Will AI Replace Jobs? What the 2028 Memo Gets Wrong

A thought experiment about whether AI will replace jobs has been circulating this year, written as a memo dated June 2028. In it, AI-driven displacement of white-collar workers sets off a deflationary spiral: unemployment at 10.2 percent, the S&P 500 down 38 percent, and what the authors call Ghost GDP — output that is produced but never circulates through the real economy as wages and spending. It is well constructed. It is also, measured against the evidence now available, wrong in an instructive way.

We assembled what the research actually shows, drawing on the World Economic Forum, the IMF, MIT Sloan, Harvard Business School, Stanford, Yale's Budget Lab, Goldman Sachs Research, Anthropic, and our own analysis of enterprise hiring. These are independent bodies using different methods on different populations. They converge on one conclusion: AI changes the composition of work far more than it reduces the demand for it.

That is not an optimistic reading. It carries a harder implication for anyone running a workforce, which I come to at the end. But the starting point has to be the numbers, because the case that AI will replace jobs is argued with numbers and deserves to be answered with them.

Will AI replace jobs? What the aggregate data says

The WEF's Future of Jobs Report 2025 draws on more than 1,000 employers representing 14 million workers across 55 economies. It projects 170 million new jobs created and 92 million displaced between 2025 and 2030 — a net gain of 78 million, roughly 7 percent of today's employment. For every job displaced, close to two are created, and more than 35 million of the additions sit in the care economy alone.

The current data does not show a crisis either. Global unemployment stood at 4.9 percent in 2025, the lowest level since 1991 on ILO figures. Yale's Budget Lab, tracking AI's labour market footprint through quarterly Current Population Survey releases, finds that the broader labour market has not experienced a discernible disruption since ChatGPT's release, and that measures of AI exposure, automation and augmentation show no sign of being related to changes in employment. Their summary is blunt: if AI is roiling the job market, the data isn't showing it.

History points the same way. Stanford's Charles Jones shows that US real GDP per capita has grown at roughly 2 percent annually for 150 consecutive years, through electrification, the transistor and the internet. David Autor's work, cited in Goldman Sachs analysis, finds that 60 percent of workers today hold occupations that did not exist in 1940. Goldman projects that generative AI will permanently raise global GDP by 7 percent.

What enterprise hiring actually looks like

Aggregate projections are useful but distant. The more immediate evidence sits in what large employers are asking for. Draup's January 2026 analysis of more than one billion global job descriptions across Fortune 500 companies shows AI capability spreading into every business function — and spreading fastest precisely where the displacement narrative expects contraction.

Year-on-year growth in AI skill mentions ran at 24.8 percent in customer support and service, 23.6 percent in sales and marketing, 23.0 percent in industrial manufacturing, 21.3 percent in financial services operations, and 19.3 percent in human resources. Information technology grew at 9 percent, not because AI is retreating there but because it arrived first and is now normalising.

The second finding is the one I would put in front of a board. Skills density — the average number of distinct skills a job description asks for — is rising across every major function. Data engineering and analytics moved from 15.22 to 19.15 skills per posting. Software engineering moved from 14.52 to 18.29. Risk, compliance and governance moved from 10.12 to 13.76. Automation simplifies roles. What we are observing is the opposite: roles are getting harder, because workers are expected to run AI tools alongside their existing expertise.

The third finding explains where the new work is. Fortune 500 demand is concentrating on operating AI rather than building it. AI governance and model risk skills grew 81 percent year on year, the fastest-growing cluster in enterprise hiring, and cost optimisation and margin protection grew 77.6 percent. These are expert human roles that exist only because AI is being deployed at scale, and they cannot themselves be automated.

Wages are rising for the workers who adapt

The displacement thesis predicts wage compression. The evidence runs the other way. The IMF's January 2026 Staff Discussion Note, covering vacancy data across six economies, documents a wage premium of 3.0 to 3.4 percent at the job-posting level in the US and UK for roles requiring new skills; in US local labour markets, each one-percentage-point increase in the new-skill share is associated with 2.3 percent higher wages and 1.3 percent higher employment.

The most striking result comes from Brynjolfsson, Li and Raymond, who studied a generative AI assistant deployed across 5,179 customer support agents. Average productivity rose 15 percent. Novice and low-skilled workers gained 34 percent. Already-expert workers gained close to nothing. The mechanism they identify is that the model disseminates the best practices of the most able workers, democratising expertise that previously took years to acquire. The study also found improved customer sentiment and higher retention.

Employers are behaving accordingly: in the WEF survey, 52 percent expect to allocate a greater share of revenue to wages against 7 percent expecting a decline.

Collaboration is winning over delegation

Anthropic's Economic Index, analysing anonymised Claude interactions in November 2025, found that augmented interactions — where a human learns, iterates and gets feedback — accounted for 52 percent of usage against 45 percent fully automated, with the augmented share rising 5 percentage points in three months.

MIT Sloan's EPOCH study analysed 19,000 tasks across 950 job types and identified five capabilities that resist automation: empathy, presence, opinion and judgment, creativity, and hope. All five are associated with employment growth, and human-intensive tasks increased between 2016 and 2024. Harvard Business School Working Paper 25-039 sharpens this: automation-prone roles see falling demand and simplified skills, while augmentation-prone roles see rising demand, complexity and pay. Our skills density data is that finding at enterprise scale.

Accenture's Julie Sweet put it well at the India AI Impact Summit in February 2026, arguing for humans in the lead rather than humans in the loop. Her company's C-suite survey across 20 countries found 78 percent of executives naming growth, not cost reduction, as AI's greatest benefit.

The constraint is workers, not jobs

Here is the uncomfortable part. Our Global AI Report puts the worldwide pool of workers with the requisite AI skills at 2.2 million, the US leading at 310,000 — a fraction of what Fortune 500 employers alone need. Separately, 73 percent of companies report difficulty building quality talent pipelines, and the half-life of technical skills has fallen below two years, with roughly 40 percent of current tech skills expected to be partially obsolete by 2027.

The defining workforce problem of this decade is therefore not that AI will replace jobs. It is that too few people are prepared to work in AI-augmented environments — and that problem lands on the CHRO. Five priorities follow:

• Redesign jobs around human-AI collaboration, allocating routine analysis to systems and reserving judgment and accountability for people.
• Build AI literacy as a baseline workforce capability rather than a specialist one.
• Plan capabilities rather than headcount, since role counts no longer describe capacity.
• Treat learning velocity as a strategic capability, because shrinking skill half-lives make periodic training inadequate.
• Expand governance talent deliberately, given that model risk and responsible AI oversight are the fastest-growing demand we observe.

What this means for the next planning cycle

Organisations that treat AI mainly as a cost-reduction technology will capture a fraction of what is available, and will do it while competitors absorb the scarce workers who know how to operate these systems. The evidence does not support waiting for clarity. It supports moving budget now toward reskilling, job redesign and governance capacity.

None of this is guaranteed. Daron Acemoglu's point stands: AI's labour market impact is a set of choices about costs, organisational design and policy, not a fixed destiny. Firms can invest in task creation or in task substitution, and the outcomes diverge sharply. The Fortune 500 hiring data suggests the largest employers are so far choosing task creation. The question for any individual enterprise is whether it is making that choice deliberately or by default.