Standard Work for Machines: Procedures an AI Can Follow

Every automation project begins with a confession nobody planned to make. The team sits down to hand a process to a tool — to specify, exactly, what happens at each step — and discovers that nobody can. The written procedure says one thing; the experienced operator does another; the exceptions, which turn out to be a third of the work, live in one person’s head and a folder of emails. Lean manufacturing learned this lesson decades ago with human beings: standard work — the current best method, written down, agreed, and followed — was never really about the paper. It was about forcing an organization to know its own process well enough to write it truly. Automation re-teaches the same lesson with higher stakes, because a person handed a vague procedure improvises around the gaps, while a tool handed one follows it into the wall — faithfully, quickly, and at scale.

The tool is the strictest auditor your procedure will ever meet

This is why automating a process is best understood as an audit that executes. Every ambiguity the humans had been silently resolving — the approval that “usually” comes from one of two managers, the form field that means different things in different months, the step everyone skips because its input never actually arrives — becomes a design decision someone has to make explicitly, possibly for the first time. Teams that treat this as an obstacle produce brittle automation and blame the tool. Teams that treat it as the point discover the automation project’s real deliverable is not the workflow; it is the first accurate map of how the work has actually been done, including the workarounds that turn out to be load-bearing and the official steps that turn out to be decoration. Writing standard work precise enough for a machine is the same writing lean demanded for people: the sequence, the standard time or condition, the check that tells you the step succeeded, and the named owner allowed to change it. The machine simply removed the option of pretending the document was finished.

Standard work is a floor, not a ceiling

The standing fear about standard work — that it freezes a process and forbids improvement — gets the lean logic backwards, and it gets the automation logic backwards too. A standard is what makes a deviation visible; a deviation, examined, is where improvement proposals come from. The tool following the standard exactly produces something a human team rarely can: a clean record of where the standard fails — every exception forced, every step where the input was wrong, every loop back for rework. That exception log is the improvement agenda, written by the process itself. The discipline that follows is identical to the old one: the owner reviews the exceptions on a rhythm, decides which are noise and which are the process telling the truth about itself, and revises the standard in writing, with a date, instead of letting private workarounds accumulate until the document and the reality are strangers again. Automate the process you have documented, learn from the exceptions it surfaces, and re-standardize. The machine did not change that cycle. It just made skipping it impossible to hide.

Expect the writing of the standard to surface one argument in every department, and welcome it: the operators who carry the exceptions in their heads are not obstacles to the project — they are its primary source material, and often its most under-credited experts. The documentation sessions that work treat them that way, walking the process at the desk where it happens, asking what the manual gets wrong, and writing the answers down with their names attached as authors rather than suspects. A standard written about people’s work without them will be quietly routed around within a quarter, by humans and then by whoever configures the tool after them. A standard written with the people who do the work has the only authority that has ever made one stick: it is true, and everyone in the room knows who made it true.

← All Insights

© 2026 Cognosco Learning. All rights reserved.