Which tasks do workers not want automated?

Refusal is patterned, and the pattern is not the one the debate expects. Asked about 844 tasks drawn from their own occupations, and prompted to weigh job loss and lost enjoyment before answering, 1,500 US workers were positive about automating 46.1 per cent of them. Where they said no, three things predicted it: they did not trust the system to be right, the task was one they enjoyed, or the task carried a decision they would be answerable for. The commercially interesting group is smaller than either the enthusiasts or the sceptics assume. It is the set of tasks a machine can already do and the people doing them do not want it to.

Consent turns out to be broader than the argument about it#

Most writing about automation and worker preference is written without asking any workers, or by asking them about AI in general, which produces an attitude rather than a decision. Shao, Zope, Jiang, Pei, Nguyen, Brynjolfsson and Yang did something narrower and more useful. They took computer-compatible tasks performed at least monthly from the US Department of Labor's O*NET database, put each task only to people who confirmed they performed it, and asked them to rate it after considering the consequences.

The result: 46.1 per cent of tasks drew a positive attitude to automation, which the authors report was given "even after explicitly considering concerns such as job loss and reduced enjoyment, as guided by our auditing framework". The distribution is not lopsided in either direction. Strong enthusiasm, a rating of four or five out of five, covered 7.11 per cent of tasks. Strong objection, two or below, covered 6.16 per cent. Almost everything real sits in the middle, which is where the interesting management questions are and where an opinion poll about AI cannot reach.

The reason people gave for wanting a task automated was consistent and, in this research's terms, revealing. "Automating the task would free up my time for high-value work" was selected in 69.38 per cent of cases, ahead of the task being repetitive at 46.6 per cent, an opportunity for quality improvement at 46.6 per cent, and stressful at 25.5 per cent. Workers are not asking to be relieved of work. They are asking to be relieved of the part of it that was never the job.

Distrust outranks fear of replacement, by roughly two to one#

Asked openly how they envisioned using AI in their daily work, 28.0 per cent of participants expressed fears, concerns or negative sentiment. A topic model of those answers puts the largest category at lack of trust in the accuracy, capability or reliability of the systems, 45.0 per cent. Fear of job replacement came second at 23.0 per cent. The absence of human qualities came third at 16.3 per cent, and the authors record what workers meant by it: "workers express specific concerns about losing a 'human touch' in their work, diminishing creative control, and the desire to maintain agency in decision-making".

That ordering deserves more attention than it has had. The public conversation about worker attitudes to AI is almost entirely a conversation about displacement. The workers being displaced-in-theory are twice as likely to name a quality problem. A person who says the tool gets things wrong is making a technical claim that can be tested, and if they are right, the organisation that overrode them has bought an error rate rather than a saving. This is the same argument the estate makes about who owns verification, arriving from the other end.

Two correlations complete the picture. Desire for automation runs negatively against concern about job loss, at a Spearman rho of -0.22, and more strongly against enjoyment of the task, at -0.28. Sector variation is wide. In Arts, Design and Media, only 17.1 per cent of tasks drew a positive rating.

The red light zone, and why investment is pointing at the wrong half of it#

Plotting what workers want against what AI experts judge feasible produces four regions, which the authors define as follows.

• Automation Green Light Zone. "Tasks with both high automation desire and high capability. These are prime candidates for AI agent deployment with the potential for broad productivity and societal gains."
• Automation Red Light Zone. "Tasks with high capability but low desire. Deployment here warrants caution, as it may face worker resistance or pose broader negative societal implications."
• R&D Opportunity Zone. "Tasks with high desire but currently low capability. These represent promising directions for AI research and development."
• Low Priority Zone. "Tasks with both low desire and low capability. These are less urgent for AI agent development."

The paper does not publish a task count for each zone, and no page should invent one. What it does publish is where the money is going. Mapping Y Combinator companies onto the same grid, 41.0 per cent of company-task mappings fall into the Low Priority Zone and the Automation Red Light Zone, with the authors noting that "many promising tasks within the 'Green Light' Zone and Opportunity Zone remain under-addressed by current investments". Two fifths of a startup cohort is building either for work nobody particularly wants automated and nobody can automate well, or for work that can be automated and the people doing it object to.

For a buyer, the Red Light Zone is the one to have a policy about. It is the only region where the technology, the business case and the workforce point in different directions at the same time, and where a deployment that clears procurement can still fail on the floor.

