What happens when AI removes the visible work?

Automation removes the routine execution and leaves the person responsible for the exceptions. Those arise less often and carry more consequence, and the practice that would have equipped someone to handle them came from the routine work that is now gone.

Lisanne Bainbridge published this in 1983, about process control. It is the oldest finding in this field and the one most often rediscovered without attribution.

The irony, in her terms

Bainbridge's formulation is that the more advanced a control system is, the more crucial the contribution of the human operator may become, and that by taking away the easy parts of the task, automation can make the difficult parts harder. Ironies of Automation (1983).

The mechanism is not subtle. The easy parts were how the operator stayed practised. They were also how the operator kept a live picture of the state of the system, which is what makes it possible to notice that something is wrong before an alarm says so. Remove them and both go, leaving a person who is asked to intervene rarely, at speed, on the worst day, in a system they have not been inside for months.

Everything written about AI and deskilling since is a restatement of this paper, usually without knowing it.

Why time saved is the wrong measure

The measurement problem follows directly. Most organisational reporting goes wrong at exactly this point.

Time saved prices the work that was removed and not the work that was added. Checking output you did not produce is work. Deciding whether to override is work, and harder work than doing the task would have been. Carrying responsibility for a result you cannot fully reconstruct is work of a kind that does not appear on any timesheet at all.

A programme can therefore report hours saved accurately while the residual job has become more demanding. Both statements are true and only one is being collected.

Task productivity and workload are different variables

This is the distinction most corporate AI discussion collapses, and it predates AI entirely.

A tool can reduce the effort a single task requires while total workload rises, because organisations respond to cheaper tasks by increasing volume, shortening deadlines, widening responsibilities, or adding verification that did not previously exist. Faster production does not entail less work. It entails cheaper units, and what happens to the number of units is a management decision rather than a property of the tool.

Anyone claiming AI reduces workload is making a claim about their organisation's response, not about the technology, and the two get reported as though they were the same finding.

Responsible de-automation

The reverse move is harder than it looks, and this is the part almost nobody plans for.

Handing work back to people restores the task. It does not restore the capability, because the capability decayed while the system was running, and the decay for cognitive work is faster than for physical. The measured rates are here. Switching the system off in a hurry, which is what usually prompts the question, is the worst possible moment to discover that.

So responsible de-automation means restoring responsibility together with the information, practice, staffing and authority needed to exercise it. Aviation is the nearest working model, because it is the one industry that treats manual practice as a scheduled requirement rather than an aspiration, at fixed intervals, before it is needed. What professions can learn from aviation.

Weick and Sutcliffe's high-reliability principles supply the organisational half: preoccupation with failure, reluctance to simplify, and deference to expertise rather than to rank. Managing the Unexpected. The last of those matters most here, because an oversight role without authority is a formality.

What this page does not establish

Bainbridge was writing about process control, where the operator monitors a physical plant with continuous feedback. Professional work with a language model is a different situation: the feedback is slower or absent, the failure is a wrong answer rather than an alarm, and the transfer of her finding is an argument rather than a measurement.

The workload claim is a conditional. No study here establishes that AI increases workload; the point is that task productivity and workload are separable and are routinely reported as though they were not.

And nobody has measured how much practice is enough to keep an oversight role real. Aviation regulates to intervals derived from its own accident history, which is a defensible basis for aviation and not evidence about anyone else.

Key sources

• Bainbridge, L. (1983). Ironies of Automation. Automatica, 19(6), 775-779. In the essential works.
• Weick, K. E. and Sutcliffe, K. M. (2001). Managing the Unexpected. Jossey-Bass, third edition Wiley 2015. In the essential works.
• Perrow, C. (1984). Normal Accidents: Living with High-Risk Technologies. Basic Books. In the essential works.

Related SuperSkills research

On oversight that is not a safeguard, human in the loop and meaningful human oversight. On the cost of verifying, the verifier's discount. On supervising work you could not do, who supervises work they cannot do. On the decay rates, how fast do skills decay. On the industry that scheduled the practice, what professions can learn from aviation.

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The ironies of automation are Bainbridge's, published in 1983, and the argument on this site is a restatement of hers rather than a discovery of its own.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.