Chapter 5 — Redesign One Role for Accountable AI Work

Chapter 5 — Redesign One Role for Accountable AI Work

Open the Role Before You Redesign It

The first step in redesigning a role is not to rewrite the job description. It is to open the role up. Chapter 4 separated its work fragments. This chapter asks what the redesigned role must reliably deliver.

A workable redesign starts with four questions.

First: what stable result does this role actually deliver? “Responsible for X” and “assists with Y” may look fine in a job description, but they do not define an outcome. Customer service is not there merely to reply to customers. It exists so problems are handled accurately, emotions are addressed, promises are not made carelessly, and the customer experience does not break down. Recruiting is not simply résumé screening; it is getting the right people into the right places in a way that can later be reviewed against outcomes. Media buying is not reporting; it is producing more stable, more explainable growth from the budget spent.

Second: what tasks should AI take on? The answer comes from the decomposition, not instinct. Repetitive execution, information organization, first drafts, standardized checks, review support, and anomaly detection are common starting points. The design choice is to protect human attention for judgment, relationships, and accountability.

Third: what judgment does the person retain? This is the question most often missed. Many workflows contain a “human confirmation” step. It looks as if a human is in the loop, but the person is only clicking a confirmation button. AI makes a recommendation, the person glances at it and approves it, hundreds of times a day, and the system retains the person’s name when something goes wrong. That is not judgment; it is endorsement.

Real human judgment requires at least three conditions: independent information, authority to overturn AI, and freedom from process punishment for doing so. Remove any one and the person cannot realistically press “reject.” Without independent information, there is nothing to rely on beyond the evidence supplied by AI. Without authority to overturn it, the workflow proceeds unchanged. If rejection requires a written explanation, a three-person meeting, and a challenge to explain why the machine might be wrong, people learn to approve with eyes closed.

Fourth: how should the accountability boundary change? When AI joins the work, a role cannot be bounded only by actions; it must be traced through results. Who gives the instruction? What does AI do? Who reviews the output? Who approves an external commitment? Who starts corrective action when a result goes wrong? Which named person is accountable for the outcome?

Try a compensation claim. A customer asks for 8,000; the system recommends 3,000 based on similar historical cases; the representative sends it; and the customer complains to the general manager’s inbox the next day. Put a name next to each question. Who instructed AI to calculate? Which data did it use? Is 3,000 a recommendation or a conclusion? Did anyone review it before it was sent? Above what amount must a supervisor approve? After the complaint, who can overturn and recalculate the claim? If even two blanks cannot be filled, the accountability boundary is unfinished. A company that can fill them spends the next review meeting asking whether the data or rules should change. A company that cannot spends it asking whether to punish the representative.

The four questions recompose what Chapter 4 separated: a stable result, AI-enabled tasks, retained human judgment, and a traceable accountability boundary.

Redesign the Role Around the Result

Consider a media-operations role. The execution layer may include table checking, data consolidation, and report preparation: comparing rate cards, reviewing historical performance by channel, pulling scattered platform data into a report, and formatting it for management. The steps are regular and repeatable. The output is information, not judgment.

The redesigned role must not stop there. It may now carry more of the professional work: interpreting whether a client’s concern reflects a misunderstanding of the data or an expectation that was misaligned from the beginning; deciding whether a shift in spending reflects a durable change or a temporary market signal; preparing a recommendation that someone can actually defend to a customer. AI can prepare material and reveal patterns. It cannot build trust on someone’s behalf or accept the consequence of a bad allocation.

Once the role is defined by result rather than activity, AI can be placed correctly. The question is no longer “Should AI replace this media-operations role?” It becomes: Which tools can take over repetitive, low-judgment work? How much time can that save? What judgment, relationship work, or quality control should the role now perform? What evidence will show that it is delivering a better result?

This is a methodological illustration, not a statistical conclusion. The balance between execution and judgment varies substantially across companies. The point is not to produce a fixed ratio. It is to see the kinds of work inside a role before deciding what the role must become.

But once repetitive work is taken over, do not leap to “Do we still need this person?” The more important question is what the organization intends that person to do with the time released. Without an answer, AI efficiency merely creates empty time and then anxiety.

The redesign should therefore name both a positive future and a boundary. The positive future says what contribution the role is being asked to make: improve the quality of client decisions, reduce avoidable exceptions, strengthen a customer relationship, or maintain a human–AI workflow that produces a reliable result. The boundary says what this role no longer owns: routine reconciliation, first-pass formatting, repetitive retrieval, or an approval that can be safely delegated. Without the first, AI looks like a threat. Without the second, the old work returns through the back door and consumes the time that was supposedly released.

