Anthropic published two new pieces of research this quarter that together paint a more textured picture of AI productivity than the headline numbers typically convey. The first estimates productivity gains across occupations using its Economic Index — a framework that tracks what tasks AI is actually being used for and how much time is being saved. The second introduces new "primitives" for understanding AI use patterns: not just whether workers use AI, but how, at what frequency, and on which task types. Together they represent an attempt to get past the 14% and 50% figures that dominate coverage and into the granular mechanics of how AI changes work.
The 14% figure comes from a 2023 MIT/Stanford study of customer service workers — still widely cited, still the cleanest controlled experiment we have. The 50% figure is from Anthropic's own internal survey, in which employees self-reported using Claude for 60% of their work and estimated a productivity boost of roughly half. Both numbers are real, but they measure different things in different contexts with different methodologies. The MIT/Stanford study measured actual output — resolved tickets per hour — against a control group. The Anthropic internal data is self-reported and measures perceived productivity in a company whose employees have both the skill and the incentive to use AI effectively.
What Anthropic's Economic Index research adds is occupational granularity. Software developers account for the largest share of total productivity gain attributable to AI — about 19% of the aggregate benefit — followed by general managers, market research analysts, customer service representatives, and secondary school teachers. The distribution is skewed toward knowledge workers who produce text and code, and away from roles that involve physical presence, real-time judgment, or interpersonal interaction in ways that AI cannot yet replicate. That pattern is consistent across multiple studies and is worth taking seriously as a map of where the productivity argument is strongest.
The coding skills research Anthropic released earlier this year adds a complication that the productivity enthusiasm tends to skip over: AI assistance boosts output but may impede skill formation. Developers who use AI coding tools complete more tasks, faster — but the research found evidence that heavy reliance on AI for routine coding tasks reduced the rate at which those developers improved their own abilities over time. If true at scale, this implies a productivity gain in the short term and a skill-atrophy problem in the medium term, particularly for junior developers whose career trajectory depends on accumulating expertise through practice.
The new "primitives" framework is an attempt to create a more stable vocabulary for this kind of research. Rather than asking "is AI productive" as a binary, Anthropic is trying to disaggregate the question: which specific task types benefit most, at what frequency does AI use need to occur to produce gains, and where does AI assistance reduce quality rather than improve it? These are harder questions to answer than the headline number, but they are the questions that actually matter for anyone trying to manage the transition — a team leader deciding how to structure work, a company deciding what roles to hire for, a policymaker deciding what retraining looks like.
The honest summary of what we know in early 2026 is that AI productivity gains are real, meaningful, and unevenly distributed. They are largest for knowledge workers producing text and code, particularly for lower-skilled workers doing routine versions of those tasks. They come with a set of second-order effects — skill formation, over-reliance, quality in tail cases — that are much less well understood. The macro projections — 1.2 to 1.8 percentage points added to annual US productivity growth — are plausible and significant if they materialise, but they rest on assumptions about adoption rates and task suitability that are difficult to verify in advance. The research is accumulating. The confidence is not yet warranted for the strongest claims being made about it.