Implementation Became the Mechanism

By June 2026, five propositions had become reasonably defensible.

First, frontier capability was improving rapidly, especially on bounded language, coding, and analytical tasks. Second, access to generative AI often raised task-level speed or quality in experiments. Third, adoption was broad but shallow: many workers had tried the tools, while only a small share of total hours and complete workflows used them. Fourth, effects were heterogeneous across the “jagged frontier”: AI could help weaker performers on well-defined tasks and harm performance where contextual judgment or verification exceeded user skill. Fifth, aggregate productivity and employment data did not yet contain an unambiguous AI break.

This left a translation puzzle. The literature could show impressive capability and credible task gains but offered much thinner evidence on worker output over time, firm value added, organizational redesign, and economy-wide TFP. July's contribution is not to close that gap. It is to identify the conversion mechanisms inside it.

1. Implementation became an estimand

The JudgeGPT experiment does something most AI evaluations do not: it randomizes not only access but the complement that tells users where and how to deploy it. Targeted training increases usage at least fourfold relative to generic technology training, moves prompts away from costly-to-verify open legal questions, and produces a 6.3% output increase without a measured quality loss. Access alone is not the relevant treatment.

This finding aligns with the non-AI garment-factory RCT in the broad radar. A communication technology creates no value until HR incentives change. The common mechanism is organizational: the effective productivity of a tool depends on whether someone owns adoption, workers know the tool's comparative advantage, and verification is assigned to tasks where it is economical.

2. Guardrails became an economic input

Chen et al. randomize patient access to a medical chatbot before outpatient visits. Only 17.1% of those offered access use it, yet the intent-to-treat effects move real care: prescriptions fall 4.6 percentage points from an 87% baseline, while diagnostic testing rises 2.7 points from a 23% baseline. Total spending is unchanged.

The conversation logs reveal the mechanism. The chatbot attaches cautions to 90.8% of traditional-Chinese-medicine mentions and 87.6% of antibiotic mentions, but gives a clean recommendation in 94.5% of diagnostic-test discussions. The system's defensive, liability-sensitive design propagates into clinical choices, especially when physicians are receptive to patient input. This is not evidence that AI improved health: the paper has no clinical outcome with which to value fewer drugs against more tests. It is strong evidence that the developer's implicit loss function can become an input into another institution's production process.

3. Durable human capital depends on augmentation, not mere assistance

Contractor and Reyes distinguish assisted output from later unaided performance. Their 0.27-standard-deviation immediate knowledge gain retains 76% after one week. But use logs reveal the boundary: students who ask for explanations preserve benefits; students who outsource text see short-run essay gains vanish when the tool is removed. AI can be a tutor or a substitute for practice. “Used AI” is too coarse a variable to tell which.

4. Realized usage produces a different map from technical exposure

AI Premium uses 380 trillion tokens of actual consumption to infer an AI factor and firm-level market betas. Technical exposure measures explain under 2% of their cross-firm variation. Market exposure is more positive for interaction and communication content and more negative for analytical and information-use skills. That is surprising only if capability maps are mistaken for equilibrium incidence. Prices reflect adoption, complementarity, competition, bargaining, and expected rents—not merely whether an LLM can perform a task.

5. More compute is not monotonically more useful

Havránek and Irsova's negative experiment matters beyond academic feedback. Authors prefer a single frontier-model pass to two multi-agent debate systems, although one uses around 30 times the tokens. AI judges almost always rank the real human referee report last. The result is narrow but sharp: inference-time complexity can lower user value, and model-based evaluation can reverse the human outcome criterion.

6. The buildout itself may contain a contest externality

The BIS AI Investment Race calibrates a dynamic winner-take-most contest among hyperscaler–lab coalitions. Its baseline produces investment at 1.51 times the efficient level, a 50% bust probability, and expected destruction of $189 billion per year. These are model outputs, not forecasts. Still, the mechanism is important: private investment can be individually rational and collectively excessive when firms race for a few dominant positions, finance specialized assets with debt, and own circular stakes in one another.

The macro evidence is not yet internally consistent—and should not be made to look so.

Scott Davis reports that the three most AI-exposed US sectors grew productivity by 3.7% annualized since early 2024 versus 1.7% elsewhere. They account for 16% of hours but 40% of US productivity gains. Across European countries, greater Claude use is associated with a stronger exposure–productivity slope. This is suggestive of realized AI gains, but explicitly correlational.

Boyle, Fernald, and Li offer a competing decomposition. US output per hour grew about 2.5% annually from 2023 to 2026Q1, one point above 2005–19. Measured TFP contributed 0.8 point to the acceleration, yet since early 2024 their utilization estimate accounts for essentially all TFP growth, leaving little utilization-adjusted improvement. In this reading, AI may have raised demand and uncertainty, causing firms to work existing inputs harder—not yet smarter.

Both can be true. AI-exposed sectors may be improving while aggregate TFP is dominated by cyclical intensity; or exposure may proxy for pre-existing sectoral advantages. The decisive evidence will be sustained, utilization-adjusted productivity divergence within industries after comparable adoption, ideally linked to firm-level workflows and value added.

1. Capability is not economically weighted output

Benchmarks assign equal or arbitrary weight to tasks. GDP weights outputs by market value; welfare also includes time, quality, variety, consumer surplus, and nonmarket services. Saturating an exam does not reveal which production bottleneck has been relaxed.

2. Adoption is usually binary, intensity is continuous

“Uses AI” can mean one query per month or a redesigned production process. The Fed note finds uptake rising with firm size but emphasizes shallow intensity. Surveys need task shares, workflow penetration, model costs, verification time, and complementary investment.

3. Quality adjustment is especially hard in services

Courts can count cases, but justice quality is multidimensional. Education can count test answers, but durable reasoning differs from polished text. In many services, nominal revenue is deflated with imperfect prices, so free or cheaper AI-enhanced quality can disappear from measured real output.

4. AI capital straddles accounting categories

Data centers and computers enter investment; imported equipment subtracts through net exports; software is partly capitalized; data, model fine-tuning, training, and workflow redesign are incompletely measured intangibles. The Fed estimates that AI-related components contributed meaningfully to GDP growth from 2025 to 2026Q1, while imports offset much of gross investment in some quarters.

5. Exposure measures answer different questions

Engineering scores ask what AI could perform. Usage logs ask what people currently delegate. Stock betas ask which cash flows or discount rates covary with AI demand. These measures should disagree. Treating one as ground truth creates false contradictions.

6. Generated measurement can contaminate downstream inference

July's bank-run archive shows the upside of LLM extraction. But AI Premium, AI-judged paper feedback, text-based court-quality scores, and generated covariates all place model error inside the estimator. Validation must be designed around the target coefficient, not generic classification accuracy.

7. Utilization is not welfare

Fewer prescriptions, more diagnostic tests, more resolved cases, or more books are intermediate outcomes. Their value depends on clinical health, legal accuracy, consumer use, and the counterfactual cost of resources. AI research will overstate progress if it counts directional behavior as improved welfare without measuring the endpoint that gives the behavior economic value.