FIELD NOTES / AI ECONOMY AUDIT
AI is advancing faster than the economy can absorb it.
A 44-source audit of productivity, business adoption, prices, jobs, infrastructure, households, and the feedback loops that would have to appear before the macro story becomes truly extraordinary.
Research note: Educational analysis only. This is not financial advice. Do your own research (DYOR) before making any investment decision.
01 / THE THESIS
The economy is running on two clocks.
On the fast clock, model capabilities improve, useful digital intelligence becomes cheaper, and chips, networking, data centers, and power infrastructure attract enormous investment.
On the slow clock, firms redesign workflows, workers learn new tools, managers decide what to trust, regulators adjust, and task-level gains become measurable output, wages, lower prices, or better products.
The distance between those clocks explains most of the apparent contradiction. AI can be genuinely transformative at the frontier while remaining difficult to isolate in aggregate economic data.
THE TRANSMISSION CHAIN
Evidence weakens as the claim gets larger.
Strong evidence exists close to the technology. Confidence falls as the claim moves from bounded tasks toward the whole economy.
-
01STRONG
Capabilities
Tested software and cyber capabilities are improving rapidly.
-
02STRONG
Infrastructure
The physical AI buildout is already macroeconomically important.
-
03STRONG
Bounded tasks
AI improves many bounded workflows, but speed, value, and accepted output are different quantities.
-
04UNEVEN
Organizations
Adoption is meaningful, but shallow within many firms.
-
05UNCERTAIN SHARE
Productivity
U.S. productivity strengthened. AI's exact contribution is unresolved.
-
06NOT ESTABLISHED
Macro transformation
Broad deflation and aggregate job loss are not established. Recursive acceleration is not established at scale.
THE RESEARCH MAP / 112 QUESTIONS
The audit began with questions, not conclusions.
The research posed 112 diagnostic questions across 13 categories. The final report consolidated their answers into 59 thematic sections, but the complete original framework is preserved here.
Do task-level gains survive review, correction, integration, security, and management overhead?
Would inflation have been materially higher without AI?
Are AI-exposed jobs weakening first through hiring and hours rather than layoffs?
Do compute, adoption, revenue, and reinvestment form a measurable feedback loop?
01 / 7 QUESTIONSUnderlying economic regime
- Do real GDP, real GDI, final sales, and gross output describe the same economy after accounting for normal revision patterns?
- Is real growth still strong per capita and per working-age adult?
- How much growth comes from private final demand versus government spending, inventories, trade, or unusually concentrated capital expenditure?
- Is growth broad across the median industry and region, or dominated by a handful of AI-related companies and locations?
- Has potential output genuinely accelerated, or is demand temporarily running above existing capacity?
- Do electricity use, freight, tax receipts, payrolls, and corporate revenue confirm the official output measurements?
- Does a state-space or dynamic-factor model identify a genuine change in the latent growth regime?
02 / 9 QUESTIONSHidden productivity growth
- Has labor productivity accelerated after adjusting for labor hoarding, changing hours, industry composition, and post-pandemic normalization?
- Has the growth-accounting residual increased after properly incorporating AI compute, software, data, and other intangible capital?
- Is productivity improving at the median firm, or primarily at a few frontier technology companies?
- Are gains occurring inside adopting firms, or because already productive firms are gaining market share?
- Do task-level time savings survive the inclusion of verification, corrections, integration, security, and management overhead?
- Are businesses producing more with unchanged resources, or temporarily maintaining output after reducing labor and increasing work intensity?
- Is there an AI implementation J-curve in which current investment raises measured costs before producing later productivity?
- Do formal structural-break tests find an acceleration without choosing the breakpoint after seeing the data?
- Is productivity itself accelerating, or merely remaining temporarily above its earlier trend?
03 / 7 QUESTIONSOutput and quality that statistics may miss
- Are GDP statistics missing internally produced AI software, model training, proprietary data, and organizational capital?
- Are free or bundled AI services creating consumer surplus that is absent from nominal expenditure?
- Are official deflators capturing improvements in speed, reliability, functionality, and output quality?
