Work as Imagined

Somewhere in your organisation there is a document describing a place that does not exist.

The policy everyone signed and no one has read since. The procedure followed to the letter exactly once a year, in the nervous fortnight before the audit. The values, framed, on a wall, and absent from every room behind it. I have written a few of these myself, with real conviction, on quiet afternoons when writing the thing felt satisfyingly like fixing it.

What happens next is so reliable it is almost a reflex. We notice the gap between what the document says and what people do, decide the message simply hasn't landed, and send it again — louder this time. A refreshed training module. A relaunch. A dashboard that renders adherence as a number a steering committee can watch, hopefully, climb. There is even a tidy name the advisory trade sells the manoeuvre under: compliance gap analysis. Find the daylight between the rule and the behaviour, and close it with more rule.

It rarely closes anything, for a reason that is more or less the whole of this essay: the gap is not a fault in the transmission. Sending the signal again, only louder, does nothing when the radio is tuned to the wrong station.

The gap is mostly information

There is a tradition that worked this out years ago, and it lives, of all places, in the study of how things go wrong.

The human-factors researcher Erik Hollnagel drew a distinction I keep coming back to, and Steven Shorrock has mapped in full: work-as-imagined versus work-as-done. Work-as-imagined is the job as the people who design it picture it: orderly, properly resourced, performed under the conditions the procedure quietly assumed. Work-as-done is the job as it actually happens, carried out by real people under real pressure, with the resources they have rather than the resources the policy hoped for, improvising past the parts that don't fit.

The two never match, and not because anyone has failed: a policy is a still photograph of something that keeps moving. The instant it is written it begins to drift out of date against a reality that won't sit still for the portrait, and the people doing the work close that distance themselves, in a hundred small adjustments no document ever sees.

So the safety-science reading, built on Hollnagel and Sidney Dekker, reaches a conclusion the gap-closing industry will not enjoy: the paper-to-practice gap is not an aberration to be stamped out. It is a permanent feature of any attempt to run complex, variable work off pre-written instructions. The useful question was never how to eliminate it, but whether the organisation can see it honestly and learn to read it.

Picture the spreadsheet. Every organisation has at least one — the unofficial workbook a team built on the side, with its own tabs and rules, because the sanctioned system was designed for a version of their work that no longer exists, if it ever quite did. The technology function calls it shadow IT and files it on a risk register, fairly enough. But that spreadsheet is frequently the most accurate description of how the work is actually done anywhere in the building: a running correction to a system that froze its assumptions the day it went live. The people who built it were not being difficult. They were supplying, in formulas and side-tabs, the judgment the official process could not hold. Dekker puts the knife in further: the conditions that "provoke the violation" are usually designed into the work long before anyone bends a rule.

Strip the disapproving vocabulary away and the gap is often the opposite of what we assumed — not defiance but intelligence, a live reading of the distance between the world the policy imagines and the world the work lives in.

And sometimes it is just rot

Here I have to turn the argument on myself, because the view I have just made so attractive has a real weakness, and ducking it would be a cheat.

Let me be plain about the boundary first, since it is the one that matters most. None of this is an argument against compliance, and none of it touches the hard obligations — a regulated duty, a safety-critical control, a requirement that exists in law. There the posture is never comply or surface the gap. It is comply, surface the gap as well, and never pretend the second excuses the first. What follows is an argument against mistaking the document for the doing. It is not, anywhere, an argument against the doing.

Because some procedures are scar tissue. The second signature on a large payment. The access review nobody is permitted to skip. The conflict-of-interest declaration for the arrangement everyone can already see. These are not the dead hand of bureaucracy smothering the wisdom of the frontline. They are what an organisation grew over a wound — the residue of the time someone improvised, sensibly, in context, with their excellent local judgment, and it went badly and expensively wrong.

The sociologist Diane Vaughan gave the pattern its name: the normalisation of deviance. A shortcut saves time and causes no apparent harm, so it is taken again, and again causes no harm, and slowly stops looking like a shortcut at all. It becomes simply how things are done here, right up until the quarter it produces a real loss, at which point everyone is astonished, although nobody has much right to be. Work-as-done drifting from work-as-imagined is not always the system being clever; sometimes it is the system coming loose, and the ignored, much-mocked procedure was the one thing holding a routine process a safe distance from the failure that eventually finds it.

