It may be accurate fiction. It may be required by law. Someone may have spent six months approving the verbs. Still, it leaves most of the work out.
Work-as-imagined is how planners believe the job happens. Work-as-prescribed is the version recorded in rules, checklists, and software. Work-as-done is what people actually do when the day gets complicated.
In The Varieties of Human Work, Steven Shorrock explains these differences. He also describes work-as-disclosed: the version people feel safe describing to managers and auditors. No single account gives us the whole job: as imagined, as done, as told.
Most managers treat the distance between policy and practice as a discipline problem. Someone failed to follow the process. Perhaps another training course will straighten things out.
Why isn’t the gap the first thing managers examine?
The gap grows with scale. The farther planners sit from the work, the cleaner the work looks. From headquarters, a queue is a number. At the service desk, it is twelve people with twelve different problems.
No procedure can predict every condition under which someone will use it. Staffing changes. Equipment breaks. Two reasonable rules collide, usually before lunch or on a Friday afternoon.
Workers adapt because they must. They borrow tools, change the sequence, ask a friend, or skip a step. The official process survives because people repair it while using it.
These repairs are called exceptions. That is a comforting word. It suggests rare events surrounding a stable and sensible center.
In many organizations, the exceptions are the operating model.
Consider “work to rule,” the labor tactic of following every official instruction exactly. Operations often slow down or stop. Strict compliance disables the system it was meant to protect.
The gap was doing more work than management knew.
This is also where the job gets learned. A worker meets a rule that does not fit and must discover its purpose. Then the worker weighs several bad choices and selects the least dangerous one.
That is judgment.
Judgment does not install. It accumulates through mistakes, stories, observation, and consequences. Much of it passes between workers without entering the official record.
The AI Governance Problem
Most AI governance begins with work-as-imagined. A committee defines approved uses. Legal writes restrictions, security assigns controls, and someone makes a slide containing several reassuring shields.
The organization now has governance-as-prescribed. Meanwhile, work-as-done has already moved on.
An employee uses a public tool because the approved system is too slow. A recruiter learns which wording changes a candidate score. A manager relies on AI summaries because reading the source material would consume the afternoon.
A policy may say a person makes the final decision. In practice, challenging the machine takes time and attracts attention. The recommendation is officially optional and socially mandatory.
Authority has moved. Accountability has not.
This is why “human in the loop” can be an empty promise. A person may appear in the workflow without having useful control. If the worker lacks time or permission to disagree, human review is theater.
The same adaptations that make ordinary work resilient can make AI work dangerous. A shortcut may help a customer today while exposing private data tomorrow. A useful prompt may become an unofficial decision rule.
None of this requires a villain. Usually, somebody is trying to clear a queue.
AI governance cannot stop at approving models. It must govern the work built around them. That work keeps changing after launch, even when the software does not.
People find new uses. Supervisors change their expectations. Saved time becomes added capacity, then a target, then the minimum acceptable output.
Yesterday’s safeguard becomes today’s ritual.
A model inventory will not reveal this. Neither will a vendor assessment. Both describe the system from a respectable distance.
Governance must enter the gap.
It must study how people use AI under ordinary pressure. It must notice when suggestions become orders and shortcuts become infrastructure. It must ask who gains from speed and who pays for mistakes.
That requires workers to describe what they really do. Punishing them for revealing workarounds teaches them to hide the information governance needs. The dashboard stays green, which is apparently the important thing.
The goal is not to close the gap. The goal is to see it, learn from it, and decide which adaptations to support, constrain, or stop.
Start with one important AI-assisted workflow. Follow five real cases from beginning to end with the employees doing the work. Record each workaround, override, delay, and private rule of thumb.
Then fix one condition in the work before writing another policy. Repeat the exercise every quarter.
Governance that cannot see work-as-done is another procedure manual.
Another piece of fiction.
=====