AI Productivity Is Turning the Federal Reserve’s Rate Model Into a Measurement Test

The latest US productivity numbers look like evidence for the optimists. Nonfarm business productivity rose at a 3.6% annualized rate in the third quarter of 2026, while unit labor costs fell 0.7%. Output expanded 3.7% and hours worked barely moved. That is the kind of combination that can support faster real growth without an equivalent increase in inflation. It is also the kind of release that can make artificial intelligence appear to be a straightforward disinflation story.

It is not. AI productivity is becoming a measurement test for the Federal Reserve because the demand effects of the technology are arriving on a different schedule from its supply effects. Data-center construction, equipment purchases, equity wealth and scarce-component demand can strengthen spending today. The broader gains from reorganizing production, diffusing tools across firms and matching workers to new tasks may take years to become visible. In the interval, the same technology can raise potential output and put upward pressure on inflation, real rates and the neutral policy rate.

Federal Reserve Vice Chair Philip Jefferson set out this tension in a November 13 speech on technology and the macroeconomy. His baseline is that new technology will lift productive capacity substantially but gradually. Yet expectations of those future gains are already increasing aggregate demand through investment and wealth. That sequence matters more for markets than the simple claim that AI will eventually make workers more productive.

The critical question is therefore not whether AI can improve productivity. Evidence at the task level already says it can. The harder question is whether the improvement has become broad, persistent and large enough to change the economy’s speed limit—and whether the Fed can recognize that change in real time without confusing an investment boom for a completed supply expansion.

AI productivity is improving, but attribution remains weak

The Bureau of Labor Statistics’ preliminary third-quarter release contains several genuinely encouraging signals. Productivity rose 3.6% at an annualized quarterly rate and 2.1% from a year earlier. Unit labor costs fell during the quarter, although they were still 2.4% higher over four quarters. Since the fourth quarter of 2019, productivity has grown at a 2.2% annualized rate, above the 1.5% pace of the previous business cycle and equal to the long-run rate since 1947.

Those numbers strengthen the case that the economy may be capable of producing more than the pre-pandemic trend implied. They do not establish that generative AI caused the improvement. Productivity is output divided by hours, and both inputs are revised. The current release incorporated updated national accounts data and revised some earlier productivity readings. A quarter can also look strong because output rebounds, hours are measured differently, low-productivity activity contracts, or firms intensify existing capital.

Jefferson’s historical comparison is useful precisely because it resists instant attribution. In the late 1990s, information-technology capital added roughly three-quarters of a percentage point to annual labor-productivity growth. But the payoff did not come only from buying computers. Firms reorganized operations around cheaper information processing, changed their production functions and eventually lifted total factor productivity outside the technology sector. The technology was visible well before economists could prove its aggregate contribution.

The same problem now appears in reverse. AI is visible in products, investor narratives and corporate budgets, but it is not yet deeply embedded across the entire economy. Jefferson cited surveys suggesting that 40% to 50% of working adults use generative AI at work, while only a modest share of users employ it daily. A Census-based measure showed roughly one-quarter of US firms using AI in at least one business function by mid-2026, with adoption near 40% in knowledge-intensive sectors and around 10% in mining and leisure and hospitality.

Adoption is therefore broad enough to matter but too uneven to justify treating every productivity surprise as an AI dividend. The Census Bureau’s research on the microstructure of AI diffusion emphasizes differences across firms, business functions and worker tasks. The gap between an employee using a chatbot and an enterprise redesigning a production system is economically large. The first can save minutes. The second can change output per hour across an organization, but it also requires data, governance, training, integration and new management practices.

This distinction protects the analysis from two symmetrical errors. The skeptical error is to dismiss AI because the aggregate statistics do not yet show a dramatic break. The optimistic error is to capitalize every task-level gain immediately as economy-wide productivity. Both overlook diffusion. A general-purpose technology can be important long before it is universal, and valuable long before its full effect is measurable.

The investment clock is running ahead of the diffusion clock

AI affects the economy through at least two clocks. The investment clock begins when firms commit capital to computing capacity, data centers, power, networking equipment and specialized components. It also includes the wealth effect from higher equity valuations and the consumption supported by expectations of future income. This clock can accelerate demand almost immediately.

The diffusion clock starts more slowly. It measures the time required for firms to find productive uses, redesign workflows, train workers, resolve reliability problems and spread successful practices beyond leading companies. The most powerful productivity gains often depend on complementary organizational capital rather than the tool alone. A coding assistant may generate output faster, for example, but the net gain is smaller if review, debugging, security and integration costs rise at the same time.

