Operational discovery is the missing input to an AI decision

Most organizations can produce an account of how a process is supposed to run. The account lives in a procedure, a system configuration, or the memory of the person who designed it. Operational discovery is the separate task of finding how that process runs when it leaves the procedure: which handoff is real, which step is a workaround, which fact is re-entered because the first system never received it.

The distinction matters once the decision is no longer “should we try a tool” and becomes “what, exactly, are we handing to a system that will act.” A tool that drafts can tolerate a vague description of the job. An agent that is supposed to own a step cannot. It needs the step.

The Agent-Ready Company reviews the early firm-level evidence on generative AI and finds a narrow result. A field experiment across 66 firms and 7,137 knowledge workers found that, among the treated workers who adopted the tool, time on email fell by about two hours a week, with no detected shift in which tasks those workers performed. A second study, linking adoption surveys to Danish administrative records, found precise nulls on earnings and recorded hours at the worker and the workplace, ruling out effects above 2 percent two years after ChatGPT’s release. The authors of that study read the null as employers absorbing the technology through task reorganization rather than through pay.

Neither result says the technology fails at the task it was given. Both say that a faster individual does not, by itself, tell the organization which work changed. Economists have made the same point about information technology since the 1990s: measured returns showed up where the firm had already built the organizational complement, not where it had only installed the system.

That complement, for an agent, is a description of the operation that is specific enough to delegate against. Without it, the firm is choosing a tool. With it, the firm is choosing a boundary: what the system may own, what a person still decides, and which exception is still undescribed.

Operational discovery is therefore not a survey of satisfaction and not an inventory of software. It is an account, tied to roles and to evidence, of the steps that actually occur. The official path and the path people use when the official path stops are both part of the account. So is the place where two roles describe the same handoff differently. That disagreement is often the finding, not a data-quality problem to be averaged away.

The paper argues that measured returns concentrate where operational knowledge was already codified, and that delegating past the boundary of what a system does reliably makes performance worse. Discovery is how an organization finds that boundary instead of assuming the procedure is the boundary. What nobody has yet measured, and what the paper says outright, is how often a firm that cannot describe its own work still proceeds to delegate it.