- What happens if AI prices are measured per successfully completed useful task rather than per token, subscription, or compute hour?
- Are firms delivering better products for unchanged prices, creating hidden real-output growth?
- Can hedonic price indexes measure declining prices for equivalent levels of intelligence or capability?
- Does an alternative quality-adjusted output index materially change estimated productivity or inflation?
04 / 10 QUESTIONSHidden AI-related deflation
- Are quality-adjusted prices falling in AI-exposed services even when posted prices remain unchanged?
- Is the cost per accepted unit of legal, coding, design, analytical, or customer-service work falling?
- Are unit labor costs declining faster in highly AI-exposed industries than in credible control industries?
- Are AI cost savings being passed to consumers, retained as corporate margins, or reinvested into greater output?
- Would inflation have been materially higher without AI? In other words, is AI producing counterfactual disinflation even while headline inflation remains elevated?
- Are falling digital-service prices being obscured by housing, healthcare, energy, tariffs, or other high-weight CPI components?
- Do producer-price declines propagate through input-output networks into consumer prices with a measurable lag?
- Are official services-price indexes inadequately adjusting for rapidly improving digital quality?
- Does matched-item or trimmed-mean inflation reveal an AI-related decline that headline indexes conceal?
- Are these price changes genuinely AI-caused rather than consequences of weak demand, imports, subsidies, competition, or base effects?
05 / 9 QUESTIONSLabor-market effects beneath unemployment
- Are AI-exposed occupations experiencing weaker hiring, hours, wage growth, or entry-level openings than comparable occupations?
- Are layoffs being concealed by attrition, hiring freezes, contractor reductions, or slower replacement hiring?
- Is stable unemployment accompanied by declining participation, reduced job-search intensity, or movement into lower-quality employment?
- Are young workers and recent graduates showing displacement before the aggregate workforce?
- Are AI-related task substitutions eventually creating complementary jobs at the firm or industry level?
- Do payrolls, vacancies, recruiter activity, freelance work, and job postings tell the same story after removing duplicate and phantom vacancies?
- Are workers receiving higher compensation for higher productivity, or are gains primarily accruing to margins and capital owners?
- Do newly created AI jobs offset displaced jobs in number, compensation, skill requirements, and geography?
- Is occupational transition happening quickly enough to prevent persistent mismatch or underemployment?
06 / 8 QUESTIONSCausal attribution to AI
- Do comparable firms with earlier AI adoption outperform later adopters after controlling for pre-adoption trends?
- Were apparent AI adopters already better managed, financed, or more productive before adoption?
- Can effects be identified using plausibly exogenous deployment timing, task exposure, compute availability, or model access?
- Do difference-in-differences and event-study estimates show parallel pre-trends and effects beginning only after deployment?
- Do randomized or staggered internal rollouts produce the same gains as observational studies?
- Do results survive controls for ordinary software investment, capital deepening, fiscal stimulus, monetary conditions, offshoring, and post-pandemic normalization?
- Are measured effects heterogeneous by task, firm size, worker skill, and implementation quality?
- What observable outcome would falsify the claim that AI caused the improvement?
07 / 9 QUESTIONSAI adoption and realized economic value
- Are enterprises progressing from experiments to recurring, mission-critical production workflows?
- Are paid usage, retention, renewals, and customer expansion improving after subsidies or introductory pricing end?
- What percentage of AI activity produces accepted, economically useful work rather than retries, monitoring, synthetic traffic, or low-value output?
- Are reported time savings converted into additional output, better quality, shorter working hours, or merely more low-value activity?
- Does realized value exceed the full cost of models, compute, integration, human oversight, compliance, and organizational change?
- Are companies obtaining revenue growth as well as cost reductions?
- Is usage expanding because customers demonstrate genuine willingness to pay, or because services remain priced below full economic cost?
- Does falling cost per unit of intelligence stimulate enough additional demand to increase total revenue and compute consumption?
- Are adoption and measurable ROI broadening beyond hyperscalers and technologically sophisticated firms?
