Anthropic made AI-work data easier to query, not more representative

On July 22, 2026, Anthropic launched the Anthropic Economic Index connector in claude.ai. Anthropic said the connector was available that day from the claude.ai connectors directory, worked in any conversation with any Claude model, and required no installation. A user can ask a question in plain language, narrow it through follow-up questions, and ask Claude to show the underlying Index data.

The easier interface does not change the population behind the data. The Economic Index describes patterns in Claude usage, so an answer about tasks, occupations, industries, or geography can help a business leader form a hypothesis. Before that hypothesis informs a test, the leader still needs to identify the population, period, product surface, and unit behind the answer, then compare the public pattern with evidence from the workflow under consideration. This article shows how to do that without turning Claude-usage patterns into workforce claims.

The connector changes access while the Index keeps the same scope

The connector adds a conversational way to explore datasets that Anthropic continues to provide free for direct inspection. Anthropic recommends starting with a broad question, drilling into specifics, and asking Claude to expose the data behind its answer.

The Economic Index, last updated June 26, 2026, covers usage indexes, augmentation versus automation, occupations and industries, use cases, geography, and filterable preview data. A leader can move from a general question about a field to the tasks and classifications that produced the answer without leaving the conversation. That convenience can also hide definitions: a fluent answer may combine a familiar occupation name, a percentage, and a trend into something that sounds more general than the source supports.

The Index describes what people do with Claude

On the Index’s occupation and task view, percentages represent the share of Claude.ai conversations associated with a task or group of tasks. Anthropic maps those tasks to the O*NET occupational classification system, but the chart does not identify the job title of the person who started each conversation. A conversation mapped to an accounting task, for example, is evidence that the task appeared in Claude usage. It is not evidence that a known accountant used Claude or that a given share of accountants use it.

Privacy protections also affect what appears. Anthropic uses privacy-preserving classifiers and suppresses cells with too few observations. On the occupation and task view, percentages may not add to 100 percent because privacy filters remove some data. Missing or incomplete totals therefore need to stay visible in any interpretation.

Geographic usage indexes have a different denominator. They compare Claude use in a place with its working-age population. An index above or below one describes relative Claude use against that population baseline. It is not a labor-market adoption rate, a share of workers using Claude, or a measure of productivity.

Some Index analyses also separate chat and Cowork, Claude Code, and Anthropic’s first-party API. Those surfaces support different interaction patterns, so a result from one cannot silently stand in for all Claude use. Our prior analysis of what the Economic Index can measure examines observed task exposure in more detail. For connector users, the immediate discipline is simpler: keep the surface and measurement unit attached to every result.

Broad questions can hide the population, period, surface, and unit

A question such as “Which occupations use AI the most?” contains several unresolved definitions. In an Index answer, “AI” may mean activity observed on a particular Claude surface. “Occupations” may refer to tasks mapped to O*NET occupation groups rather than verified user job titles. “Most” might mean the largest share of observed conversations, a geographic usage index, or another metric selected by the query.

Time matters for the same reason. The Index page and its connected datasets have update dates and sample periods. A current answer should state which release and observation window it uses, especially if the question asks whether behavior changed. Comparing two periods is useful only when their methods, surfaces, and units are comparable.

Before acting on an answer, ask Claude to restate it with the following details visible:

• the observed population, such as Claude.ai conversations or sampled Claude users;
• the data period and Index release;
• the product surface or surfaces included;
• the unit being counted or classified;
• the denominator used for any percentage or index;
• any privacy suppression, missing cells, or method changes that limit comparison.

These details are part of the finding. If Claude cannot identify one of them from the available source, record it as unknown instead of filling the gap with an assumption.

One useful question starts with a local decision

An operations leader considering a narrow finance workflow could begin with this question:

Example

In the latest available Economic Index data, what does Claude usage show about tasks associated with bookkeeping, accounting, and auditing clerks, especially the balance between augmentation and automation? State the population, observation period, Claude surface, unit, denominator, and privacy limitations before interpreting the result.

This question names a work area and asks for a distinction that could affect test design. Augmentation means a person and Claude collaborate on the task; automation means the user delegates the task more fully under Anthropic’s classification. Neither label establishes that a business can run the task accurately, safely, or economically in its own systems.

The first answer should narrow the next question. If it identifies a task relevant to the business, ask Claude to show the rows behind that finding, define each column, explain the filters and classification hierarchy, identify suppressed or missing values, and link to the source table or dataset. Then check whether the prose summary matches the displayed unit and denominator. A percentage of Claude.ai conversations should remain a percentage of Claude.ai conversations when it reaches the decision memo.

The underlying rows matter because adjacent measures can answer different questions. A task’s share of conversations, its augmentation-versus-automation classification, and a geographic usage index do not describe the same population or outcome. Seeing the source fields makes those differences harder to blur.

A public usage pattern becomes useful when it meets local workflow evidence

When the source rows show a meaningful pattern around a task your team performs, the result can justify examining the task locally. It cannot tell you how often your team performs it, how much review it requires, which exceptions dominate, or whether Claude can work with your source systems and policies.

The local comparison should use evidence the team already owns or can collect. Start with the workflow’s volume, elapsed time, error and rework patterns, review burden, source quality, system handoffs, and consequence of a bad result. Check a representative sample of ordinary cases and exceptions. These observations establish whether the public pattern corresponds to a costly or repeatable problem in your operation.

If the task and local evidence align, define a small test that keeps consequential actions under review. Measure the workflow before and during the test with the same units. Record which cases passed, needed edits, failed, or stayed manual, along with the time people spent checking the output. As our guide to what a first AI pilot should leave behind explains, the point of a pilot is to produce enough evidence for the next decision.

A positive result may support a wider test. A high review burden, weak source data, or frequent exceptions may support revising the workflow or leaving it manual. Both outcomes are more useful than treating a public usage pattern as a forecast about local staffing.

The connector is decision-grade at the hypothesis stage

The connector can support question formation, source discovery, and a decision to investigate one task. It can help a leader compare task categories, inspect whether usage appears more augmentative or automated, and find the data behind an answer. Those are bounded uses because the source directly contains the relevant classifications.

It cannot establish how Claude is changing employment, productivity, hiring demand, or wages across the labor market. Anthropic says this directly in the announcement: the Index reflects Claude usage rather than the labor market as a whole. The June 2026 Economic Index report adds that usage data carries limited information and that economic impact will ultimately appear in aggregates such as employment and productivity, not usage logs alone.

Survey findings require their own boundary. Anthropic’s Economic Index survey samples Claude users, is not representative of the general population, may be affected by who chooses to respond, and excludes infrequent users from the linked analysis. If a connector answer draws on survey data, the answer should say so and keep that respondent population separate from conversation-level usage data.

A connector result is therefore insufficient on its own for a hiring freeze, headcount reduction, staffing model, automation budget, or expected-return claim. Those decisions require local workflow evidence and, for labor-market conclusions, representative economic measures beyond Anthropic’s product data. A precise query does not widen the source population.

Use the Economic Index connector to form a hypothesis and refine the question until its evidence and limits are visible. Use local workflow evidence to decide whether to run or expand a small test. Hiring, staffing, automation, and investment decisions remain with the team that owns the work, the consequences, and the evidence.