Work-as-Imagined vs Work-as-Done: Why Process Interviews Describe a Job Nobody Does (2026)

Short answer: Work-as-imagined is how managers, designers and process documents believe a job is done. Work-as-done is what people actually do to get it done. They are almost never the same, and the gap is not laziness or dishonesty: it is the adjustments people make to cope with real conditions. For researchers the key point is a third variety, work-as-disclosed - what people say about their work, shaped by who is asking. A standard process interview collects work-as-disclosed, and a product built on it is designed for a job nobody does.

The terms come from resilience engineering, the branch of safety science associated with Erik Hollnagel, and were extended by the human factors specialist Steven Shorrock. They are among the most useful ideas a product or UX researcher can borrow, because so much B2B research is, underneath, a study of how work gets done. This guide explains the four varieties of work, the evidence that the gap is large and systematic, the interview techniques that get closer to work-as-done, and how AI-native tools like Koji change what a researcher can realistically observe.

The four varieties of human work

Shorrock's 2016 essay "The varieties of human work" sets out four related but different things we call "the work":

Shorrock notes that work-as-done is achieved by adjustments, variations, tradeoffs, compromises that are necessary to meet demand, and that when people describe their work, the message is tailored to the purpose or objective of the message and, more or less deliberately, to what is thought to be palatable, expected and understandable to the audience. That second point describes a research interview exactly.

Why the gap exists

Hollnagel, Wears and Braithwaite set out the core argument in From Safety-I to Safety-II: A White Paper (2015): It is an unspoken assumption that work can be completely analysed and prescribed and that Work-As-Imagined therefore will correspond to Work-As-Done. The assumption is wrong because Work-As-Imagined is an idealized view of the formal task environment that disregards how task performance must be adjusted to match the constantly changing conditions of work and of the world.

Their most important claim for researchers is about why work succeeds: Things do not go right because people behave as they are supposed to, but because people can and do adjust what they do to match the conditions of work. If that is right, the adjustments are not noise around the process. They are the process, and a product that removes the ability to adjust can make the work worse while matching the documented procedure perfectly.

Work-to-rule: the proof everyone already knows

There is a well-known way to show that work-as-prescribed is not how work gets done. In a work-to-rule, employees follow every official rule and procedure exactly, and output slows sharply. Unions use it as a form of industrial action precisely because it works. If following the documented process exactly is enough to slow an organisation down, the documented process is not what normally keeps it running. Every process document your participants show you describes a way of working that, followed literally, would not deliver their results.

The evidence: what happens when a system is built for work-as-imagined

The best-documented case comes from hospital medication safety. Barcode medication administration systems were designed around an imagined workflow: scan the patient's wristband, scan the medication, give the dose. Ross Koppel and colleagues studied how nurses actually used them (Journal of the American Medical Informatics Association, 2008, 15(4):408-423). They identified 15 types of workaround, grouped as omitted steps, steps done out of order and unauthorised steps, arising from 31 different causes. Nurses overrode the system's alerts for 4.2% of patients charted and 10.3% of medications charted.

The workarounds were mostly not carelessness. They were responses to conditions such as wristbands that would not scan, medications without barcodes, urgent situations and a system that assumed conditions the ward did not have. The designers had modelled work-as-imagined; the nurses were doing work-as-done. Much of the resulting risk came from the gap between the two.

Most B2B software faces the same pattern on a smaller scale. The buyer describes the process as imagined; the product team builds against the process as prescribed; the users adapt with side spreadsheets, sticky notes, copy-paste and a colleague who "just knows". Each of those adaptations is a finding that a standard interview is unlikely to surface.

Why normal days matter more than incidents

Hollnagel's white paper uses a simple number to change where you look. If something fails 1 time in 10,000, one should expect things to go right 9,999 times out of 10,000. Yet most investigation, and most research, focuses on the rare failure. The authors argue that we should look at the many cases where things go right and try to understand how that happens.

This has a direct consequence for interview design. The standard "tell me about a time it went wrong" prompt (the critical incident technique) produces failure-shaped data, which is useful but skewed. Work-as-done lives mostly in the ordinary cases that participants find too boring to mention. A study that only asks about pain points will describe the exceptions and miss the everyday adjustments that make the job work.

How work-as-disclosed misleads interviewers

Work-as-disclosed is not lying. Three pressures shape it, and each is predictable.

The audience effect

People describe their work in the way that seems safe for the listener. Shorrock notes that people choose what they want to say to regulators and may hold back because they fear resources will be withdrawn or sanctions applied. A participant talking to a vendor, or to someone they think reports to their manager, will tend to describe the prescribed process, because describing the workaround admits to breaking a rule.

Invisible expertise

Much of work-as-done is automatic. The experienced user no longer notices the three-step workaround they do fifty times a day, so it never comes up when they are asked "how do you do X?"

The generalisation trap

Asked how they "usually" do something, people give a tidy summary: a small piece of work-as-imagined about their own job. The mess only appears when you ask about a specific recent instance.

Interview techniques that get closer to work-as-done

Ask about the last time, not the usual way

"Walk me through the last time you did this, starting from what triggered it" produces a specific episode, with its interruptions, workarounds and tools. "How do you usually do this?" produces a summary. This is the same principle behind the Mom Test and switch interviews in JTBD: specific past behaviour beats general description.

Ask directly about workarounds and side systems

Make the workaround safe to discuss: "Most teams have a spreadsheet or a trick that isn't in the official process - what's yours?" Normalising it lowers the audience effect. Ask to see the side spreadsheet. It is often the most valuable artifact in the study.

