Research · Part 3 of 6
The state of play: large gains in tasks, small gains in the economy
Adoption is near-universal and value is concentrated in a small minority. The best evidence says the constraint is organisational, not technical.
Adoption is near-universal, value is concentrated in a small minority, and the best evidence says the constraint is organisational rather than technical. Four findings define the state of play in late 2026.
1. Adoption has outrun value
McKinsey's 2025 survey of 1,993 respondents in 105 countries found 88% of organisations using AI in at least one function, but only 7% reporting it fully scaled and 31% in the scaling phase (IDM summary). The 6% it classes as high performers, those attributing 5% or more of EBIT to AI, were nearly three times as likely as others to have fundamentally redesigned individual workflows.
MIT Project NANDA's widely quoted claim that 95% of organisations saw no measurable return from generative AI points the same way (Virtualization Review). It should be read with care: it rests on 52 interviews and 153 conference survey responses, and it does not define "return".
2. Large gains in tasks, small gains in the economy
Controlled studies find large effects at the level of the task, while studies of whole labour markets find small ones. The gap between the two is the clearest evidence that transformation, not access, decides value.
| Study | Setting and design | Finding | Evidence type |
|---|---|---|---|
| Brynjolfsson, Li and Raymond | 5,000 customer support agents, staggered rollout | 14% more issues resolved per hour; largest gains for novices, little for the most experienced | Peer-reviewed (QJE, 2025) |
| Dell'Acqua et al. | 758 BCG consultants, randomised | Inside the model's capability: 12.2% more tasks, 25.1% faster, over 40% higher quality. Outside it: 19 points less likely to be correct | Field experiment, working paper |
| Humlum and Vestergaard | About 25,000 workers in 7,000 Danish workplaces, linked to payroll data | Time savings of about 3%; effects on earnings and hours ruled out above 1% | Working paper, administrative data |
| Acemoglu | Task-based macroeconomic model | Total factor productivity gain of no more than 0.66% over ten years | Model, peer-reviewed (Economic Policy, 2025) |
Two mechanisms explain the gap. The first is the "jagged frontier": AI helps sharply on some tasks and harms on others, so gains depend on someone redesigning which tasks go to AI and how output is checked. The second is the J-curve: returns arrive only after complementary investment in process, skills and data (Brynjolfsson, Rock and Syverson, 2021). Saved time that is not redeployed into redesigned work leaves no trace in output or pay.
3. Why initiatives fail
The documented causes of failure are about framing, foundations and organisation, not models. RAND's interviews with experienced practitioners identified five root causes (Asia Financial, on RAND, 2024):
- the problem AI is meant to solve is misunderstood or poorly communicated;
- the data needed to train an effective model is missing;
- teams chase the newest technology instead of the user's problem;
- the infrastructure to manage data and deploy models is lacking;
- AI is applied to problems too hard for it to solve.
Recent evidence adds two more: work is not redesigned around the tool (McKinsey's high-performer finding), and systems do not learn or retain context, so users reject them for high-stakes work (NANDA's "learning gap"). A further cause is inferred rather than measured: governance and benefits tracking are added only when production is attempted, which is when most pilots stall.
4. Agentic AI raises the stakes
Agents are moving from experiment to use faster than the controls around them. In 2025, 62% of organisations were at least experimenting with agents and 23% were scaling them (IDM summary). Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 for cost, unclear value or weak risk controls, estimates that only about 130 of thousands of "agentic" vendors are genuine, and expects at least 15% of day-to-day work decisions to be made autonomously by 2028 (Gartner, June 2025).
Already 51% of McKinsey's respondents report at least one negative consequence from AI, most often inaccuracy (30%). When systems act rather than advise, those failures become actions. Decision rights, evaluation and the ability to withdraw an agent's authority are now part of transformation itself.
Sources
- Acemoglu, D. (2025). The simple macroeconomics of AI. Economic Policy; NBER Working Paper 32487.
- Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at work. Quarterly Journal of Economics.
- Brynjolfsson, E., Rock, D. and Syverson, C. (2021). The productivity J-curve. American Economic Journal: Macroeconomics.
- Dell'Acqua, F. et al. (2023). Navigating the jagged technological frontier. Harvard Business School Working Paper 24-013.
- Humlum, A. and Vestergaard, E. (2025). Large language models, small labor market effects. Becker Friedman Institute Working Paper 2025-56.
- Gartner: over 40% of agentic AI projects will be cancelled by end of 2027
- IDM: McKinsey State of AI 2025 (secondary report of the survey)
- Virtualization Review: MIT NANDA, The GenAI Divide (secondary report)
- Asia Financial: RAND, root causes of AI project failure (secondary report)