Who captures the AI productivity dividend — and does your organisation even know?

Every AI productivity conversation has a gap at its centre, and the gap is measured rather than imagined. When the Federal Reserve Bank of Atlanta surveyed nearly 750 corporate executives, it found perceived gains consistently outrunning measured ones: for 2026, mean reported AI-driven labour-productivity growth was 3.0 per cent against an implied 1.8 per cent; for 2025 the split was 1.8 against just 0.6.¹ Leaders, in other words, are working from a figure roughly double the one the data supports.

Follow the hours and the gap explains itself. Generative-AI users save an average of 5.4 per cent of their work time, about 2.2 hours a week for someone on a 40-hour schedule, a figure that falls to roughly 1.4 per cent once non-users are counted.² The same researchers are candid that workers may simply pocket the saved time as leisure rather than producing more — which is exactly how a dividend can accrue entirely to the employee and stay invisible to the organisation.² None of this is a fringe effect. Adoption is mainstream: 35.9 per cent of US workers had used generative AI by December 2025, and the early consequences show up as task reallocation and quality improvement rather than headcount cuts.³ The dividend is being paid into accounts nobody is reading.

The dividend has a default destination

Who keeps the value is not a neutral question, and left alone it answers itself. CEPR's analysis finds AI-related innovation tends to reallocate income towards capital, with medium- and high-skilled workers facing the sharpest cuts to their income share, driven by wage compression rather than lost jobs.⁴ The aggregate prize is real but modest. The World Economic Forum estimates AI is already saving around 1.6 per cent of total work hours, an early productivity boost of roughly 1.3 per cent if that time is redeployed.⁵ A small pie, then, with a strong gravitational pull towards one side of the table.

The point is not to settle whether AI's gains belong to labour or to capital. It is to make a narrower question answerable inside a single organisation: where has the dividend actually landed? That is a measurement problem dressed as a fairness problem, and most firms are trying to solve it with no instrument at all.

Negotiating blind

When they reach for an instrument at all, they reach for the wrong one. Leaders systematically underestimate AI use, putting heavy users at roughly 4 per cent of staff when the real figure is closer to 13 per cent; only 20 per cent of executives expect AI to handle 30 per cent or more of their work within a year, against 47 per cent of employees.⁶ A leadership that misjudges adoption threefold is in no position to negotiate a fair split of anything. Activity-tracking tools will not rescue it, since counting hours logged or keys pressed measures presence, not value flow. What reveals where freed time has gone is workstyle data: focus-block patterns, collaboration ratios and output cadence, read before and after AI adoption.¹²

The strongest objection is that there may be little to capture. Aggregate gains are small, early, and quite possibly a J-curve lag before output materialises; the Atlanta Fed itself puts part of its paradox down to delayed revenue realisation.¹⁵ On that reading, any drive to claw back employee time savings is premature and self-defeating, because it threatens the trust and voluntary adoption that produce the gains in the first place. The honest position holds two truths at once: aggregate gains are modest, and the gap between individual efficiency and organisational output is nonetheless real, and widening, in low-visibility hybrid settings. The constructive response is shared visibility and an explicit dividend compact. Not quotas, and not a race to extract every freed hour.

That compact rests on three moves: reset expectations openly, redefine roles around AI-augmented capacity, and share the dividend deliberately rather than by default. Each depends on the same precondition — seeing the dividend in the first place. So the question for any leader confident that AI is paying off is simple, and uncomfortable: if the gains are real, where exactly did they land, and could you prove it?

Footnotes

1. Baslandze, S., Edwards, Z., Graham, J. R., McClure, T., Sparks, M., Meyer, B., Waddell, S. R., & Weitz, D. (2026). Artificial intelligence, productivity, and the workforce: Evidence from corporate executives (Working Paper No. 2026-4). Federal Reserve Bank of Atlanta. https://www.atlantafed.org/research-and-data/publications/working-papers/2026/03/25/04-artificial-intelligence-productivity-and-the-workforce-evidence-from-corporate-executives ↩ ↩² ↩³

2. Bick, A., Blandin, A., & Deming, D. J. (2025). The impact of generative AI on work productivity. Federal Reserve Bank of St. Louis. https://www.stlouisfed.org/on-the-economy/2025/feb/impact-generative-ai-work-productivity ↩ ↩² ↩³

3. International Center for Law & Economics. (2026). AI, productivity, and labor markets: A review of the empirical evidence. https://laweconcenter.org/resources/ai-productivity-and-labor-markets-a-review-of-the-empirical-evidence/ ↩

4. Minniti, A., et al. (2025). AI and the distribution of income between capital and labour. CEPR / VoxEU. https://cepr.org/voxeu/columns/ai-and-distribution-income-between-capital-and-labour ↩

5. World Economic Forum. (2026, January). Chief economists have clear ideas about AI productivity gains. https://www.weforum.org/stories/2026/01/the-where-and-when-of-ai-making-us-more-productive-according-to-experts ↩ ↩²

6. Ott, A. (2026). Superagency in the workplace: Empowering people to unlock AI's full potential [summary of McKinsey report]. LinkedIn. https://www.linkedin.com/posts/annaott_superagency-in-the-workplace-empowering-activity-7308389338604601344-JQgg ↩