What icetana's own FAQ and product pages say its self-learning analytics layer does not do, written for control room, campus and remote guarding teams running due diligence. Every constraint below carries the documentation behind it, the operational impact, and the workaround.
Every constraint below comes from icetana's own current product pages and FAQ answers at icetana.ai, checked on August 17, 2026. Where icetana 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 one thing this page corrects against older material is worth stating up front: the FAQ line about external devices now reads differently from the wording quoted in older write-ups. icetana styles its own name in lower case and this page follows that.
The learning side is documented down to the operating system and the camera count per server. What happens after the alert is documented too, and it ends with a person.
A constraint list is only useful next to an honest account of the product.
icetana states the model begins learning at once, delivers alerts within 24 hours and is fully learned in a week, with no manual configuration step because the baseline is learned per camera rather than defined by a person. Events reach the Livewall within three seconds. On an estate where nobody has time to write rules per camera, that is the whole pitch and it is stated in units you can hold the vendor to.
Ubuntu Server 22.04 LTS or Red Hat Enterprise Linux 8, an NVIDIA RTX 4000 Ada recommended for optimal compute power per PCIe slot with a full compatibility list published behind it, 1 Gbps network interfaces as a minimum and a ceiling of up to 400 cameras per server for the Safety and Security product. On storage icetana is equally direct: your data remains securely stored on your servers, and icetana states it cannot access that data without explicit permission.
icetana states it is largely camera agnostic and compatible with most existing systems, names Milestone, Genetec and Nx Witness as supported video management platforms, and states it can run independently without one. Adding it is a layer decision rather than a platform migration, which keeps the cost of trying it low and means a failed trial costs a fortnight rather than a rollout.
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.
That is icetana's own FAQ wording. The same answer states that icetana sends alerts to the Livewall interface, where operators can view and respond to events in real time, and no speaker, strobe or talk-down behavior appears on the pages this page read. The documentation describes what reaches the wall, not what happens at the fence.
On a staffed control room this is the design working as intended: the operator is the next step and the wall is where the work happens. On a campus whose gatehouse closes at 18:00, or a remote yard nobody watches at 03:00, a detection is worth exactly as much as the person available to act on it.
Keep the response capability on the platform that owns the cameras, or with a central station, and let icetana do the triage it is built for. Where a site is unstaffed after 18:00, ask which system is expected to answer an event at 03:00 and write the name of that system into the evaluation, because icetana's own answer is that a person is watching the wall.
icetana publishes that the model learns each camera's own baseline rather than being told what to look for, and its named events are shaped accordingly: unusual location, time of day, direction, speed and count, alongside loitering, fire and smoke, trip and fall, aggressive behavior and hazard detection, with Forensic Quick Find, license plate recognition and facial recognition.
The model will surface a forklift in an aisle it never uses, and it will stay quiet about a forklift moving too fast down the aisle it always uses, because that is not rare. A requirement stated as a named condition on a named zone, such as protective equipment on one line or dwell at one dock door, is a different question from the one an outlier model is answering.
Write your five questions down and sort them into two columns: rare things you want surfaced, and named conditions you want checked. Run a two-week baseline on one camera and count how many of the second column appear. That number, not a feature list, tells you whether the architecture fits.
icetana reads from a video management system or directly from cameras, with Milestone, Genetec and Nx Witness named, and publishes the infrastructure it needs: Ubuntu Server 22.04 LTS or Red Hat Enterprise Linux 8, an NVIDIA RTX 4000 Ada recommended, 1 Gbps network interfaces minimum and up to 400 cameras per server. It also states that Facial Recognition or License Plate Recognition need extra hardware on top of that ceiling.
A 1,200-camera campus is three servers with RTX class GPUs, patched and refreshed by your team, on top of a video management system that keeps its own license, its own renewal date and its own support contract, and adding plate recognition adds hardware again. Comparing an analytics license against a platform that also records is not a comparison.
Put the recorder, the servers and the analytics on one quote line and model three years, not one. Ask which of the three grows when the camera count does, and confirm the GPU specification against the exact camera count and resolution rather than the published ceiling.
The estates icetana names as industries are remote guarding, mall management, hotels, safe cities, education and cultural properties, with retail reached through mall management rather than as a chain of its own. Manufacturing and warehousing do not appear. That is the ground the marketing claims rather than a statement about what the model can see.
A plant or a distribution operator evaluating icetana is asking it to do a job its own pages do not claim, which usually shows up as a reference call the vendor struggles to arrange rather than as a technical failure. The first pilot then becomes the proof, at your cost and on your timetable.
Ask for a reference in your own vertical and at your own camera count before the pilot is scoped. If there is not one, run the trial on your hardest site rather than your easiest, and agree in advance what a pass looks like in your units: events per shift, false positives per camera per day, and how many of your named questions were answered.
