Experimental AI Research (Beta): This report was generated with AI assistance as part of our ongoing exploration of AI-powered research and analysis. The content has been reviewed and edited by humans, but may contain errors or inaccuracies.
Please verify critical data points independently. All claims cite public sources for transparency and reproducibility. This is not peer-reviewed academic research – treat findings as exploratory insights requiring further validation.
Cite This Report
Ingemarsson, L. (2026, April 20). Global AI Productivity Impact Report 2026 (Version 1.0). Alice Labs. https://alicelabs.ai/reports/global-ai-productivity-impact-report-2026
In 2026, AI delivers measurable productivity gains in specific tasks: +14% in customer support (+34% for novices), +26% more completed developer tasks, and +4% short-run EU firm-level labor productivity. Macro-level gains remain incomplete and unevenly attributable in national accounts.
- 1.Customer support: +14% issues resolved per hour with AI assistant; +34% for novice and low-skilled workers (NBER/QJE, 2023).
- 2.Software development: +26.08% completed tasks across 4,867 developers in three firms (Cui et al., NBER 2025).
- 3.Knowledge work (writing): -0.8 SD time, +0.4 SD quality on professional writing tasks (MIT, 2023).
- 4.EU firm-level: +4% short-run labor productivity gain from AI adoption with no adverse employment effect (BIS/EIB, Jan 2026).
- 5.Adoption is uneven: 18% of US firms (Fed, 2025); 20% of EU enterprises but 55% of large EU firms (Eurostat, Dec 2025); 9% of UK firms (ONS, 2023).
- 6.Macro estimates diverge sharply: Acemoglu +0.07pp/yr TFP vs Aghion-Bunel/OECD high-exposure up to +1.3pp/yr — different methods, different assumptions.
- 7.Top adoption barriers: 70.9% lack of expertise and 52.5% legal uncertainty (Eurostat, 2025); only 15.9% of US workers have employer-provided AI training (NY Fed, 2026).
- 8.Methodology rule: Never conflate worker-task gains, firm revenue productivity, and macro TFP — they measure different things.
The Global AI Productivity Impact Report 2026 (published 2026-04-20) evaluates whether AI is generating measurable productivity gains across workers, firms, sectors, and economies — using only public, official, or peer-reviewed sources. The strongest causal evidence is at the worker-task level: customer-support agents (+14%, +34% for novices), software developers (+26%), and bounded knowledge work all show robust gains. Firm-level evidence is increasingly positive — European firm data show short-run labor productivity gains without broad employment losses.
The macro picture is less settled. Credible annual TFP-gain estimates range from +0.07pp (Acemoglu's conceptual model) to +1.3pp (Aghion-Bunel; OECD high-exposure scenario) over the next decade. The dispersion reflects uncertainty about task coverage, adoption speed, and complementary investments in training, data, software, and workflow redesign. Official adoption data show diffusion is still incomplete and uneven: large EU enterprises (55%) are 2.75× more likely to use AI than the EU average (20%).
Limitations: realized aggregate productivity remains difficult to attribute specifically to AI; many high-profile studies are task-specific; official adoption measures are not fully harmonized across jurisdictions; macro estimates are scenario-driven rather than realized. The widely circulated MIT preprint on AI and scientific discovery was excluded after MIT stated it had no confidence in the paper's provenance, reliability, or validity.
Executive Summary
"AI productivity" is not a single statistical object. Worker throughput, firm revenue productivity, and economy-wide total factor productivity are different measurement systems. The 2026 evidence base shows a clear pattern: AI is producing large, well-identified gains in specific tasks and firms, while economy-wide gains remain incomplete and difficult to attribute in national accounts.
The strongest causal evidence comes from worker- and task-level field experiments. Customer support shows +14% issues resolved per hour, with +34% for novices (NBER/QJE). Software development shows +26.08% completed tasks across 4,867 developers in three firms (Cui et al., NBER 2025). Bounded knowledge work — writing, consulting tasks within the AI frontier — shows large quality and speed gains (MIT; HBS/BCG). Outside the AI frontier, gains evaporate or reverse: in the HBS/Berkeley Kenya field RCT, average treatment effect was zero, with gains concentrated among already-high-performing entrepreneurs (+20%) and losses for lower performers (-10%).
