Anthropic-affiliated researchers have analyzed millions of real Claude conversations to map exactly which economic tasks the chatbot performs and for which occupations, producing the first large-scale, task-level dataset published for any of the three leading AI assistants. The findings, paired with a separate Anthropic Economic Index preprint documenting uneven geographic and enterprise adoption, give Claude a data-backed claim to a seat alongside ChatGPT and Gemini at the top of the AI assistant market. For companies deciding which tool to standardize on, the research offers the earliest direct evidence of how one major assistant is actually used in daily work, and it raises a pointed question: where is the equivalent data for the other two?
Why the three-way race matters for enterprise decisions right now
Businesses across industries are locking in AI assistants for internal workflows, from drafting code to summarizing legal filings. That consolidation is happening fast, and it favors whichever platforms can demonstrate real traction in high-value professional tasks. The Anthropic-linked research paper titled “Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations,” published on arXiv, does exactly that for Claude. It maps conversation data against occupational categories, showing where the assistant concentrates its usage. Neither OpenAI nor Google has released a comparable primary dataset of millions of conversations coded by task and occupation for ChatGPT or Gemini.
That asymmetry creates a testable hypothesis. Claude’s documented concentration in analytical and coding occupations suggests it could retain enterprise users at higher rates than Gemini once all three platforms publish equivalent primary logs. Without matching transparency from OpenAI and Google, procurement teams are left comparing marketing claims against one system’s actual usage record. The gap between verified task-level data and general market positioning is where the real competitive tension sits.
A second preprint, the Anthropic Economic Index report on uneven geographic and enterprise AI adoption, sharpens the picture. According to Anthropic authors on arXiv, enterprise uptake clusters in tech-heavy regions, and adoption rates vary sharply across industries. That geographic concentration means the race is not uniform. Companies in software-dense metros are choosing tools faster, and the assistants that show up strongest in those early-adopting corridors gain a compounding advantage as contracts renew and teams build workflows around a single platform.
What millions of Claude conversations reveal about real AI work
The core contribution of the first paper is granularity. Rather than relying on surveys or self-reported usage, the researchers analyzed millions of actual Claude conversations through a task and occupation lens. The methodology links each conversation to specific economic activities, producing a map of which jobs and which tasks within those jobs are being augmented or performed by the assistant. Software engineering, data analysis, and legal research appear as areas of heavy use, consistent with Claude’s reputation as a tool favored by developers and knowledge workers.
The task taxonomy used in the study allows for distinctions between high-level ideation, routine drafting, debugging, data cleaning, and more specialized analytical work. By tying these activities to occupational categories, the authors show not only that AI is present in knowledge work but which slices of that work are most likely to be delegated. Early evidence suggests that repetitive documentation, boilerplate code, and first-pass research are especially likely to be handled by the assistant, while final decisions and domain-specific judgment remain with human workers.
The Anthropic Economic Index preprint adds an institutional layer. Its sampling approach and privacy-preserving methods are described in enough detail for independent scrutiny, a level of methodological openness that neither ChatGPT nor Gemini has matched with comparable published research. The report’s finding of uneven adoption, both by geography and by enterprise type, aligns with what procurement consultants have observed anecdotally but had not previously been able to confirm with primary conversation data.
Together, the two papers create a feedback loop for Anthropic’s positioning. The task-level data shows where Claude is strong. The adoption data shows where it is growing. For enterprise buyers evaluating the top three assistants, this combination is more actionable than benchmark scores on standardized tests, which all three systems perform well on but which tell little about how a tool behaves inside a real organization’s daily operations.
The practical consequence for readers making purchasing or workflow decisions is direct. If a team’s work falls within the occupational categories where Claude shows dense usage, the published evidence supports a shorter evaluation cycle. Teams can move more quickly from pilot projects to full deployment, confident that similar organizations are already using the tool for analogous tasks. If the work sits outside those categories, the absence of equivalent data from ChatGPT and Gemini makes comparison harder, not easier. Buyers in that position are operating on brand reputation and informal trials rather than published task-level records.
Missing data from OpenAI and Google leaves the ranking incomplete
The strongest unresolved question in the three-way race is whether ChatGPT and Gemini usage patterns would tell the same story, a different one, or a more favorable one for those platforms. OpenAI has published model capability benchmarks and safety evaluations, and Google has released technical reports on Gemini’s architecture. But neither company has opened millions of real user conversations to academic-style task and occupation coding. Until they do, the claim that these three assistants form a clear top tier rests on market share estimates, app download figures, and enterprise contract announcements rather than on matched primary evidence about what users actually do with each tool.
The two Anthropic-affiliated preprints also carry a built-in limitation. They describe Claude’s usage, not the broader market. A system that attracts a particular user base, say, developers and analysts, will naturally show concentration in those occupations. Whether that concentration reflects a genuine product advantage or simply early-adopter demographics is a question the current data cannot answer on its own. Comparable logs from ChatGPT, which has a large general consumer base, and Gemini, which benefits from deep integration with Google’s productivity tools, would be needed to interpret Claude’s profile in relative rather than absolute terms.
There is also the issue of selection bias on the enterprise side. Organizations that opt into Anthropic’s tools may differ systematically from those that standardize on OpenAI or Google, in risk tolerance, regulatory exposure, or industry mix. If, for instance, more heavily regulated firms gravitate toward a particular vendor, their usage patterns will skew toward compliance, documentation, and audit support tasks. Without parallel datasets from the other providers, it is impossible to know whether Claude’s observed task distribution is unique or simply one instance of a broader pattern common to all advanced assistants.
For now, this leaves enterprise decision-makers with an incomplete leaderboard. Claude can point to detailed evidence about how it is used, where, and by whom. ChatGPT and Gemini can point to scale, integration, and brand familiarity, but not to similarly granular public data. The resulting asymmetry may push some buyers toward the one system they can study in depth, even if they ultimately deploy multiple assistants side by side.
How enterprises can act amid asymmetric transparency
In the absence of matched datasets from all three major assistants, enterprises face a practical challenge: how to make durable platform choices under uncertainty. One emerging strategy is to treat the Claude findings as a baseline, assuming that any assistant adopted at scale will concentrate first in tasks that are repetitive, text-heavy, and relatively low risk. Procurement teams can then run targeted pilots with alternative tools in the specific workflows that matter most to them, using Claude’s documented task mix as a reference point rather than a definitive map.
Another approach is to build internal telemetry that mirrors the Anthropic studies. By instrumenting prompts, outputs, and occupational metadata within their own organizations, companies can develop private versions of the task and adoption maps described in the preprints. That data, even if never shared publicly, can inform renegotiations with vendors, guide training and change-management efforts, and highlight gaps where additional automation tools or process redesigns might be needed.
Ultimately, the Anthropic-affiliated research raises the bar for what counts as credible evidence in the AI assistant market. It demonstrates that large-scale, privacy-preserving analysis of real work conversations is both technically and ethically feasible. Until OpenAI and Google release comparable studies, the three-way race will remain partly speculative, with one competitor competing on documented usage and the others competing on promises. For enterprises deciding which assistant to bet on, that difference in transparency is no longer a minor detail; it is part of the product.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
