--- id: ECON-MEAS-001 title: "Measuring the AI Economy" alternative_names: - "AI Economic Statistics" - "AI Adoption and Investment Measurement" category: "Economy & Policy → Measurement" volume: IX chapter: 54 entry_type: "Standard Article" level: 3 status: "Draft" last_reviewed: 2026-10-15 version: "0.1.0" prerequisites: - "ECON-COMP-001" related_entries: - "ECON-IND-001" - "ECON-LABOR-001" - "ECON-INFRA-001" - "CAP-BENCH-001" --- # Measuring the AI Economy ## 1. Concise Definition Measuring the AI economy means estimating, with stated definitions and units, how much AI is used, invested in, produced and paid for, and what it changes. The main data families are firm surveys of adoption, usage data from AI providers, investment and funding databases, national accounts and labor statistics, and physical indicators such as compute and electricity. ## 2. Overview & Significance Almost every headline about "AI adoption" or "AI investment" depends on a definition and a data source. Figures for the same concept can differ severalfold because of wording, sampling, weighting and scope. This article describes the main sources, what each measures, and its known limits, using dated figures only where they illustrate a measurement point. It is not a table of current statistics; see `FACT-001`. ## 3. Historical Development Official statistics were not designed to isolate AI: software and hardware spending, research and development, and intangible capital are reported in broad categories. Industry trackers (annual AI indexes, funding databases) filled the gap from about 2017. After the public release of general-purpose chat assistants in 2022, statistical agencies added AI questions to business surveys: the US Census Bureau's Business Trends and Outlook Survey (BTOS) added AI items, changed the wording in November 2025, and began a second AI supplement on 17 November 2025 (U.S. Census Bureau, 2026). Providers began publishing aggregated, privacy-preserving usage analyses (Handa et al., 2025; Chatterji et al., 2025). ## 4. Technical / Conceptual Explanation ### 4.1 Firm adoption surveys The BTOS is a biweekly, nationally representative survey of US businesses. The Census Bureau's report of 26 May 2026, covering data collected from 14 December 2025 to 3 May 2026, states that overall AI use hovered between 17% and 20% (19.8% at 3 May 2026), with the question "whether the business used AI in the past two weeks" and, from November 2025, the wording "in any business function" rather than "in producing goods or services." Use was 37% among firms with at least 250 employees and below 20% among firms with four or fewer (U.S. Census Bureau, 2026). The wording change means series before and after November 2025 are not directly comparable, a standard measurement break. A Census-based working paper using the 2026 AI supplement reports that from November 2025 to January 2026, 18% of firms used AI in at least one function (32% weighted by employment), rising to an expected 22% within six months; most adopters use it in few functions; 23% of firms report worker-level use (41% employment-weighted); two-thirds use AI to augment tasks, while employment reductions are rare (2%) (Bonney et al., 2026). The gap between 18% of firms and 32% of employment shows that firm-weighted and employment-weighted adoption answer different questions. ### 4.2 Provider usage data Usage logs from AI providers show what users do but only for that provider's users. Handa et al. (2025) mapped Claude.ai conversations to US Department of Labor O*NET tasks with a privacy-preserving pipeline; the abstract reports that software development and writing tasks together account for nearly half of usage, about 36% of occupations use AI for at least a quarter of their tasks, and about 57% of usage suggests augmentation versus 43% automation. Chatterji et al. (2025) classified a representative sample of ChatGPT conversations from November 2022 to July 2025, reporting that about 10% of the world's adult population had adopted ChatGPT by July 2025 and that non-work messages rose from 53% to over 70% of usage; two authors are OpenAI employees, as the paper discloses. Neither is a measure of the economy-wide population, and the classification of "augmentation" versus "automation" is the authors' coding. ### 4.3 Investment and funding data Aggregators of private-investment deals count funding to companies classed as AI companies. The Stanford AI Index 2026 economy chapter reports that global corporate AI investment more than doubled in 2025 and that private investment rose 127.5% (Stanford HAI, 2026). I could not read the chapter's data appendix, so I do not state its dollar totals or its source database; secondary summaries give conflicting or unverifiable totals [VERIFICATION NEEDED]. Funding data differ from capital expenditure: a funding round is a transfer of equity, whereas expenditure on chips and buildings (see [[ECON-INFRA-001]](../vol09_economy/ECON-INFRA-001_energy_data_centers_and_semiconductor_supply_chains.md)) appears in company filings. ### 4.4 National accounts and productivity Official productivity statistics measure output per hour with a lag and cannot attribute change to AI without further identification. Effects at task level (see [[ECON-LABOR-001]](../vol09_economy/ECON-LABOR-001_ai_productivity_and_labor_markets.md)) do not map directly to aggregate series. Free or low-priced AI services create consumer surplus that GDP does not capture; estimates of such surplus exist but vary and are not adopted here. ### 4.5 Physical indicators Training compute in floating-point operations, accelerator shipments, and data-centre electricity consumption provide quantities rather than values. The IEA's estimate of about 415 TWh for data centres in 2024 is a scenario-based estimate with acknowledged uncertainty (IEA, 2025; see [[ECON-INFRA-001]]). ## 5. Types, Components & Classifications | Data family | Unit | Typical