AI at work: what 105,000 Copilot chats reveal

The chart shows the ten largest work activities in the source data. Decisions and problem solving lead at 27.5%, followed by finding information at 13% and recording information at 11.7%. The ten displayed percentages sum to 86.1% of classified goals. They use the original denominator, rather than being rescaled to sum to 100%.

Citation's reading of this distribution is that an AI workflow deserves attention before the final document appears. Consider a research brief: choosing a question, comparing explanations and deciding which evidence matters are all part of the work. Counting only the finished pages would leave those decisions out of view.

A useful way to apply the chart is to separate a writing assignment into decisions and deliverables. Before asking for a draft, define the reader, the question the piece needs to answer and the evidence needed to support that answer. These are editorial choices, and each one gives the eventual draft a clearer purpose.

For example, an instruction to write about remote work leaves several decisions implicit. An instruction to compare two explanations for a change in employee retention makes the research task more precise. The writer can then ask which sources address the same population and period, where their definitions differ and which conclusion the evidence can reasonably support.

This is an interpretation of how the categories could inform a workflow. The chart itself does not measure whether a particular prompting approach produces better writing.

A search result can help a writer discover a report, but the original report is where the meaning of a statistic needs to be checked. A percentage needs a denominator. A change needs a starting point and a period. A comparison needs categories that actually measure the same thing.

Before reusing any figure, keep a small evidence note alongside it: what was measured, who or what was included, when it was observed and where the original result can be found. Keeping these details close to the draft makes it easier to notice when an attractive headline goes beyond its evidence.

Our suggested workflow has four steps. It is a practical application of the categories, rather than a procedure tested by this dataset.

The point is to make the work inspectable. A colleague reviewing the draft should be able to follow a claim back to the evidence and understand how the conclusion was reached. That is useful whether AI helps with one step or several.

Microsoft's published methodology describes 105,000 conversation samples from one week in February 2026, covering North American commercial segments and excluding education. Conversations with multiple activities contribute fractional counts. We rank the 31 activities in the downloadable data by their underlying shares and display the ten largest, preserving the published rounded percentages. The table has no total row because the chart does not include every activity.

Treat this as a snapshot of the measured Copilot activity. It cannot establish a trend over time or represent every AI tool, industry or country. A larger bar describes a larger share of classified goals, not a larger productivity benefit.

Explore the original Microsoft report and its downloadable chart data for the category definitions and methodology.