The AI Productivity Boom Is Still Waiting for Its Measurement Moment

The AI Productivity Boom Is Still Waiting for Its Measurement Moment

The most important fact about the artificial-intelligence cycle is not that investment is large, adoption is fast, or equity markets have capitalized the story aggressively. It is that the aggregate productivity data have not yet delivered the clean acceleration that would confirm the strongest macro version of the AI thesis. Headline total factor productivity growth is running near 1.5%, which looks respectable on the surface. But the utilization-adjusted measure associated with Susanto Basu, John Fernald, and Miles Kimball is below zero. That distinction matters. It says the economy may be producing more output because firms are running labor and capital harder, not because the same bundle of inputs is being transformed into output with materially greater efficiency.

That is not a bearish verdict on AI. It is a timing problem, a measurement problem, and an organizational-capital problem. General-purpose technologies rarely enter the national accounts as a clean vertical line. Electricity, the internal-combustion engine, computers, enterprise software, cloud computing, and mobile internet all required complementary investment before they changed the production frontier at scale. Firms had to redesign workflows, build new infrastructure, train workers, write new procedures, change procurement systems, reorganize management layers, and decide which old processes to abandon. AI is moving through the same passage. The tools can be impressive at the task level while the macro data remain unimpressed.

The danger for investors is that markets often price the endpoint before the economy has done the work required to reach it. In a useful AI boom, productivity should eventually show up as more output per hour, better margins from process redesign, faster product development, cheaper service delivery, and a higher return on intangible capital. In a less useful boom, spending rises first, utilization rises second, and measured efficiency arrives late or not at all. The difference between those paths determines whether AI is a durable supply-side shock or merely an expensive capital cycle with pockets of genuine innovation.

This article takes the productivity disappointment seriously without treating it as conclusive. The right question is not whether AI is fake. The right question is what evidence would distinguish an early-stage general-purpose technology lag from a capital-intensive investment boom whose economic returns are being overstated. That requires looking at utilization-adjusted productivity, intangible investment, diffusion, measurement, market pricing, labor-market reorganization, and the policy environment that either accelerates or slows adoption.

Headline TFP Is Not the Whole Story

Total factor productivity is supposed to capture the part of output growth not explained by measured labor and measured capital. In growth accounting, output growth can come from more hours worked, more capital per worker, better-quality labor, or a residual improvement in how inputs are combined. That residual is TFP. It is not pure technology, because it also contains measurement error, organizational change, economies of scale, resource reallocation, and sometimes plain statistical noise. Still, it remains one of the best macro indicators for whether an economy is becoming more efficient.

The headline number near 1.5% is therefore encouraging only up to a point. It suggests that broad productivity is not collapsing, and it gives optimists room to argue that the economy is adapting well despite high rates, post-pandemic distortions, and volatile investment composition. But headline TFP can be flattered when firms are using existing labor and capital more intensively. If factories run longer shifts, offices ask salaried employees to absorb more work, trucking fleets log more miles, or data centers operate closer to capacity, output rises even if the underlying production function has not improved.

That is why the Basu-Fernald-Kimball utilization adjustment is so important. The BFK framework tries to separate true efficiency growth from cyclical variation in how intensely inputs are used. It reflects a basic insight from macro production theory: measured inputs do not fully capture effort, workweek intensity, capacity utilization, and the shadow utilization of capital. When demand is strong, firms can squeeze more from the existing capital stock and workforce. Measured output rises, but it may not represent a permanent technology shift. When utilization-adjusted productivity is below zero while headline TFP is positive, the composition of growth becomes less flattering.

The implication is straightforward. Recent output gains may be coming from an economy working harder rather than working smarter. That is not sustainable as a long-run growth model. You can extend hours, raise effort, increase occupancy, sweat equipment, and absorb administrative tasks for a while. But unless process efficiency improves, fatigue, maintenance, wage pressure, quality problems, and bottlenecks eventually appear. A true productivity boom should reduce the amount of input pressure required to produce a given unit of output. The current aggregate data have not yet clearly shown that.

Why Utilization Adjustment Matters for the AI Debate

AI is often discussed as if a new tool instantly changes the economy's production frontier. That is not how macro data behave. A model that helps a programmer write code, an analyst summarize documents, a designer test variations, or a call-center worker answer questions may raise task-level productivity. But aggregate productivity depends on whether those task gains change the firm-level and economy-wide allocation of labor, capital, and output. The utilization adjustment helps identify whether the economy is seeing genuine efficiency or simply more intense deployment of existing resources.

