Working with Claude: collaboration and delegation

Anthropic's January 2026 Economic Index reports that 52% of sampled Claude.ai conversations were classified as augmentation and 45% as automation. The remaining 3% fit neither category. The sample covers 1 million conversations from November 13 to 20, 2025.

Augmentation includes learning, task iteration and validation. Automation includes direct task delegation and feedback loops. These are classifications of interaction patterns. They describe how an exchange proceeds, rather than whether the result is correct or whether a person's job has been replaced.

A research project contains decisions with different levels of ambiguity. Renaming a set of section headings can have a precise target. Choosing which explanation best fits conflicting evidence needs a judgment about the sources and the question. Treating both as an instruction to finish the task hides that difference.

Citation's reading is that the chart offers a useful vocabulary for planning involvement. At the beginning of a task, identify what is already settled and what is still open. A clear format, known source material and explicit acceptance criteria can make a bounded assignment easier to review. An unresolved argument may benefit from several rounds of questions and revision.

For example, ask for alternative outlines when the structure is uncertain, then explain which tradeoffs matter. Once an outline is chosen, a narrower instruction can help turn checked notes into a section. The final review still belongs to the person responsible for the work, regardless of how many turns the conversation took.

A short exchange can produce an accurate result or miss a crucial requirement. A long exchange can improve a draft or keep repeating the same misunderstanding. Conversation shape alone cannot decide between those outcomes.

A useful delegation brief therefore includes a completion check. For a table, name the required columns and explain how missing values should appear. For a summary, identify the source text and the questions it must answer. For a citation, check the underlying publication details rather than accepting an entry because its formatting looks familiar.

For a literature review, start by keeping the research question and source list in view. Ask for a proposed comparison, inspect whether the papers address comparable populations and revise the structure where they do not. This creates a visible record of why the argument changed.

When the remaining work becomes mechanical, narrow the request. A checked set of notes can support a first draft, but the reviewer should still follow each important claim to its source. Record substantive changes to the argument so that a collaborator can distinguish evidence decisions from wording preferences.

This workflow is an editorial suggestion. The reported percentages do not test it, establish a preferred number of exchanges or tell a team which responsibilities to delegate. They provide context for asking those questions more precisely.

The sample includes Claude.ai Free, Pro and Max conversations. The report also examines API traffic separately; those observations are excluded here. Anthropic uses model classification, so the categories should be understood as measured patterns rather than exact accounts of every user's intention.

Explore Anthropic's overview and its full Economic Index report for the collaboration definitions and sampling approach.