You can't automate what you can't define

Every company is now asked the same question: where will you deploy AI agents? Budgets are set, pilots are launched, vendors are shortlisted. The harder question comes first, and it is almost always skipped.

What, exactly, should the agent do?

An agent needs a job description. Not an aspiration, not a role title — a precise and true account of the work: the real path a task takes, the exceptions that pull it off that path, the friction that makes it slow or expensive. Almost no organization has one that is true. What they have instead is documentation.

The whiteboard version of work

Ask how invoice exceptions get handled, or how schedules get coordinated across teams, and you will be handed an org chart, a process document, maybe a training deck. These artifacts describe intent. They record how the work was designed, usually years ago, by people who no longer run it.

The actual work lives somewhere else. It lives in the calls, the tickets, the messages — the thousands of small interactions where someone notices the standard path won't fit this case and improvises. Over time the improvisations become the process. The documentation stays where it was.

This gap is not a failure of discipline. It is the normal condition of any operation that has been running for more than a few years. The problem is what happens when you hand the documented version to an agent and tell it to get to work. You are not automating the operation. You are automating the fiction of the operation, and the fiction breaks on first contact with a real Tuesday.

What a definition of work actually is

A definition of work is a discovered account of how work actually runs — computed from the interaction record itself, not asserted by anyone.

It is discovered: the categories of work emerge from what people actually say and do, in the operation's own vocabulary, rather than being imposed from a template. It is ranked: each category carries its measured volume, its cost, its friction, so you know not just what the work is but what it is worth. And it is living: it is recomputed as the operation drifts, because operations always drift.

This is different from a job description, which describes a person rather than the work. It is different from an SOP, which describes the intended path and is silent on the exceptions — and in mature operations, the exceptions are the job. It is also different from process mining on system logs, which can only see what systems record: the timestamps and status changes, not the phone call that explains why the status changed. Most of the real work happens in conversation, and conversation is exactly what system logs cannot see.

The structure is discovered by geometry, not asserted by a whiteboard. That sentence hides real machinery, but the machinery is the research note's job, not this one's.

Why this matters now

Agents are amplifiers. They take whatever account of the work you give them and execute it faster, at scale, without the human habit of quietly correcting for reality. Point an agent at an accurate definition of work and you compound your understanding. Point it at the whiteboard version and you compound the error — quickly, and with confidence.

There is a second problem, quieter but just as expensive: prioritization. Without measured friction, automation targets get chosen by visibility. The task someone complained about in the last leadership meeting gets an agent; the task that silently consumes thirty percent of a team's week does not, because nobody ever named it. A definition of work replaces anecdote with measurement. You automate what is expensive, not what is loud.

The prerequisite everyone is skipping, in other words, is not better agents. It is a true account of the work for agents to run on.

The box just opened

For the entire history of business operations, the interaction record was a sealed box. Every call, every ticket thread, every message exchange landed in it, and almost none of it was ever examined. The seal was not policy. It was cost. Listening to everything meant paying people to listen to everything, and no operation could afford that. So organizations sampled — a QA review here, a survey there — and ran on the whiteboard version because it was the only version they could afford to hold.

That constraint is gone. The cost of transcribing speech and of turning language into computable form has fallen on the order of a hundredfold in a few years. It is now cheaper to analyze every interaction than it once was to sample a fraction of them. The lid is off, whether or not any particular organization chooses to look.

What comes out of the box is uncomfortable. The real path differs from the documented one, sometimes sharply. The exceptions turn out not to be exceptions — they are the job. Friction that was always felt as a vague drag now has names, categories, and dollar amounts attached. And once these things are measured, they cannot be unmeasured. The fiction does not survive contact with its own data. A leadership team that has seen the real map cannot go back to managing from the drawn one.

But the myth ends the way it always did. At the bottom of the box is hope: the first honest map of how the organization actually runs. Not the org chart's guess and not the consultant's interview notes — the operation's own record, structured and ranked. That map is precisely the ground truth agents need, arrived just as agents became worth deploying.

The timing argument is simple. The box opens for everyone; the costs fell for the whole economy at once. The only choice left is whether you open it deliberately, on your own operation, and act on what you find — or wait until the map gets drawn for you.

Observe, learn, execute

This is the sequence Karez is built around. Observe: capture the signal from the channels where work actually happens. Learn: discover the structure in it — the real categories, ranked by volume and friction. Execute: deploy agents on that structure, with a job description that is finally true.

The order is the point. Execution without definition is speed without direction. For how the definition is actually computed — how scattered conversations become a ranked map of work — see the research note.