So "read the gap as a signal" curdles, with very little effort, into "never enforce anything" — negligence in the robes of humility. The genuinely hard part, the part no framework retires for you, is telling the clever gap from the rotting one. No document does that work.

Why the louder signal fails anyway

Set the rotting cases aside for a moment, because even where the gap is benign, sending the policy again, louder, usually makes things worse.

This is the reflex The Great Unravelling was about: meeting complexity with more control. Tighten the procedure, add the checkbox, escalate the oversight, and the weave pulls so taut that intelligence stops moving through it. People stop telling you how the work is really done, because the honest answer has been reclassified as misconduct. The gap doesn't close. It goes quiet, and a quiet gap is the one kind you can no longer manage.

There is an older cost underneath. Jeffrey Pfeffer and Robert Sutton named it a generation ago in The Knowing-Doing Gap: the smart talk trap, our habit of mistaking the production of a plan for the doing of the thing. Writing the policy feels like solving the problem; the document gets confused with the deed. Hewlett-Packard's Lew Platt is supposed to have sighed that the difficulty was never knowing what to do, but getting the company to actually do what it already knew.

It is the same shape I went after in An AI performance of progress: the artefact photographs as achievement while the place underneath it doesn't move an inch. A policy can be a performance too, and a handsomely bound one. A rule reaches the document layer of an organisation. It does not, by itself, reach the layer where the work actually changes, which is never the technology and never even the data, but a structure of commitments between people who hold the authority to make them. You can broadcast a rule to a thousand inboxes by lunchtime. A commitment has to be made, by a particular person who could have said no. No document has ever reached that layer, and no dashboard ever will.

The newest twist: work-as-generated

Everything to this point is old. The gap I have been describing has been opening for as long as organisations have written things down, and it has always opened from the same direction — from below, where people improvise under a policy that some other person at least sat down and wrote.

Lately it has begun opening from above as well.

The policy itself is increasingly produced by machine, generated faster than anyone authors it and a great deal faster than anyone absorbs it. Ask, and the full apparatus of governance arrives in seconds: the structure, the headings, the confident finished prose, the ten tidy principles. The form of a policy, on demand. And the form, as anyone who has survived a relaunch knows, was never the part that was missing.

So between Hollnagel's two terms a third has wedged itself in: work-as-imagined, work-as-done, and now work-as-generated — the rule that no mind ever actually passed through. Imagined by no one, finished on arrival. The gap now widens from both ends at once: the done-work drifting off below as it always has, and the imagined-work floating up and away above, composed by something that has never set foot in the building. In the middle accumulates a stack of documents that no human truly wrote and no human truly follows.

The symptom is already measurable. The discipline meant to keep this very technology on a leash has opened a vast space between using it and governing it: by 2026, roughly 83% of organisations report using AI tools while only about a quarter have governance equal to managing them. The thing sold as the great closer of gaps has opened one of the widest gaps going, and chosen to do it in its own house.

In fairness, there is a version of all this that helps. The same tools, pointed the other way, can read how work is genuinely done rather than how it is imagined, and hand you the real gap to look at honestly for once. But that requires wanting to know your policy is being ignored, which is rarely why the tool was bought. And reading the gap well — telling the clever divergence from the rotting one — still needs the seasoned judgment no model carries. The machine can finally show you where the imagined and the done have parted company; it cannot tell you which of those partings to trust. That was never the document's job, and it is not the machine's.

So, no framework

I am not going to give you four steps, partly because I don't believe there are four, and partly because — as established — I have a drawer of policies of my own, written in great earnest and consulted by no one since, least of all me.

The work is not closing the gap. The gap is permanent; it is the standing cost of governing living work with fixed paper, and the bill never stops arriving. The work is staying in honest conversation with it: writing policy to be answered rather than merely obeyed, treating the daily divergence as the most useful management information you own rather than the most embarrassing, and keeping enough hard-won judgment in the room to tell a clever gap from a dangerous one. None of that photographs well. None of it makes the quarterly assurance pack, which is roughly how you know it matters.

So before the next charter, the next framework, the next document that will describe a place that does not exist: it might be worth wandering down to where the work is actually done and asking the people doing it what they have quietly changed. They will tell you — if they trust that you won't write them up for it.

That conversation is the whole job. The rest is very tidy paper.