Ask a room where to use AI and you will get a catalogue.
Search. Summarization. A policy chatbot. A better way to find the document nobody can locate. Those are honest answers to the question. They describe the tools the room has been shown and the parts of work that are easiest to name.
Ask the same room where the work breaks and the language changes. What comes back is rework, waiting, missing authority, unexplained returns, parallel records that disagree, and exceptions resolved outside the formal process. The people did not become more insightful between questions. The second question finally asked for the knowledge they uniquely hold.
Two labels before I build on that, because the claim opening this essay is the one a careful reader should test hardest. The contrast between the two questions is a personal observation: a recurring pattern in professional practice, unnumbered and not attributable to any organization, and no study I can cite measures it. What I draw from it is inferred. The structural weight of the argument sits on the sociotechnical literature below, not on my experience, and the observation is about the question rather than about the people answering it. Someone who answers a placement question with a tool request has answered it correctly.
This essay is about why the question matters, whose knowledge it retrieves, and what rights attach to what happens next. The thesis in one sentence: the people closest to work hold indispensable evidence about where it breaks, while the people accountable for the whole hold different constraints, and AI redesign works only when those positions are combined through explicit decision rights, starting with evidence about breakage rather than a request to place a tool.
The second half of that sentence is load-bearing. Without it, this is a better interview technique for extracting employee knowledge. With it, this is a co-design method.
I wrote Superworkers, Not Replacements as an argument that AI amplifies individual capability rather than replacing people. This essay asks a different question: who has standing to redesign the amplified work, and how do the people affected participate?
Where the process map stops being true
Part 2 of this series described what happens when amplified output meets an unchanged acceptance rate. That constraint occurs somewhere specific in the flow, and the people at that point can describe it.
Your process map is a theory. The people doing the work know where it stops being true. They know which required field cannot be completed when the request arrives, which approval is really a negotiation, which exception has no owner, and which workaround keeps the customer from feeling the seam. Much of that knowledge was never written down because writing it down was not the work.
Breakage is not a feeling. It is a set of events the work produces: rework, waiting, re-explanation, workarounds, and unresolved exceptions. Those events are observable, describable, and owned by the people who encounter them. They can be reported without asking anyone to become a technologist. And the answer can legitimately be a process, authority, data, incentive, or interface defect rather than an AI opportunity. A method that can only return AI is not diagnosis. It is procurement.
Work is jointly social and technical
The idea that a technically superior method can fail because it disrupts the social structure the prior method depended on is as old as the sociotechnical tradition itself. Trist and Bamforth showed this in a 1951 study of coal mining: mechanizing the work system while severing the social arrangements that made it function reduced the performance of the whole, even though the technical method was superior. I am paraphrasing rather than quoting, and I am not suggesting that a field study of mid-century mining directly predicts what AI will do. The parallel is narrower and more useful than that. A technical improvement that treats work as only technical will sometimes make things worse, and the mechanism that explains it has been understood for seventy-five years.
Cherns formalized the design principles that follow from that insight, across two papers in 1976 and 1987. Four of them matter here. Specify only what must be fixed, so the people doing the work retain the latitude to handle what cannot be anticipated. Control variance near where it appears, rather than through distant rules written before the variance was understood. Place boundaries so they do not sever the information a role needs. Treat design as continuing rather than as something that finishes when the system ships. Those principles are doctrine rather than measured causal effects, and I am using them as such.
The place where this gets uncomfortable is participation. Coch and French reported in 1948 that workers who participated in designing changes to their own work showed better adaptation than those on whom equivalent changes were imposed. That canonical study is also one of the most contested in the participation literature. Gardner in 1977, and Bartlem and Locke in 1981, argued that the compared groups differed in pay rate, task difficulty, supervision, and training, confounding the participation variable. The result cannot carry a prescription by itself.
The prescription in this essay rests on the broader sociotechnical logic, on the identity and algorithmic-management literature below, and on explicit decision rights. I am not building the argument on a single contested experiment, and I want that visible before the experiment's defenders or critics arrive in the comments.
Witness and architect
Here is the model, and it is mine. No source in the evidence base states it in this form.