Suppose a company introduces AI tools into customer service. Employees answer more tickets per day because the system drafts responses. At first, management may keep staffing unchanged, reduce backlog, and ask the same team to handle more channels. Output rises. But if the firm also spends heavily on software licenses, consultants, data cleaning, compliance review, model monitoring, and internal training, the net productivity effect may be modest. The visible productivity inside one task can be offset by new costs elsewhere. In the aggregate data, that may look less like a revolution and more like an investment phase.

A second example is software development. AI coding tools can speed up routine generation, documentation, testing, and search. Yet software organizations often respond by expanding the scope of projects, increasing review needs, adding security checks, and building more ambitious systems. Faster code creation does not automatically mean faster shipped value. If the bottleneck shifts from typing code to defining requirements, integrating systems, validating outputs, and maintaining reliability, the measured productivity gain is delayed. AI improves one margin while exposing another.

That is why the utilization-adjusted signal deserves attention. If the economy were already converting AI adoption into broad efficiency gains, one would expect to see stronger evidence after stripping out cyclical intensity. Instead, the adjusted measure being below zero suggests the boom remains incomplete at the macro level. AI may be useful, but its usefulness is still embedded in a transition where firms are investing, experimenting, and absorbing organizational cost before the efficiency dividend becomes visible.

General-Purpose Technologies Arrive With a Lag

The optimistic interpretation is the classic general-purpose technology lag. Paul David's work on electricity remains the canonical example: factories did not become radically more productive simply because electric motors existed. Early factories often replaced steam engines with electric motors without redesigning the production process. The full payoff required distributed power, new plant layouts, continuous-flow production, and managerial experimentation. The technology was real long before the productivity statistics fully reflected it.

Information technology followed a similar path. Robert Solow's famous observation that computers were visible everywhere except in the productivity statistics captured the frustration of the late 1980s. The later acceleration in U.S. productivity during the 1990s and early 2000s suggested that the delay was not proof of failure. It reflected the time needed for hardware, software, networks, supply chains, logistics, and business processes to become complementary. Erik Brynjolfsson and Lorin Hitt later emphasized that computerization paid off most when firms also invested in organizational capital.

AI fits this pattern unusually well. The models are powerful, but the production system around them is young. Many firms are still deciding which use cases are reliable, which data can safely be used, which workflows should be automated, which decisions require human review, and which risks are acceptable. The early phase is noisy because it contains pilots, failed deployments, vendor churn, duplicate tools, governance costs, and training overhead. Those activities consume resources before they generate durable output.

This means the negative utilization-adjusted reading should be read as a warning, not a final answer. It tells us the macro proof has not arrived. It does not tell us that proof cannot arrive. The better analogy may be a construction site: spending, disruption, and reorganization are visible before the completed building begins producing services. The relevant investment question is whether the construction is creating a productive asset or simply expanding cost without discipline.

The Complementary-Capital Test

The key test for AI is complementary capital. A general-purpose technology becomes economically important when other forms of capital are rebuilt around it. For AI, those complements include data infrastructure, cloud architecture, enterprise software integration, cybersecurity, model governance, worker training, redesigned job descriptions, procurement standards, legal review, and managerial measurement systems. Chips and data centers are only the most visible layer. The harder layer is the organizational redesign that lets firms convert model capability into lower unit cost or higher-quality output.

This is why the AI capex cycle can be both necessary and dangerous. Investment in compute, power, networking, and software is required if AI is going to scale. But high investment does not guarantee high productivity. In a Solow-style framework, capital deepening raises output when workers receive more or better capital. TFP rises only when the economy combines inputs more efficiently. If the boom merely gives firms more expensive capital without changing processes, returns can disappoint. If it enables new production methods, TFP should eventually improve.

Investors should therefore distinguish between gross adoption and productive adoption. Gross adoption asks whether employees have access to AI tools, whether companies have signed vendor contracts, and whether executives mention AI on earnings calls. Productive adoption asks whether cycle times fall, error rates decline, working capital improves, customer acquisition costs drop, research productivity rises, and revenue per employee increases after accounting for new costs. The second set is harder to measure but much more important.

The BFK-adjusted weakness implies that the second set has not yet dominated the macro data. Firms may be building the complements, but the payback is not broad enough to overwhelm utilization noise. That is normal for a young technology wave, but it should keep analysts disciplined. AI does not deserve credit for productivity that has not yet appeared. It deserves an option value on future productivity, with the size of that option depending on evidence of complementary-capital formation.

