What Should the Work Become?

A few months ago, I published the Four-Question Role Screen. I was trying to get past the broad predictions about AI and jobs and give management teams a straightforward way to examine the work inside their own companies. What makes up the job? Would cheaper output create more demand? Does automating the whole role make economic sense? What is competition forcing you to do? [1]

I still think those questions are useful and straightforward to implement. But looking back at the argument, there is something I need to clarify, and something else that I didn’t take far enough.

I wrote that when AI takes over some tasks in a complex job, the worker becomes more valuable on what remains. I should have written can become more valuable. [1]

The focus effect that Joshua Gans and Avi Goldfarb describe depends on what happens to the remaining work. More time can improve it. Lower costs can attract more customers. But neither guarantees the same headcount. As Gans explains in his July essay, The Last Ten Per Cent, people with the expertise to do the remaining work can command higher pay even as fewer people are needed. [2]

That is why I’m not comfortable stopping at “AI frees people for higher-value work.” What work? How much of it exists? Who needs it?

Then there is the question underneath all of that. We are starting with the jobs the company has today. What if AI changes what the company should be doing in the first place?

The organization is part of the question

There is a lesson from electrification worth going back to. In the early transition, a factory could replace its steam engine with electric power while keeping the shafts and belts that drove the machinery. It had changed the power source without changing much of the factory. [3, 4]

Individual electric motors opened up a different possibility. Machines could be arranged around the flow of production rather than the old power system. Electrification did bring energy savings, but Warren Devine’s research identifies the more important benefit as greater output from the same capital and labor. Realizing the fuller gains involved changing how production was organized. [3, 4]

I wonder whether we are carrying too much of the old organization into our thinking about AI. We take the existing departments, put better tools in each one, and ask everyone to be more productive. That may help. It may also leave us with the same unnecessary handoffs, just happening faster.

A 2026 working paper by Mert Demirer and colleagues makes this practical. When AI can handle several consecutive steps, those steps may be combined instead of passed between people. The sequence and the handoffs affect the economics. The boundaries between jobs become part of the decision. [5]

But even that model holds the underlying production steps fixed. It examines how to reorganize them, not how entirely new steps emerge. Useful, but it leaves another question open. [5]

What else could the business do?

Take a hypothetical architecture practice. Suppose AI reduces the effort involved in first-pass drawings and revisions. The practice could spend more time on difficult design problems. It could take on smaller projects that previously couldn’t support its fees. Or, with no additional demand, it might need fewer people. Producing more design options would not help if it simply left clients and senior architects with more reviewing to do.

Now suppose the practice asks a different question. Could it use design models and information about how a building actually performs to offer ongoing advice after people move in? Energy use, space, changes as the owner’s needs evolve. It would be testing a continuing service rather than just completing the existing project faster.

For this practice, that could mean a new service or a new group of customers, not necessarily a new occupation. It might require different expertise, responsibilities and commercial terms. The point is to start with a customer problem, not find something to occupy the hours left over.

David Autor and his colleagues examined this in their March 2026 paper, What Makes New Work Different from More Work? Using historical Census records and more recent data through 2023, they found that new occupational roles are associated with specialized expertise and persistently higher pay. New work can create demand for different capabilities, not just more people doing the old job. [6]

That does not tell us which AI-related occupations will emerge, or whether they will arrive in time for displaced workers. Nor does a successful new service necessarily require many additional employees. AI could perform much of it from the beginning. A growth opportunity is worth pursuing without treating it as a promise about jobs.

There isn’t a fixed destination

This is where I would qualify another phrase in my earlier piece: “the parts AI can’t touch.” I would not use that as a permanent description of analysis, coordination or exception handling. [1]

Look at what OpenAI and Anthropic are now trying to demonstrate: systems that use software and work through longer problems involving several steps. They report gains, but those are the developers’ evaluations. METR’s separate testing also shows progress on longer tasks, mainly in software and related technical work. That still doesn’t establish that a model can take over a normal working day inside your company. [7, 8, 9]

I take the direction seriously without treating it as a timetable. A role should not be considered secure just because it contains several complicated tasks. And moving someone from processing information into analyzing it is not, by itself, a durable people plan.