source | Main bias | | :--- | :--- | :--- | :--- | | Firm adoption | % of firms; % of employment | Census BTOS | Wording and weighting; self-report | | Task/usage | Share of conversations | Provider logs | Single-provider population | | Investment | USD, deals | Funding databases, AI Index | Classification of "AI company" | | Capital spending | USD | Company filings | Mix of AI and non-AI use | | Productivity | Output per hour | National statistics | Lag; attribution | | Physical | FLOP, TWh, units | IEA, Epoch AI | Modelling assumptions | ## 6. Applications & Concrete Examples Two worked illustrations of definitional sensitivity: (1) the 18% (firm-weighted) versus 32% (employment-weighted) adoption figures from the same Census supplement; (2) the Census BTOS 19.8% (any business function, past two weeks, survey of businesses) versus ChatGPT's "about 10% of the world's adult population" (individuals, global, one product). These are different quantities and must not be compared. ## 7. Advantages, Limitations & Unresolved Questions - *Strengths*: government surveys are representative and transparent; provider logs are large and granular. - *Limitations*: self-reported use is ambiguous (drafting an email versus integrating a model in a process); wording changes break series (Census, November 2025); provider data are not representative; investment figures depend on classification. - *Unresolved*: how to price free AI services; how to separate AI from general software and cloud spending; how to measure agentic and embedded AI that users do not recognise as AI. ## 8. Current State of Knowledge As of 2026-10-15 the Census BTOS reports roughly one in five US businesses using AI in the past two weeks (the May 2026 release), concentrated in large firms and in information and finance sectors (Information 39.7% and Finance and Insurance 33.9% at 3 May 2026; Retail about 14%), per the Census report. A secondary-source figure for August 2026 was not confirmed from a Census release and is not used [VERIFICATION NEEDED]. The AI Index chapter states that productivity gains are measurable in narrow tasks while macro-level evidence is early (as relayed by secondary search summary; PDF text not read) [VERIFICATION NEEDED]. ## 9. Future Trajectory & Strategic Outlook Observed: statistical agencies are expanding AI questions and experimental supplements. No forecast is registered here. Good measurement practice for any new figure: state the definition, population, unit, date and weighting, and prefer primary datasets. ## 10. Related Entries & Knowledge Graph - **Foundational Concepts**: - [[ECON-COMP-001] The Economics of Compute and Scaling](../vol09_economy/ECON-COMP-001_compute_economics.md) - **Related Methods & Technologies**: - [[CAP-BENCH-001] AI Evaluation, Benchmarking & Contamination](../vol04_capabilities/CAP-BENCH-001_ai_evaluation.md): measuring capability rather than economic use. - [[ECON-LABOR-001] AI, Productivity and Labor Markets](../vol09_economy/ECON-LABOR-001_ai_productivity_and_labor_markets.md) - [[ECON-IND-001] The AI Industry Structure](../vol09_economy/ECON-IND-001_the_ai_industry_structure.md) - [[ECON-INFRA-001] Energy, Data Centers and Semiconductor Supply Chains](../vol09_economy/ECON-INFRA-001_energy_data_centers_and_semiconductor_supply_chains.md) ## 11. References & Foundational Literature 1. U.S. Census Bureau. (2026, May 26). *Large Firms With at Least 20 Employees Biggest AI Users* (Grundy, A., Breaux, C., & Khatiwoda, D.). https://census.gov/library/stories/2026/05/ai-use-businesses.html. [BTOS AI-use rates, wording change, firm-size and sector breakdowns, data 14 Dec 2025 to 3 May 2026]. 2. Bonney, K., Breaux, C. L., Dinlersoz, E., Foster, L. S., Haltiwanger, J. C., & Pande, A. A. (2026). *The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks*. NBER Working Paper 35141. https://www.nber.org/papers/w35141. [Firm-weighted versus employment-weighted adoption; augmentation versus employment reductions]. 3. Handa, K., Tamkin, A., McCain, M., Huang, S., Durmus, E., Heck, S., Mueller, J., Hong, J., Ritchie, S., Belonax, T., Troy, K. K., Amodei, D., Kaplan, J., Clark, J., & Ganguli, D. (2025). *Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations*. arXiv:2503.04761. https://arxiv.org/abs/2503.04761. [Provider-log mapping to O*NET tasks; authors are Anthropic-affiliated: preprint]. 4. Chatterji, A., Cunningham, T., Deming, D. J., Hitzig, Z., Ong, C., Shan, C. Y., & Wadman, K. (2025). *How People Use ChatGPT*. NBER Working Paper 34255. https://doi.org/10.3386/w34255. [ChatGPT usage composition and adoption to July 2025; OpenAI-affiliated authors disclosed]. 5. International Energy Agency. (2025). *Energy and AI*. https://www.iea.org/reports/energy-and-ai. [Example of a physical indicator and its stated uncertainty]. 6. Stanford Institute for Human-Centered AI. (2026). *AI Index Report 2026, Chapter 4: Economy*. https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_4_economy.pdf. [Industry tracker of investment and adoption; PDF text not extractable in this session, so only the headline growth statement relayed by a search summary is used]. ## 12. Evidential Classification Matrix | Section / Claim | Evidence Category | Supporting Source | Confidence / Status | | :--- | :--- | :--- | :--- | | BTOS rates, wording change, firm-size gradient | Established (official statistics) | Ref 1 | High | | 18% vs 32% adoption, augmentation vs reduction | Preliminary (working paper) | Ref 2 | Moderate | | Provider-log shares | Preliminary (company-affiliated preprints) | Refs 3, 4 | Moderate-Low for representativeness | | AI investment growth statement | Preliminary | Ref 6 (summary only) | Low-Moderate | | Comparability of different adoption measures | Established (definitional) | Refs 1, 4 | High | ## 13. Revision History - **v0.1.0** (2026-10-15): Initial draft. Citations checked against publisher pages where marked; not independently fact-checked.