Performance measures have to change with the role. A role that is still assessed only by the number of actions completed will be pulled back toward volume even if AI can produce the actions faster. The new measure may include outcome quality, customer wait, exception resolution, avoided rework, or the usefulness of knowledge captured for the next cycle. The measure should be modest enough to use in ordinary management, but it must reflect the result for which the redesigned role is accountable.

Knowledge also has to travel with the redesign. If prompts, source judgments, corrections, and exception patterns remain in a personal account, the role has not become a repeatable organizational capability. The employee may be faster, but the organization has not become smarter. The process owner and knowledge maintainer do not need to turn every action into documentation. They need to preserve the recurring decisions and boundaries that a new colleague would otherwise have to rediscover under pressure.

The redesign should be visible in the role’s ordinary operating rhythm. What does the person look at each day that they did not look at before? Which exception can they now resolve without a hidden escalation? What weekly review brings together quality, waiting, and customer effect? What does the person add to the guidance after a decision is overturned? These questions keep role redesign from becoming an impressive workshop followed by the old work under a new title.

The role may not become larger or more senior. It may become narrower and more reliable. A person may spend less time preparing material and more time ensuring that a small set of consequential cases receives the right judgment. Another may stop performing routine reconciliation and become responsible for keeping source rules, review criteria, and corrections usable for the entire workflow. The relevant test is not whether every role becomes more glamorous. It is whether each changed role has a real result, sufficient authority, and a contribution that the organization can recognize.

Make Process Ownership Explicit

This is where process ownership becomes necessary. The process owner is accountable for the result of the human–AI workflow. This person is not merely standing beside AI and watching it work. The responsibility includes the workflow’s boundaries, operating quality, exception handling, and feedback from results. The process owner must be able to say when it may run automatically, when human intervention is required, and when the work must stop for a retrospective.

In factory terms, process ownership resembles line supervision. The supervisor does not tighten every screw, but the quality of what comes off the line remains within that person’s remit. They know which machine to stop, where to add an inspection, and who investigates a batch problem. AI entering a workflow adds a line that never sleeps. It still needs a person accountable for the process around it.

A role that once meant “complete the task” may therefore become “ensure that a human–AI system reliably delivers the result.” That is not one more tool skill. It is a change in the nature of the role.

Role redesign is not the phrase “familiar with AI.” It means revising four dimensions together: the deliverable result, the AI tasks, human judgment, and the accountability boundary. If those four dimensions do not move, AI is only a new attachment to an old role. When they are clear, the role can shift from a person performing many actions to a person using AI to deliver stable results.

Rewrite One Role in an Hour

Start with one high-frequency role. Do not begin by asking whether it can be cut. Bring together the accountable executive, the chief human resources officer (CHRO), the chief information officer (CIO), and the role’s direct manager. Put the four questions on a whiteboard.

In the first round, write facts, not wishes. What result does the role deliver now? What does the person actually do each day? What has AI already taken over? Which actions do employees resent most? Which outputs, if wrong, would make customers, the organization, or the finances bear the cost?

Only in the second round write the redesign. Which actions go to AI? Which judgments must stay with people? Which exceptions must escalate? Who is the process owner? Should the role’s performance measure move from “how many actions were completed” to “what result was delivered”? What rule, exception pattern, or correction must become organizational memory?

If the meeting produces only a sentence saying that the role should improve its AI capability, it was wasted. If it produces a human–AI accountability table, an outcome definition, and one operating change to test, the role has begun to be redesigned.

Turn the Redesign into an Operating Test

The workshop is not the redesign. It is the point at which the organization becomes specific enough to run one.

Choose one bounded change and run it in ordinary work. It might be that AI prepares the first draft of a response while a person retains authority over the final promise. It might be that AI organizes evidence for a recurring review while a manager decides the exceptions. It might be that a role stops producing a routine report and begins maintaining the review criteria and correction record that make the report reliable. The change should be narrow enough that the team can observe it, but real enough that someone outside the project feels the result.

Before the test begins, state the result that matters and the failure that cannot be tolerated. If the redesigned role is supposed to reduce customer wait, record where the wait begins now and who can change the next handoff. If it is supposed to improve the quality of a recurring decision, state what evidence a reviewer needs in order to reject an AI recommendation. If it is supposed to release capacity for relationship work, make that new work visible in the role’s calendar and measure. Otherwise routine execution will refill the time as soon as the pilot ends.

