What Cobalt AI's own documentation says its monitoring intelligence layer does not do, written for security operations leaders deciding whether the bottleneck is the alarm queue or the systems underneath it. Every constraint below carries the documentation behind it, the operational impact, and the workaround.
Every constraint below comes from Cobalt AI's own platform, handlers and integrations pages, read on August 14, 2026. Cobalt's site is client rendered and returns nothing to a plain fetch, so it was read in a browser rather than summarized from a press write-up. Where Cobalt publishes nothing, this page says so rather than guessing. Spot AI sells a competing platform, and one framing needs correcting before any constraint is read: Cobalt is not a video management system with a thin analytics layer, it is a reasoning layer that deliberately sits over systems somebody else records with, and its constraints follow from that choice rather than from a gap.
Cobalt is explicit that it sits above the systems already installed. That is what makes it fast to deploy, and it is also why a problem in the layer below stays a problem after it arrives.
A constraint list is only useful next to an honest account of the product.
For video, Eagle Eye Networks, Milestone XProtect and Genetec, plus any RTSP or RTSPS stream. For access control, eight platforms including LenelS2 OnGuard, Genetec, Avigilon, Avigilon Alta, S2 NetBox, Brivo and Acre Feenics. Then Okta for identity, ServiceNow and Salesforce for ticketing, and Slack, phone, SMS and email. A security team can qualify or disqualify its own stack in five minutes rather than in a discovery call.
Up to 94% of events resolved without human escalation, an average handler response time under 15 seconds, and more than 6,000 hours of operator time returned per month at a single enterprise deployment, alongside Salesforce, FedEx and Ally Financial as named customers. Numbers a vendor publishes are numbers a buyer can hold it to in a pilot.
Handlers are enumerated per event type, each following a chain Cobalt names as Detect, Reason, Act and Resolve: door forced open, door held open, tailgating and unauthorized entry, gun detection, person on ground, person detection after hours, crowd forming, fence climbing, loitering, rapid egress, runaway alarm detection and automated alarm root causing. Cobalt also states the design goal plainly, which is to make human operators more effective rather than to replace them.
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.
Cobalt states it is built to sit over the infrastructure you already operate, ingesting from cameras, access control, video systems and sensors with no infrastructure overhaul required, and that it supports edge, cloud or hybrid deployment. It does not describe recording video itself.
The host platform's retention, its search, its camera compatibility and its bandwidth profile all remain yours. If the recurring problem is that a camera did not see something, or that the footage was gone before the claim arrived, more reasoning above the record will not move it, and you are funding two products to reach one outcome.
Write down the last ten incidents that went badly and mark which layer failed each time: the camera, the record, or the queue. Where the queue is the answer, this is the cheapest possible fix because nothing underneath has to change. Where the record is the answer, fix that first and add reasoning afterwards.
The automated actions Cobalt publishes are service tickets, access control updates, Slack, SMS and guard dispatch, with escalated events arriving carrying the full reasoning chain, video context and a recommended course of action. No speaker, talk down, strobe or horn behavior appears on the pages read.
At three in the morning the sequence ends with somebody being told rather than with something happening in the yard. For a site whose real requirement is that an intruder leaves, the reasoning that decided the alarm was genuine is only as valuable as the response it can reach, and here the response is a person traveling.
Cobalt's own answer is guard dispatch and its Command Center managed service, so price the guard response next to the software and check the response time against what the site actually needs. Where deterrence has to happen before anybody arrives, keep a product on the perimeter that can act, and let Cobalt reason over it.
Three video platforms are named, plus any RTSP or RTSPS camera stream. No camera manufacturer or model list appears on the pages read, and what a camera needs in order for a given handler to work on it is not publicly specified. The documentation covers which systems it connects to, not which lenses those systems have to be pointing.
That decides which site can be the pilot, and it is not obvious in advance. Fence climbing and person on ground depend on framing, distance and lighting as much as on the reasoning above them, so a handler that works well on a modern camera covering a car park may behave differently on a 2015 dome watching a loading bay from thirty meters.
Choose the pilot site from the platform side first, meaning one already running Eagle Eye Networks, Milestone XProtect or Genetec, then ask for handler-by-camera qualification in writing for the cameras at that site. Prove the two handlers that justify the purchase before the rollout list is agreed.
All thirteen published handlers are door, perimeter, weapon or person events. No safety or operations handler appears on the pages read: nothing for personal protective equipment, forklift proximity, dock dwell or standard-operating-procedure adherence.
The platform is aimed precisely at a security operations center, which is a strength for that buyer and a boundary for everybody else. If EHS and operations want answers from the same feeds, that is a second product, and the arithmetic of paying for two belongs in the decision rather than in next year's budget.
Ask whether a handler outside the published list is a product or a services engagement, what it costs to build, and what it costs to keep working afterwards. Scope the safety and operations requirement separately and price both together, so the comparison is between one platform and two rather than between two feature lists.
