The channel partner is becoming a forward-deployed engineer
The hard part of putting an agent on a camera is the knowledge sitting in three people's heads at the plant. Someone has to get it out, and I think that someone is the integrator.
Before a video agent is worth anything to a customer, we need three things from them. A definition of what good looks like on that line, in that store, at that door. A digital twin, meaning images of their site tagged in their own vocabulary. And an output spec: what should happen when the agent sees something, whether that is a bar on a chart, a report, an alert or a row in a data lake.
Almost none of that exists on paper. It lives in a few people's heads, usually the plant manager, a line lead and whoever has been there longest, and getting it into a form a machine can use is the work customers actually pay for.
Who owns what
This is how we draw the stack now.
At the bottom sits infrastructure: the recorder in the closet, the cameras it talks to, the access-control feeds, the camera-health telemetry. Above it, inference: raw video understanding, the models, the record of every run. Both of those are ours. The model layer itself sits below all of it, with the frontier labs, and neither we nor our partners play there.
Above the line is the orchestration and artifact layer. It captures the three artifacts, iterates the output against reality until it is production grade, maintains it as the process changes, and eventually exposes it headlessly, through APIs, to whatever system the customer runs. That layer belongs to the channel.
The line matters because of where stickiness lives. If a customer leaves the partner, the orchestration layer and the run history go with the partner. If a customer leaves us, they lose the interpretation history our models have built on their sites. Both sides hold something real, in different places.
Forward-deployed, by a different name
Palantir made the phrase forward-deployed engineer famous: an engineer who lives at the customer and builds the thing that makes the platform useful there. That is the job I have just described, and I believe it belongs to the channel.
The sequence we are running with partners is explicit. Our engineers lead the first deployments and the partner shadows. Then the partner leads, with us a support call away. The partners taking this seriously are hiring engineers and applied-AI product managers to do it, and getting their reps on real accounts, the kind where a head of manufacturing has already signed off on phase two and the controls team is in the room.
One thing we learned the hard way: charge for the build. A paid build phase with metrics agreed up front gets treated as work, and the customer's people show up to the meetings. We tried doing it for free early on and it got treated like a demo.
Why the data compounds
AI models improve with more high-quality data far more than with clever pre-training. The three artifacts are exactly that data, so every deployment a partner does well makes the platform smarter for the next one. Running cable never had that property.
There is also a mundane unlock hiding in the infrastructure layer. Our recorder already reports camera health and image quality, free. A partner who can show a customer which cameras are failing and which lenses are fogging has a reason to visit every quarter, and a refresh conversation backed by numbers. That quarterly visit is where the orchestration work gets sold.
Who is already doing this
Cobalt, an integrator group built from 21 acquisitions, is one of the leaders. They are leading from the front on this model and working closely with us to build it, down to the whiteboard above. Cobalt has company. Many of the large partners in our channel have reached the same conclusion on their own: they have to become AI deployment specialists for their customers, or watch that work go to someone who will.
The integrator's margin has come from hardware and labour for thirty years. I think the next thirty years of margin come from judgement, from knowing a customer's process well enough to encode it, which is a harder job than pulling cable and a better one. It is also the only way agents get into a million buildings, because we cannot visit them all ourselves.