The direct manager has a particular responsibility here. A role cannot be redesigned only in a workshop attended by senior executives. The manager allocates the work, recognizes the changed contribution, sees when the old task returns through an exception, and notices whether the employee has enough authority to exercise judgment. If the manager continues to reward speed, availability, and raw volume while the executive asks for quality and learning, the employee receives the old message. The role will revert even if the new description is excellent.

The first review should ask four plain questions. Did the result improve? Did AI actually take over the intended task, or did hidden rework simply move elsewhere? Did the person retain real judgment and authority, or become a rubber stamp? What knowledge, correction, or boundary should be carried into the next cycle? The answers may show that the role design needs another iteration. That is not a failure of the method. It is how a role becomes credible through operating evidence rather than a promise on paper.

This is also where the process owner and knowledge maintainer meet. The process owner watches the whole result: quality, exceptions, customer or business effect, and the point at which the process must be paused. The knowledge maintainer turns recurring correction into usable guidance. One person may hold both responsibilities in a small workflow, but the work must not disappear between them. A system that improves only because one skilled employee is constantly present has not yet produced a redesign the organization can rely on.

When the operating test works, the organization can decide whether to extend it, adjust it, or embed it in the role’s normal rhythm. When it does not, the organization has learned where the role, workflow, or authority boundary is still wrong. Either outcome is better than declaring success because the job description now contains the word AI.

The manager should make the test visible to the person whose role is changing. Hidden redesign is a poor management practice even when the intended result is sensible. Explain what work is leaving the role, what contribution is expected to replace it, which judgment the person is now expected to exercise, and how the organization will evaluate that contribution. If the organization cannot describe a credible future contribution, it has not completed a role redesign; it has merely identified a task that can be automated.

This conversation is not a promise that every employee will prefer the new role or that every role will have the same future. It is a requirement of honest operating design. People cannot supply the context, correction, and relationship work that make AI useful if they are expected to infer the organization’s intentions from an unexplained reduction in routine work. The CHRO, direct manager, and accountable executive do different parts of this work, but none can outsource it to the tool or to a revised job description.

The review also gives the organization a chance to detect a common failure: the person nominally retains judgment but has lost the conditions needed to exercise it. They may no longer receive the source information that makes a recommendation interpretable. They may lack authority to stop an output. Their workload may have increased so sharply that review becomes a formality. Or they may be measured against a target that rewards approval rather than correction. When this happens, the redesign has shifted risk onto the person without giving that person a viable role in controlling it. The remedy is not to demand more diligence. It is to restore information, authority, time, and a clear escalation path.

At the end of the first operating test, record a short decision: continue as designed, continue with a stated correction, narrow the use, or stop. Record why. This decision becomes part of the organizational memory for the next role redesign. It prevents each manager from rediscovering the same boundary alone, and it gives later teams a concrete example of how the organization distinguished a helpful AI task from a trustworthy human–AI result.

Questions of headcount belong in Chapter 10, after the organization has established whether the work is stable, where accountability has moved, what knowledge must remain available, and what it intends to do with the capacity released. A role redesign is not a shortcut around those decisions.

Before closing the workshop, ask three decision questions. What work will stop next week? What result will the redesigned role be expected to protect instead? What authority, information, and review time must move with the role for that expectation to be fair? Write the answers beside the role, not in a generic transformation plan. They turn a plausible redesign into a testable management commitment.

Then choose a review date and the evidence that will be examined. It may be a customer wait, an exception pattern, a quality check, a decision record, or a measure of rework. The point is not to create a universal scorecard. It is to establish whether the role is actually delivering the result it was redesigned to deliver. If the evidence is unclear, the organization should revise the role, the workflow, or the authority boundary before it turns the initial change into a staffing conclusion.

Tool: Four Questions for Role Redesign

Use this table this week. Choose a role whose day-to-day work you cannot clearly describe, or the one employees complain about most. Let the people who do the work fill it in; then review what they wrote. If the four lines cannot be filled, nobody has understood the role well enough to redesign it.

A job is not the only unit of analysis. Once you open the role and make the deliverable result, AI tasks, human judgment, and accountability boundary explicit, you are no longer merely speeding up an old job. You are redesigning a human–AI system that can deliver stable results.