08 / 9 QUESTIONSAI capital expenditure and capacity
- How much AI capital expenditure represents additional real compute and infrastructure versus rising equipment, electricity, and construction prices?
- Is installed capacity being utilized productively, or accumulating as unfinished projects, inventory, or underused equipment?
- Are useful workloads, revenue, and cash flow increasing proportionally with deployed compute?
- What is the incremental return on AI capital after energy, networking, maintenance, depreciation, and financing costs?
- Are depreciation schedules realistic given rapid hardware obsolescence?
- Is capital expenditure funded through operating cash flow, durable customer commitments, debt, equity, subsidies, or circular financing arrangements?
- Are power, grid, permitting, memory, networking, and skilled-labor bottlenecks delaying the conversion of spending into output?
- Would a demand slowdown expose genuine overcapacity, or could the infrastructure be profitably repurposed?
- Does AI capital complement the existing capital stock, or prematurely render portions of it obsolete?
09 / 8 QUESTIONSHouseholds, demand, and distribution
- Is consumption being sustained by real income growth, wealth effects, credit expansion, or declining savings?
- Do median households show the same resilience as aggregate consumer spending?
- Are delinquency, revolving-credit usage, and debt-service burdens deteriorating beneath healthy headline consumption?
- Are AI benefits raising median real incomes and purchasing power, or primarily increasing profits and asset values?
- Is the labor share of income rising or falling in highly AI-exposed industries?
- Are productivity gains reducing prices for lower-income households or concentrating benefits among higher-income users?
- Is AI increasing business formation and competitive entry, or strengthening incumbent market power?
- Are regional, educational, and occupational inequalities widening as AI activity clusters geographically?
10 / 9 QUESTIONSMonetary, fiscal, and financial conditions
- Are higher real interest rates reflecting stronger productivity expectations, inflation risk, fiscal borrowing, term premia, or increased demand for investment capital?
- Has AI raised the neutral real interest rate by creating more profitable uses for capital?
- Is AI currently inflationary through investment demand and bottlenecks before becoming disinflationary through productivity?
- Have estimates of potential output, the output gap, and noninflationary employment become biased by omitted AI supply effects?
- Are financial conditions restrictive for ordinary businesses while remaining accommodative for large AI firms?
- Is credit creation flowing toward productive capacity or increasingly speculative and leveraged projects?
- Are widening credit spreads identifying genuine project risk that equity markets overlook?
- Is fiscal policy reinforcing AI investment while crowding out other private activity?
- Would future rate cuts reflect successful AI-driven disinflation or economic weakness caused by displacement and declining demand?
11 / 7 QUESTIONSInternational and sectoral controls
- Do countries with greater plausibly exogenous AI access subsequently achieve stronger productivity than comparable countries?
- Are U.S. gains still exceptional after controlling for fiscal policy, demographics, energy, exchange rates, and industrial composition?
- Do highly AI-exposed industries outperform credible low-exposure placebo industries?
- Are productivity gains appearing in countries adopting AI without sharing the U.S. technology-investment cycle?
- Are domestic gains offset by imported hardware, outsourced labor, or foreign intellectual property?
- Is AI increasing global productive capacity or merely shifting profits and market share between countries?
- Are export controls and geopolitical fragmentation slowing diffusion enough to alter aggregate outcomes?
12 / 9 QUESTIONSAcceleration and singularity-like dynamics
- Are AI-assisted research, engineering, and software-development cycles shortening measurably with each model generation?
- Is AI contributing to the design of improved models, algorithms, chips, data centers, or scientific instruments in independently validated ways?
- Are capability improvements translating into economic output with progressively shorter lags?
- Do compute efficiency, capability, adoption, revenue, and reinvestment form a measurable self-reinforcing feedback loop?
- Is the elasticity of economically useful output to compute increasing rather than diminishing?
- Are scientific discoveries per researcher, research dollar, and unit of time accelerating?
- Are bottlenecks disappearing, or merely shifting from compute to power, data, reliability, regulation, and physical execution?
- What measurable thresholds in productivity, automation, scientific output, and iteration speed would distinguish an industrial revolution from a singularity-like regime?
