essay March 10, 2026
Frontier operations
Notes on operating at the boundary where human judgment and machine capability meet.
Most of what gets called “AI strategy” inside large companies is still downstream of pre-AI assumptions. Someone built the org chart, the operating model, the platform landscape, and the success metrics assuming people would do the work. Then someone else added an AI tool. This is not AI strategy. This is automation with extra steps.
The work I’ve started calling frontier operations is the work of designing for the actual boundary — the place where what humans do best and what machines do best meet and have to be choreographed rather than just glued together.
A few principles I’m testing:
The boundary moves. Anything that’s a hard human task today might be a soft one in eighteen months. You don’t design around capability snapshots; you design around capability trajectories. The architecture has to make the boundary itself a variable.
Judgment compounds. The places where human judgment matters most are the places where it compounds — pattern recognition that improves with exposure, taste that sharpens with reps, trust that accrues. Frontier operations protect and concentrate the work that builds these. They don’t eliminate human work; they curate it.
Process is the bottleneck, not tooling. Most organizations could ten-x the value of their existing AI capability by changing what they ask people to do, in what order, with what handoffs. The new tools are the easy part. The redesigned operating model is the hard part, and the work nobody wants to do.
Speed is a tax on quality, but only sometimes. The interesting question isn’t “human or machine” but “where does this benefit from a deliberate pace and where does it benefit from being instant?” Frontier operations make this an explicit choice rather than a default.
The case for this kind of work, in a large organization, is harder to make than it looks, because the people who can authorize it are the people whose mental models were formed by the old operating model. The argument has to be made in their language, with their evidence standards, on their timelines. That’s where most of the actual difficulty lives.
I’ll keep writing about this as it sharpens.