Capability Is Not Capacity
The central problem in the AI transition is not whether frontier systems are becoming more capable. It is whether firms and economies can reorganize production around those capabilities. This chapter traces that conversion through tasks, complementary inputs, organizations and institutions, so the reader can diagnose why a powerful model may produce delayed or uneven macroeconomic gains rather than an immediate productivity surge.
That distinction matters after an investment boom. A data center, chip plant or cloud contract can be built on the expectation that AI will transform work. But the asset becomes productive only when it is connected to tasks, people, processes and customers. The distance between a benchmark and an economy is where most of the uncertainty lives.
From model capability to productive tasks
AI can raise output through several channels that are related but not interchangeable. It can automate a task that a person previously performed. It can assist a worker within a task, making the worker faster or more accurate without removing the task. It can also help produce new ideas, including by accelerating research and development. These channels have different effects on employment, investment and the distribution of gains.
Task automation expands the set of activities that can be performed with less human labor. Within-task assistance raises the productivity of the people who remain involved. AI-assisted research can improve the production of ideas, but the resulting benefit may arrive only after experimentation, validation and investment. A system that performs well on a narrow cognitive task does not automatically reorganize an entire production process. Aggregate output depends on how many tasks are affected, how important those tasks are, and whether other inputs can expand alongside them.
This is why the share of tasks affected matters more than an impressive performance result in isolation. If AI improves one step in a process but energy, materials, physical presence or coordination remain binding, total production may rise only modestly. The IMF notes that low elasticities of substitution, non-automatable complementary tasks, infrastructure and diminishing returns can limit economy-wide gains. Very high productivity in a subset of tasks does not guarantee very high productivity for the economy as a whole.
There is also a difference between substituting for labor and expanding what labor can do. Substitution may reduce the labor needed for a task, but the resulting capacity can be wasted if demand, equipment or other stages of production do not expand. Complementarity can raise the value of workers who know how to use the system, but it can also increase the importance of scarce expertise. The same model can therefore alter the composition of output and wages without producing a uniform increase in every worker's productivity.
The relevant aggregate question is not “How good is the model?” but “How much of the production system can change because of the model?” If only a narrow task is improved, the effect may remain local. If many tasks across connected industries are reorganized, the effect can spread through investment, prices and employment. But broad reorganization also increases the need for physical capital, reliable inputs, standards and coordination. Scale magnifies both the possible return and the number of interfaces at which the process can fail.
This distinction also changes how we think about research automation. AI may help researchers search, test and combine ideas, creating a feedback loop in which better tools improve the process that improves the tools. But the IMF discussion does not treat an intelligence explosion as inevitable. Diminishing returns and the need to validate results can weaken the feedback. Research output still has to become reliable knowledge, usable technology and deployed capital. The mechanism can be powerful without being unlimited.
The BIS reports early evidence of sizeable micro-level gains from generative AI, particularly in parts of non-routine cognitive work, and cites a median estimate of about 0.5 percentage points of annual total-factor-productivity growth across a wide range of macro estimates. That number should be read with its denominator and uncertainty intact. It is a median estimate of a possible macro effect, not a measured gain that every country has already received. The same source notes that micro gains may take time to appear in aggregate productivity as firms reorganize and resources move among them.
The evidence also has a level-of-analysis boundary. A worker may complete a task faster while the firm spends more time checking output, changing its workflow or resolving new errors. A firm may improve its own productivity while demand shifts toward a bottleneck elsewhere in the supply chain. A country with many exposed occupations may appear highly prepared while lacking the power, capital or institutions needed to deploy the technology. Micro evidence is therefore valuable, but it cannot be summed into a macro result without considering reallocation and complements.
The delay is sometimes described as a productivity J-curve. New technology can require complementary investment, training and organizational redesign before its measured output benefits exceed the disruption and setup costs. In the AI case, the curve may differ by sector. A software team may integrate an assistant relatively quickly; a hospital, factory or public agency may face physical, legal and accountability constraints that make deployment slower. The relevant unit is not “AI adoption” in the abstract but the path by which a particular set of tasks is connected to a particular production system.
The conversion chain shows why capability and capacity can diverge. Frontier capability is an input to the process, not its final output. Tasks have to be selected and redesigned; complementary inputs have to be available; managers and workers have to coordinate; customers and regulators have to accept the new arrangement. A failure at any interface can delay the payoff without proving that the underlying technology has no value.
The adoption system around the model
The first interface is physical. AI systems require electricity, computing hardware, data centers and networks. Some applications also require robots, machinery or workers who remain physically present. The IMF identifies energy and grid capacity, data-center infrastructure, capital accumulation and non-automatable tasks as limits on aggregate productivity. These constraints can coexist with rapidly falling inference prices. It may be cheaper to generate one more model response while still being difficult to build the power, equipment or organization needed to use millions of responses reliably.
The physical constraint can change the timing of the macro payoff. If power or specialized hardware is scarce, firms may spend heavily to secure future capacity, lifting investment and input prices before the applications using that capacity are ready. If the constraint is later relieved, adoption can accelerate. If it persists, the economy may retain a large investment impulse without receiving an equivalent increase in output. The same spending therefore has a different meaning depending on whether it is opening a bottleneck or merely bidding up a scarce input.