- What evidence would distinguish genuine recursive acceleration from rapid but conventionally financed capital deepening?
13 / 11 QUESTIONSStatistical safeguards and falsification
- Were the hypotheses, exposure definitions, comparison groups, and time windows specified before examining the outcomes?
- How many relationships were tested, and do findings survive false-discovery or family-wise-error corrections?
- Are results robust across data sources, transformations, seasonal adjustments, endpoints, and revision vintages?
- Do uncertainty intervals incorporate serial correlation, clustered shocks, errors in variables, and model uncertainty?
- Do placebo dates, industries, and outcomes reproduce the supposed AI effect?
- Do the models improve genuinely out-of-sample forecasts relative to simple historical baselines?
- Are nonlinearities and structural breaks supported by formal tests rather than visually selected trends?
- Could survivorship, publication, vendor, selection, look-ahead, or reverse-causality bias explain the findings?
- Are effect sizes economically meaningful as well as statistically significant?
- Does Bayesian model comparison favor an AI-regime explanation over ordinary-cycle, fiscal, monetary, and post-pandemic explanations?
- What future evidence would materially reduce, not merely increase, confidence that AI is transforming the economy?
02 / THE MACRO SNAPSHOT
The economy is not sending one clean signal.
Private demand is stronger than headline GDP. Productivity is strong over a year but weak in the latest quarter. Hiring is soft, inflation remains elevated, and household buffers are thin.
| Dimension | Latest reading | Why the counter-reading matters |
|---|---|---|
| Growth, Q2 2026 | 1.5% real GDP, annualized | 3.9% real private final sales to domestic purchasers |
| Production vs. income, Q1 | 2.1% GDP | 1.2% GDI; their average was 1.7% |
| Labor, June | +57,000 payrolls | 4.2% unemployment; 61.5% participation |
| Prices, June | 3.7% PCE inflation, year over year | 3.3% core PCE remained above target |
| Households, June | 2.7% saving rate | A limited buffer if employment or prices weaken |
| Productivity, Q1 | 0.3% quarterly annualized | 2.8% year over year |
| Distribution, Q1 | 53.7% labor share | Lowest recorded value since the series began in 1947 |
Sources: BEA GDP, BEA income and spending, BLS employment, and BLS productivity. Caveat: advance and monthly estimates are revisable and do not identify an AI effect.
03 / THE PRODUCTIVITY DETECTIVE STORY
Productivity strengthened. Causality did not arrive with it.
Recent growth is stronger than the previous business cycle. That is an important fact. It is not proof that generative AI caused the change.
04 / ADOPTION DEPTH
Adoption is broad enough to matter and shallow enough to bottleneck.
Counting a firm as an adopter does not reveal how many functions, workers, or consequential decisions actually use AI.
AMONG ADOPTING FIRMS
Capability is not adoption. Adoption is not intensity. Time saved is not automatically output. Tokens are not accepted work.
05 / THE DEFLATION PARADOX
Digital intelligence gets cheaper while the system around it stays scarce.
The correct denominator is not price per token. It is the full cost of an accepted, correct, secure, integrated economic task.
Strongly deflationary
Model capability per dollar has improved rapidly, but measured capability and complete task cost depend on inference budget, failures, review, integration, security, subscriptions, and task difficulty.
Locally inflationary
Advanced chips, memory, networking, grid capacity, electricity, metals, land, cooling, construction, financing, and specialized labor remain scarce.
Estimated annual addition to 2022–25 GDP quantity growth when free ad-supported digital content is treated as barter output.
- Not an official GDP revision
- Not entirely attributable to AI
Falling prices in selected technology categories predate generative AI. Current aggregate data do not isolate meaningful AI-caused consumer-price deflation.
- Category weights are small
- Attribution remains unresolved
06 / WORKERS AND HOUSEHOLDS
The gains are not reaching everyone at the same speed.
Recent indicators show a pronounced early-career divergence in AI-exposed occupations, while all-age effects remain modest and causal attribution remains unresolved.