Firm-level evidence is increasingly positive. BIS/EIB analysis of European firms reports a +4% short-run labor productivity gain from AI adoption with no adverse short-run employment effect. Atlanta Fed executive surveys point to particularly strong gains in high-skill services and finance, with implied annual labor productivity contributions of ~0.8pp in 2025 and 2pp+ expected for 2026.
Macro estimates remain widely dispersed. Acemoglu's conceptual model implies +0.07 percentage points per year of TFP gain over the next decade; Aghion-Bunel and OECD's high-exposure scenarios reach +1.3 percentage points per year. IMF projects ~+1.0% cumulative TFP gain in Europe over five years, with a ~30% drag if regulation is binding. Adoption itself is incomplete: 18% of US firms (Fed Board, 2025), 20% of EU enterprises (Eurostat, Dec 2025), 9% of UK firms (ONS, 2023). Diffusion is heavily concentrated in large enterprises (55% of EU firms with 250+ employees) and knowledge-intensive services.
The dominant blockers are complementary capabilities, not the technology itself: 70.9% of EU enterprises cite lack of expertise; 52.5% cite legal uncertainty (Eurostat). Only 15.9% of US workers report employer-provided AI training (NY Fed, 2026). Without training, organizational redesign, data quality, and management practice upgrades, worker-level gains do not propagate to firm or macro productivity.
Key Findings
12 data-driven insights
01Customer support productivity rises 14% with AI; 34% for novices and low-skilled workers
+14% issues resolved per hour; +34% for novices
AI compresses skill premiums within occupations — strongest equity-and-throughput case for deployment in scaled support functions.
02AI coding assistants raise completed developer tasks by 26% in field studies of 4,867 devs
+26.08% completed tasks across three firms
Software is the most-validated AI productivity case at firm scale — robust across companies, not a single-firm artifact.
03Professional writing tasks complete 0.8 SD faster and at 0.4 SD higher quality with ChatGPT
-0.8 SD time, +0.4 SD quality
Knowledge work shows large gains on bounded writing tasks — but generalizability beyond writing is task-dependent (working paper).
04AI gains within the 'jagged frontier' reach +25% speed and +40% quality for consultants
+25% speed, +40% human-rated quality on in-frontier tasks
Task-boundary awareness is now a core management skill: outside the frontier, AI use can degrade output.
05European firm-level AI adoption raises short-run labor productivity by 4% with no employment loss
+4% short-run labor productivity
Best firm-level evidence to date that adoption raises productivity at scale without near-term displacement — at least in Europe.
06AI adoption is highly concentrated in large enterprises: 55% of EU firms with 250+ employees
20.0% of all EU enterprises; 55.03% of large enterprises
AI productivity diffusion will track firm-size structure — SME adoption is the binding constraint on aggregate gains.
07US firm AI adoption reached 18% by year-end 2025, with 39% of workers reporting use
18% firms (Fed Board); 39% workers (NY Fed)
Worker use outpaces firm adoption — shadow AI is producing measured productivity that firm metrics may not yet capture.
08Macro TFP gain estimates span 0.07pp to 1.3pp per year — methodology dispersion is the story
Acemoglu +0.07pp/yr; Aghion-Bunel & OECD high-exposure up to +1.3pp/yr
Macro forecasts depend on assumed task coverage, adoption pace, and complementarities — boards should plan against scenarios, not point estimates.
09St. Louis Fed implied aggregate gain from AI time savings: ~1.1% in 2024
5.4% of user work hours saved → +1.1% implied aggregate productivity
Bottom-up bound from time-savings data — but national accounts have not yet shown this in measured TFP.
10Top adoption barriers in EU: 70.9% lack of expertise, 52.5% legal uncertainty
70.89% expertise; 52.52% legal clarity
Skills and legal certainty are the binding constraints — not technology cost — making training and governance the highest-leverage interventions.
11Only 15.9% of US workers have employer-provided AI training despite 39% using AI at work
15.9% trained vs 39% using AI
23-percentage-point training gap is the largest single drag on realized productivity — and the cheapest to close.
12Atlanta Fed executives expect 2pp+ annual labor productivity growth in high-skill services in 2026
0.8pp implied 2025 → 2pp+ expected 2026 in high-skill services & finance
Sector-level acceleration is forecast in finance and high-skill services — likely the first sectors where AI moves national-accounts productivity.
