Staying valuable in the age of AI

Your job is being unbundled into tasks rather than deleted, and value is moving to whichever tasks still require expertise. That single reframing changes the career question from an unanswerable one, will AI replace me, to an answerable one: which of my tasks will AI take, and does removing them raise or lower the expertise required by everything left? The economics on this is unusually clear, and asymmetric. When automation strips out the less expert parts of a job, the people who remain get paid more. When it strips out the expert parts, wages fall even as employment rises, because the job is now open to more people. So learning to use the tools is table stakes, commoditising by the month, and no answer at all. What works is deliberately moving your working time towards the parts of your role that machines make more consequential rather than less, and being able to prove you can do them.

What Autor's work on tasks tells you#

The most useful study for anyone thinking about their own career is not about AI at all. Autor and Thompson, in work published in the Journal of the European Economic Association in 2025, analysed four decades of task data across 303 US occupations and asked what happens to the value of the labour that remains when tasks are automated. Their answer is that it depends entirely on which tasks go. Automation that removed the inexpert tasks from an occupation raised wages and reduced employment: the job became harder and the people doing it became more valuable. Automation that removed the expert tasks lowered wages and raised employment: the job became easier and more people could do it. The same technology, in two occupations, produces opposite outcomes for the humans, and which one you get is determined by what is left rather than by how much was taken.

The best current data on generative AI specifically is a caution against panic and against complacency simultaneously. Humlum and Vestergaard linked large-scale adoption surveys to administrative labour records in Denmark, covering around 25,000 workers across 7,000 workplaces in eleven exposed occupations including accountants, journalists, legal professionals, software developers and marketers. Two years after the launch of ChatGPT they found precise null effects on earnings and hours, ruling out effects larger than two percent. Nothing had happened to pay. But underneath that flat surface the structure of work had already moved: employers were absorbing AI through task reorganisation, new tasks in content generation, AI oversight and AI integration were widespread, and adopters were transitioning into higher-paying occupations. Their own phrase for it is that technological change reshapes work well before it surfaces in earnings or hours. If you are waiting for the salary data to tell you what is happening, you are reading the slowest available indicator.

Two further findings matter for where value sits. Brynjolfsson, Li and Raymond, studying 5,172 customer-support agents, found AI assistance raised productivity by fifteen percent on average, by thirty percent for the newest and least experienced staff, and barely at all for the most skilled. AI raises the floor far more than the ceiling, which compresses the visible gap between a novice and an expert and makes it much harder to tell them apart from the output. And Dell'Acqua and colleagues, in the 2023 experiment with 758 consultants that produced the jagged technological frontier, found that on a task just outside the model's competence, consultants using AI did worse than consultants with none, because they trusted confident output they should have challenged. Knowing where the frontier runs turns out to be worth more than being able to operate inside it.

Employers, asked directly, say something consistent with all of this. The World Economic Forum's 2025 Future of Jobs report names analytical thinking as the most valued core skill and identifies skills gaps as the single biggest barrier to transformation over the next five years.

How far the Danish result travels#

Denmark is a high-trust, high-wage, heavily unionised labour market with strong employment protection, and two years is early. Null effects there do not license a confident forecast for a US technology firm or a UK professional-services partnership, and the authors do not claim otherwise. The Autor and Thompson data run from 1980 to 2018, which means their model is derived from earlier waves of automation. It is the best conceptual tool available for thinking about which tasks matter. It does not measure generative AI.

The entry-level picture is genuinely contested rather than merely uncertain, with reasonable people reading the same hiring data in opposite directions; that argument is set out separately in will AI replace entry-level jobs. And the honest limit on all of it is that these studies measure short-run performance and earnings. What they cannot yet measure is how the capability of a workforce changes over five or ten years of habitual AI use, which is the horizon on which the advice below actually pays off or fails.

Unbundling, and what it does to a role#

The frame that does the work is unbundling. AI is not coming for jobs as units; it is shredding roles into tasks, swallowing some, stretching others, and leaving a smaller human core under pressure to justify itself. I wrote about this in The Great Unbundling of Work in May 2025, and the sentence people kept returning to was that your job is not disappearing, it is dissolving. The consequence nobody plans for is hollowing rather than unemployment: expertise that took twenty years to build being commoditised out from under someone while their job title stays exactly the same.

Put Autor and Thompson next to that and you get the practical question, which is uncomfortable enough that most people avoid it. Look at your week as tasks rather than as a role. Which of them could a capable model do adequately today? Now ask the second question, the one that actually determines your position: once those are gone, is what remains harder than your job is now, or easier? If harder, AI is making you more valuable and you should accelerate it. If easier, AI is opening your job to a much larger pool of people and no amount of tool fluency will protect you. It is an uncomfortable exercise and considerably more useful than a list of future-proof skills.

It is also why I think the standard advice is weak. "Learn to use AI" describes a capability that is being deliberately engineered to require less skill every quarter, and building a career on operating an interface that is getting easier is a losing position. The durable version is different in kind: be the person who can tell when the output is wrong, who can decide what should have been asked, and who can be accountable for the result. That is the verifier's role, and organisations systematically underpay it because verification looks like checking rather than like producing. That mispricing will not last. It is the arbitrage available to anyone paying attention. I made the same argument in the Observer in July 2026 about the next power class: it will not be the people who build the models.