Named enterprise customers are published, including Salesforce, FedEx and Ally Financial. A SOC 2 report, an ISO certificate or an equivalent statement does not appear on the pages read. The documentation covers who uses the platform, not what an auditor was shown.
This is an absence of published detail rather than a finding, and it still matters here more than it would elsewhere, because the platform reads badge logs and video together and holds the reasoning about both. A security review will ask, and the answer has to arrive as a document rather than as a customer logo.
Request the certification pack in week one, alongside the data-processing terms covering how badge and video data are held and for how long. Enterprise customers of that size will have asked the same questions, so the material almost certainly exists even though it is not on a public page.
The same five constraints in one view, sized to paste into an evaluation document.
Swipe the table sideways to see every column.
Cobalt AI data comes from its own platform, handlers and integrations pages, read on August 14, 2026. Gaps are marked as not publicly specified.
If the bottleneck is volume rather than coverage, if the access control system generates more events than anyone can read, and if the video platform underneath is sound and mid-contract, Cobalt is aimed exactly there and publishes numbers for it. Layering rather than replacing is the lowest-disruption way to add judgment to a stack you cannot change this year, and the named integration list makes it one of the fastest platforms here to qualify against a real enterprise environment.
Spot AI is the record and the reasoning in one platform, which is why its constraint list looks different. Video arrives from any ONVIF or RTSP IP camera, with legacy analog through the Intelligent Video Recorder, and full-resolution video stays on that recorder in the building while only event metadata crosses the network. Retention, search, case history and the detections are properties of the same system, so an answer that depends on footage from six weeks ago does not depend on a platform underneath that somebody else chose.
The detections reach past the security queue. 15+ pre-trained Video AI Agents span security, safety and operations, covering after-hours intrusion, vehicle break-in, personal protective equipment, forklift near-miss, falls and hazard-zone crowding, and Iris builds a detection that is not on the list in natural conversation rather than through a services engagement. Response happens at the scene through talk down, strobes and horns on standard speakers, so the sequence can end with something happening rather than with a ticket.
Before comparing feature lists, decide which layer is actually failing: the camera, the record, or the queue. Only one of those is fixed by reasoning.
None of that makes Spot AI the right answer for every buyer. Cobalt names eight access control platforms it reads directly, where Spot AI reaches access control through integrations, and it publishes operating statistics such as 94% of events resolved without escalation and a handler response under 15 seconds that Spot AI does not publish in the same form. For a security operations center mid-contract on XProtect or Genetec, adding reasoning above the stack is a far smaller change than replacing the system of record, and that is a real advantage rather than a consolation. Spot AI's newest agents ship through a beta program rather than blanket general availability.
A live pilot on your cameras answers in a week what a spec sheet cannot.
Customer-reported outcomes from named Spot AI customers.
All Star Elite brought cash shrink down from about 6% to 1% across its store estate, and cut investigation time by more than half.
Cambridge City cut footage search from two hours to 30 seconds after consolidating seven municipal locations on one platform.
Storage Asset Management was alerted to a break-in around one in the morning, and law enforcement arrived while the incident was still in progress.
"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, all documented: it is a layer over systems you already run, so it inherits their retention, search and compatibility; the published automated actions are workflow actions rather than anything at the scene; no camera manufacturer list or handler-by-camera requirement is published; all thirteen handlers are security and access-control shaped; and no security certification appears alongside the named enterprise customers.
Yes, and that is the point of the product. Cobalt names Eagle Eye Networks, Milestone XProtect and Genetec for video plus any RTSP or RTSPS stream, and eight access control platforms including LenelS2 OnGuard, Genetec, Avigilon, Avigilon Alta, S2 NetBox, Brivo and Acre Feenics. It states it layers over the infrastructure you already operate with no overhaul required, and supports edge, cloud or hybrid deployment.
It reasons first and then takes a workflow action. Its own published example takes a door forced open alarm, cross-references the video against badge logs, identifies a faulty request-to-exit sensor as the root cause, opens a service ticket and clears the event. The published action set is tickets, access control updates, Slack, SMS and guard dispatch. Speaker, talk down, strobe and horn behavior is not publicly specified.
No SOC 2 report, ISO certificate or equivalent statement appears on the pages read for this article, which is an absence of published detail rather than a finding. Given the platform reads badge logs and video together, a review should request the certification pack and the data-processing terms in week one, including how long the reasoning chain and its video context are retained and where they are held.
It depends on which layer is failing. If the video platform underneath is sound and the queue is the problem, Cobalt is aimed exactly there. If the constraint is what the cameras notice, how long the record survives, or whether anything happens on site before somebody arrives, a platform that owns the record fits better: Spot AI runs 15+ pre-trained Video AI Agents on any ONVIF or RTSP camera, keeps full-resolution video on the Intelligent Video Recorder, and deters through standard speakers.