The Job Characteristics Model Explains Why AI Gains Stall

Many of you in our profession will know the Hackman and Oldham Job Characteristics Model. It is worth going back to, because the problem it was built to explain has returned in a new form.

In the late 1960s and early 1970s, organizations aggressively optimized work to improve efficiency. Large employers such as AT&T, General Electric, the airlines, and other complex operations broke jobs into narrowly defined tasks. Productivity often improved on paper. At the same time, engagement, responsibility, and pride in work eroded. Managers blamed incentives or attitudes. The deeper issue was structural.

J. Richard Hackman, trained at Yale and later a professor at Harvard, observed this from inside those organizations. His conclusion was that people disengaged not because they lacked motivation, but because job design had stripped away ownership, discretion, and visibility into outcomes. Greg Oldham, working closely with him in the early to mid-1970s, brought the empirical rigor, focusing on how to measure what people actually experience while working. Their breakthrough — the job characteristics model, published in 1976 and consolidated in the 1980 book Work Redesign — showed that jobs motivate indirectly, by shaping three psychological experiences: meaningfulness, responsibility, and knowledge of results. Autonomy, end-to-end ownership, and feedback were not soft factors. They were structural requirements for sustained performance and adaptability.

The same dilemma, now with better tools

Fast forward to the AI era and we face the same dilemma. When AI strives for efficiency, and we do not consciously enhance human agency alongside it, we will face the same issues. The mechanism is familiar. Work gets subdivided in pursuit of throughput; the parts of a job that gave a person ownership and a view of the result get absorbed elsewhere; the measured numbers improve for a while and then stop improving, and nobody can say precisely why.

What makes this moment different is the level at which the subdivision happens. AI does not eliminate jobs. It automates and accelerates tasks and subtasks. Roles persist, but their scope, skill requirements, and time allocation change as task composition shifts. A title on the org chart can look entirely stable while the work behind it has been reorganized underneath.

That reorganization has a predictable direction. As AI compresses routine, technical, and analytical work, the hardest-to-automate tasks — judgment, coordination, exception handling, and accountability — become the primary constraints on productivity and outcomes. Human effort and compensation increasingly concentrate on those activities. Yet most organizations have not fully internalized this shift. They are still budgeting, planning, and describing roles as though the composition of work inside them were fixed.

That is the gap I want to draw attention to, because it is a design gap rather than a technology gap. Nothing about the technology forces an organization to leave a role thinner than it found it. That outcome is chosen, usually by default, when the efficiency case is approved and the question of what the job becomes afterwards is never put on the agenda.

What a redesign looks like inside a hospital

I was recently involved in a hospital work redesign project, which was very rewarding. The setting is clinical, but the principles apply well beyond it, which is why I think it is useful for HR leaders to see. For each role, we developed simulation snapshots.

Looking at a role rather than a headcount changes the conversation in a way that is hard to appreciate until you have done it. A role described as a title invites a binary question: does it survive or not. A role described as its constituent work invites a much more useful one: which parts of this change, which parts do not, and what is the resulting job. The second question is answerable, and it produces a design rather than a forecast.

In imaging, AI reduces time spent on image processing, quality checks, scheduling, and administration, improving throughput and turnaround times. It would be easy to stop there, declare a productivity win and move on — and leave radiologists with a faster, thinner, less coherent version of the job they had before.

The real value emerges only when human work is intentionally redesigned around what remains uniquely human. Radiologists increasingly add value through judgment-heavy tasks such as synthesizing imaging with clinical context, resolving ambiguity and edge cases, prioritizing and escalating critical cases, framing diagnostic narratives, advising clinicians, assessing downstream risk, and owning the consequences of decisions. That is a different job from the one the schedule used to describe, and a more demanding one.

Why the job characteristics model points to those tasks

These tasks restore task significance, autonomy, and accountability — exactly the psychological conditions Hackman and Oldham identified as essential for motivation and performance. That is the part I would ask leaders to sit with. The work AI is worst at absorbing turns out to be the work that carries the conditions people need in order to stay engaged with a job at all.

It is a fortunate alignment, and it is not self-executing. A redesign that simply subtracts the automatable tasks and leaves whatever remains in place will tend to fragment the residue into a queue of exceptions with no ownership attached to it. Each individual exception may require judgment; the job as a whole can still fail to give anyone discretion over a whole piece of work or sight of how it turned out. The judgment work has to be deliberately assembled into a role. Nothing assembles it automatically.

This is also why the redesign question cannot be deferred until after deployment. By the time the automation is live, the remaining work has already fallen into whatever shape the workflow happened to leave it in, and changing that shape afterwards means re-opening decisions that everyone considers settled.

Where the design lag actually shows up

The core lesson is consistent across decades: productivity gains stall not when technology fails, but when work design lags behind technology. In the earlier wave, the lag sat between the drive for efficiency and the structure of the jobs it produced. Today it sits between deployed AI capability and the way roles, workflows, spans of control, and performance measures are still written.

You can usually see the lag before it shows up in output. Job descriptions describe a task mix that no longer matches how people spend their week. Performance measures still count volume in an area where volume has been automated. Career paths assume a progression through routine work that AI has largely absorbed, which quietly removes a rung people used to climb. Accountability for an outcome sits with someone who no longer performs most of the steps that produce it. None of these register as failures. They register as friction that nobody can attribute.

What to do before the efficiency case is signed off

My own view is that the sequence has to run in a particular order, and that most of the value is lost when it does not. Decompose each priority role into its tasks and get an honest read on where time actually goes. Work out which of those tasks AI will compress. Then, before approving the efficiency case, write down what the role becomes: which judgment, coordination, exception-handling, and accountability tasks it now owns end to end, what discretion the person holds, and how they will see the result of their decisions.

Two tests are worth applying to the redesigned role. Does the person still own a whole piece of work, rather than a fragment of one? And can they see the outcome their judgment produced, without waiting for a downstream report to tell them? If the answer to either is no, the design will underperform its business case regardless of how good the technology is — and the shortfall will be attributed to adoption, or change management, or the tool.

The job characteristics model did not treat motivation as a benefit of well-designed work. It treated motivation as a consequence of structure. The organizations that get the most out of AI will be the ones that treat human agency the same way: as something built into the design of the role, decided at the same table and on the same day as the automation itself.