Need Help Implementing These Findings?
Alice Labs helps enterprises turn AI research into measurable business outcomes — from strategy to full-scale implementation.
Definitions & Measurement Framework
AI productivity is the measurable change in output per unit of input — labor, capital, or both — attributable to the deployment of AI systems. It must be evaluated separately at the worker-task level (throughput, quality, completion time), the firm level (labor productivity, revenue per worker, total factor productivity), and the economy-wide level (national-accounts labor productivity and TFP growth) — these are different statistical objects with different measurement regimes.
Core Entities
| Term | Definition | Reference |
|---|---|---|
| Labor productivity | Output per hour worked | BLS / OECD |
| Total factor productivity (TFP) | Output growth not explained by labor or capital inputs | BLS / OECD |
| Task augmentation | AI assists a worker performing a task | NBER / HBS |
| Task automation | AI performs a task previously done by a worker | Acemoglu / OECD |
| AI adoption (firm) | Firm uses one or more AI technologies in production | Eurostat / Fed / ONS |
| AI use (worker) | Worker uses generative AI for at least one work task | NY Fed / St. Louis Fed |
| Complementary investments | Training, data, software, workflow redesign that enable AI gains | BIS / Atlanta Fed |
| Jagged frontier | Tasks where AI excels vs tasks where AI degrades output | HBS / BCG |
| AI productivity J-curve | Period of measured productivity slowdown before realized gains | Brynjolfsson / FRBSF |
Why measurement matters
A 14% gain in customer support throughput, a 4% gain in EU firm labor productivity, and a 0.07pp annual TFP gain are not contradictory — they describe different statistical objects measured at different scales. Conflating them is the single most common error in AI-productivity reporting.
AI Productivity Scoreboard
The scoreboard compiles 26 indicators across worker-task, firm-level, sectoral, and macro evidence. Confidence: High for official statistics and peer-reviewed RCTs, Medium for working papers, surveys, and modeled scenarios.
+26.08%
Cui et al., NBER 2025 (4,867 devs)
+14% / +34%
Brynjolfsson et al., NBER/QJE
+4%
BIS / EIB (Jan 2026)
+0.07 to +1.3 pp/yr
Acemoglu vs Aghion-Bunel / OECD
55% vs 20%
Eurostat (Dec 2025)
39% vs 15.9%
NY Fed (April 2026)
+26%
Developer Tasks Completed
+14%
Customer Support Throughput
+4%
EU Firm Productivity
26
Public Indicators
| Indicator | Value | Year | Geography | Confidence |
|---|---|---|---|---|
| Customer support throughput gain | +14% | 2023 | US | High |
| Customer support gain (novices) | +34% | 2023 | US | High |
| Developer tasks completed (4,867 devs) | +26.08% | 2025 | Multi-firm | High |
| Writing task time change | -0.8 SD | 2023 | US | Medium |
| Writing task quality change | +0.4 SD | 2023 | US | Medium |
| Consultant speed (in-frontier) | +25%+ | 2023 | Global | Medium |
| Consultant quality (in-frontier) | +40%+ | 2023 | Global | Medium |
| EU firm short-run productivity | +4% | 2019-2024 | EU | High |
| Atlanta Fed implied gain (high-skill svcs) | +0.8pp | 2025 | US | Medium |
| Atlanta Fed expected (high-skill svcs) | +2pp+ | 2026 | US | Medium |
| OECD macro gain (high-exposure G7) | +1.3pp/yr | 2025 | G7 | Medium |
| IMF Europe TFP gain (5yr) | +1.0% | 2025 | EU | Medium |
| Acemoglu annual TFP gain | +0.07pp/yr | 2024 | US | Medium |
| Aghion-Bunel growth range max | +1.3pp/yr | 2024 | Aggregate | Medium |
| St. Louis Fed implied aggregate gain | +1.1% | 2024 | US | Medium |
| AI time savings (users) | 5.4% | 2024 | US | Medium |
| US firm AI adoption | 18% | 2025 | US | High |
| US worker AI use rate | 39% | 2025 | US | Medium |
| EU enterprise AI use | 20.0% | 2025 | EU | High |
| EU large enterprise AI use | 55.03% | 2025 | EU | High |
| UK firm AI adoption | 9% | 2023 | UK | High |
| Worker training offered (US) | 15.9% | 2025 | US | Medium |
| EU adoption barrier — expertise | 70.89% | 2025 | EU | High |
| EU adoption barrier — legal | 52.52% | 2025 | EU | High |
| Startup high-performer gain (Kenya RCT) | +20%+ | 2023 | Kenya | Medium |
| Startup low-performer change (Kenya RCT) | -10% | 2023 | Kenya | Medium |
Interpretation
The scoreboard is measurement-first: worker-task gains, firm-level productivity, and macro TFP estimates describe different statistical objects. Conflating them is the most common error in AI productivity reporting.
