An IMF analysis published in November 2025 argued that Europe's gains from artificial intelligence would depend partly on market integration, adoption incentives, and complementary policy choices. The conditional form of the argument deserves attention. It distinguishes a technical capability from an economy's capacity to convert that capability into sustained improvement. A demonstration can show that a task is performed faster; it cannot, by itself, establish how organisations will redesign work, how demand will respond, or who will capture the resulting value. The distance between demonstration and transformation is populated by institutions rather than empty time.
Consider an accounting firm whose staff can draft routine reports in half the previous time. The immediate saving is clear under the conditions of the trial. But the firm may need additional review, better data management, revised responsibility for errors, and training that temporarily reduces output. These complementary investments can be indispensable while remaining less visible than the software. If the firm buys access without making them, disappointing results do not necessarily disprove the tool's capability. They may reveal that the task isolated in the demonstration was only one component of a production process whose other components remained unchanged.
The opposite mistake is equally possible: attributing every organisational improvement to the new technology. A firm introducing software may simultaneously improve management, standardise procedures, and select unusually motivated staff for a pilot. The observed gain then combines several changes. The question is not whether the package worked, but which inference the evaluation supports. A successful pilot establishes less about universal adoption when participation, supervision, and task selection differ substantially from ordinary operation. Scaling is not simply multiplication. It changes the distribution of users, the difficulty of cases, and the surrounding constraints.
At the level of an individual task, productivity usually concerns output relative to inputs. At the level of an organisation, however, output must be defined and quality accounted for. Producing twice as many reports is not an improvement if recipients must spend longer detecting errors or extracting relevant conclusions. Costs can migrate across departmental boundaries while the originating department celebrates efficiency. A narrow measure may record success because the burden has moved outside its field of view. The economic unit chosen for evaluation therefore affects whether an apparent gain survives a more comprehensive accounting.
There is also a distinction between substitution and complementarity. A tool may replace some activities while increasing the value of others. If drafting becomes cheaper, careful problem formulation, verification, and communication may become more important. This does not imply that every displaced worker will automatically move into the complementary role. Skills, bargaining power, recruitment practices, and geographic constraints shape the transition. An aggregate possibility of higher output is compatible with severe disruption for particular groups. Treating the aggregate as a sufficient answer to distributional concerns confuses the size of the potential surplus with the conditions under which people can share it.
Adoption itself is heterogeneous. A large firm can spread fixed implementation costs across many transactions, while a smaller enterprise may find the same system uneconomic. Conversely, a flexible small organisation may redesign procedures more quickly than a large institution burdened by incompatible systems and fragmented authority. There is no universal ranking independent of context. What appears as technological reluctance may reflect rational uncertainty about integration costs, liability, or the durability of a supplier. Calling all delay resistance to progress prevents analysts from distinguishing avoidable inertia from a defensible refusal to purchase an incomplete solution.
The market structure surrounding the technology can influence its wider benefits. If improvements reduce production costs but competition is weak, much of the gain may become higher margins rather than lower prices or higher wages. If entry becomes easier, the same technical change may intensify competition and redistribute value towards users. Neither outcome is contained in the algorithm alone. Ownership, contracts, switching costs, and access to complementary resources mediate the result. A discussion that attributes the distribution of benefits entirely to technical performance overlooks the rules through which performance is converted into claims on income.
Expectations can create another divergence between financial and productive outcomes. Investors may price assets on the basis of anticipated future gains before those gains appear in measured output. Such anticipation can finance useful experimentation, but it can also produce valuations that depend on unusually favourable scenarios. A rising valuation is therefore not direct evidence that the promised productivity has already materialised. Nor does disappointment in a valuation prove that the underlying technology has no productive use. Financial claims and technical capabilities operate on related but distinct timelines, with different sensitivities to uncertainty and competition.
Measurement becomes especially difficult when the output changes rather than merely becoming cheaper. A service may offer faster access, greater customisation, or entirely new functions that older measures capture imperfectly. This possibility should encourage methodological care rather than unlimited optimism. Unmeasured benefits are not a licence to count every promised advantage as real. Analysts must distinguish a plausible measurement gap from a convenient explanation for absent evidence. The proper response is to improve definitions and gather alternative indicators, while acknowledging that some innovations can remain valuable even when their full contribution is not immediately visible in a standard aggregate.
Infrastructure can also become binding. Reliable electricity, connectivity, suitable data, and trained personnel determine whether a capability can be deployed at scale. Investment in these complements may support other activities, but it also competes for resources. A productivity narrative that treats physical inputs as frictionless can exaggerate the ease of diffusion. Conversely, an infrastructure constraint can motivate beneficial investment rather than permanently prevent adoption. The relevant question is how quickly, at what cost, and with what distributional consequences the constraint can be relaxed. A bottleneck is a mechanism to analyse, not a final verdict.
Public policy consequently faces more than a choice between enthusiasm and prohibition. It can support skills, interoperability, competition, reliable infrastructure, and evaluation while assigning responsibility for harmful outcomes. These interventions can complement innovation by reducing uncertainties that otherwise deter useful adoption. They can also be poorly designed, entrench incumbents, or impose disproportionate compliance costs. The policy task is empirical and institutional: identify which obstacle matters, specify the intended mechanism, and test whether the intervention improves the relevant outcome. Invoking innovation does not excuse scrutiny, just as invoking risk does not settle the case for restriction.
The productive significance of artificial intelligence will therefore emerge through a sequence of contingent decisions rather than a single moment when machines become impressive enough. Task performance matters, but so do organisational redesign, demand, competition, and the capacity to absorb adjustment. The conditional argument is not timid; it is more exacting than technological inevitability. It requires asking what has to be built around the machine and what evidence would establish that the building has succeeded. A productivity revolution is an alteration in how an economy produces value. A machine capable of contributing to that alteration is an opportunity, not the completed event.
Source notes
- [6] IMF, How Europe Can Capture the AI Growth Dividend (20 November 2025). Dated context: gains depend on adoption and complementary institutional conditions. Open primary source