Codifying expertise at scale

Give the most capable model in the world a real operations job and it will underperform a competent human who has done it for a year. Not because the human is smarter, but because the human knows things the model was never told — the conventions, the judgment calls, the definition of good that no one ever wrote down. Intelligence is not the missing ingredient. Expertise is.

Closing that gap is the central technical problem of putting AI to work in operations. We approach it as one discipline: codifying expertise at scale.

Models are general. Operations is specific.

A frontier model carries a vast, general prior about the world. An operations desk runs on something narrower and deeper — the specific way this company, in this industry, handles this work. Which exceptions matter. What “resolved” means here. Which rules are real and which are merely habits. A general model has no access to any of it, and improvises in the gap, confidently and often wrongly.

Expertise is tacit by default

The reason this knowledge never made it into software is that it is tacit. The best operator on a desk cannot fully tell you how they do their job. They can do it, and they can react to a case in front of them, but the rules live below the level of articulation. Ask them to write a manual and you get a thin sketch that misses everything that actually matters. This is why two decades of process documentation and RPA never captured it.

How we codify it

We treat expertise the way an apprentice would: by watching the work, not by reading a description of it. We sit with the people who run the desk, walk through real cases, and surface the decisions — including the ones they make without noticing. Then we structure that into capabilities: explicit, tested representations of how a specific slice of work is done well, with the edge cases and the definition of a good outcome attached.

The output is not a prompt or a policy document. It is a durable, inspectable unit of know-how an operator can run — and that a person can read, question, and correct. Expertise becomes an asset the company owns, rather than a risk that walks out the door when someone retires.

Why this scales

Two reasons. First, much of the craft repeats across companies in the same industry; a capability captured once becomes a strong starting point everywhere, so each new deployment begins far ahead of zero. Second, the last mile — the part specific to one operation — is captured through the same observe-and-structure process, and then the operator is fine-tuned to that company. Shared foundation, bespoke edge.

This is what turns a one-off integration into a method. The marginal cost of putting an operator on the next desk falls every time we codify another piece of the field’s expertise.

Codified, not frozen

Operations change. Carriers, regulations, customers, and tools all move, and expertise that is written down once and never revisited rots. Because capabilities are explicit and versioned, they can be updated deliberately as the work changes — reviewed, corrected, and improved the way any important asset is maintained. Codified is not the same as frozen.

The future of operations work depends on this layer. Bigger models will keep arriving. They will only become operators when the expertise they lack is captured from the people who have it — accurately, at scale, and kept current.