Five weeks rewrote how workers handed off work to their AI teammate

On 2 November 2026, researchers published the result of a five-week study in which workers gradually rewrote how they passed work to an AI teammate. They did not start trusting the agent more because the agent performed better in week one. They started trusting it differently because they had watched it fail in recoverable ways, corrected its memory of their preferences, and learned which task shapes it handled without supervision. By week five, their handoffs looked nothing like their handoffs in week one, and a trust score taken at week one would have measured a relationship that no longer existed.

That finding sits directly beside a meta-analysis of 420 human-AI collaboration papers, also published on 2 November. Across those studies, human-AI pairs beat the average solo human performer but fell short of the best solo human performer. The gap is not a curiosity. A product team that routes a task to an agent is accepting a floor somewhere between average and best, and the routing decision carries the most weight for work where the best solo human would have produced something the average solo human would not. That is the decision your product is making every time a worker delegates without enough history to know where the agent sits in that range.

The MIT ask-or-act model, published the same day, adds a third dimension. An agent deciding whether to ask a clarifying question before acting has to weigh more than the cost of interrupting one worker on one task. The model shows that the number of agents running in parallel shapes whether a question is cheap or expensive to ask across the whole system. A worker with one background agent finds an interruption tolerable. A worker whose attention is already split across several agents running simultaneously finds the same question far more costly, even if the question itself is identical.

Taken together, the three findings describe a single interaction problem: the handoff is not a moment, it is a relationship with a timeline, and the timeline changes what is appropriate to measure, route, and ask at any given point. A trust score at week one is a measure of novelty. A routing decision made without knowing the agent's track record on this task shape is a guess. An interruption that felt cheap in isolation becomes expensive when a worker is already managing three agents in the background.

If you are designing a delegation flow, build the history into the handoff surface itself. Show the worker how many tasks of this type the agent has completed, how many it escalated, and how many the worker revised after receiving them. Surface that record before the worker decides whether to send the task or set the interruption threshold. A worker who can see five weeks of pattern in thirty seconds is making a routing decision; a worker who cannot is making a guess, and your product is treating the two as equivalent.