Microsoft Research described the task of evaluating Microsoft Copilot at the scale of large organizations. For such checks, test data reflecting the complexity of real companies are needed.
Using real corporate data for this purpose is not possible: they contain sensitive information. In the study's title, synthetic data are contrasted with real data as a path to a more substantial assessment of confidence in Copilot's performance.
The available description provides no details about a specific methodology for creating synthetic sets, test results, or comparisons with other systems. The source is presented as a single material from Microsoft Research, and its evidentiary basis is limited to metadata and a brief synopsis.
editorial commentary
Why it matters
A likely practical implication is to continue searching for ways to test corporate AI without access to sensitive datasets. The nearest observable signal will be the published details of the methodology and results of comparative evaluations. Substantial uncertainty remains: the provided synopsis does not show how well synthetic data reflect real organizational operation.