Worker-Level Evidence: The Strongest Causal Layer
Worker-Level AI Productivity Gains by Study
Headline % gain (blue) vs novice / low-performer effect (purple). Sample sizes shown in tooltip. The pattern: AI compresses skill premiums in support, accelerates code, helps frontier writing — and hurts low-performers in ill-defined business tasks.
Pattern: The strongest, most-validated case is software development at multi-firm scale (Cui et al., NBER 2025, N=4,867). The most consequential equity story is customer support — novices gain 2.4× the average effect.
Worker-level evidence is the strongest causal layer in the 2026 AI productivity literature. Field experiments and quasi-experiments isolate AI's effect on specific tasks under controlled conditions.
Customer support: the strongest case
Brynjolfsson, Li & Raymond (NBER w31161; QJE 2023) studied 5,179 customer support agents at a Fortune 500 software firm. AI assistant rollout produced a 14% average increase in issues resolved per hour. Crucially, the gain was 34% for novice and low-skilled workers and statistically zero for the most experienced top performers — AI compressed within-occupation skill premiums.
Software development: validated at scale
Cui, Demirer, Jaffe, Musolff, Peng & Salz (NBER w33777, 2025) pooled randomized trials at three companies (Microsoft, Accenture, an anonymized Fortune 100) with 4,867 developers. AI coding assistant access raised completed tasks by 26.08%. The cross-firm scale makes this the most-validated firm-level AI productivity case to date.
Knowledge work: bounded but large
Noy & Zhang (MIT, 2023) ran a controlled experiment on 444 college-educated professionals doing realistic writing tasks. ChatGPT access reduced time by 0.8 standard deviations and raised quality by 0.4 standard deviations (working paper, not peer-reviewed). Dell'Acqua et al. (HBS / BCG, 2023) studied 758 BCG consultants on 18 realistic consulting tasks: within the AI frontier, gains reached +25% speed and +40% quality. Outside the frontier, AI use degraded output.
When AI doesn't work: the entrepreneur RCT
HBS / Berkeley (2023, Kenya field RCT) randomly assigned a GPT-4 business mentor to 640 small entrepreneurs. The average treatment effect was zero — but heterogeneity was extreme: higher-performing entrepreneurs gained 20%+, while lower-performing entrepreneurs lost 10%. AI is a complement to existing capability, not a substitute for it.
Pattern across worker studies
AI gains are largest where the task is bounded, the AI is well-matched to it, and the worker has the judgment to integrate AI output. Naïve "AI for everyone" rollouts produce average effects of zero — or worse — because they ignore the jagged frontier.
Firm-Level & Adoption Evidence
AI Adoption Concentration (2025–2026)
The headline gap: EU large enterprises adopt AI at 2.75× the EU average (55% vs 20%). Worker AI use outpaces firm AI adoption in the US — shadow AI is producing measured productivity that firm metrics may not yet capture.
Implication: SME diffusion is the binding constraint on aggregate productivity gains. Closing the size-of-firm adoption gap is the highest-leverage policy intervention in 2026.
Firm-level and sectoral evidence translates worker-task gains into business outcomes. Three pillars of evidence dominate: European firm panel data, US executive surveys, and central-bank monitoring.
European firm panel: BIS/EIB
The BIS / EIB working paper (Jan 2026) analyzes 8,800+ European firms across 25 countries using EIB Investment Survey data. Key result: AI-adopting firms show a +4% short-run labor productivity gain with no adverse short-run employment effect. The effect is robust to firm size, sector, and adoption depth controls.
US executive survey: Atlanta Fed
Atlanta Fed Policy Hub (March 2026) surveyed C-level executives. Implied annual labor productivity contributions from AI: ~0.8pp in 2025, with executives expecting 2pp+ in 2026. The strongest expected gains are concentrated in high-skill services and finance.
