AI has made its way into businesses faster than it has made an impact on their profit and loss accounts. The most compelling results do not come from the spectacular automation of entire workstations, but from less flashy areas: searching for knowledge, drafting initial versions of documents, onboarding staff and handing matters over between departments.
In 2025, 20.2 per cent of companies in OECD countries were using AI, compared with 8.7 per cent two years earlier. However, adoption remains heavily dependent on scale: 52 per cent of large firms were using the technology, compared with just 17.4 per cent of small firms. This is not yet a productivity revolution. It is a rapid diffusion of tools, the results of which are unevenly distributed.
A 2025 McKinsey study highlights the disconnect between usage and value. Nearly nine in ten organisations reported regular use of AI, yet almost two-thirds had not begun to scale it across the entire organisation. Thirty-nine per cent of respondents reported some impact on EBIT, and in most cases this did not exceed 5 per cent of the result. Companies have implemented the technology, but have rarely restructured their operations around it.
Time is lost between tasks
The most significant productivity gains occur where AI shortens the path to the right information.
In a survey of 5,179 customer service consultants, a generative AI assistant increased the number of cases resolved per hour by an average of 14 per cent. Among new and underperforming staff, the improvement was 34 per cent. Experienced consultants gained little. The model did not take over the process. It disseminated the methods used by the best agents and shortened the time it took for new staff to reach full proficiency.
In specialist work, the effect depends on the type of task. BCG consultants using GPT-4 completed typical analytical tasks 25.1 per cent faster, finished 12.2 per cent more of them, and received higher quality ratings. However, when presented with a problem beyond the model’s capabilities, they were 19 per cent less likely to provide the correct answer than those working without AI. A smoothly phrased suggestion only speeds up work until it begins to replace one’s own assessment of the situation.
Less coordination, not fewer people
Another type of efficiency gain was revealed by an experiment carried out at Procter & Gamble. It involved 776 specialists tackling real-world problems related to product development. A single person supported by AI achieved a level comparable to that of a two-person team working without it.
The technology also bridged the gap between functions. Research and development specialists proposed solutions more grounded in market realities, whilst commercial staff took better account of technical constraints. AI did not eliminate the need for collaboration. It reduced the cost of translating the same problem between teams with different areas of expertise.
It is precisely in coordination that a large part of digital inefficiency lies: in waiting for consultation, searching for the owner of the information, re-entering data and correcting a document whose context has been lost between systems.
AI can also slow things down
METR studied experienced developers working in their own, well-known open-source repositories. With access to AI tools, they took an average of 19 per cent longer to complete tasks, although prior to the experiment they had predicted a 24 per cent increase in speed. Time was consumed by generating responses, checking code and correcting results. Significantly, even after completing their work, the participants still believed that AI had sped them up.
More recent observations from 2026 suggest that current models may already be helping developers more frequently. However, METR did not specify the precise extent of the improvement, as the results were skewed by participant selection and a change in the way multiple agents were used simultaneously.
Productivity cannot be measured by the number of licences
The flaw in many implementations lies in the measurement itself. The number of active users, prompts and agents in operation describes the consumption of the technology, not its value.
Much more is revealed by case resolution times, the number of corrections, the proportion of issues resolved on first contact, the length of an employee’s onboarding period, and the number of handover instances between teams. In a McKinsey study, this re-engineering of workflows was the factor most strongly associated with the impact of generative AI on EBIT. Despite this, only 21 per cent of organisations using this technology reported a fundamental change to at least some of their processes.
Automation remains an attractive narrative because it is easy to showcase. Productivity, on the other hand, is less spectacular. It arises when information is received sooner, cases are less likely to be sent back for correction, and specialists do not waste hours reconstructing context that the company already possesses.