The Market Is Pricing a Future Supply Shock

Equity markets have largely priced AI as a future supply shock. Semiconductor companies, cloud platforms, data-center infrastructure, power equipment, software vendors, and selected industrial suppliers have benefited from the belief that AI will raise the economy's productive capacity and create large pools of new profit. That belief is not irrational. If AI can automate cognitive tasks, compress research cycles, improve software productivity, personalize services, and optimize logistics, it can raise both growth and margins.

But the market has priced the probability-weighted future faster than the national accounts have confirmed it. That gap is not unusual. Financial markets discount expected cash flows, while productivity statistics record realized output. Markets should move first if investors have credible information about future technology diffusion. The risk is that market prices can become too confident about timing, magnitude, and distribution. A real technology can still be a bad investment if the cash flows arrive later than expected or accrue to a narrower set of firms.

The distinction between beneficiaries and adopters matters. The early profits in a technology boom often accrue to infrastructure providers. During the railroad boom, track, steel, and land speculation moved before the full transportation productivity dividend matured. During the internet buildout, networking equipment and telecom capacity were overbuilt even though the internet itself became enormously valuable. In AI, the chip and data-center layer can be a profitable bottleneck for a while, but the broader productivity thesis requires downstream firms to use that infrastructure profitably.

That is why utilization-adjusted productivity is a useful reality check for market narratives. If AI is only shifting profits toward a few capital suppliers while the rest of the economy bears adoption costs, the macro effect will look different from a broad productivity revolution. If AI diffuses into thousands of firms and changes workflows, the adjusted productivity data should eventually improve. Markets can anticipate that, but they cannot indefinitely substitute anticipation for evidence.

Labor Intensity Is Not Productivity

The source of recent output gains matters because labor intensity has limits. If firms meet demand by pushing existing teams harder, measured output may rise in the short run. Salaried employees may absorb more tasks, service workers may handle more interactions per hour, managers may delay hiring, and teams may use AI tools to keep up with a heavier workload. This can look productive from a distance, but it may be closer to effort extraction than efficiency.

True productivity improvement changes the slope of the tradeoff. It allows workers to produce more value without proportionally more stress, hours, or capital wear. It reduces rework, shortens handoffs, lowers error rates, improves scheduling, and eliminates low-value tasks. AI should be judged by whether it changes those margins. A chatbot that lets a team process more tickets is useful. A redesigned service architecture that prevents tickets, resolves common problems automatically, and frees workers for complex cases is productivity.

The difference also matters for wages and inflation. If output gains depend on higher effort, workers eventually demand compensation or leave. Unit labor cost pressure can reappear. If output gains come from genuine efficiency, firms can share gains through wages, margins, and lower prices with less inflation pressure. That is why productivity is central to the macro policy debate. It determines whether strong growth is compatible with disinflation.

AI bulls should want the utilization-adjusted measure to turn decisively positive. That would suggest the economy is no longer merely squeezing more from existing inputs. It would indicate that workflow redesign, capital investment, and knowledge diffusion are raising the effective production frontier. Until then, the more honest statement is that AI has increased expected productivity, not proven realized aggregate productivity.

Measurement May Be Understating Some AI Gains

There is a legitimate counterargument: the data may be missing part of the benefit. Productivity measurement is notoriously difficult in services, software, health care, finance, education, and digital consumer products. If AI saves time in writing, search, translation, coding, research, customer support, or internal analysis, some of the welfare gain may not show up as measured output. A worker may produce better work, faster responses, or more optionality, while GDP records little change.

Quality adjustment is especially hard. If an AI-assisted analyst reviews more scenarios, catches more errors, and produces better decisions, the output may be more valuable even if the measured quantity of reports is unchanged. If a doctor uses AI to improve triage, the benefit may appear as reduced waiting time or better outcomes rather than more measured health-care output. If a small business uses AI to create marketing material that previously would not have existed, part of the gain may be consumer surplus rather than recorded revenue.

The digital economy has always created this problem. Free search, maps, open-source software, messaging, and online information improved welfare far beyond their direct price. AI may deepen that wedge. A household using AI for tutoring, planning, translation, or personal administration may experience real gains that national accounts undercount. Therefore, weak measured productivity is not the same as zero social value.

