In February 2026, the San Francisco Federal Reserve published a letter titled "The AI Moment? Possibilities, Productivity, and Policy." The question mark in the title is doing a lot of work. After more than three years of generative AI deployment at scale — across knowledge work, software development, customer service, and content production — aggregate productivity statistics have not moved in ways that are clearly attributable to AI. For an economy that has seen near-trillion-dollar annual investment in AI infrastructure, this is a puzzle worth taking seriously.

The argument: the gap between AI's visible impact at the firm and task level and its invisible impact in aggregate statistics is not evidence that AI isn't working. It reflects genuine measurement problems, adoption lags, and structural constraints that historical technological transitions suggest will persist for years. The more important policy and investment question isn't "why don't we see it yet" but "what determines whether we ever see it, and when."

The paradox has a historical precedent

The current situation echoes what economist Robert Solow famously observed in 1987: "You can see the computer age everywhere except in the productivity statistics." For roughly a decade after computers became widespread in business, aggregate productivity growth remained sluggish. Then, in the mid-1990s, it accelerated sharply — driven by complementary investments in business process redesign, workforce retraining, and organizational restructuring that took time to accumulate.

The historian of technology Paul David documented the same pattern with electrification. Electric motors were commercially available by the 1880s. The productivity gains from electrification — which required factories to be redesigned around the new power source rather than simply substituting electric motors for steam — didn't appear in aggregate data until the 1920s. A 40-year lag.

Neither analogy is perfect for AI. Generative AI diffuses differently than capital equipment, affects a broader range of cognitive tasks, and operates in an economy with different sectoral structure. But the basic mechanism — that general-purpose technologies require complementary reorganization before their productivity effects materialize in aggregate — is well-established. Aghion and Bunel's 2024 review of AI and growth models formalizes this intuition and suggests the current moment is consistent with early-phase GPT diffusion.

What the current evidence actually shows

Micro-level evidence of AI's impact is substantial and largely positive. Field experiments find significant productivity gains for specific tasks: GitHub Copilot studies show 55% faster task completion for certain coding tasks; McKinsey's work on call center agents finds similar-magnitude gains. The NBER working paper on labor market impacts documents real changes in task composition within occupations exposed to AI.

The macro picture is different. The OECD's 2025 analysis finds that most broad productivity measures have not shown clear AI-attributable acceleration. The IMF working paper on AI and productivity in Europe reaches a similar conclusion, while noting that early adopters show stronger effects than laggards — suggesting the aggregate signal is being diluted by uneven adoption.

The Penn Wharton Budget Model projects 1.5% cumulative GDP growth from AI by 2035, nearly 3% by 2055. Goldman Sachs economists estimate AI could add 0.3 percentage points to annual productivity growth over the next decade — real but modest, well within the range of previous general-purpose technology effects. KPMG's projections are more bullish, projecting up to $2.84 trillion in US GDP impact by 2030, though these rely on aggressive adoption assumptions.

The spread in these projections isn't primarily about disagreement over AI's technical capabilities. It's about assumptions regarding adoption speed, complementary investment, and adjustment costs.

The measurement problem is real

Part of the gap between micro and macro evidence may be measurement failure rather than absence of effect. GDP accounting is built around measuring output of goods and services at market prices. Several categories of AI-generated value are structurally difficult to capture:

Quality improvements that aren't priced. If AI makes a software product faster and more reliable without a price increase, GDP doesn't see the improvement. If AI helps a doctor diagnose faster, the visit still bills at the same rate. Many of the most significant AI productivity effects may be showing up as quality improvements in outputs that aren't reflected in price indices.

Consumer surplus. When AI tools are provided free or at marginal cost below their user value, the surplus accrues to users without appearing in GDP. The ECB's analysis from March 2026 flags this explicitly: conventional productivity metrics almost certainly understate AI's economic contribution because of the consumer surplus problem.

Compositional shifts. AI may be shifting the composition of output toward activities whose value is harder to measure — insight generation, creative work, scientific research — rather than simply making existing measured activities faster.

None of this means the aggregate data is wrong. It means the aggregate data is measuring something specific, and that something may systematically miss categories where AI's impact is concentrated.

The structural constraints on European adoption

The divergence between US and European AI adoption is worth examining as a natural experiment in what constrains diffusion. The IMF working paper identifies three structural factors that slow adoption in Europe relative to the US:

SME prevalence. AI adoption requires upfront investment in tools, training, and process redesign. Large firms with scale economies can amortize these costs more easily. Europe's industrial structure — with a higher share of small and medium enterprises — creates structural friction.

Capital market depth. Frontier AI development requires capital concentration at a scale that European financial markets have historically not provided. This doesn't just affect who builds AI; it affects which firms can absorb the complementary investments needed for adoption.

Regulatory uncertainty. The EU AI Act creates compliance requirements whose full scope is still being interpreted. Uncertainty about what's permissible raises the effective cost of adoption for early movers. Whether the regulatory framework ultimately produces better outcomes is a separate question from its near-term effect on adoption timing.

These aren't arguments against European regulation. They're arguments for understanding that policy design shapes who captures AI's productivity gains and when — and that coordination between innovation policy and adoption incentives matters as much as the regulatory framework itself.

What determines whether productivity gains materialize

The historical evidence on general-purpose technology diffusion suggests three factors dominate:

Complementary investment. Electrification required factory redesign. Computers required IT infrastructure, software, and process re-engineering. AI appears to require analogous complementary investments: data infrastructure, workflow redesign, and capability development in the humans working alongside AI systems. Firms that invest in these complements see the gains. Firms that deploy AI tools without changing underlying processes see much less.

Human capital adjustment. Productivity gains from GPTs tend to concentrate in workers and firms that adapt their skill sets to work alongside the technology rather than against it. The ADP research on AI employment effects suggests that early-career workers in high-AI-exposure jobs are seeing employment contractions precisely because they haven't yet accumulated the complementary skills that make AI augmentation rather than substitution.

Market competition. Productivity gains that accrue to a small number of firms may not diffuse to aggregate output growth if competitive pressure is insufficient to force adoption across industries. The economics of AI — high fixed costs, network effects, data advantages — create concentration dynamics that could slow diffusion relative to more competitive technologies.

The question worth sitting with

Whether AI produces a macroeconomic productivity surge of the magnitude that some projections suggest depends less on AI's technical trajectory than on whether the complementary ecosystem develops at pace. The technology's capabilities are not the binding constraint. The binding constraint is organizational, institutional, and political.

The current absence of a clear aggregate productivity signal is not evidence that the projections are wrong. It's consistent with being in the early phase of a diffusion process that historical precedent suggests takes longer than initial enthusiasm implies. The Dallas Fed's 2025 analysis reaches this conclusion explicitly: the productivity gains are likely real and will materialize, but probably more slowly and unevenly than optimistic near-term projections suggest.

What should policy do in the meantime? The San Francisco Fed letter offers a measured answer: invest in measurement infrastructure to capture AI's actual economic effects better, support the complementary human capital investments that determine who benefits, and resist the temptation to either hype or dismiss a technology whose full economic impact will be measured in decades, not quarters.

The question mark in that title is the appropriate epistemic posture. This is a real moment. Whether it's the AI moment depends on choices that haven't been made yet.

References