AI workforce strategy

Two facts, both well evidenced, and the gap between them is where workforce strategy actually lives. Adoption has been extraordinarily fast: by late 2024, around forty percent of working-age American adults were using generative AI and nearly a quarter of employed people had used it for work in the previous week, making work adoption as rapid as the personal computer and overall adoption faster than the internet. And the labour-market effect, measured properly, is so far close to nothing: linked survey and administrative data covering some 25,000 Danish workers found precise null effects on earnings and hours two years after ChatGPT launched. Nothing had happened to pay. But underneath, the structure of work had already moved, through task reorganisation and entirely new tasks in AI oversight and integration. That is the whole strategic problem in one sentence. The work is being reshaped now and the numbers most executives are watching will not show it for years. A workforce strategy built on waiting for the productivity data is a strategy to be late.

What the evidence shows

On adoption, Bick, Blandin and Deming ran nationally representative US surveys and found the speed unusual by historical standards. As of late 2024, nearly forty percent of the population aged 18 to 64 used generative AI in some form, twenty-three percent of employed respondents had used it for work at least once in the previous week and nine percent used it every working day. They also found something that ought to temper the transformation rhetoric: between one and five percent of all work hours were being assisted by generative AI, with reported time savings equivalent to about 1.4 percent of total work hours. Enormous reach, thin penetration into the actual hours.

On outcomes, Humlum and Vestergaard produced the most rigorous available reading by linking large-scale adoption surveys to administrative labour records in Denmark, across roughly 25,000 workers in 7,000 workplaces in eleven exposed occupations including accountants, journalists, legal professionals, software developers and marketers. Two years after the launch of ChatGPT, using difference-in-differences, they estimate precise null effects on earnings and recorded hours at both worker and workplace level, ruling out effects larger than two percent. What moved instead was the structure of work: employers absorbed AI through task reorganisation, new tasks appeared in content generation, AI oversight and AI integration, and adopters transitioned into higher-paying occupations. Their own summary is the line every executive team should hear: technological change reshapes work well before it surfaces in earnings or hours.

On who gains, Brynjolfsson, Li and Raymond, studying 5,179 customer-support agents, found productivity up fourteen percent on average, thirty-four percent for the newest and least experienced staff, and barely moving for the most skilled. AI raises the floor far more than it raises the ceiling. On where value ends up, Autor and Thompson, analysing four decades of task data across 303 US occupations, showed that what matters is which tasks are removed: automation that stripped out the less expert tasks raised wages, and automation that stripped out the expert tasks lowered them. The same technology, applied to two different jobs, produces opposite consequences for the people in them.

And on the design of human-machine work, the most important recent result is a warning against the default. Vaccaro, Almaatouq and Malone, in a 2024 meta-analysis of 106 experimental studies and 370 effect sizes, found human-AI combinations performing significantly worse on average than the better of human or AI alone, with losses concentrated in decision-making tasks. Putting a person in the loop is not a control. It is a design choice that can subtract.

Employers, asked what they need, are consistent. 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 largest barrier to business transformation over the next five years.

Where the evidence is uncertain

Denmark is a high-trust, high-wage, heavily unionised labour market with strong employment protection. Null wage effects there do not license a confident forecast for a US technology company or a UK professional-services partnership, and the authors do not claim they do. Two years is also early for a general-purpose technology; the historical pattern for electricity and computing was a long lag between adoption and measured productivity, which is an argument for patience rather than for complacency.

The adoption surveys are self-reported and were run in 2024, so the figures are already conservative. The Autor and Thompson task data runs to 2018, making their framework a lens for thinking about generative AI rather than a measurement of it. And the meta-analysis draws on studies published between 2020 and mid-2023, before the current model generation, which probably shifts the balance of competence further towards the machine rather than away from it.

The largest uncertainty is the one nobody can resolve yet. No dataset measures what happens to organisational capability over five or ten years of AI-assisted work, because five years have not passed. That is the horizon on which the argument below either proves prudent or proves overcautious, and anyone claiming to know which is guessing.

The SuperSkills interpretation

Most organisations do not have an AI workforce strategy. They have a licence count and a training day, and they have mistaken the two for a plan. I call the resulting behaviour usage theatre: activity that produces the evidence of transformation, dashboards of seat adoption and prompt volumes, without producing any change in how work is designed. It is measurable, reportable and almost entirely disconnected from value, which is exactly why it is popular.

The Humlum finding is the one to build strategy around, because it separates two things that get conflated. Adoption is nearly free and is happening anyway. Redesign is expensive, slow, political, and is the only thing that converts adoption into either value or damage. Organisations that stop at adoption get the tool without the benefit. Organisations that redesign the doing without redesigning the learning get the benefit and accumulate capability debt, which is the loss of human knowledge, skill and judgement that builds when work is automated faster than the ways people learn through doing it are rebuilt. That debt does not appear on any dashboard, because the outputs still look fine, right up until a decision arrives that the AI cannot make and nobody in the room has been trained to.