Still, investors cannot stop at the measurement defense. Financial assets are priced on monetized cash flows, not only unpriced consumer surplus. If AI creates benefits that are real but hard to monetize, the macro welfare story may be stronger than the equity story. If AI creates enterprise value, it should eventually appear in revenue per employee, margins, pricing power, lower cost growth, or faster product cycles. Measurement uncertainty should make us humble, not careless.

Diffusion Is Uneven Across Firms

Aggregate productivity can remain weak even when frontier firms are improving quickly. The reason is diffusion. A few leading firms may deploy AI effectively because they have clean data, strong engineering cultures, cloud-native systems, and managers who can redesign processes. Many other firms operate with legacy software, fragmented databases, legal uncertainty, thin IT budgets, and workflows that are hard to automate. The average productivity number depends on how quickly frontier practices spread.

This unevenness is one reason market indexes can look stronger than the real economy. Large technology firms and AI infrastructure suppliers may show visible revenue growth and margin strength while small and mid-sized firms face adoption frictions. Public-market capitalization then becomes concentrated in the firms closest to the AI buildout, even though most employment sits elsewhere. Aggregate TFP will not surge until the benefits move beyond the winners at the top of the distribution.

The literature on superstar firms and intangible capital is relevant here. When technology raises scale economies, the best firms can pull away. They use data, software, and network effects to expand margins and market share. That can raise profits without proportionately raising economy-wide productivity if weaker firms stagnate. A stock-market AI boom can therefore coexist with an ambiguous macro productivity picture for a surprisingly long time.

For policymakers and investors, the key variable is diffusion velocity. Are AI tools becoming cheaper, easier to integrate, safer to govern, and more usable by ordinary firms? Are workers learning how to redesign tasks around them? Are vendors solving the boring integration problems, not just showcasing impressive demos? The productivity boom depends less on the maximum capability of frontier models than on the median firm's ability to use them repeatedly and safely.

The Portfolio Implications Are Conditional

The investment implication is conditional optimism. AI is a real option on higher productivity, but current aggregate data do not justify treating that option as already exercised. Investors should continue to own exposure to firms with durable AI bottlenecks, credible demand, strong balance sheets, and the ability to convert capex into high-return assets. But they should be wary of paying for productivity acceleration in companies whose margins, free cash flow, or customer economics do not yet show it.

For equities, the question is not simply whether a company uses AI. Almost every company will claim to use AI. The better questions are whether AI reduces unit costs, increases revenue per employee, improves customer retention, compresses development cycles, or expands addressable markets. A credible AI beneficiary should be able to explain the mechanism. Vague language about transformation is not enough. The productivity thesis must eventually become operating leverage.

For bonds, the signal is also complex. If AI ultimately raises productivity, it can reduce inflation pressure for a given level of demand. But if it also raises expected growth, capital demand, and investment spending, the neutral real rate may be higher. That means AI can be good for real growth without recreating the low-rate world of the 2010s. The transition phase can even be bond-unfriendly if capex demand is large and productivity gains are delayed.

For credit, the distribution matters. AI may help large, profitable firms widen margins while pressuring labor-intensive service companies, outsourcing models, and firms with weak technology stacks. It can improve some borrowers' cash-flow resilience while increasing disruption risk for others. The productivity data will not tell that story at the issuer level. Analysts need to map AI exposure to costs, customers, labor mix, software architecture, and competitive position.

What Would Confirm the Productivity Boom

Several signals would make the AI productivity thesis more convincing. The first is sustained improvement in utilization-adjusted TFP, not only headline TFP. The second is broadening across industries, especially outside the technology sector. The third is evidence that revenue per employee and margins are improving after accounting for AI-related spending. The fourth is a decline in error rates, cycle times, backlog, and administrative overhead in measurable enterprise processes.

The fifth signal is labor reallocation rather than simple labor squeezing. If AI raises productivity, workers should gradually move from routine information-processing tasks toward supervision, exception handling, client work, product design, data stewardship, and higher-value judgment. That transition can be disruptive, but it is healthier than simply asking the same workforce to process more work with the same organizational design. Productivity comes from redesign, not just acceleration.

The sixth signal is falling integration cost. Early AI deployments are expensive because firms need consultants, data work, governance, security, and workflow redesign. Over time, the technology becomes more productive if the cost of deployment falls and reusable patterns emerge. Cloud computing followed that path: what once required custom infrastructure eventually became standardized services. AI needs a similar move from bespoke experimentation to repeatable production systems.