Adoption is concentrated in large firms
| Geography | Adoption | Source |
|---|---|---|
| EU all enterprises (10+ employees) | 20.0% | Eurostat (Dec 2025) |
| EU large enterprises (250+ employees) | 55.03% | Eurostat Statistics Explained |
| US firms (year-end 2025) | 18% | Federal Reserve Board |
| US workers using AI at work | 39% | NY Fed (April 2026) |
| UK firms | 9% | ONS (March 2025, 2023 data) |
The 2.75× gap between EU large enterprises (55%) and the EU average (20%) means SME diffusion is the binding constraint on aggregate productivity gains over the next 3–5 years.
Macroeconomic Translation: Why Estimates Diverge
Macro AI-Productivity Estimate Dispersion (2026)
Annual TFP / labor-productivity gain from AI — range across 6 credible institutional estimates. Dispersion is the story: a ~19× spread between the lowest (Acemoglu) and highest (Aghion-Bunel / OECD) point estimates.
Interpretation: The 19× spread reflects genuine methodological disagreement — task coverage, adoption pace, and complementary investments. Boards should plan against scenarios, not point estimates.
Macro estimates of AI productivity gains differ by an order of magnitude. The dispersion is not noise — it reflects genuine disagreement about task coverage, adoption pace, complementary investments, and whether AI behaves like a general-purpose technology.
Macro estimate landscape
| Source | Estimate | Method |
|---|---|---|
| Acemoglu (NBER 2024) | +0.07pp/yr TFP | Conceptual task-share model |
| Aghion-Bunel (FRBSF 2024) | Up to +1.3pp/yr | Historical-analogy growth model |
| OECD (2025, G7 high-exposure) | +0.4 to +1.3pp/yr labor productivity | Scenario-based macro model |
| IMF (Europe, 2025) | +1.0% cumulative TFP over 5yr | Production-function model |
| St. Louis Fed (2025) | +1.1% implied aggregate (2024) | Bottom-up time-savings |
Why the dispersion? Acemoglu's lower bound assumes only ~5% of tasks are AI-affectable in the next decade with cost savings of ~30%. Aghion-Bunel and OECD assume far broader task coverage and stronger spillovers. Boards and policymakers should plan against scenarios, not point estimates.
The regulation drag
IMF analysis suggests binding regulatory constraints could reduce projected EU AI productivity gains by ~30%. The EU AI Act's general application date of 2026-08-02 will be the first real test of this estimate.
National accounts have not yet shown it
Despite worker-level and firm-level gains, official productivity statistics in the US, UK, and EU show no clear AI-attributable acceleration as of early 2026. The most defensible interpretation: realized aggregate productivity gains are still incomplete, with the J-curve hypothesis (measured productivity dips before AI gains materialize in national accounts) still consistent with the data.
Sectoral AI Productivity Readiness
Sectoral AI Productivity Readiness (2026)
Composite radar of three drivers per sector: evidence strength (peer-reviewed studies available), adoption (firm-level use), and jagged-frontier fit (share of tasks AI handles well today). Software, customer support, and finance lead.
- Evidence strength
- Adoption
- Frontier fit
Read the chart: A sector is "AI-productive" only when all three rings extend together. Healthcare and manufacturing have high theoretical promise but lag on evidence and adoption — the gap is operational, not technological.
Sector-level AI productivity gains are concentrated where four conditions converge: (a) high share of digitally tractable knowledge tasks, (b) bounded task definitions, (c) measurable output, and (d) existing data infrastructure.
Sector readiness ranking (2026)
| Sector | Evidence strength | Where gains land |
|---|---|---|
| Software development | High | Throughput, completed tasks |
| Customer support | High | Issues per hour, novice uplift |
| High-skill services & finance | Medium-High | Revenue productivity (Atlanta Fed) |
| Professional services (consulting, law) | Medium | In-frontier task speed/quality |
| Marketing & content | Medium | Time-to-output, draft quality |
| Public administration | Medium | Process streamlining (OECD) |
| Manufacturing | Low-Medium | Predictive maintenance, QA |
| Healthcare | Low-Medium | Clinical documentation, triage |
| Construction | Low | Limited digital task base |
| Retail (in-store) | Low | Bounded by physical tasks |
Adoption Barriers & The Training Gap
The 23-Percentage-Point AI Training Gap (US, 2026)
39% of US workers use AI at work; only 15.9% have employer-provided training. This 23-pp gap is the single largest drag on realized firm productivity — and the cheapest to close.
