Anthropic estimates current AI models could add 1.8% annual US labor productivity growth

TL;DR

Anthropic’s study of 100,000 real Claude.ai conversations shows that AI reduces task completion time by roughly 80%, which, when extrapolated to the whole US economy, could raise labor productivity growth by about 1.8% per year over the next ten years – roughly double the recent growth rate.

Methodology: Privacy‑preserving transcript analysis

Anthropic used its CLIO privacy‑preserving analysis system to sample 100 k anonymized Claude.ai conversations across Free, Pro, and Max tiers. For each conversation it generated two estimates with Claude:

Time without AI – the number of hours a competent professional would need to complete the task unaided.
Time with AI – the total minutes the user spent interacting with Claude (including reading, thinking, typing, and implementing suggestions).

Claude produced these estimates via structured prompts (see Appendix). The resulting task‑level times were then mapped to O*NET occupations and matched with May 2024 Occupational Employment and Wage Statistics (OEWS) data to compute implied labor costs.

Validation of Claude’s time estimates

Self‑consistency

Across 1,800 consented conversations, Claude’s estimates were highly stable under prompt variations, with log‑scale correlations of r = 0.89–0.93.

External benchmark on software‑engineering tickets

A separate benchmark compared Claude Sonnet 4.5’s estimates (title + description only) to developers’ own estimates and actual tracked times for 1,000 JIRA tickets:

Source	Spearman ρ	Log‑scale Pearson r
Developers (self‑estimate)	0.50	0.67
Claude Sonnet 4.5 (no examples)	0.44	0.46
Claude Sonnet 4.5 (10‑example prompt)	0.39	0.48

Claude’s directional correlation is only slightly weaker than developers’, though it tends to compress the distribution (over‑estimating short tasks, under‑estimating long ones).

Task‑level productivity findings

Average time savings – Across the sample, Claude reduces task duration by ≈ 80% (median 84%).
Task cost – The median conversation corresponds to a professional labor cost of $54 (based on OEWS wages).
Occupational variation – Management and legal tasks have the longest estimated human times (~~2 h) and highest implied costs ($119‑$133). Food‑preparation, installation/maintenance, and transportation tasks are shorter (~~0.3‑0.5 h) and cheaper (~$8).
Extreme examples – Curriculum‑development tasks estimated at 4.5 h were completed in 11 min (≈ 96% saving, $115 implied cost). Invoice‑writing tasks saved ~87% of time.

From task‑level gains to economy‑wide impact

Anthropic applied Hulten’s theorem, weighting each task’s productivity gain by:

Claude‑estimated fraction of occupational time spent on that task, and
The occupation’s share of the total US wage bill (OEWS 2024).

Assuming universal adoption of current‑generation AI for the observed tasks, the model yields a 1.8% annualized increase in US labor productivity. With a labor share of 0.6, this translates to an ≈ 1.1% annual rise in total factor productivity (TFP) – a level not seen since the early 2000s.

Occupational contributors

The top five occupations driving the projected gain are:

Software developers (≈ 19% of the total gain)
General & operations managers (~6%)
Market‑research analysts & marketing specialists (~5%)
Customer‑service representatives (~4%)
Secondary‑school teachers (~3%)

Sectors such as restaurants, healthcare delivery, construction, and retail contribute little, largely because few of their tasks appear in the Claude sample.

Potential bottlenecks and uneven gains

While many tasks show 80‑95% time reductions, others exhibit modest or negligible savings (e.g., diagnostic‑image checks ~20%). Tasks that receive little AI acceleration could become relative bottlenecks, reshaping how workers allocate time within occupations.

Limitations of the analysis

Estimate accuracy – Claude’s predictions are imperfect and lack real‑world validation beyond the software‑ticket benchmark.
Task taxonomy – Mapping to O*NET abstracts away tacit knowledge, inter‑task dependencies, and non‑task activities (e.g., post‑chat validation).
Adoption assumptions – The 1.8% figure assumes universal, immediate adoption of current models, ignoring diffusion lags and heterogeneous uptake.
Scope of data – The dataset only reflects Claude.ai usage, which may be biased toward tasks users deem AI‑friendly.
Structural effects – The model does not capture organizational restructuring, capital investment, or broader innovation that historically amplify productivity gains.

Implications and future work

Upper‑bound estimate – The 1.8% annual gain should be viewed as an upper bound for current‑model effects, not a forecast. Faster model improvements could raise the bound, while slower adoption could lower it.
Economic Index tracking – Anthropic will continue to compute these estimates as part of its Economic Index, enabling longitudinal observation of how task‑level savings evolve with model capability and adoption.
Policy relevance – Even a modest uplift in productivity could influence wage growth, inflation dynamics, and labor‑market policy, underscoring the need for timely measurement.
Research agenda – Future work should (1) validate model‑based time estimates against larger real‑world datasets, (2) incorporate post‑chat validation effort, and (3) model organizational restructuring effects.

“These productivity gains come from making existing tasks faster to complete. Historically, transformative productivity improvements came from fundamentally reorganizing production, not merely speeding up old tasks.” – Anthropic research note

Appendix: Prompt snippets used for time estimation

(Full prompts are reproduced in the source document; key excerpts are shown here for reproducibility.)

Human‑time estimation prompt – asks Claude to reason about required hours, outputting <answer> tags.
Interaction‑time estimation prompt – asks Claude to estimate user minutes spent, considering reading, typing, and implementation.
Software‑development benchmark prompt – provides task title/description and requests a raw hour number.
Task‑allocation prompt – requests a JSON mapping of O*NET task IDs to weekly hours per occupation.