The seventh signal is a stronger link between AI capex and downstream output. Data centers and chips are inputs. The macro payoff requires better software, better services, better logistics, better research, better health care, better manufacturing, and better public administration. If the input boom grows faster than downstream monetization, margins and returns will eventually be questioned. If downstream output begins to accelerate, the early investment phase will look justified.

What Would Falsify It

The thesis would weaken if utilization-adjusted TFP stays negative while AI investment remains extraordinary. That combination would suggest that the economy is adding cost and intensity without improving underlying efficiency. It would not mean every AI use case is useless, but it would challenge the idea that AI is already a broad macro productivity shock. Markets would need to reprice timing and breadth.

A second warning sign would be rising capital intensity without rising returns on invested capital. If firms spend more on compute, software, and automation but fail to lift margins or growth, AI capex would look less like productive investment and more like competitive necessity. Companies sometimes invest because they fear falling behind, not because the investment has clear positive net present value. That kind of arms race can be valuable for suppliers while disappointing for adopters.

A third warning sign would be excessive concentration. If a small number of infrastructure providers capture most of the economic surplus while customers struggle to generate returns, the aggregate productivity effect may be smaller than the market expects. A true general-purpose technology should eventually create broad user surplus, not only supplier margins. The internet ultimately did both, but the path included a large capital-market shakeout.

A fourth warning sign would be organizational resistance. AI can automate tasks, but it also creates governance, trust, legal, and cultural challenges. Workers may resist tools they perceive as surveillance or job threat. Managers may deploy systems poorly. Regulators may impose necessary but costly controls. Data may be too fragmented or sensitive. These frictions do not destroy the technology, but they slow the conversion of technical capability into measured productivity.

The Accounting Identity Behind the Narrative

A useful way to discipline the AI debate is to begin with a simple production function. If output is represented as Y = A F(K, L), then K is capital, L is labor, and A is the efficiency with which the economy combines them. AI can affect all three terms, but each channel has a different market implication. If AI mainly raises K because firms buy more chips, servers, power contracts, and software, then measured output may rise through capital deepening. If AI mainly changes L by allowing workers to handle more tasks per hour, the effect may first appear as labor augmentation. If AI truly raises A, then the economy is combining capital and labor more efficiently. The strongest macro claim requires the third channel.

The current data are uncomfortable because the third channel is not yet clearly visible. Headline TFP near 1.5% sounds like A is improving. The utilization-adjusted weakness says some of that improvement may be a mirage created by unusually intense input use. That distinction matters for valuation because capital deepening and TFP have different cash-flow signatures. Capital deepening requires upfront spending, depreciation, financing, and often higher fixed costs. TFP improvement can raise margins and growth without the same proportional capital burden. If AI remains mostly capital deepening, the winners may be narrower and the macro payoff slower. If it becomes true TFP, the upside broadens.

There is also a denominator problem. Firms are spending heavily on AI-related capital and operating expense today, while many benefits are prospective. During the buildout, the capital stock rises before output fully responds. That can temporarily depress measured productivity or returns on capital, even if the project is rational. But a temporary denominator problem has to end. At some point, the new capital must produce more saleable output, lower cost, or higher quality. Otherwise the investment was not delayed productivity; it was overinvestment.

This is why the distinction between gross domestic product and gross domestic experimentation matters. A society can run thousands of pilots, demos, proofs of concept, internal tools, and vendor migrations. Those activities are economically necessary, but they are not the same as final output. Some are learning investments; some are dead ends. The national accounts will not credit every experiment as productivity. They will credit the subset that changes production at scale. Investors should do the same.

The same logic applies inside companies. A management team may report millions of hours saved through AI assistance. That number is not meaningless, but it is incomplete. Saved hours become productivity only if they translate into fewer hours required for the same output, more output with the same hours, higher-quality output that customers pay for, or faster innovation that produces future revenue. If saved hours are reinvested into meetings, monitoring, duplicated review, or broader but unfocused projects, the accounting benefit evaporates. The firm feels busier and more modern, but the production function has not shifted much.

This is the core reason utilization-adjusted productivity should be part of the AI dashboard. It forces the debate away from anecdotes and toward the source of output growth. Are we seeing more capital? More intensity? Better input quality? Or a higher efficiency residual? The most bullish long-run case is still plausible, but the burden of proof belongs there. Technology narratives become financially dangerous when the residual is assumed rather than measured.

