The mechanically interesting part is which tool did the work. Anthropic Insights, previously named Clio, is not an instrument built for outsiders; it is the privacy-preserving system Anthropic's own teams use to analyse patterns across millions of Claude conversations [2]. The pilot therefore did not establish that independent measurement of AI usage is possible in principle. It established that it is a query away on the vendor's own analysis stack, for whoever the vendor admits [1].
Then the arithmetic. Each team worked with roughly 250,000 Claude.ai or Claude Code conversations from April and May 2026 [4]. The Stanford SALT Lab reports that over half of conversations involved a person delegating a consequential task, defined as work that affects others or is hard to undo, which cuts against prior research suggesting people keep that sort of work for themselves [6]. Applied to the sample, that is upwards of 125,000 conversations in two months [7]. Separately, in nearly three-quarters of conversations the human set direction while Claude assisted, and usually adapted the output rather than using it verbatim [8]. The complement is on the order of 62,000 conversations where that description did not hold [9].
What the published results do not let anyone do is cross those two findings. The cell a board actually wants is consequential work delegated without the human directing it, and the aggregate release does not produce that cell. The next honest step for anyone in legal or finance is to note that professional guidance was where consequential delegation peaked, legal and financial questions in particular [c6b], and that people vary in how much they understand what Claude produced even when they were the ones steering [c8b].
Anthropic's stated reason for opening the door is that real-world interaction data sits inside a handful of labs, leaving outsiders to choose between the labs' own analyses of real usage and their own analyses of public datasets that skew casual [10]. The pilot buys a third option on the lab's terms. Those terms: review rights confined to user privacy, information that would help people violate usage policies, Anthropic's confidential information, and research accuracy [11], with researchers otherwise free to publish findings that are inconvenient for Anthropic [12]. Three of those four categories are narrow. "Research accuracy" is the one with room in it, and it is the category a lab would reach for if a finding were both awkward and contestable. Nothing in the account says it was used that way. The point is that it exists and is where pressure would land.
For operators the governance consequence is simple to state and awkward to answer. Anthropic says this is the first time external researchers have run public independent studies on an AI company's own usage data [5], the aggregate data from each project is being released [13], and Anthropic has run a further privacy audit of what was shared and is inviting expressions of interest for future work [14]. Category-level facts about how employees delegate consequential work are now produced by parties whose access is granted by the model vendor, not by the employer. The material covers Claude.ai and Claude Code conversations [4]; it does not say what happens to enterprise-tier traffic, which is the question worth putting to a vendor in writing.