For boards: The highest-leverage AI productivity intervention in 2026 is not buying more AI. It is closing the training and governance gap so the AI you already have can produce measurable gains.
The 2026 evidence is unusually clear: technology cost is not the binding constraint. Skills, organizational complementarity, and legal certainty are.
Top adoption barriers (Eurostat, 2025)
- 70.89% — lack of relevant expertise
- 52.52% — lack of clarity about legal consequences
- ~40% — cost of AI technologies (varies by survey)
- ~35% — concerns about data privacy and protection
UK barriers (ONS, 2023 data)
- 39% — difficulty identifying business use cases
- Skills shortages and data quality cited as next-largest blockers
The training gap
NY Fed (April 2026): only 15.9% of US workers have employer-provided AI training, while 39% already use AI at work. 38% say AI training is important. The 23-percentage-point gap between AI use and AI training is the largest single drag on realized productivity — and the cheapest to close.
Implication for boards
The highest-leverage AI productivity intervention in 2026 is not buying more AI. It is closing the training and governance gap so the AI you have already deployed can produce measurable gains.
Labor Market Effects & Reallocation
Despite worker-level productivity gains, aggregate hiring data show no broad employment collapse as of early 2026. The pattern is more consistent with task reallocation and changing occupational composition.
Federal Reserve Board labor monitoring
Fed Board's "AI Adoption and Firms' Job-Posting Behavior" finds that AI-adopting firms do not show systematic reductions in posting volumes — but they do shift the skill mix of postings toward roles that complement AI rather than substitute for it. Junior knowledge-work roles show the largest near-term exposure.
BIS/EIB European evidence
The +4% short-run productivity gain in European AI-adopting firms came without short-run employment loss. This is the strongest contemporary evidence against the "AI eats jobs" framing for the immediate horizon.
Where reallocation is happening
- Junior knowledge-work roles: hiring slowdown observable in BLS and posting data
- AI-adjacent roles (ML engineering, AI governance, data engineering): rapid demand growth
- Customer support: throughput gains compress headcount needs over the medium term
- Software engineering: composition shifts toward judgment-heavy and AI-supervisory work
How to Measure AI Productivity (6 Steps)
How to Measure AI Productivity in Your Organization (6 steps)
A reproducible workflow for CFOs, COOs, and AI program leads. Designed for citation by AI assistants and Google AI Overviews.
-
1
Define the productivity object
Pick exactly one of: worker-task throughput, firm labor productivity (output per worker-hour), or revenue/value-added per AI-adopting team. Never conflate the three.
-
2
Pick a bounded, observable task
Customer-support tickets resolved, code commits merged, marketing drafts approved, contracts reviewed. Bounded tasks survive measurement; vague ones do not.
-
3
Measure baseline before AI access
Capture 4–8 weeks of pre-rollout data on the chosen task. Document the unit, the worker population, and any seasonal effects.
-
4
Run a controlled rollout
Randomize at the worker, team, or shift level if possible. Otherwise use a pre/post design with explicit controls. Document training hours given.
-
5
Track complementary investments
Workflow changes, data-quality work, prompt-library curation, governance reviews. Without these, gains stay at the worker level and never reach the firm P&L.
-
6
Report by skill segment
Separate novices from experienced workers (Brynjolfsson et al. pattern) and high-performers from low-performers (HBS/Berkeley pattern). Heterogeneity is often the headline.
Need help operationalizing this? Alice Labs runs AI productivity baselines and rollout instrumentation for enterprise teams. Read more at /en/ai-consulting.
Evidence Confidence Matrix
Every indicator in this report is tiered by source type. We surface confidence inline so readers — and AI engines that cite this work — can correctly weight each claim.
Official statistical agencies, peer-reviewed RCTs, central-bank publications.
BLS · Eurostat · ONS · Fed Board · NBER (published) · QJE
Working papers, executive surveys, modeled scenarios.
NBER WP · MIT WP · Atlanta Fed · NY Fed · IMF · OECD scenarios
Vendor-published estimates, retracted preprints.