The second interface is organizational. Firms must decide which tasks to automate, which workers to augment and where human review remains necessary. That is not a one-time purchase. It can involve redesigned workflows, new responsibility lines, training, quality control and the accumulation of experience about failure. The IMF reports that many firms face organizational inertia, compliance burdens and managerial risk aversion. These are not merely attitudes standing outside the economy. They affect the speed at which capital becomes productive and the amount of output a given technology can support.
Reliability concerns illustrate the mechanism. If a model occasionally produces an error that is cheap to correct, a firm may use it with light supervision. If the same error creates legal, safety or reputational costs, the firm may require extensive review or avoid the application. The model's technical capability has not changed, but the economic value of deployment has. The relevant price is therefore not only the fee for using the system; it includes supervision, verification, liability and the cost of changing the surrounding process.
Organizational adaptation also determines whether gains spread beyond early adopters. A large firm may have enough data, technical staff and managerial slack to experiment, while a smaller firm faces a higher fixed cost of integration. A regulation that is manageable for the large firm may be prohibitive for the small one. The result can be a paradox: policy intended to reduce risk improves reliability for the firms that can comply but slows diffusion elsewhere and reinforces concentration. The IMF identifies this possibility without treating it as inevitable; the effect depends on design and institutional capacity.
The third interface is institutional. Regulation can protect safety and trust, but poorly calibrated or excessively complex rules can slow diffusion and favor dominant firms that can absorb compliance costs. The IMF also emphasizes that public trust, worker expectations and social narratives influence adoption. A rule that is technically workable may still fail if affected communities do not regard the deployment as legitimate. Conversely, a deployment that proceeds quickly under weak oversight may create backlash or losses that slow later diffusion.
Skills are part of the same system. The value of an AI tool depends on whether workers can judge its output, combine it with domain knowledge and adapt their tasks. Formal education is not the only relevant input; expertise and practical context matter. Where skills do not match the new task mix, adoption can be slower or produce lower-quality results. A country may therefore possess access to the same model as a more prepared country while receiving less productivity from it because the complementary labor and organizational capacity differ.
These interfaces explain why the IMF scenario exercise changes diffusion rather than capability. Both of its scenarios assume rapid advances in frontier systems and, over a five-year horizon, technical capacity to perform a wide range of cognitive and manual tasks. The baseline-diffusion path then introduces friction: regulation, infrastructure limits, skills gaps, social resistance and organizational constraints keep adoption uneven. The runaway-diffusion path assumes that those frictions are much weaker, allowing automation to spread broadly and quickly. The technology is held broadly similar; the economic outcome changes because absorption changes.
The comparison is useful precisely because it is not a forecast. It isolates a causal question: if capability improves rapidly, how much of the macroeconomic effect depends on the speed and breadth of deployment? Under gradual diffusion, productivity gains may arrive in pockets while existing institutions have more time to adjust. Under rapid diffusion, output and productive capacity may rise faster, but labor displacement, concentration and fiscal pressure can also arrive sooner. Neither path is guaranteed, and the workshop's closed-door judgments are not a probability distribution. They show what policymakers would need to prepare for if adoption moved at different speeds.
State capacity is part of the absorption mechanism, not merely a response after the fact. Governments need information about where AI is being deployed, which tasks are changing, how financial exposures are accumulating and which groups are bearing the adjustment. They also need tax administration, social-protection delivery, data governance and supervisory technology capable of operating in a changing environment. Weak capacity can produce a self-reinforcing slow-diffusion equilibrium: lower readiness delays adoption, delayed adoption limits learning and investment, and the resulting weaker growth leaves fewer resources for readiness.
This capacity is not just a national score that can be read off a dashboard. It is a set of operating abilities: collecting timely data, coordinating agencies, enforcing standards, delivering support and updating rules as the technology changes. The IMF warns that data gaps on deployment and labor impact constrain real-time policy calibration. Without those measurements, a government may see investment and market valuations but miss whether tasks are actually changing, whether concentration is rising or whether losses are accumulating in particular households and sectors.
The reverse is not automatically benign. High readiness can attract capital and accelerate adoption, but if it is concentrated in a small number of firms or regions, it can widen inequality and increase systemic dependence. Faster diffusion can make a country look more productive while making its labor market, fiscal system or financial sector more exposed to a rapid adjustment. Capacity therefore has two meanings: the ability to use AI productively and the ability to govern its consequences. A country needs both.
The practical conclusion is modest but important. Falling model prices, impressive benchmarks or a fast-growing investment budget are evidence of capability and commitment, not complete evidence of economy-wide productivity. The stronger claim—that AI has raised potential output broadly and durably—requires evidence that tasks are being reorganized, complementary inputs are available, organizations are adapting and adoption is producing reliable output. New evidence on actual deployment, task-level performance, firm reorganization, energy use and the timing of investment returns would change the judgment more than another isolated frontier benchmark.
The investment cycle therefore has an unresolved middle. Capital has been committed, but capacity has to be assembled through tasks, infrastructure, institutions and trust. That middle is also where the economy begins sending mixed signals: construction and valuations can rise before productivity, while adoption can be delayed even as capability advances. The next question is how to read those signals together rather than allowing one headline measure to stand in for the whole transition.