Every exec I've met this quarter has shown me a chart that says their knowledge workers are faster since the LLM landed on their desk. Nobody has walked me through what, specifically, the AI did on its own to move the chart.
The polite question is what the AI contributed. The useful one is what the AI did that nobody asked for.
A head of product in Amsterdam, in July
A head of product at a Series B B2B SaaS firm walked me through her quarterly adoption review. Her senior PM's drafting time on PRDs was way down since the team had moved to Claude. The dashboard showed exactly what her CEO had asked for.
I asked her to open the last PRD her senior PM had shipped and read the chat with me. Not the output. The transcript. Step by step. Who prompted each move.
She did. Her PM opened with a goal. The AI picked which competitors to benchmark. The AI chose the rollout tier. The AI picked the success metric. Her PM accepted all three and wrote the summary.
She stared at the screen for a while. Her senior PM had not picked the metric and she had not noticed.
What KAIST and CMU just measured
Yoonsu Kim and colleagues at KAIST and Carnegie Mellon published What Did the AI Take On? on October 5. They gave 24 frequent LLM users three knowledge-work tasks: pitch-deck ideation, an electricity-generation explainer, a cover-letter revision. They collected 992 annotated reasoning steps for who initiated each one.
About 48.6% of the steps were AI-initiated. Nearly half of what the model contributed was work the user had not requested. Of that half, Serendipitous Delegation accounted for 30.4%: the user liked it in retrospect. Unintended Delegation took another 18.1%: the user did not want it. Regretted Delegation took 7.8%: the user wished they had done it themselves.
The direction gets sharper when you slice by outcome.
The AI initiated 70.0% of the steps users marked as unwanted, versus 41.1% of wanted steps. For minor steps the split was 75.8% AI-initiated to 44.4% for critical steps.
The AI is running most of the trivial work, which is probably fine. It is also running most of the work the user would not have chosen. And for 63.9% of steps the users wanted the AI to proceed without checking in. The appetite for delegation is real. The user's picture of what they already delegated is wrong.
Your dashboard counts outputs. Who initiated the step inside the output is a different quantity, and your people cannot report it until you walk them back through their own chat. Faster is the easy answer. Who took the step is the one you still owe the room.