A witness holds direct evidence relevant to a decision. At the point of work, that includes what was redone, what waited, what had to be re-explained, which formal step was bypassed, how an exception was actually resolved, and what quality looks like before it can be reduced to a metric. Much of this knowledge is tacit. A practitioner who can recognize a bad case before stating a universal rule for it is holding evidence the design must preserve. The correct response is not to dismiss it as anecdote or to assume a process map already contains it.
Witness evidence also exists at the enterprise level. Strategic commitments, portfolio dependencies, regulatory exposure, resource constraints, and consequences another function cannot see are all evidence the design requires.
An architect has authority and accountability for selecting what the work becomes. The role requires assembling the relevant witness positions, making trade-offs among admissible alternatives, explaining the choice, assigning ownership, and remaining answerable for the consequences.
The model is intentionally symmetric. Employees hold irreplaceable local evidence and can hold architectural authority. Executives witness strategic and institutional constraints but do not possess detailed work evidence by title. An executive asked to design a local exception flow without the people who run it is making the same category error as an employee asked to choose an enterprise AI architecture from a product demonstration. The distinction is about which knowledge a position holds, not about which people are capable.
The domain practitioner-builder I described earlier is one person who holds both roles. "She is a domain native who has taken a second craft." The witness evidence and the architectural authority reside in the same person because the domain knowledge required to verify the work is the same knowledge required to redesign it. That convergence is not available in every case, but where it exists it is the strongest form of the model.
The question error
"Where should we use AI?" asks the respondent to infer the sponsor's preferred technology, select a target from their own work, anticipate effects on role and identity, make cross-functional trade-offs, and endorse a design whose constraints have not been disclosed. A tool request is a reasonable answer to that question. It reflects the options made visible by the question itself.
This is not a claim about fear or incapacity. I am not going to tell you that employees answer defensively, or that they resist change, or that they cannot see the bigger picture. The design error belongs to the question, not to the people answering it.
Selenko and colleagues, in a 2022 theoretical review, distinguished two channels through which AI affects workers: what work they do and who the work lets them be. That is a function channel and an identity channel, and the identity channel operates independently of employment threat. Reassuring someone that their job is safe does not address an objection rooted in professional identity, craft, or the meaningful exercise of judgment. A placement question engages the identity channel at the same time it asks for a technical recommendation. The combination changes what the question is actually asking.
Separately, Kellogg, Valentine and Christin documented in a 2020 peer-reviewed review that algorithmic systems exert organizational control through six mechanisms: restricting, recommending, recording, rating, replacing, and rewarding. Workers contest that control rather than passively accepting it. I am not claiming that every AI deployment is algorithmic management. I am pointing out that AI can alter the control structure of work, not just the tools, and that a placement question does not surface those stakes.
And Parasuraman and Riley named the design failure directly. In human factors, "abuse" describes automation introduced by designers and managers without due regard for consequences to human performance. The failure mode is located in the people who frame and deploy the system, not in the people whose work changes. Part 5 uses that taxonomy more fully. Here it anchors the anti-blame position: if the exercise produces bad evidence, ask what the exercise got wrong before you ask what the workers got wrong.
What each role holds
The witness holds evidence, the architect holds accountability, and asking one to supply the other's vantage point is a design failure with a predictable output.
Witnesses hold direct evidence and protected validation rights; architects hold accountable choice and must receive evidence, constraints, and concurrence. Equal columns compare what each role holds, can answer, must receive, and should be asked. A final band shows the category error created when an AI-placement question collapses the roles.
Witness
Architect
Holds
Witness
Direct evidence relevant to the decision, including local or enterprise constraints
Architect
Accountability and authority for selecting among admissible designs
Can answer
Witness
What broke, what was redone, what waited, what was re-explained, what moved off-system
Architect
Which breakage is being addressed, which design is chosen, what trade-off is accepted, and who owns the outcome
Must receive
Witness
Protection, validation rights, attribution by consent, and a traceable response
Architect
The relevant witness record, disclosed constraints, risk boundaries, and co-design concurrence
Asked well
Witness
“Describe recent work that had to be redone and what made it move again.”
Architect
“Which admissible design addresses this evidence, and why is it the accountable choice?”
Employees and executives can occupy either column. The columns describe positions in a decision, not people.
Asked badly — “Where should we use AI?” collapses witness evidence, technical design, risk acceptance, and accountable choice into one prompt.