At the same time, waiting for the technology to settle would be a poor reason to do nothing. Use what works, with clear limits. Keep testing the assumptions behind what comes next.

Before the next staffing decision

I would still run the four questions. The scores help organize a management discussion; they are not measured probabilities of job loss. Then I would look beyond the individual role and follow one piece of work through the business. Where does time actually go, including review and correction? Which handoffs are necessary? What improves for the customer when the work changes? [1]

Alongside that, examine something the business has not been able to offer at a viable cost. Talk to customers. Try a limited version. A new service will not arrive with the same evidence of demand as an established one. It needs a test with clear conditions for expanding or stopping, not a requirement to prove its market before anyone tries it.

Keep a record of the model, tools and operating conditions used in the assessment, and what improvement would change the decision. Retest when there is credible evidence of it. A new model announcement is a reason to investigate, not an instruction to change the payroll.

The people plan needs equal attention. In The Invisible Interface, I argue that some early-career work is also how people learn judgment. If we remove it, we need another way for people to practice, get feedback and learn to recognize when the system is wrong. Making someone responsible for checking an answer does not give them the expertise or authority to challenge it. [10]

Some roles may shrink before new opportunities develop. The new work may require different skills, or fewer people than hoped. Employees deserve a clear account of that, not reassurance built around a job nobody has defined.

But I would be equally wary of a plan that calculates the labor savings and never asks what the company could now do differently. Before deciding how many people the business needs, I’d want to see both the case for changing the existing work and a serious test of what else customers might pay the company to do.

References

[1] Glorikian, Harry. “The Four-Question Role Screen.” April 2026. Canonical online version. https://glorikian.com/four-question-role-screen/

[2] Gans, Joshua. “The Last Ten Per Cent: How Task-Level AI Exposure Fits into a Jobs Forecast.” Agglomerations, Economic Innovation Group, July 15, 2026. https://agglomerations.eig.org/p/the-last-ten-per-cent

[3] Devine, Warren D., Jr. “From Shafts to Wires: Historical Perspective on Electrification.” The Journal of Economic History 43, no. 2 (1983): 347–372. https://doi.org/10.1017/S0022050700029673

[4] Wright, Gavin. “Natural Resource Industries.” Historical Statistics of the United States, Millennial Edition Online. Cambridge University Press. See the discussion of electrification and unit drive. https://hsus.cambridge.org/HSUSWeb/toc/showEssayPathIdDb.ESS.01.html

[5] Demirer, Mert, John J. Horton, Nicole Immorlica, Brendan Lucier, and Peyman Shahidi. “Chaining Tasks, Redefining Work: A Theory of AI Automation.” Working paper dated February 22, 2026; arXiv version 1, June 14, 2026. See Section 3, footnote 6. https://arxiv.org/abs/2606.15960v1

[6] Autor, David, Caroline Chin, Anna Salomons, and Bryan Seegmiller. “What Makes New Work Different from More Work?” Working paper, March 15, 2026. MIT Economics: full working paper

[7] OpenAI. “GPT-6 Astra: A New Generation of Intelligence.” September 2026. Developer-reported capabilities and evaluations; accessed September 8, 2026. https://openai.com/index/gpt-6-astra/

[8] Anthropic. “Introducing Claude Fable 5.1 and Claude Mythos 5.1.” September 2026. Developer-reported capabilities and evaluations; accessed September 8, 2026. https://www.anthropic.com/claude-fable-and-mythos-5-1

[9] METR. “Time Horizon 1.1,” January 29, 2026, and “Task-Completion Time Horizons of Frontier AI Models.” See the methods and limitations; accessed September 8, 2026. https://metr.org/blog/2026-1-29-time-horizon-1-1/ https://metr.org/time-horizons/

[10] Glorikian, Harry. The Invisible Interface. Ideapress Publishing, 2026. Chapter 10, “What We Owe Each Other,” particularly apprenticeship and paths to mastery. Publisher: The Invisible Interface