No SOC 2, ISO 27001 or NDAA statement appears on the pages this page read, next to unusually detailed deployment documentation and an explicit position that your data remains securely stored on your servers. icetana is an Australian company listed on the ASX. This is an absence on public pages rather than a finding about the product.
A correctional facility, a government estate or a university procurement office will ask for the pack before the technical evaluation finishes, and an absence on a website turns into a two-week delay rather than a decision. The on-premises data position helps here, because most of the residency questions answer themselves.
Request the documents by name early: the current certification or audit report, the data-handling and sub-processor description, the penetration test summary and the support-access policy that sits behind the statement that icetana cannot access your data without explicit permission.
The same five constraints in one view, sized to paste into an evaluation document.
Swipe the table sideways to see every column.
icetana data comes from icetana's own product pages and FAQ answers at icetana.ai, checked on August 17, 2026. Gaps are marked as not publicly specified.
If operators are already watching a wall and the problem is that there is too much of it, most of the constraints above stop applying, because the person the design depends on is already there. The baseline is learned per camera rather than configured, first alerts land inside a day, the requirements are published so servers can be sized in advance, and the video and the data stay on infrastructure you own. For a campus, a mall or a remote guarding operation, that is a short path to value.
Spot AI is built around named questions rather than statistical ones, which is why its constraint list looks different. 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, so a condition on a zone is checked because somebody asked for it, not because it happened to be rare. Iris builds anything not on that list in natural conversation in about eight minutes.
The path also continues past the screen. AI Security Guard reads the context and then answers within seconds through talk down, strobes or horns on standard speakers already on site, so a yard nobody is watching at 03:00 still gets a response. Any ONVIF or RTSP IP camera connects as it is and legacy analog comes in through the Intelligent Video Recorder, which also holds the full-resolution video in the building while only event metadata crosses the network.
The architecture question is simpler than the feature lists suggest: are you buying triage for people who are already watching, or an answer for hours when nobody is.
None of that makes Spot AI the right answer for every buyer. icetana publishes server requirements, operating system versions, a GPU compatibility list and a per-server camera ceiling that Spot AI does not publish at all, so an infrastructure team that wants to size from a document will find more of it on icetana's side. icetana also keeps everything on servers you already own, which is a cleaner answer for an estate with a hard residency rule than Spot AI's hybrid design, where event metadata does cross the network. Spot AI is a video AI platform rather than a layer you drop over an existing video management system, so adding it is a platform decision rather than a small one. 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.
Cambridge City cut footage search from two hours to 30 seconds after consolidating seven municipal locations on one platform.
Liberty-Perry School District resolves an incident in about five minutes, after evaluating ten systems before choosing Spot AI.
Blackmon Oil runs a vehicle loitering filter to keep parking lots clean and safe at its around-the-clock stores, with a smaller overnight crew.
"Spot AI has saved our risk management officer a tremendous amount of time, and increased our efficiency in incident reporting."
Five, all documented: icetana's own FAQ states it does not send alerts directly to external devices, so the path ends on the Livewall; the model learns what is unusual per camera rather than checking a named condition on a named zone; it is an analytics layer, so the recorder and the servers stay on the bill; the industries it names are campus and control room shaped rather than industrial; and no certification statement appears beside its otherwise detailed deployment documentation.
Yes. icetana states it is largely camera agnostic and compatible with most existing systems, with resolution and frame-rate requirements that vary by feature. It integrates with Milestone, Genetec and Nx Witness, and its FAQ states it can also operate independently without a video management system underneath it. That makes it a layer you add rather than a platform you migrate to, which keeps the cost of trying it low.
The event appears on the icetana Livewall within three seconds for an operator to view and respond to in real time. icetana's FAQ states plainly that it does not send alerts directly to external devices, and no speaker, strobe or talk-down behavior appears on the pages this page read. So the response depends on somebody watching, which is the single most important thing to establish for a site that is unstaffed overnight or at weekends.
The certification pack, by name, because no SOC 2, ISO 27001 or NDAA statement appears beside the deployment detail: the current audit report, the data-handling and sub-processor description, the penetration test summary and the support-access policy behind the statement that icetana cannot access your data without explicit permission. Ask separately for the GPU specification against your own camera count rather than the published ceiling of 400 cameras per server.
It depends on whether a person is available. If the sites that cost you money are the ones nobody watches at 03:00, or the questions have moved from what is unusual to a named condition on a named zone, a platform such as Spot AI fits: 15+ pre-trained Video AI Agents run across the cameras already installed, Iris builds a custom detection in about eight minutes, and the answer runs through talk down, strobes and horns on standard speakers. If the control room is staffed and the camera count is the problem, icetana is well matched to it.