On January 15, 2026, Anthropic published its fourth Economic Index report. The headline change is the introduction of “economic primitives” as a classification scheme, along with a first set of numbers: 56% of Claude conversations are classified as augmentation, 41% as automation, with the remainder mixed.
What Economic Primitives Are
The classification places each conversation on a simple axis: is the AI helping a person do the work better (augmentation), or is it completing a task a person would otherwise execute (automation)? Compared to the vague question of whether “AI will replace jobs,” this binary breaks the issue into measurable units. The 56/41 split is the first large-scale reading.
Primitives also make the measurement repeatable. Once each conversation maps to a small set of categories, successive reports can track movement in the mix rather than re-litigating definitions each quarter — which is exactly what a fourth-edition series needs to stay useful.
How to Read the Numbers
- Augmentation still leads: the mainstream use of Claude today is helping people accomplish tasks; “the AI does the whole job” is the smaller share
- But 41% is not low: automation is a genuine second curve, not an edge case — direct counter-evidence to the assumption that AI is mostly an assistive tool
- Roughly 3% mixed: the classification’s own gray zone, a reminder that augmentation and automation often sit on a continuum rather than a toggle
The split also reframes the displacement debate. A conversation being automated does not tell you whether the person moved on to higher-value work or simply did less — the data measures what the AI did, not what the human did next. That distinction is where job impact actually lives, and it is invisible to this method.
What the Methodology Can and Cannot Do
The report’s value and its limits both come from the data source. The value: it measures real conversations rather than surveys or speculation, and as the fourth edition it is starting to accumulate a comparable time series. The limits are equally clear:
- The sample is Claude’s user base, which is not a portrait of economy-wide AI adoption
- Classification relies on interpreting task content, so borderline cases inevitably involve judgment
- Conversation data cannot capture the organizational decisions and process changes behind adoption
Treat it as a reliable directional thermometer; treat it as a precise labor-market census and you will misuse it.
The Lesson for Enterprise Adoption
For organizations deploying AI, these numbers suggest a useful self-examination: where does our internal usage fall on this split? Organizations skewed toward augmentation should invest in training and workflow design — making people better at using AI. Processes skewed toward automation need acceptance criteria and exception handling, because when something goes wrong, no human sits in the loop. The two distributions imply completely different management approaches, and most organizations likely contain both without ever having measured them. Public benchmarks like this one make the measurement normal; the organizations that internalize it early will manage the transition with fewer illusions.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.