot AI ships: Spot AI raises the detection and expects your own process to carry it from there. Spot AI also states camera compatibility by protocol rather than publishing a coverage rate. A constraint list earns its keep by lining each platform's shape up against your starting point.
A live pilot on your plant cameras answers in a week what a spec sheet cannot.
Customer-reported outcomes from named Spot AI customers.
Staccato runs context-aware protective equipment rules by zone and person type across an 800-acre campus, supporting its ISO certification work.
Silver Bay Seafoods reported protective equipment compliance and safety standards up 10 to 15% across 22 locations and up to 800 seasonal employees.
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.
"We chose Spot AI knowing our surveillance system could become an extension of our safety team. We now have someone on that team whose job is to study the video and flag cases for review."
Five, all documented. Nothing happens on the floor at the moment of an event, because the published output is an alert feeding task assignments, follow-ups and coaching. The published scope is safety and operations rather than security. The outcomes depend on somebody running the workflow every week at every site. The conditions behind the over 95% of existing IP cameras figure are not published, and no analog route appears. And the accuracy figure, the customer percentages and the retention and export terms all lack the detail an evaluation needs.
Yes, and the claim is one of the strongest in this category: Voxel states it works with over 95% of existing IP cameras and that a site is live within 48 hours of installation using its existing camera infrastructure. The constraint is the remainder rather than the rate. No resolution, frame rate, codec or mounting condition is published, so the cameras that fall outside cannot be identified in advance, and the figure is stated about IP cameras with no analog route named.
Not in the published set. Ergonomics, material handling, vehicle interactions, protective equipment compliance, pedestrian zone violations and near-miss detection are what the pages name, and intrusion after hours, loss, tailgating and evidence packs for an investigation do not appear anywhere. That is focus rather than a shortcoming, and it is a scope question to settle with whoever is funding the cameras before a rollout rather than after.
It is published without the context needed to test it: no baseline, no sample, no observation window and no statement of which detections it covers, so it cannot be compared with another vendor's number or used as an acceptance threshold. The same applies to the customer percentages, one of which is explicitly a first-30-days figure. Treat all of them as evidence the program works somewhere, and agree in writing what accuracy means for your pilot before it starts.
It depends on whether the cameras also have to cover security and whether the response has to happen on site. For a coaching-led EHS program with a team to run it, Voxel publishes more named customer evidence than almost anything else here. For an estate that needs safety, operations and security on the same feeds with an automated response, a platform such as Spot AI fits: 15+ pre-trained Video AI Agents on any ONVIF or RTSP camera plus legacy analog through the Intelligent Video Recorder, with talk down, strobes and horns on standard speakers.