AI & strategy

The AI ROI Paradox: Why Adoption Isn't Showing Up in the Bottom Line Yet

By Abderrazzak Jirari

· 7 min read

Rows of servers in the CERN data centre
Photo: Hugovanmeijeren

In 1987, Nobel laureate economist Robert Solow wrote that you could see the computer age everywhere but in the productivity statistics (Solow, 1987). Nearly forty years later, the line applies almost word for word to artificial intelligence.

Adoption is massive. According to McKinsey's annual survey published in August 2026, nearly nine in ten respondents say their organisation regularly uses AI in at least one business function (Tinkoff et al., 2026). Yet the share of respondents who attribute any effect at all on earnings before interest and taxes (EBIT) to AI stands at 37%, essentially unchanged from a year earlier (Tinkoff et al., 2026).

How can a technology this widely adopted leave so little trace in results? And, above all, what do the rare organisations that earn a real return do differently?

What the data say: a gap that is not closing

McKinsey's survey of 1,719 respondents in 97 countries, conducted in May and June 2026, shows real movement: 44% of respondents say AI is scaling across their organisation, up from 38% a year earlier (Tinkoff et al., 2026). But so-called "high performers," which attribute at least 5% of their EBIT to AI and describe its impact as significant, account for only about 6% of respondents, a share unchanged since 2025 (Tinkoff et al., 2026).

Other studies point the same way, with harsher numbers. A 2025 report from MIT's NANDA initiative concludes that only 5% of integrated generative AI pilots produce significant value, with the vast majority showing no measurable effect on results (Challapally et al., 2025). That widely repeated 95% figure should, however, be read with caution: it rests on 52 interviews, 153 survey respondents and an analysis of 300 public deployments (Legal.io, 2025). It is a signal, not a representative measure.

On so-called agentic AI, Gartner forecasts that more than 40% of projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls (Gartner, 2025). It also calls out "agent washing," the rebranding of existing assistants or chatbots as agents, and estimates that only about 130 of the thousands of vendors claiming agentic capabilities are genuine (Gartner, 2025).

Canada is not immune, but it is behind. The share of Canadian businesses that used AI to produce goods or deliver services rose from 6.1% in the second quarter of 2024 to 12.2% in 2025, then to 19.2% in 2026 (Do et al., 2026). Adoption has tripled, yet 40% of businesses still consider AI irrelevant to their operations (Do et al., 2026).

First finding: individual gains are real

It would be wrong to conclude that AI does not work. The best experimental studies show clear gains at the task level. In a study published in the Quarterly Journal of Economics, Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed more than 5,000 customer support agents: access to an AI assistant increased the number of issues resolved per hour by 14% on average, with gains concentrated among novice and less experienced workers (Brynjolfsson et al., 2025).

The paradox, then, does not lie in the technology. It lies in the passage from the task to the organisation.

Second finding: gains dissipate before they reach the bottom line

The most illuminating study of that passage comes from Denmark. Anders Humlum and Emilie Vestergaard linked surveys of 25,000 workers, in 7,000 workplaces and 11 exposed occupations, to administrative data on earnings and hours worked (Humlum & Vestergaard, 2025). Their result: users report average time savings of only 2.8% of work hours, far from the gains of 15% or more observed in controlled experiments, and the effect on earnings and hours is statistically zero (Humlum & Vestergaard, 2025).

Where does the saved time go? According to the authors, more than 80% of it is reallocated to other work tasks, including new tasks created by AI itself, such as editing generated text (Fortune, 2025). The gain exists, but the organisation absorbs it without turning it into a measurable result.

Canadian data confirm this measurement trap. A Statistics Canada analysis finds that businesses using AI show labour productivity 16.8% higher than others (Li & Liu, 2026). But once firms' prior productivity and complementary capabilities (data analytics, cloud computing, training) are taken into account, the gap falls to 5.1% (Li & Liu, 2026). In other words, much of the advantage already existed: it is often already high-performing firms that adopt AI. The authors also find no significant short-term relationship, as gains depend on organisational changes that take time (Li & Liu, 2026).

Third finding: the J-curve, or the invisible cost of transformation

Economic history offers a key. Brynjolfsson, Rock and Syverson showed that general purpose technologies, such as electricity or computing, require large intangible investments: process redesign, training, new skills, data (Brynjolfsson et al., 2021). Accounting measures these investments poorly, so measured productivity first stalls or even dips before rising: the J-curve.

This reading calls for patience, not complacency. Daron Acemoglu, also a Nobel laureate, estimates that AI's aggregate productivity gains over ten years may be much more modest than the rhetoric promises, on the order of 0.7% in total factor productivity, because only a fraction of tasks are both exposed to AI and profitable to automate (Acemoglu, 2024). Between promise and skepticism, the relevant question for a leader is not "will AI transform the economy?" but "is my organisation making the investments that get it through the bottom of the curve?"

What the organisations that benefit do differently

McKinsey's data are most useful here, because they compare the 6% of high performers directly with everyone else (Tinkoff et al., 2026):

  • nearly three-quarters of them have fundamentally redesigned their workflows around AI, compared with about one-quarter of the others;
  • they pursue growth and innovation, not just cost reduction;
  • they are twice as likely to have defined processes for measuring AI's impact;
  • they are twice as likely to have senior leaders visibly committed to AI initiatives.

Conversely, organisations that simply insert an assistant into unchanged processes get what Humlum and Vestergaard measured: a little time saved, scattered across the organisation, invisible in the results.

Five decisions to escape the paradox

Based on these data, five decisions seem to separate organisations that earn a return from those that pile up pilots.

  1. Set the outcome before choosing the tool. Define a precise business indicator (processing time, cost per case, error rate) and measure its baseline before any deployment. High performers are twice as likely to do so (Tinkoff et al., 2026).
  2. Redesign the process, not just its tooling. The gain comes from reconfiguring the workflow, not from adding an interface (Tinkoff et al., 2026).
  3. Decide in advance how freed-up time will be used. Without an explicit decision, saved time dissolves into other tasks (Humlum & Vestergaard, 2025). It must be deliberately reallocated: more volume, better service or a new activity.
  4. Invest in complementary capabilities. Data, cloud and training explain much of the productivity gap (Li & Liu, 2026). In Canada, only 32% of businesses using AI have trained their existing employees (Do et al., 2026).
  5. Reserve agents for cases where the value is clear. Gartner recommends pursuing agentic AI only where it delivers demonstrable value or return, and being wary of rebranded products (Gartner, 2025).

Conclusion: a management problem, not a technology problem

The AI ROI paradox is not proof that the technology fails. It is the symptom of organisations that adopted a tool without changing how they work. Individual gains are documented, organisational gains remain rare, and the gap between the two is a matter of management: choice of use cases, process redesign, measurement, training.

For Quebec businesses and public bodies, still early in their adoption, that is good news. They can learn from the pioneers' mistakes and enter AI through the right door: the outcome to achieve, not the tool to deploy.

(The author)

Portrait of Abderrazzak Jirari

Abderrazzak Jirari

Strategy, Digital and AI Consultant

Trained in political science (from an honours bachelor's to a master's at Université de Montréal), I'm now completing a graduate diploma (DESS) in private- and public-sector management at HEC Montréal.