The Enterprise Dashboard Investors Should Demand

Investors need a practical dashboard for separating AI productivity from AI theater. The first metric is revenue per employee, adjusted for acquisitions and mix. If AI allows a firm to scale output without scaling headcount, revenue per employee should improve over time. This is not a perfect measure, because pricing, outsourcing, and product mix can distort it. But persistent improvement after the adoption period is a useful signal that the firm is getting operating leverage from knowledge work.

The second metric is gross margin and service cost per unit. Many AI use cases are supposed to reduce the cost of support, documentation, sales enablement, claims processing, compliance review, coding, testing, and internal analytics. If those use cases are real, the cost line should eventually show it. Management commentary should connect the tool to the cost pool. A vague claim that AI is everywhere is weaker than a specific claim that average handling time, engineering cycle time, claims leakage, fraud review cost, or customer onboarding expense has fallen.

The third metric is capital efficiency. AI infrastructure is not free. Companies that build or rent large compute capacity need to show that incremental gross profit justifies the spend. For hyperscalers and infrastructure suppliers, the relevant question is utilization, pricing durability, depreciation life, and customer payback. For adopters, the relevant question is whether AI spending is replacing other cost or adding a new layer. A tool that saves labor but adds equal software, review, and governance cost may still be strategically necessary, but it is not yet a margin story.

The fourth metric is error-adjusted throughput. AI can increase speed while also increasing review burden or operational risk. Productivity should be measured after quality control. A legal team that drafts documents faster but spends the same total time reviewing hallucinations has not achieved the advertised gain. A bank that automates credit memos but adds model-risk checks may improve consistency, but the net time saving must be measured honestly. Error-adjusted throughput is more useful than raw output volume.

The fifth metric is organizational redesign. Firms that simply bolt AI tools onto old workflows may report enthusiasm without productivity. Firms that redesign approval chains, data access, job scopes, internal controls, and customer journeys have a better chance of changing unit economics. This is where management quality becomes decisive. The technology is general, but adoption is local. Two firms can buy the same model and produce very different economic results because one changes the process and the other only changes the interface.

The sixth metric is diffusion beyond the expert class. If AI remains useful mainly to top engineers, top analysts, and sophisticated users, it can raise frontier productivity without transforming the median firm. A broader macro boom requires tools that average workers can use safely, repeatedly, and within clear governance. Training, interface design, workflow integration, and trust are therefore not soft issues. They are central to productivity realization.

This dashboard also helps avoid the common mistake of treating layoffs as automatic productivity evidence. Reducing headcount can improve margins in the short run, but it may reflect cyclical cost cutting rather than AI-driven efficiency. The stronger signal is sustained output, quality, and growth after the cost base changes. If firms cut workers and service quality deteriorates, the productivity claim is weak. If firms cut routine work, redeploy talent, maintain quality, and accelerate output, the claim is stronger.

Policy and Macro Feedback Loops

The productivity lag also has policy consequences. Central banks cannot responsibly set policy on the assumption that AI has already raised the speed limit of the economy. If utilization-adjusted productivity is weak, then strong demand may still create inflation pressure. The Federal Reserve and other central banks can acknowledge upside productivity risk, but they need realized evidence before treating faster growth as noninflationary. A premature assumption of AI-driven supply expansion would repeat a classic policy error: easing into a demand boom because a future supply response is expected.

At the same time, policymakers should not smother the experimentation phase. If AI is a general-purpose technology, society needs investment in power, data centers, broadband, cybersecurity, education, and public-sector modernization. It also needs competition policy that prevents bottlenecks from becoming permanent toll roads while still allowing large fixed-cost investments to earn returns. The policy target should be diffusion with accountability: make adoption easier, but require governance, security, and transparency where stakes are high.

Labor policy is equally important. The transition will not be painless if AI changes task demand. A productivity boom that simply weakens worker bargaining power can generate political backlash and slow adoption. A better equilibrium combines automation with training, mobility, credential flexibility, and wage gains where productivity truly improves. This is not only a fairness argument. It is also an efficiency argument. Workers who understand and trust the tools are more likely to redesign tasks around them rather than resist them.

Energy policy is another feedback loop. AI infrastructure requires electricity, grid capacity, cooling, and physical construction. If power constraints bind, the AI investment boom can raise costs and delay deployment. If grid expansion and generation investment keep pace, infrastructure becomes an enabler rather than a bottleneck. That means the AI productivity story is partly a real-assets story. The digital economy still depends on physical capacity.

