Don’t automate a mystery.

A good name can outrun a usable plan.

Earlier this year OpenClaw swept the world. Moltbook, agents talking to other agents, all of it. Like every other company, Airspace started asking what agentic solutions looked like in our field, and “AI Operator” quickly became a powerful idea here: an agent that could watch operational work, understand what needed to happen, and eventually do it.

The excitement was fair. The upside is obvious: more tasks handled automatically, closer monitoring of every shipment, better communication without a person touching the order, and more time for our operations team to spend on the genuinely hard problems. At one point someone jokingly proposed buying twenty Mac minis and running an operator on each of them. Ambition was comfortably ahead of reality.

Our AI team was enthralled and alarmed in roughly equal measure. An AI operator would be great. It would also certainly fail, they said in no uncertain terms, unless we were careful about how we got there.

This one is different from most of my product stories. Usually the job is to find the problem nobody has named yet. Here the problem was already clear, and I was as taken with the idea as everyone else. What I needed was not a better read on the problem. It was a credible picture of how to actually get there, and that came from our head of AI.

He painted it in three parts, grounded in what the data could actually support: a knowledge base organized around each order, a ladder of trust an AI climbs from watching to suggesting to acting, and eventually an autonomous agent at a few well-chosen points. Supervisors, the people who run the operations teams, would still decide what work should happen at all. Those are not generic-agent requirements. They are workflow requirements.

So my role shifted. I spent my time helping him direct his vision to a realistic surface where he could gain leverage, then move the project into reality through a political organization.

The system already knew more than it could show.

While our head of AI had the recipe for success, I did know some things he didn’t. Thanks to my deep knowledge of Airspace’s product, I was struck by how similar his vision sounded to something Airspace already had: a task workflow system. It was not yet an elegant product surface, but it allowed supervisors to create automatically triggered workflows and showed operators tasks at important moments. Some steps were already automated. Most were still manual.

That was the opening. Instead of creating a separate all-encompassing agent or an entirely new system, we could start by taking small steps to improve the existing task system. The task system could be enhanced in a few ways. Instead of static, hand-built workflows, we could allow Ops Supervisors to be aided by AI. We imagined supervisors uploading real Standard Operating Procedures (SOPs) agreed on with customers, and working with a special agent in the platform to break them down into ideal task workflows. With those in place, not much different from today, we could pick special points to implement and test agentic flows. We called those special points AI Nodes: each one attached to a single step of a workflow, with one job it could be trusted to do.

A prototype that made the next move concrete.

I worked with the AI team to build the supervisor and operator surfaces in Product Kit, the prototyping toolkit we developed and referenced in the Product Lab story. The working prototypes dramatically reduced confusion about the ultimate form the surfaces should take. Instead of debating a distant general-purpose operator, the team could see a supervisor refining workflow rules and an operator receiving the right kind of support at the right time.

Nothing has shipped yet, but the team has made significant progress. The prototypes earned the effort real backing: executive-backed initiatives on the roadmap, a dedicated developer for the AI team, and a scope narrowed to the first few nodes and the feedback mechanism. Implementation can now begin with small, trustworthy capabilities rather than a promise that one agent will somehow do everything.