There is a trap on the other side. It catches ambitious people early in a career. AI lets you produce work that looks like it carries fifteen years of judgement when it carries eighteen months, and the gap does not show until a decision arrives that the model cannot make. I call this synthetic seniority. It is a fast way to get promoted and a slow way to become unemployable, because you arrive in a senior role having skipped the reps that were supposed to prepare you for it. Staying valuable over a career and looking valuable this quarter are different projects, and AI has made the second one dramatically easier without touching the first.

The argument on the other side#

David Autor's Applying AI to Rebuild Middle Class Jobs (2024) is the strongest counter to the caution on this page. His argument is that AI, unlike earlier automation, can extend expert decision-making to people who do not yet hold the expertise, potentially restoring middle-skill work rather than hollowing it further. If that holds, the career advice inverts: lean into the tool, because it is the thing raising what you are able to do.

My reading is that Autor's mechanism and the argument here are compatible under exactly one condition. It is the condition nobody is measuring. Extended expertise has to be acquired rather than borrowed. Where the support builds judgement over time, the person genuinely climbs. Where it substitutes for judgement that never forms, they are exposed the moment the situation moves outside what the model handles. Both futures are available from the same technology, and which one you get is a design question rather than a forecast.

Two further sources sharpen the practical version. Cui and colleagues, across three randomised experiments with 4,867 developers at Microsoft, Accenture and a Fortune 100 firm, found completed tasks rose 26 percent, with the largest gains among the least experienced, though the estimate carries a standard error of 10.3 percent and should be quoted with it. And Agrawal, Gans and Goldfarb's Prediction Machines gives the cleanest economic statement of why judgement appreciates: AI reduces the cost of prediction, and when prediction becomes cheap, the value of its complements rises. Judgement is the complement.

Unbundle your own role first#

Unbundle your own role before someone else does. Write down everything you actually did last week as discrete tasks, then mark each one: the machine can do this now, the machine will do this soon, this needs a human. Most people have never seen their job written this way and find the exercise unsettling. The discomfort is part of it.

Ask the expertise question, task by task. For each thing AI takes, does the remaining work get harder or easier? Move deliberately towards the harder side. That may mean taking on the ambiguous, judgement-heavy, politically awkward work that nobody wants. That is the work that keeps its value.

Learn where the frontier runs in your own domain. Not in general. In your work. Know the specific categories of problem where the model sounds most confident and is most often wrong, because that knowledge does not transfer, does not commoditise, and is what the jagged-frontier result says people lack.

Make your judgement legible. If output quality no longer proves capability, you need another way to demonstrate it: decision logs, recorded reasoning, the annotated draft showing what you prompted and what you changed and why. This is how you avoid being priced as though the model did it, because increasingly the assumption will be that it did.

Keep taking the reps you could now skip. The capability that carries your value should be exercised unaided at intervals, deliberately, for the same reason pilots fly manual approaches. See using AI without dependency for the practice and how humans learn with AI for the evidence underneath it.

Development of the idea#

The Great Unbundling of Work (25 May 2025) set out the tasks-not-jobs argument and the identity crisis underneath it. Knowledge Is No Longer Power developed the argument about where the human edge moves once information is free. In the Observer, The Next A.I. Power Class Won't Build the Models (17 July 2026) made the case for judgement as the scarce asset, and in Entrepreneur UK I argued in July 2026 that AI does not create bad decisions, it exposes them faster. It is developed in full in SuperSkills (Kogan Page, 2026).

Key research and primary sources

• Agrawal, A., Gans, J. and Goldfarb, A. (2018). Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Review Press, updated edition 2022.
• Cui, Z. K., Demirer, M., Jaffe, S., Musolff, L., Peng, S. and Salz, T. (2025). The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. Management Science, 2025.
• Susskind, R. and Susskind, D. (2015). The Future of the Professions: How Technology Will Transform the Work of Human Experts. Oxford University Press, updated edition.
• Autor, D. and Thompson, N. (2025). Expertise. NBER Working Paper 33941; published in the Journal of the European Economic Association, 23(4), 1203-1271.
• Humlum, A. and Vestergaard, E. (2025). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777, revised March 2026.
• Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889-942. Earlier version NBER Working Paper 31161.
• Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier. Harvard Business School and BCG working paper.
• World Economic Forum (2025). The Future of Jobs Report 2025.
• Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking. Microsoft Research and Carnegie Mellon, CHI 2025.

Related SuperSkills research#

On which capabilities to build, human skills in the age of AI and what stays human. On the career traps, synthetic seniority, the verifier's discount and the missed reps. On the pipeline and early careers, will AI replace entry-level jobs and the missing rungs. On the organisational version of the same problem, capability debt. On the prior question of exposure itself, see will AI replace my job. On why tool fluency is the wrong thing to build a career on, see why "learn to prompt" is weak career advice. The graded evidence is in the evidence base. See which jobs are safest from AI.

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. This work draws on research across more than 200 organisations in 30 countries over seven years. Findings are attributed to the studies that produced them and kept separate from the interpretation, which is the author's. The unbundling of work and the task-versus-job distinction are widely discussed in labour economics and are not his coinages; synthetic seniority, the verifier's discount, the missed reps and capability debt are part of the SuperSkills lexicon. This is a living reference, reviewed and updated as significant new evidence appears.

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