Hiring changes
Early effects may arrive through fewer openings, lower entry-level recruitment, attrition, contractor changes, or task reassignment rather than mass layoffs.
Several shocks overlap
Rates, remote work, technology overhiring, venture retrenchment, offshoring, tax changes, and slower labor-force growth complicate attribution.
Augmentation dominates
Representative Census data show augmentation is far more common than AI-related headcount reductions among current adopters.
07 / THE BUILDOUT AND STOCKS
The buildout is the clearest signal. Returns will still disperse.
Infrastructure spending proves conviction and current demand. It does not guarantee an attractive return for every operator, supplier, project, or stock.
The physical buildout is real
Compute, memory, networking, data centers, power, cooling, and grid infrastructure are already macroeconomically important.
Scarce complements can benefit first
Providers of constrained inputs may capture value before AI gains become visible in median household income or broad consumer prices.
Every project earns its cost of capital
Utilization, financing, depreciation, grid access, obsolescence, geography, market power, and overcapacity can separate winners from losers.
A more useful stock-research checklist
- 01
Accepted work Does the product create correct, usable output after review?
- 02
Full-cost ROI Do customer gains survive integration, errors, compliance, and support?
- 03
Utilization Is installed capacity becoming productive revenue fast enough?
- 04
Capital structure Can cash flow absorb capex, depreciation, leases, debt, and dilution?
- 05
Durability Does the advantage survive cheaper models, new hardware, and customer insourcing?
Important: This discussion is educational and general. It does not recommend any security or strategy and is not financial advice. Markets can move against even a correct technology thesis. Verify the data, study the company-specific fundamentals, and DYOR.
08 / THE RECURSIVE LOOP
Evidence of an AI-R&D feedback loop is emerging. Self-sustaining acceleration is not established.
Faster models and larger buildouts are visible. Independent measurements of AI-assisted R&D are emerging in narrow settings, but a self-sustaining loop from AI-assisted R&D to faster AI progress and broad economic acceleration is not established.
09 / THE EVIDENCE LEDGER
What the audit can say with confidence.
| Grade | Proposition |
|---|---|
| Strong | Frontier capabilities are advancing rapidly in tested software and cyber domains; transfer to messy real-world work remains uneven. |
| Strong | AI improves productivity on many bounded tasks, but speed, value, and accepted output are different quantities. |
| Strong | AI infrastructure spending is macroeconomically important. |
| Moderate to strong | U.S. productivity has genuinely strengthened. |
| Moderate to strong | Official statistics miss some digital welfare, quality, and produced data. |
| Plausible | AI already contributes positively to aggregate productivity. |
| Plausible | Recent evidence suggests localized early-career effects in exposed occupations; breadth and causality remain unresolved. |
| Weak | AI explains most of the recent aggregate productivity improvement. |
| Weak | AI is already causing measurable economy-wide consumer-price deflation. |
| Weak | AI is already causing aggregate employment decline. |
| Not established | The economy is in a measurable recursive singularity loop. |
10 / WHAT WOULD CHANGE THE VERDICT
Five signals worth watching.
A stronger conclusion should require stronger evidence, especially as the claim moves from useful tools to economy-wide transformation.
- 01
Broad total factor productivity
Measured gains after separating AI capital deepening from genuine efficiency.
- 02
Full-cost adopter productivity
Output and quality after integration, review, security, training, and failure costs.
- 03
Validated scientific output
More reliable discoveries per researcher or R&D dollar, not more drafts or tokens.
- 04
Cost per accepted economic task
A quality-adjusted measure that prices the complete workflow rather than raw inference.
- 05
AI-assisted acceleration of AI research
Independent evidence that AI meaningfully shortens the loop that produces better AI.
11 / METHOD
How the claims were graded.
Four research files, 44 sources, official data first, then representative microdata, experiments, established papers, company evidence, scenarios, and commentary.
Separate the layers
Capability, adoption, intensity, output, revenue, capital spending, and accepted work were treated as different variables.
Search for counterevidence
Alternative explanations such as rates, remote work, pandemic reallocation, offshoring, and pre-existing trends were tested.