Which is why I think readiness scores measure the wrong thing. Counting tools, pilots, policies and training completions tells you about activity. It tells you nothing about whether the organisation has decided where human judgement must remain, whether juniors can still build capability, or whether anyone is accountable for a decision the machine now shapes. That argument is set out in the AI readiness lie, and the underlying distinction is drift versus design: most organisations arrive at their AI configuration through a thousand small decisions nobody quite made, and then describe the result as a strategy.

Autor and Thompson give the strategy its actual content, and it is a more uncomfortable brief than most workforce plans contain. For every role you are changing, ask whether removing the automatable tasks makes what remains harder or easier. Where it makes the job harder, you have created a more valuable role and you now need fewer, better, better-paid people, and a way to develop them. Where it makes the job easier, you have commoditised a role your organisation may depend on, and you should expect wages, standards and retention to follow. Nobody enjoys running that analysis. It is the difference between a workforce strategy and a communications plan.

The four decisions a real workforce strategy makes

Not a maturity model. Four decisions, each of which has a right answer specific to your organisation, and each of which is currently being made by default if it is not being made deliberately.

• Where human judgement must remain, and why. A written list of the decisions a person has to own, with the reasoning. An organisation that has never written this list has already answered by default, one busy afternoon at a time.
• How people will still learn. If juniors no longer do the work through which judgement was built, name what replaces it. This is the single most neglected decision in AI adoption and the most expensive to reverse. See how humans learn with AI and the missing rungs.
• Which way each role is moving. Apply the expertise test role by role. Harder means fewer and better; easier means commoditising. Both need a plan, and they are different plans.
• What you will measure. Adoption metrics are the ones you already have and the ones that mean least. Add at least one measure of capability without the tool, and one measure of how often humans genuinely disagree with machine output. Both go down when things are going wrong, which is why nobody wants them.

What to do

Redesign work, not just access. Licences are the cheap part and the part that does nothing on its own. The Danish evidence is that value, where it appears, comes through task reorganisation and new tasks, which are management decisions rather than procurement decisions.

Put a named executive on capability, not just on adoption. Somebody senior should be answerable for whether the organisation can still do the things it has automated. In most companies nobody owns this, which is precisely why it erodes.

Protect the development pathway explicitly. Decide which work juniors still do unaided and defend it in the redesign, because it will not survive a productivity review otherwise. It is slower this quarter and it is the only source of your 2032 leadership.

Do not put a human in the loop and call it governance. Specify who, at what point, with what authority to stop the process, and measure the disagreement rate. See human and AI decision making.

Stop waiting for the productivity number. It is the slowest indicator available and it will move long after the decisions that determine your position have been taken.

Development of the idea

In CEOWORLD in July 2026 I set out the drift versus design argument directly, including the four postures organisations take, the Sleepwalkers, the Programmed, the Stuck and the Designers. In Irish Tech News in July 2026 I made the usage-theatre case that you are not adopting AI, you are paying for it. The Box of Amazing essay The Architecture of Drift develops the drift versus design matrix, and The Great Unbundling of Work (25 May 2025) sets out the task-level redesign the strategy depends on. The full framework is in SuperSkills (Kogan Page, 2026).

Key research and primary sources

• Bick, A., Blandin, A. and Deming, D. J. (2024). The Rapid Adoption of Generative AI. NBER Working Paper 32966.
• Humlum, A. and Vestergaard, E. (2025). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777, revised March 2026.
• Autor, D. and Thompson, N. (2025). Expertise. NBER Working Paper 33941; published in the Journal of the European Economic Association, 23(4).
• Vaccaro, M., Almaatouq, A. and Malone, T. (2024). When combinations of humans and AI are useful: a systematic review and meta-analysis. Nature Human Behaviour, 8, 2293-2303.
• Brynjolfsson, E., Li, D. and Raymond, L. (2023). Generative AI at Work. NBER Working Paper 31161; published in the Quarterly Journal of Economics, 2025.
• World Economic Forum (2025). The Future of Jobs Report 2025.

Related SuperSkills research

On measuring the right things, the AI readiness lie and usage theatre. On the underlying posture, drift versus design and why AI transformation is not a change-management problem. On the capability consequences, capability debt, the missing rungs and how humans learn with AI. For the HR-specific version, the CHRO guide to AI. The executive version of this is how leaders should respond to AI.

About this research

Rahim Hirji is the author of SuperSkills: The Seven Human Skills for the Age of AI (Kogan Page, 2026) and the 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. Drift versus design, capability debt, usage theatre and the missing rungs are part of the SuperSkills lexicon. This is a living reference, reviewed and updated as significant new evidence appears.