A five-point scale that replaces the automate-or-not question#

The instrument underneath all of this is the Human Agency Scale, which the authors introduce as "a shared language to quantify the preferred level of human involvement". Its five levels, in their wording:

• H1. "AI agent handles the task entirely on its own."
• H2. "AI agent needs minimal human input for optimal performance."
• H3. "AI agent and human form equal partnership, outperforming either alone."
• H4. "AI agent requires human input to successfully complete the task."
• H5. "AI agent cannot function without continuous human involvement."

The authors are careful about how to read it: "Importantly, higher HAS levels are not inherently better, different levels suit different AI roles." H1 and H2 favour automation approaches; H3 to H5 favour augmentation. Across 104 occupations, H3 was the dominant worker-desired level in 47 of them, which is 45.2 per cent. Editors were the only occupation where workers predominantly wanted H5. Experts put only mathematicians and aerospace engineers there.

A scale of this kind does something a binary cannot. It lets an organisation say that a task is being automated to H2 rather than automated, and it makes the difference between H2 and H3 a decision somebody has to make and own rather than an emergent property of a procurement. That is the distinction the estate's delegation boundary map was built for, now with a published measurement behind it.

A quarter of the ratings match, and the mismatches run one way#

Each of the 844 tasks carries two Human Agency Scale ratings, one from the workers who do it and one from a panel of 52 AI researchers and practitioners. They match on 26.9 per cent of tasks. On 47.5 per cent, the worker wants more human involvement than the expert judges technically necessary. The authors summarise it plainly: "workers generally prefer higher levels of human agency than what experts deem technologically necessary".

Where the gap is widest tells you when it will bite. The authors report that "disagreements are most pronounced in the lower HAS range", and that five of the ten occupations with the largest divergence are also occupations experts rate as predominantly H1, which is full autonomy. Friction of this kind concentrates in the roles that current capability assessments say need no human at all, rather than spreading evenly across the economy.

Almost every organisational decision about what to automate is made using the expert number and never the worker number, because the expert number is the one a vendor supplies. Where the two diverge by two scale points, somebody in that organisation is going to be asked to hand over a task they believe requires them, on the authority of an assessment they have never seen.

Enjoyment is a capability signal, not a perk#

The enjoyment correlation, at -0.28, is the finding most likely to be dismissed. Read as sentiment, it says people want to keep the nice bits, which is not a reason to design an operating model around them. Read against the rest of this research, it says something harder to wave away.

The tasks people report enjoying in professional work are disproportionately the ones with a judgement in them: the diagnosis rather than the write-up, the argument rather than the formatting, the decision about what the client actually needs. Those are also the repetitions that build the capability the organisation is paying for, which is the argument of missed reps and the mechanism behind capability debt. An automation programme that follows enjoyment ratings downward is, by accident, following capability formation downward too.

There is a shape to what that produces, and I named it in 2025 as zombie work: the threat after automation is not being removed but being hollowed out in place, with the judgement and the creativity taken out of a role while the person is still required to be present, converting decision-makers into approvers. WORKBank is the first dataset that lets you see it forming before it forms, because it asks people task by task rather than about their job as a whole. A role can lose every H4 task it had and keep its title, its headcount line and its salary band, and no report in the organisation will show the change.

The paper's own signal points the same way. Comparing tasks by their required human agency against the wages associated with their core skills, the authors find "traditionally high-wage skills like analyzing information are becoming less emphasized, while interpersonal and organizational skills are gaining more importance", and describe this as an early signal rather than a measured shift. What is being repriced is information handling. What is being asked for is the part of the job that involves other people.

Preferences are evidence, and they are not a veto#

A page that treated worker preference as decisive would be making the same mistake in reverse as the deployments it criticises. Three counterweights belong here, and two of them come from the authors themselves.

Workers may be wrong about the technology. The paper states that "domain workers may still lack full awareness of the evolving capabilities and limitations of AI agents", and mitigates it partially by requiring at least ten respondents per occupation and by pairing every rating with an expert one. Workers may also be wrong about their own interests: a preference to keep a task tells you what somebody wants, and nothing about whether the organisation or the person benefits from their keeping it.

Workers may not be answering honestly. The authors say so directly: "some workers may also withhold honest feedback due to concerns about job security or surveillance". Anyone running this exercise inside their own organisation should assume that effect is larger than it was on Prolific, Upwork and LinkedIn, where the respondent had no employer reading the answers.

And the scale of the thing being decided is smaller than the discourse implies. The US Census Bureau's 2026 AI supplement found that among firms using AI, 66 per cent use it solely to augment tasks, and that AI-related employment decreases occurred in 2 per cent of firms. Most organisations are not yet in the position this page describes. They are close enough to it to decide their policy before they are.