This is why the current phase can be inflationary even if the long-run technology is disinflationary. Jefferson argued that the construction and equipping of data centers are driving rapid business investment and putting pressure on resource utilization and prices. Demand is concentrated in areas where supply has struggled to keep up, including memory and logic chips. The eventual response—more capacity and broader productivity—should relieve that pressure, but “eventual” is doing important work.

The pattern is familiar from other capital booms. Spending occurs before capacity becomes productive. Engineers, grid connections, transformers, land, cooling systems and advanced chips are required during the buildout. Their scarcity can lift costs in the affected sectors. Imports satisfy part of the equipment demand, limiting the direct contribution to domestic GDP, but imported capital can still deepen the domestic capital stock and raise labor productivity later.

For investors, this means the AI cycle cannot be read solely through semiconductor revenue or headline capital expenditure. The macro question is whether incremental spending is moving from narrow capacity creation toward broad output gains. That transition should appear as better productivity across more industries, less concentration in equipment bottlenecks, stable or falling unit labor costs, and stronger output without a matching rise in hours or inflation.

The distinction also explains why a strong technology narrative can coexist with a restrictive monetary stance. Jefferson said that September PCE inflation was 3.6% and core PCE inflation 3.1%, both well above the FOMC’s 2% target. He described the economy as expanding at a 3.4% pace in the third quarter, with strong business investment and resilient consumption. If technology enthusiasm raises demand before it raises aggregate supply, the Fed cannot treat the promise of future efficiency as present disinflation.

That sequencing is consistent with the broader lesson in Block2Learn’s analysis of Applied Materials’ AI demand, margins and export risk. Capacity suppliers can benefit from the buildout while the wider economy is still paying the adjustment cost. It also complements the discussion of FactSet’s AI strategy and workflow moat: durable value depends on integration into trusted workflows, not merely access to a model.

Higher productivity can lift the neutral rate instead of triggering cuts

The most consequential part of the Fed’s AI debate concerns the neutral real interest rate, commonly called r-star. Neutral is the rate consistent with full employment and stable inflation when temporary shocks have faded. It cannot be observed directly. Policymakers infer it from a changing mix of growth, inflation, labor, financial and savings data.

The popular intuition says that higher productivity reduces costs, lowers inflation and gives the Fed room to cut. That is possible, but incomplete. A persistent acceleration in productivity can make investment more profitable, increasing firms’ demand for funds. It can raise expected future income, encouraging households to consume more and save less. Reconciling stronger investment with lower desired saving may require a higher equilibrium real interest rate.

Jefferson made the policy implication explicit: if neutral rises, a given federal funds rate becomes more accommodative than policymakers thought. A policy rate that looked restrictive under the old productivity regime may be only mildly restrictive—or even neutral—under a stronger one. The Fed could then ease too soon if it interprets faster growth as temporary demand resilience rather than a higher-return economy with a higher r-star.

The opposite risk exists as well. If firms overinvest, adoption disappoints, or labor displacement weakens consumption, the neutral rate may not rise as expected. A high nominal rate could become more restrictive as inflation falls and real borrowing costs increase. Jefferson cautioned that the historical relationship between productivity growth and real interest rates is noisy. Productivity is one force among many, including demographics, fiscal policy, global savings, risk demand and the supply of safe assets.

This uncertainty changes how to interpret bonds. The long end of the Treasury curve reflects expected short rates, inflation risk, term premium and real growth. A convincing AI productivity acceleration could reduce long-run inflation pressure while raising real growth and the neutral rate. Those effects pull nominal yields in opposite directions. The result need not be a secular bond rally. It could instead be higher real yields, a firmer term premium and lower breakeven inflation.

That is one reason cross-asset hedges can fail during regime changes. Block2Learn’s analysis of why bonds and gold can fall together shows how higher real yields can pressure assets that investors expect to diversify each other. An AI-driven rise in potential growth would not reproduce every feature of an inflation shock, but it could still keep real discount rates higher than a simple disinflation narrative implies.

Equities face a related split. Higher productivity can support margins and long-run earnings, especially for firms that convert tools into lower costs or better products. Yet higher real rates reduce the present value of distant cash flows. The companies most exposed to optimistic terminal assumptions may be vulnerable even when the macro technology thesis is correct. The winners are more likely to be businesses that demonstrate measurable workflow gains, pricing resilience and disciplined capital intensity than those that merely attach AI language to existing operations.

