AI does not simply make existing work faster. It changes who contributes what, where decisions are made, and how responsibility travels through a process. If the work itself is never redesigned, AI is added on top of a system that was built for a different division of labor.
Automating a task inside an unchanged process usually relocates the effort rather than removing it. The step is faster, and the review, the correction, the coordination and the exception handling around it grow to compensate. People experience more tools and the same workload.
Task-level thinking also hides the parts of work that matter most: how a decision is reached, who is accountable for it, what happens when something goes wrong, and how the organization learns from the result. Those live between tasks, not inside them.
AI agents make this sharper. An agent that can take multi-step action inside a process is not a faster tool, it is a new participant. Participants need defined scope, escalation paths, oversight and someone who owns the outcome. None of that is decided by automating a task.
Work design begins with the system in which value is created, not with the tasks that are easiest to change. Once you can see how the work actually flows, the leverage points become visible, and the allocation between humans and AI becomes a design decision rather than a technical default.
The design sequence
Judgment under ambiguity, decisions with real consequences, relationships, ethical calls, and anything where accountability must rest with a person who can explain it.
Synthesis, drafting, retrieval, pattern recognition and preparation of the inputs to a decision. AI often improves a decision most by improving what the decision-maker sees.
Stable, rule-based, high-volume steps where variation is not valuable and correctness can be verified. Automation and AI are different design choices with different failure modes.
The interesting cases are rarely fully human or fully machine. Most valuable redesigns define a working rhythm: AI proposes, a person decides; AI drafts, a person shapes; AI monitors, a person intervenes.
Every design needs a defined answer for uncertainty, exceptions and disagreement: what triggers a handoff to a person, how quickly, and with what context attached.
Work design should produce feedback. What was corrected, what was overridden, what failed. Without that loop, the design cannot improve and no one can tell whether it is working.
Redesign the work first, then decide what AI should do inside it. The reverse order produces tools that people route around.
Start from the whole flow rather than the task. Map how work moves today, including the informal steps, then identify the few points where a change would alter the outcome. Redesign those points as a working relationship: what AI prepares, what the person decides, what evidence travels with it, and what happens when the case is unusual.
Allocate by the nature of the work, not by what is technically possible. AI is strong where language, synthesis and pattern are involved and where a draft is a useful starting point. Humans are essential where consequences are significant, where context is unwritten, where trust is being built, and where someone must be answerable for the result.
Treat an agent as a participant with a defined scope, not as an interface. That means explicit boundaries on what it may act on, visibility into what it did, a person who owns its outcomes, and a clear escalation path. The design question shifts from “can it do this?” to “under what conditions should it, and who notices when it should not have?”
Where the cost of being wrong is high, where the situation is genuinely novel, where competing values must be weighed, and where a person has to stand behind the decision to someone else. These are not places where AI has no role. They are places where AI informs and a human decides.
Roles shift from producing output to framing problems, evaluating quality, and deciding. That is a real change in what expertise means day to day, and it needs to be named. Responsibility should not become diffuse simply because a system contributed to the work: accountability still needs a named owner.
Automation asks how a step can be done without a person. Redesign asks what the flow should look like now that a new capability exists. The second question can eliminate steps entirely, reorder them, move a decision earlier, or make something possible that was never attempted. Automation optimizes the existing design; redesign questions it.
By designing for it deliberately. Keep people in positions where they choose rather than only approve, make it easy and legitimate to disagree with a system output, preserve enough hands-on exposure that expertise does not erode, and be explicit that overriding AI is a valid professional act rather than a deviation.
Teams know what the system contributes and what remains theirs, which removes a great deal of quiet uncertainty.
Who decides, who is consulted and who is accountable is defined before the tool goes live rather than after an incident.
Exceptions reach a person quickly and with context, instead of being absorbed silently by a process that cannot handle them.
Speed is measured alongside error, rework and consistency, so improvements are real rather than displaced.
Work is designed so people still practice the judgment the organization will need in three years.
Overrides and corrections become visible signals that improve the design instead of disappearing into individual workarounds.
If you are working through how a journey, a team or a process should function with AI in it, I am glad to think it through with you. This work also runs as a design sprint, and I speak on the topic.
Or write directly to sara@bermudez-t.com.