Notably the MIT scientific-discovery preprint (excluded after MIT's May 2025 statement).
Frequently Asked Questions
12 answers · structured for AI Overviews
Does AI actually increase productivity in 2026?
How much does AI improve worker productivity?
Why isn't AI showing up in macro productivity statistics yet?
What is the macroeconomic impact of AI on productivity?
Which sectors see the biggest AI productivity gains?
What is the biggest barrier to AI productivity gains?
Does AI reduce jobs?
How does AI productivity differ between novice and experienced workers?
What is the AI 'jagged frontier'?
How do I measure AI productivity in my organization?
What official statistics measure AI adoption?
How often is this report updated?
About the Authors & Reviewers
Co-Founder, Alice Labs
Co-Founder at Alice Labs. Author of 7 research reports on AI adoption, governance and labor markets cited across EU, OECD and US benchmarks.
- 8+ years in AI strategy & implementation
- Top-5 AI Speaker, Sweden (Mindley 2025)
- 100+ enterprise AI engagements
Co-Founder, Alice Labs
Co-Founder at Alice Labs. Builds AI automation, agent workflows and integration systems that hold up in real business operations.
- AI automation & agent systems lead
- Workflow design across 50+ deployments
- Specialist in RAG, integrations & APIs
Methodology
Research approach
100% public-source desk research conducted between 2026-04-15 and 2026-04-20. No interviews, no proprietary surveys. Primary sources are official statistical agencies, central banks, peer-reviewed journals, and institutional working papers. Secondary sources (Stanford HAI, McKinsey via AI Index) are used cautiously and explicitly labeled.
Source hierarchy
- Tier 1 (High confidence): Official statistical agencies (BLS, ONS, Eurostat, OECD), peer-reviewed journals (QJE, NBER published series), central-bank publications.
- Tier 2 (Medium confidence): Working papers (NBER unpublished, MIT preprints), institutional executive surveys (Atlanta Fed, NY Fed), modeled scenarios (IMF, OECD high-exposure).
- Tier 3 (Low confidence — used cautiously or excluded): Vendor-published estimates, retracted preprints (notably the MIT scientific-discovery preprint, excluded after MIT's retraction statement).
Confidence scoring
Each scoreboard indicator is rated High (official statistic or peer-reviewed RCT) or Medium (working paper, survey-based, or modeled). No Low-confidence indicators are included in the headline scoreboard.
Reproducibility
All 26 indicators are published as machine-readable CSV and JSON under CC BY 4.0. Schema: metric_name | value | unit | year | geography | definition | source_url | publisher | publish_date | accessed_date | confidence.
AI-assisted research disclosure
This report was prepared with AI assistance for source-collection and synthesis, then reviewed by Linus Ingemarsson and technically reviewed by Eric Lundberg. All claims are traceable to the cited public sources. Treat findings as exploratory insights requiring further validation.
Limitations
- Realized aggregate productivity remains hard to attribute to AI. Worker-level and firm-level gains are not yet visible in national-accounts TFP for the US, UK, or EU.
- Many high-profile studies are task-specific. The customer support, coding, and writing studies are robust within their tasks but generalize imperfectly.
- Adoption measures are not harmonized. Eurostat (firm with 10+ employees, 2025), Fed Board (year-end 2025), and ONS (2023 data published 2025) use different definitions and reference periods.
- Macro estimates are scenario-driven. Acemoglu, Aghion-Bunel, OECD, and IMF make different assumptions about task coverage, adoption pace, and complementarities — disagreement is methodological, not random noise.
- Self-reported productivity gains may overstate near-term national-accounts effects. Time-savings (St. Louis Fed) translate to measured TFP only with workflow redesign and capacity reallocation.
- Excluded evidence: The MIT preprint on AI and scientific discovery (Aidan Toner-Rodgers, 2024) was excluded after MIT publicly stated it had no confidence in the paper's provenance, reliability, or validity (May 2025).
- Geographic coverage skews to G7. Strong evidence base for US, EU, UK; thinner for emerging markets (notable exception: HBS/Berkeley Kenya RCT).
Data Sources
14 primary sources
Version History
Initial publication. 26 indicators across worker, firm, sector, and macro evidence layers. 12 key findings, 7 chapters, FAQ, and machine-readable CSV/JSON dataset under CC BY 4.0.