Verkada's September 17 platform announcement drew a precise boundary around Catalyst, its new connection to third-party AI platforms: it would roll out as read-only. That boundary matters. It does not, however, mean the resulting workflows have no physical consequences. A report that changes a patrol priority or creates a maintenance ticket can influence what people do without directly operating a lock or camera. Verkada's September 17 platform release
The revealing change is therefore not autonomous control of a building. It is the movement of security information into the tools through which work is organized. Our thesis is that the first contest over security agents may be about who interprets and routes events, rather than who has direct permission to actuate equipment.
The control room no longer owns every interface
Security platforms have traditionally asked operators to enter a dedicated console, retrieve footage and then explain the result elsewhere. That creates a handoff between the security system and the business. A facilities manager might need a separate report; an IT team might receive a manually written description of an offline device.
Catalyst uses Model Context Protocol to connect approved external AI tools to Command data. Verkada's own examples include operational reports and a ticket when a camera or sensor loses connectivity. The release names Claude, ChatGPT and Gemini as possible destinations. It describes a new interface to information, not a blanket grant of control to every assistant. Verkada's September 17 platform release
Reading becomes operational when people act on it
Consider a hypothetical assistant asked which loading-bay cameras need attention before the night shift. If it selects the wrong devices, misses an unavailable stream or omits the age of its information, a technician can make the wrong visit. The assistant did not move a camera or unlock a door. It still helped allocate scarce operational time.
That distinction changes the evaluation. A useful agent must preserve the site, device, time and uncertainty associated with its answer. The user needs a route back to the underlying event rather than an authoritative-sounding summary alone. Permission checks must also follow the requesting user's scope; a convenient interface should not quietly broaden access to sensitive observations.
Verkada's separate July NVIDIA announcement described accelerated computer-vision work, including more capable search and event understanding. Those developments improve the supply of machine-readable observations. Catalyst addresses a different part of the chain: making platform information usable by tools outside the traditional console. Neither announcement establishes that every resulting automated workflow is accurate. Verkada's July NVIDIA announcement
The platform prize is the handoff
The strategic opportunity is to become the place where an observation becomes an assigned action. If an external assistant produces a useful report from security data, the value can extend beyond the team that originally bought the cameras. Facilities, loss prevention and operations can all have questions about the same physical environment.
That could expand demand for security infrastructure. It could also place pressure on the platform vendor to make permissions, event definitions and exports more consistent. A fragmented device estate is difficult for a human to interpret; an AI interface does not remove the fragmentation merely by hiding it behind natural language.
Verkada's open-platform update is the relevant product reference for how these integrations are exposed. Buyers should verify the supported objects, access boundaries and availability in their own deployment. An announced integration and a working organization-specific workflow are different milestones. Verkada's open-platform update
One announcement contains different kinds of autonomy
The September release also describes a separate Skydio integration that can pass camera coordinates to dispatch a drone after a specified alert. That should not be confused with Catalyst's read-only starting point. One is a defined integration into a response process; the other exposes information to outside AI tools. Combining both under a single claim that an AI assistant can now run the building would overstate the evidence. Verkada's September 17 platform release
This is why an interface map matters more than a broad autonomy label. Which component observes? Which recommends? Which requests an action? Which is allowed to execute it? The answers identify where an error could propagate and where an operator can intervene. They also reveal which supplier owns the most consequential handoff.
The counterargument is ordinary software integration
Much of this work can already be done with APIs, dashboards and conventional automation. An AI interface may add little where the workflow is stable and precisely defined. It can even be less predictable than a simple rule that creates a ticket whenever a device stops reporting.
The stronger use case is a question that spans several pieces of context and would otherwise require repeated manual investigation. The test is whether the assistant reduces that work without obscuring evidence. A persuasive comparison would measure completion time, incorrect reports, missed exceptions and human corrections against the existing workflow. Adoption alone cannot establish improvement.
Watch the reports people rely on
The first important results will be ordinary: whether an operator can trace a summary to an event, whether a ticket identifies the correct site, and whether access remains properly scoped. If those foundations hold, read-only agents can materially reshape security operations by changing how information reaches decisions. If they do not, connecting more powerful models will simply distribute confusion faster.
References
Questions, news tips or corrections? Contact [email protected].