Labor reallocation determines whether the supply gain is clean

Productivity growth is an aggregate benefit, but transitions are not aggregate experiences. A firm can produce more with fewer hours while displaced workers struggle to move into complementary roles. If retraining and job creation are fast, productivity can lift income and output with limited unemployment. If displacement outruns adaptation, the economy can experience higher productivity and higher structural unemployment at the same time.

This distinction matters for the Fed’s dual mandate. Monetary policy can support cyclical demand; it cannot directly fix a persistent mismatch between worker skills and employer needs. Cutting rates in response to structurally higher unemployment could overheat the sectors that are already expanding without restoring the jobs that technology eliminated. Holding rates too high during a cyclical slowdown, however, would compound the adjustment.

The evidence is likely to be noisy. Aggregate unemployment can remain stable while occupational churn rises. Payroll growth can broaden even as particular technical roles contract. Wage growth can cool because displaced workers accept lower-paid jobs, or remain firm because firms compete for scarce complementary skills. Participation, job openings, unemployment duration, industry-level hours and wage dispersion may reveal more than the headline unemployment rate alone.

An IMF working paper using AI patent data adds a useful long-run perspective. It estimates that AI innovation raised output per worker by 0.8% to 1.2% across OECD countries over 2000–2017 and could raise aggregate labor productivity by as much as 3.8% in the long run. The gains were larger where professional and managerial work was more prevalent, while worker mobility and the time required to learn complementary tasks shaped the payoff.

Those estimates should not be treated as a forecast for the next quarter. Patent-based measures capture innovation and diffusion imperfectly, and the study predates the current generative-AI wave. Its value is conceptual: productivity gains depend on complements. Skills, mobility, management and institutional adaptation determine how much technical possibility becomes measured output.

For markets, the labor path separates a benign supply expansion from a more difficult reallocation shock. The benign version combines higher productivity, rising real wages, stable employment and slowing unit labor costs. The difficult version combines capital deepening with weak labor share, persistent mismatch and uneven consumption. BLS reported the labor share at 53.0% in the third quarter, the lowest in a series beginning in 1947. One observation does not establish an AI effect, but it is an indicator worth watching as the investment cycle matures.

Three scenarios for the AI productivity regime

Scenario one: diffusion catches the investment boom

In the constructive case, enterprise adoption broadens beyond a few functions, workflow redesign reduces error and review costs, and new capacity relieves component bottlenecks. Productivity remains near or above 2% on a sustained basis, unit labor costs cool, and output grows without an equivalent increase in hours. Real wages improve because workers share in the efficiency gain, while unemployment stays close to full-employment levels.

The Fed can then tolerate stronger real growth without interpreting it as inflationary. Inflation converges toward 2%, but the neutral rate may settle above its pre-AI estimate because investment demand and expected income are higher. Policy easing is possible as inflation falls, though the terminal rate may be higher than markets once assumed. Equities benefit from stronger earnings, but valuation expansion is constrained by firmer real yields. Credit performs best where cash flow validates the capital spending.

Scenario two: investment runs ahead of usable output

In the overheating case, capital expenditure remains strong while diffusion stays concentrated. Data centers, power systems and chips absorb resources faster than the technology lifts economy-wide supply. Wealth effects support consumption, component prices remain firm, and inflation proves sticky. Productivity data may still print well in individual quarters, but gains are volatile, revised or narrow.

The Fed treats demand and inflation as the dominant signals. Rates stay restrictive for longer, and any assumed decline in neutral is postponed. Long-duration assets face pressure from higher real yields and term premium. Suppliers to the buildout may continue to report strong demand, but customers face a rising hurdle rate and greater scrutiny of returns on invested capital. The cycle becomes vulnerable to cancellation, delays or consolidation if monetization fails to catch up.

Scenario three: disruption arrives before productivity

In the adverse transition, firms use AI to reduce labor in specific tasks, but organizational bottlenecks prevent a broad output surge. Unemployment rises in exposed occupations, wage dispersion increases and consumption weakens. The Fed must decide whether the labor deterioration is cyclical or structural. A cyclical reading supports easing; a structural reading implies that lower rates cannot restore the previous job mix and may instead revive inflation elsewhere.

This scenario is not necessarily deflationary. Lower household demand can coexist with supply bottlenecks, fiscal responses or expensive infrastructure. Markets may initially price rapid cuts, then reverse if inflation remains above target. Corporate outcomes diverge sharply between firms with genuine productivity improvements and those using cost reduction to mask weak demand.