Co-design with actual rights
What follows is a prescription, and it is mine. No source in the evidence base specifies this allocation of rights. It is my synthesis of the sociotechnical doctrine above, the identity and algorithmic-control literature below, and operating practice, and it should be read as a proposal to argue with rather than as a finding to cite. It is also downstream of the essay's actual claim. The argument is about the question. The rights model is what stops a better question from becoming a better extraction technique.
The method is not "ask employees, then let leadership decide." It is a sequence of decisions with different rights attached. Advice is not consent, concurrence is not consensus, and accountability is not unilateral authority.
Employees and affected-worker representatives bring local and tacit work evidence, validate whether the witness record is faithful, co-author design criteria, participate in experiments, and hold concurrence rights over monitoring, individual performance use of telemetry, material job-content changes, workload effects, and professional-discretion boundaries. Where law, collective bargaining, works councils, professional standards, or policy provide stronger rights, those rights govern and this model does not diminish them.
Leaders name the outcome, disclose enterprise constraints, provide time and resources, select among admissible alternatives, and remain accountable for the resulting operating model and benefit allocation.
Risk owners state applicable obligations and risk boundaries, accept or reject control design within their delegated domain, and can stop an experiment that crosses a stated boundary.
Designers and technologists translate evidence into alternatives, make technical limits visible, build the smallest reversible intervention, and maintain the integrity of the evidence-to-design trace. They do not own the business objective or affected-worker rights.
The experiment is a governed decision instrument, not a neutral prototype. It has a hypothesis, declared purpose, scope, admissible data and actions, evaluation criteria, stop conditions, a review path, and explicit prohibition on undisclosed reuse of telemetry. Its results can legitimately produce a non-AI intervention or a decision not to proceed. Negative findings are valid outcomes.
The minimum conditions for using the word co-design under this model: a named decision and a named accountable leader exist before participation begins. The sponsor states what is genuinely open to change. Affected groups are mapped, including people downstream. Participants validate the witness record before it becomes requirements. Employee evidence is traceable into design criteria and the final decision record. Participation is resourced as work, with protected challenge and no retaliation. The experiment is reversible, purpose-limited, and stoppable by more than the sponsor.
If the sponsor is unwilling to publish the internal decision record, grant the stated rights, or protect participants from consequence, the exercise should not be called co-design.
The placement survey did not fail because the team lacked imagination
A shared-services team is asked where AI could help. The answers are familiar: better search, meeting summarization, a policy assistant. Each is plausible. None is dishonest. They are the visible forms of AI the question made available. (This example is illustrative. It describes a recurring pattern, not an internal event.)
The same team is later asked to describe recent work that had to be redone and what made the work move again. The account is different. An intake field requires information that is not available at intake. A review returns work without exposing the reason to the requester. A reconciliation persists because parallel records disagree and authority over the disagreement is unclear. Experienced practitioners keep the flow moving through context that is not present in the formal procedure.
The co-design group validates that witness record before proposing a solution. Employees define what a complete case looks like in practice and which judgment must remain with the role. The accountable leader names the outcome and the authority boundary. The risk owner defines what cannot be inferred or actioned automatically. The designer proposes a bounded intervention that assembles missing context and exposes the reason for a return without rating the employee. The experiment can be stopped by participants, the leader, or the risk owner, and its telemetry cannot be used for individual performance evaluation.
The placement survey did not fail because the team lacked imagination. It answered the question it was given. The breakage exercise found a different object because it asked for evidence before architecture and attached rights to what happened next.
The labor objection, at full strength
Here is the strongest version of the objection this essay must answer, and I am not going to weaken it.
Calling workers witnesses can become a managerial technique for appropriating their knowledge while withholding authority. The organization asks people to reveal the workarounds, judgment, relationships, and exception practices that make the process function. Designers codify that tacit knowledge. Management selects the system. The resulting automation can intensify work, expand monitoring, narrow professional discretion, or make the knowledge-holder easier to replace. A decision log saying the workers were heard does not change who owns the system or who captures the gain. Real participation means workers and their representatives help set the objective, choose what will be measured, define unacceptable uses, evaluate the experiment, and decide whether it scales. Anything less is extraction with a workshop around it.
