In January 2025 we sided with the sceptics. Consultancies were forecasting AI windfalls of $7 trillion to $25.6 trillion. We backed MIT economist Daron Acemoglu's estimate that only about 5% of tasks could be profitably handed to AI within a decade. By his reckoning, that was worth roughly a 1% lift to GDP and a 0.7% gain in productivity. Our piece, Beyond the Hype: AI Economist Reveals Modest Economic Potential of Artificial Intelligence, treated those figures as a corrective to breathless optimism.
The verdict is split, and we should say so plainly. On the macro number we were right, and the method behind it has held up well. Where we went wrong was in what we drew from it: we let "modest" signal that change inside companies would be slow, and in large, knowledge-heavy firms it has been anything but. The aggregate figure is still small, while the reorganisation underneath it is already well advanced.
The forecast held. The schedule did not.
The sharpest test of Acemoglu's framework came from the Federal Reserve Bank of St. Louis. There, economists Bick, Blandin and Deming used worker-reported adoption and time savings rather than predicted task exposure. They estimate a potential productivity gain of 1.2% based on current generative AI use, which they note is close to Acemoglu's 0.7%. They then point out that Acemoglu "assumed these productivity gains may take a decade to materialize."
So the task-based method passes its first contact with real usage data, and its magnitude looks about right, but its decade-long runway does not. An estimate built to unfold over ten years is already matched, even slightly exceeded, by what current usage implies.
The strongest objection to revising anything goes like this. Usage surveys measure enthusiasm, time savings are self-reported, and a static task-exposure model exists precisely to discipline the excitement of any given year. Two years of adoption data, the argument runs, cannot overturn a ten-year forecast. We accept most of that. We are not revising the 1% figure, and we would still bet against the trillion-dollar projections. But the objection answers a question managers are not asking. A forecast of national productivity describes an average across every firm, including the many that have barely started, and says nothing about how fast change arrives where adoption is concentrated, which is where most of our readers work. The decade-long forecast was a statement about the economy, and too many leaders read it as a schedule for their own company.
The Census Bureau's latest AI supplement to its business survey shows how lopsided that concentration is. Between November 2025 and January 2026, 18% of US firms used AI in a business function, or 32% weighted by employment. Among very large firms in information, professional services and finance, the share reached 50% to 60%. Yet 57% of users had integrated AI into three or fewer business functions, which makes adoption broad and shallow in the aggregate and deep only in particular places.
The decade-long forecast was a statement about the economy, and too many leaders read it as a schedule for their own company.
That pattern explains why both camps feel vindicated. A small national number is compatible with all of the following happening at once:
- Half or more of very large finance, information and professional services firms using AI in a business function, while fewer than one in five firms overall do.
- Entry-level hiring in exposed roles falling well behind, while overall employment barely moves.
- Executives reporting productivity gains that revenue figures only partly confirm.
- Individual teams restructuring around AI in a handful of functions while the rest of the company works as before.
Our original piece treated implementation cost and organisational adaptation as the main brake. In much of the economy they still are. In the firms where it matters most for our readers, that brake has largely come off.
The welfare warning had the right shape and unproven details
The part of our original piece most open to new evidence was its distributional warning. Following Acemoglu, we argued that GDP could grow while welfare fell, with lower-educated women among those most exposed. Evidence since then supports the general concern but not yet its specifics.
Stanford's Digital Economy Lab, using ADP payroll data, finds no evidence of widespread job displacement. Employment of workers aged 22 to 25 in AI-exposed occupations, however, is now 19% below where it would be had it kept pace with less-exposed peers. Experienced workers show no comparable gap. The gap has widened since August 2025, and it works mainly through reduced hiring rather than increased separations. Researchers from the Atlanta and Richmond Feds and Duke, surveying executives, describe AI as substituting for routine clerical work while complementing analytical and managerial roles.
The cost, then, is landing on people doing easy-to-learn tasks and people trying to get in the door. That is consistent with what we warned. But the payroll data speak to age and entry, not gender or education, and we found no credible post-2025 evidence on lower-educated women specifically, so that claim remains a hypothesis. We should also own that our original supporting citations, an unspecified Brookings study and an unnamed McKinsey report, did not carry the weight we put on them.
What "modest" should have meant
We closed the original piece by predicting "incremental, thoughtful integration that genuinely enhances human professional capabilities." At the level of the economy, that is still a fair description. At the level of a large bank or consultancy, it was the wrong frame. A task that slips from a junior analyst to a model does not feel incremental to the graduate who is never hired to do it, or to the manager who discovers, three years later, that there is no one ready for promotion.
For leaders, the practical consequence is that a small macro forecast is no reason to defer organisational decisions. The questions that matter are local and immediate. Which tasks are leaving which roles? What happens to the training ground that entry-level work used to provide? Are reported gains showing up in revenue or only in self-assessment? None of these wait for GDP statistics to move.
We were right on the number, wrong on the clock and half-right on who pays. "Modest" was the correct word for the economy and the wrong instruction for a company. The national figure will probably stay small for years, and by the time it moves, the decisions behind it will already have been made, firm by firm, by leaders who each called their own change incremental. Those decisions concern who gets hired, which skills get built and which tasks stay human.