Fiscal policy also enters the equation. Governments are eager to tax AI winners, subsidize domestic capacity, protect workers, and regulate risk. Each objective has logic, but the mix matters. Subsidies can accelerate domestic capability, or they can misallocate capital. Regulation can protect society, or it can freeze experimentation. Taxes can fund adjustment, or they can discourage investment. The productivity outcome depends on whether policy improves the complement system or fragments it.

The macro feedback loop is therefore circular. AI investment affects demand, rates, labor markets, power markets, and fiscal choices before it fully affects measured productivity. Those macro conditions then affect AI adoption. Higher rates raise the hurdle for long-duration investment. Tight labor markets make automation more attractive but also raise implementation costs. Power shortages slow data-center expansion. Political backlash changes regulation. The productivity path is not just a technology path; it is an economic equilibrium path.

Scenario Thinking for the Next Cycle

The base case should probably be delayed payoff rather than immediate disappointment or instant revolution. In this scenario, utilization-adjusted productivity remains noisy for a while, AI capex stays strong, infrastructure suppliers earn high profits, and adopters accumulate organizational learning. Over several years, reusable patterns emerge, integration costs fall, and productivity gradually improves. This would resemble the historical path of earlier general-purpose technologies. It would justify some market enthusiasm, but not every valuation.

The bull case is faster diffusion. Models become more reliable, enterprise software embeds AI cleanly, data governance improves, and firms redesign workflows more aggressively. Utilization-adjusted TFP turns positive, revenue per employee rises broadly, margins improve outside technology, and inflation pressure eases despite solid growth. In that world, AI would validate the supply-side thesis and could support a broader equity advance. The beneficiaries would expand from infrastructure suppliers to adopters with strong processes.

The bear case is capital overbuild. Firms spend heavily because competitors are spending, but use cases remain narrow, governance costs stay high, and end-user willingness to pay disappoints. Infrastructure capacity grows faster than profitable demand. Utilization-adjusted productivity stays weak, depreciation rises, and margins come under pressure. The technology would still matter, but the investment cycle would have pulled too much capital forward. That would look less like AI failure than like a familiar financial-cycle mistake around a real innovation.

A fourth scenario is concentrated productivity. AI works extremely well for a small number of firms with data scale, engineering talent, and distribution, but it diffuses slowly elsewhere. Equity indexes may remain supported because the largest companies keep winning, while aggregate productivity improves only modestly. This scenario is plausible and important because it separates market-cap-weighted outcomes from economy-wide outcomes. It can be bullish for index leaders while still disappointing for the median worker and firm.

Scenario thinking keeps the debate honest. The current data do not force one conclusion, but they do reject certainty. Below-zero utilization-adjusted productivity is not consistent with declaring that the macro productivity boom has already arrived. It is consistent with a technology transition whose payoff is still being built. The investment job is to assign probabilities, monitor evidence, and avoid paying the same dollar for capability, adoption, and realized productivity as if they were identical.

The Practical Read

The practical read is that AI remains a powerful productivity option, but the option has not yet paid out in the aggregate data. Headline TFP near 1.5% is not enough to declare victory when the Basu-Fernald-Kimball utilization-adjusted measure is below zero. The economy may be getting more output partly by working existing labor and capital harder. That is different from producing substantially more from the same inputs because the production system has become genuinely more efficient.

This distinction should make investors more disciplined, not more cynical. AI can still become a general-purpose technology on the scale of electricity, computers, or the internet. But if that happens, the evidence should gradually move from demos and spending announcements into firm-level operating metrics and economy-wide efficiency statistics. The lag is normal. The proof is still required.

The best framework is therefore staged. Stage one is capability: models become useful. Stage two is adoption: firms buy tools and run pilots. Stage three is complement formation: data, workflows, governance, and job designs change. Stage four is productivity: output rises relative to fully measured inputs, not merely relative to visible labor hours. The current economy appears somewhere between stages two and three. Markets are already pricing a version of stage four.

That gap between priced future and measured present is where both opportunity and risk live. If utilization-adjusted productivity turns positive and broadens across sectors, AI-led assets may deserve much of their premium. If the data remain weak, the market will have to separate genuine productivity compounders from companies selling or buying expensive hope. The right stance is not disbelief. It is evidence-based patience: respect the technology, respect the lag, and keep asking whether the economy is finally working smarter rather than simply working harder.