Against the thin version of this essay, that lands. Consultation without consequence is corrosive. A witness-only role can reproduce the exact separation between conception and execution that participatory design exists to challenge. Attribution is not authority. A retrospective note that feedback was "considered" is not a decision right.
The response is not to abandon the witness/architect distinction. It is to make the distinction symmetric and to refuse to assign the roles by hierarchy. Employees can be architects, individually or through representative structures, where they share accountability and decision rights. Leaders are witnesses, not sole architects, on matters where they hold only strategic context.
And the constructive answer is the rights model above. It gives affected employees more than voice: validation rights over the witness record, co-authorship of design criteria, concurrence on monitoring and workload and material job-content changes, participation in experiment evaluation, protected challenge and stop channels, and a traceable response when evidence is not adopted.
If the sponsor will not grant meaningful rights, protect participants, separate experiment telemetry from individual performance management, disclose the intended use of the knowledge collected, and publish an internal decision record, the exercise should not run under a participation label. On that version of the facts the critics are simply right.
The Monday test
Run a breakage-and-rights review on one recurring flow. Not all of them. One.
Ask the people who perform, receive, govern, and depend on the work to describe a recent instance that had to be redone or unstuck. Ask what they noticed, what the formal process missed, what they did, and who bore the consequence. Do not mention AI during this evidence round.
Create a witness record and return it to the participants for validation. Mark which knowledge is explicit, which remains tacit, which groups are absent, and which evidence should not be attributed individually.
Only then ask for alternatives. Require an AI option and non-AI options to pass the same outcome, risk, workload, professional-integrity, and downstream-impact criteria. Before any experiment, fill in the rights table: who validates the evidence, who decides the objective, who concurs on monitoring and job-content effects, who accepts risk, who builds, who can stop the test, and who owns the final decision.
The output is not a list of use cases. It is a validated breakage record, an explicit rights map, an admissible set of alternatives, and a named accountable decision. If you can do that honestly for one flow, the organization will learn more about AI readiness in a week than a use-case survey returns in a quarter.
What comes next
The breakage record from this exercise will contain coordination work: waiting, chasing, reformatting, re-explaining, routing, and reconciling. That does not make the person performing it the problem, and it does not decide what happens to the role. It identifies work whose economics changed.
Part 4 separates the coordination work that compresses from the management, judgment, development, and consequence that do not, and asks which organizational shape leadership is actually choosing. The five translations that make work agent-legible can be read as a breakage inventory rather than a tool catalogue, and a non-AI finding remains a valid result.
Sources
- Trist, E. L., and K. W. Bamforth. "Some Social and Psychological Consequences of the Longwall Method of Coal-Getting." Human Relations 4, no. 1 (February 1951): 3–38. Publisher. Paraphrased, not quoted.
- Cherns, Albert. "The Principles of Sociotechnical Design." Human Relations 29, no. 8 (August 1976): 783–792. Publisher. Revisited in Cherns, Albert. "Principles of Sociotechnical Design Revisited." Human Relations 40, no. 3 (March 1987): 153–161. Publisher. Cited as design doctrine, paraphrased.
- Coch, Lester, and John R. P. French Jr. "Overcoming Resistance to Change." Human Relations 1, no. 4 (November 1948): 512–532. Publisher. Cited with its critiques: Gardner, Godfrey. "Workers' Participation: A Critical Evaluation of Coch and French." Human Relations 30, no. 12 (1977): 1071–1078. Bartlem, Cheryl S., and Edwin A. Locke. "The Coch and French Study: A Critique and Reinterpretation." Human Relations 34, no. 7 (1981): 555–566.
- Selenko, Eva, Sarah Bankins, Mindy Shoss, Joel Warburton, and Simon Lloyd D. Restubog. "Artificial Intelligence and the Future of Work: A Functional-Identity Perspective." Current Directions in Psychological Science 31, no. 3 (June 2022): 272–279. DOI.
- Kellogg, Katherine C., Melissa A. Valentine, and Angèle Christin. "Algorithms at Work: The New Contested Terrain of Control." Academy of Management Annals 14, no. 1 (January 2020): 366–410. DOI.
- Parasuraman, Raja, and Victor Riley. "Humans and Automation: Use, Misuse, Disuse, Abuse." Human Factors 39, no. 2 (June 1997): 230–253. DOI.
Operate. Publish. Teach.