Interview both sides of the gap and treat the difference as a finding

Talk to the people who design or manage the process and the people who do it, separately, with the same questions. When the answers differ, that is not inconsistent data to be averaged away. The gap between the two accounts is often the single most useful thing the study finds.

Ask about ordinary days, not only failures

Balance "tell me about a time it went wrong" with "tell me about yesterday". Hollnagel's point applies directly: the adjustments that make work succeed are what a new product must preserve.

Reduce the audience effect

Guarantee anonymity in writing, recruit through a neutral channel, and do not have the account manager moderate. There is also evidence that the interviewer's identity matters: in a 2014 study in Computers in Human Behavior, Lucas and colleagues found participants disclosed more to a virtual interviewer they believed was operated by a computer than to one they believed a person was operating.

Observe where you can

Contextual inquiry and day-in-the-life studies see work-as-done directly. Interviews reach work-as-disclosed. Use observation for the highest-stakes workflows, and interviews at scale for breadth.

How Koji helps researchers reach work-as-done

The main practical limit on studying work-as-done has been cost. Observation is slow, and running enough interviews to cover every role, site and shift is expensive when every session needs a human moderator. Koji changes that trade-off.

• AI-moderated interviews at scale. Koji runs text or voice interviews with dozens or hundreds of participants in parallel, so you can interview doers and designers of a process across every team instead of the five people a manager nominates.
• A neutral interviewer. Koji's AI interviewer has no reporting line, no quota and no reason to hear the official version. For topics where the audience effect is strong, such as workarounds and rule-bending, a neutral channel can make it easier to say what actually happens.
• Consistent "last time" probing. Koji's customizable AI consultants can be set up to always ask for a specific recent instance and to follow up on vague "usually" answers, so every session is pushed from generalisation towards episode.
• Structured questions to quantify the gap. Pair open conversation with Koji's six structured question types (open_ended, scale, single_choice, multiple_choice, ranking, yes_no). A yes_no on "do you use a tool outside the official system for this?" asked of every participant turns anecdotes into a measured rate. See the structured questions guide.
• Automatic thematic analysis split by role. Koji's analysis links each theme to participant quotes, so you can compare what managers say about a workflow with what operators say, and put the gap in front of stakeholders with evidence.
• Real-time reporting. Findings build up as interviews finish, so you can spot an emerging workaround after the first few sessions and add a question about it for the rest.

A traditional process study built on in-person shadowing might cover a handful of people over several weeks. With Koji, a team can run a broad study of work-as-disclosed across the whole organisation in days, then target observation at the workflows where the gap is largest.

Common mistakes

• Treating the SOP as the ground truth. Work-as-prescribed is one input. It is often the least accurate description of what happens.
• Interviewing only the buyer. The buyer holds work-as-imagined. If the product will be used by someone else, interview them too.
• Asking "how do you usually..." It invites a tidy summary. Ask about the last specific time.
• Labelling workarounds as user error. A workaround is evidence of a condition the design did not anticipate. Design for it or remove its cause; do not just prohibit it.
• Averaging away the gap. When managers and operators disagree, report both accounts and the difference between them.
• Studying only failures. Most of the work happens on normal days. Ask about those too.

Frequently asked questions

What is the difference between work-as-imagined and work-as-done?

Work-as-imagined is how people who design, manage or buy a process believe the work happens. Work-as-done is what the people doing it actually do, including all the adjustments needed to cope with real conditions. The terms come from resilience engineering and the work of Erik Hollnagel.

What are the four varieties of human work?

Steven Shorrock describes four: work-as-imagined (what we think happens), work-as-prescribed (what procedures and documents specify), work-as-disclosed (what people say or write about their work) and work-as-done (what actually happens). Each is useful, and each can mislead in its own way.

Why do people's descriptions of their work differ from what they do?

Partly because much skilled work is automatic and goes unnoticed, partly because people summarise instead of recounting specific episodes, and partly because they tailor what they say to the audience. Describing a workaround can feel like admitting to breaking a rule, so the official version is safer to say.

How is this different from the say-do gap?

The say-do gap usually refers to preferences and intentions, such as saying you would buy something and then not buying it. Work-as-imagined versus work-as-done is about how a process is actually carried out. The gap is often not visible even to the person doing the work, and it exists because real conditions demand adjustment, not because people misreport their tastes.

What interview questions reveal work-as-done?

Ask about the last specific time the person did the task, walked through step by step. Ask directly about workarounds, side spreadsheets and unofficial tools, framed as normal. Ask about an ordinary recent day, not only about failures. Then ask to see the artifacts they mention.

How does Koji help study work-as-done?

Koji runs AI-moderated interviews at scale with a neutral interviewer, consistently probes for specific recent episodes, and pairs conversation with structured questions that measure how common workarounds are. Its thematic analysis can compare accounts by role, making the gap between managers' and operators' descriptions visible with supporting quotes.

Related Resources

• Structured Questions in AI Interviews - measuring how often workarounds happen, not just that they exist
• Contextual Inquiry - observing work-as-done directly in its setting
• Day-in-the-Life Research - studying ordinary days rather than only incidents
• Stated vs. Revealed Preferences - the related gap between what people say they want and what they choose
• Jobs to Be Done Interviews - episode-based interviewing that gets past general descriptions
• Customer Journey Map vs Service Blueprint - mapping the backstage work where adjustments usually hide