The indicators that can resolve the measurement test

No single release can establish an AI productivity regime. The most useful approach is to monitor a dashboard that distinguishes output gains from capital intensity and temporary composition effects.

First, productivity should be persistent across revisions and time horizons. Quarterly annualized numbers are volatile. Year-over-year growth, multi-year averages and the relationship between output and hours are more informative. A sustained improvement with stable hours is stronger evidence than one rebound quarter.

Second, unit labor costs should decelerate without a collapse in real compensation. Falling unit costs can reflect genuine efficiency, but a healthy diffusion path should eventually support real wages. If the labor share keeps falling while profits rise, the aggregate productivity gain may be socially and politically harder to sustain.

Third, adoption should broaden by industry and business function. Moving from experimentation in marketing or software toward operations, logistics, customer service, scientific work and regulated decision-making would show that AI is becoming organizational capital. Reliability, auditability and integration matter as much as model capability.

Fourth, bottleneck inflation should ease. Memory, logic chips, power equipment and construction inputs provide a real-time test of whether capacity is catching demand. Persistent scarcity indicates that the investment clock is still dominating. Falling relative prices alongside rising usage would signal that the supply response is arriving.

Fifth, labor reallocation should remain fluid. Job openings, hiring rates, unemployment duration, participation and occupational wage changes can distinguish temporary churn from persistent mismatch. A rise in unemployment concentrated among exposed skills would require different interpretation from a broad demand-driven downturn.

Finally, market measures should be read in combination. Higher real yields with stable or lower inflation compensation may indicate stronger expected potential growth or a higher neutral rate. Higher nominal yields driven by inflation compensation would suggest that demand and bottlenecks still dominate. Equity breadth can show whether gains are moving beyond infrastructure suppliers toward users that translate AI into revenue or lower costs.

Block2Learn assessment

AI productivity is real at the task level, plausible at the firm level and not yet proven as the dominant cause of the aggregate productivity improvement. That hierarchy is the right starting point. It avoids dismissing the technology while refusing to turn a promising mechanism into a completed macro regime.

The strongest current evidence is the combination of rising productivity, falling quarterly unit labor costs, broad but uneven adoption and a powerful investment cycle. The strongest caution is timing. Expectations are already affecting spending, valuations and scarce inputs, while the broad supply payoff remains gradual. That gap can keep inflation and real rates higher before it makes the economy cheaper to run.

The Fed’s measurement problem is therefore deeper than deciding whether one productivity print is strong. Policymakers must infer whether trend growth, structural unemployment and the neutral rate are changing together. Each variable is unobservable in real time and each changes the interpretation of the others. Faster growth can mean excess demand, higher potential supply or both. Higher unemployment can mean cyclical weakness, structural mismatch or both. A high policy rate can be restrictive under one estimate of neutral and insufficiently restrictive under another.

For investors, the durable thesis is not “AI lowers rates.” It is that AI changes the composition and timing of growth. The investment phase favors capacity providers but can tighten resource constraints. The diffusion phase favors firms with trusted workflows and measurable operating leverage. The mature phase may deliver lower unit costs and higher output while leaving the equilibrium real rate above the old regime.

That is a more demanding framework than a simple productivity boom, but it is also more useful. It links corporate adoption to labor data, inflation, the yield curve and valuation. Most importantly, it identifies what would falsify the thesis: productivity gains that fade after revisions, adoption that remains narrow, capital spending that fails to produce cash flow, or labor displacement that overwhelms new demand.

Conclusion

The third-quarter data have strengthened the case that US productivity is improving. They have not solved the attribution problem. AI productivity may be contributing, but the macroeconomic payoff depends on diffusion, organizational change and labor reallocation. Those processes move more slowly than capital spending and investor expectations.

That timing gap is why the technology can support future disinflation while complicating monetary policy today. Investment and wealth effects can lift demand; bottlenecks can raise prices; stronger expected returns can lift the neutral rate. Later, broader adoption can raise supply and reduce unit costs. The Federal Reserve must decide which phase is dominant without the benefit of clean real-time measurement.

The practical signal is not one spectacular quarter. It is a durable combination of stronger output, contained hours, cooling unit labor costs, broader adoption, easing bottlenecks and successful worker reallocation. If those pieces align, the economy’s speed limit has moved higher. If they do not, the AI boom may remain a capital-spending shock with a productivity promise attached.

For a structured path through productivity, inflation, interest rates and valuation, continue with the Block2Learn Learning Path.

Read. Think. Learn. Block2Learn.