Label causal reach
Official snapshots, working papers, experimental accounts, observational studies, and synthesis were not presented as equivalent.
Preserve uncertainty
Five material corrections were incorporated before publication, including narrower language on deflation, causality, and adoption depth.
Data note: The research cutoff is July 30, 2026. Advance GDP, monthly employment, inflation, productivity, and seasonal estimates can be revised. This article should be read as a dated evidence audit, not a timeless forecast.
Source policy: Changing claims use evidence published or substantively updated from May 1 through July 30, 2026. Live official series are fixed to their latest available vintage at the cutoff. The April Census adoption-depth study and the Federal Reserve project database originally published in December 2025 remain as clearly dated exceptions because no newer source reproduces their measurements; the Federal Reserve database was updated July 16, 2026.
12 / SOURCE LIBRARY
The full 44-source trail.
Primary and official sources appear first within each category. Every entry shows its publication or substantive-update date. Older exceptions are labeled directly.
01–10 Macroeconomy and productivity
- BEA: GDP, Advance Estimate, Q2 2026
- BEA: Personal Income and Outlays, June 2026
- Federal Reserve: July 2026 Monetary Policy Report, Part 1
- Federal Reserve: July 2026 Monetary Policy Report summary
- Federal Reserve: Industrial Production
- BLS: Productivity and Costs, Q1 2026 revised
- BLS: Nonfarm business-sector productivity and costs, Table 2
- BLS: Productivity tables and data
- FRED: Nonfarm Business Labor Productivity, OPHNFB
- OECD: Compendium of Productivity Indicators 2026
11–22 Labor, adoption, and realized value
- BLS: Employment Situation, June 2026
- Stanford Digital Economy Lab: AI Economic Indicators
- Federal Reserve Governor Barr: AI, Living Standards, and Inequality
- OECD: Employment Outlook 2026
- New York Fed: Remote Work and Younger Workers
- ILO: Generative AI and Labour Markets in ASEAN
- Census: AI Use Among U.S. Businesses
- Census: Microstructure of Business AI Use
- NBER Digest: Global Evidence on Business Use of AI
- BEA: AI Expectations and Outcomes
- ILO: The Aggregation Paradox of AI
- METR: Technical-Worker Productivity Survey
23–32 Buildout, capabilities, and hidden output
- Federal Reserve: The AI Buildout and the Economy
- UK AISI: Frontier AI Trends Report
- METR: Frontier Risk Report
- UK AISI: Why Agent Evaluations Must Account for Test-Time Compute
- Federal Reserve: Estimating Aggregate Data-Center Investment
- BEA: Free Digital Content and AI Impacts on Growth
- BEA: Digital Economy Data Hub
- BLS: Artificial Intelligence and Productivity
- UK AISI: Frontier AI Cloud-Misconfiguration Case Study
- METR: Task Substitution and Uplift
33–44 Prices, finance, and long-horizon studies
- BLS: Consumer Price Index, June 2026
- BEA: Cost and Price Patterns in AI-Intensive Industries
- Federal Reserve: Measurement of Software Inflation
- METR: Expenditure Horizon
- Federal Reserve Governor Cook: AI Opportunities and Risks
- UK AISI: Cheating Behavior in Frontier Model Evaluations
- Federal Reserve: April 2026 Senior Loan Officer Survey
- Bank of England: July 2026 Financial Stability Report
- BIS: Progress and Peril
- IMF: World Economic Outlook Update
- METR: Metrics of Agent Ability
- IMF: Aggregate Gains from AI and Their Distribution
THE BOTTOM LINE
A real revolution can still be early, uneven, and difficult to measure.
AI capabilities are advancing quickly. The infrastructure commitment is immense. Bounded workflows show real gains. Productivity has strengthened. Yet the organizational, distributional, and macroeconomic transmission remains incomplete.
The most useful stance is neither denial nor inevitability. It is disciplined attention to where the signal is already strong, where it is still weak, and which measurements would genuinely change the conclusion.
Educational content only. Not financial advice. Do your own research (DYOR). Research cutoff: July 30, 2026.