Running the disagreement instead of overriding it#

• Rate the tasks, not the tools. The unit that produces a usable answer is the task somebody actually performs. Ask the people who perform it, one task at a time, and ask them to consider the consequences before they answer, which is the design choice that makes the 46.1 per cent figure meaningful rather than reflexive.
• Collect both numbers. A worker-desired level and a capability assessment. One number tells you nothing, because the finding here is the gap between them.
• Treat a two-point gap as an agenda item, not an obstacle. Where a worker says H4 and an assessment says H1, one of two things is true: the assessment is optimistic about a task with judgement in it, or the worker is protecting something. Both are worth an hour of somebody's time before a rollout, and neither is visible afterwards.
• Separate the trust objection from the preference objection. Distrust of accuracy was the single largest concern at 45.0 per cent. It is testable. Run the task both ways on real cases and publish the error rate. An objection that survives the test is a finding; one that does not is resolved, and cheaply.
• Check what a role has left. Count the H4 and H5 tasks in a job before automating its H1 and H2 ones, and check the count again a year later. A job that has lost all of them has been hollowed out whatever its grade says, and the people supervising that work will be the ones who notice last.

Where this sits in my own argument#

"This Is Zombie Work" (2025) put the position that the risk after automation is being hollowed out in place rather than removed, and that naming it is the precondition for reclaiming any agency in the role. "Invisible Work" (2025) made the adjacent case that the checking and quiet correction holding an organisation upright have never appeared on any measure of output, which makes them the easiest to cut. Both were arguments from observation. WORKBank supplies the measurement they lacked, and it points the same way: the tasks people defend are disproportionately the ones no reporting system records.

Attribution note. The Human Agency Scale, WORKBank and the four zone names are Shao and colleagues' terms, not mine. Zombie work is mine, first published in 2025. Missed reps is mine. Capability debt I have used since June 2025 and make no claim of first use on, since the phrase is in independent use elsewhere. Augmentation and automation are established vocabulary and belong to nobody.

Preference is not a prediction about skill#

It does not claim that worker preference predicts anything about capability retention. WORKBank measures what people want and what experts think possible. Both are stated positions rather than outcomes, and nothing in it measures what happens to a person's skill after a task is automated. The link drawn above between enjoyment and judgement-bearing work is this research's interpretation. The paper does not test it.

It does not claim the figures generalise beyond the sample. The audit covers 104 occupations, which the authors describe as "a subset of the 287 computer-using occupations identified with the O*NET database", recruited through Prolific, Upwork and LinkedIn between January and May 2025. It is US-only, and a preprint rather than a peer-reviewed article.

It does not claim the capability assessments are correct. They come from 52 AI researchers and practitioners, with inter-annotator agreement reported as a Krippendorff's alpha of 0.539 for automation capability and 0.511 for the agency level, which the authors publish rather than hide. Expert panels have been wrong about capability timelines in both directions.

It does not claim the picture is stable. The authors say their snapshot "reflects the present state of generative AI and agentic systems as of early 2025" and that future iterations will be needed. A ratings exercise run today would produce different capability numbers and, quite possibly, different desires.

Key sources

• Shao, Y., Zope, H., Jiang, Y., Pei, J., Nguyen, D., Brynjolfsson, E. and Yang, D. (2026). Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce. arXiv:2506.06576, v3 revised 1 February 2026. Read at arxiv.org/abs/2506.06576.
• Bonney, K., Breaux, C., Dinlersoz, E., Foster, L., Haltiwanger, J. and Pande, A. (2026). The Microstructure of AI Diffusion. US Census Bureau, CES Working Paper 26-25.
• Autor, D. and Thompson, N. (2025). Expertise. Journal of the European Economic Association, 23(4).
• US Department of Labor. O*NET OnLine, the occupational task database the audit is built on.

Related SuperSkills research#

On the allocation itself, the delegation boundary map and what stays human. On the consequence for capability, missed reps, capability debt and how humans learn with AI. On the oversight that survives automation, who supervises work they cannot do and the invisible work of oversight. On agents specifically, AI agents and human judgement and who manages AI agents.

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. Every figure on this page was read in the paper itself rather than in reporting of it. Two numbers that circulate with this study are absent here on purpose: there is no published count of tasks per zone, and no named list of the occupations with the highest and lowest automation desire, because those sit in appendices that were not part of the text read for this page. The zone definitions and the scale levels are quoted rather than paraphrased, because paraphrasing a five-point scale is how it stops meaning anything.

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