The Hidden Cost of AI: The Mental Load Behind Multi-Tasking Development
Parallel AI tasks look like free capacity, until you notice the real limit was never compute. It was how much a person can hold in their head at once.
Fleet Work Team
Story points assumed a predictable, human-paced kind of work. AI breaks that assumption, and estimation hasn't caught up yet.
Story points were never really about time, but they worked because human effort is fairly consistent. A three-point ticket for one senior engineer looks roughly like a three-point ticket for another. That consistency is what let planning poker produce numbers a team could plan a sprint around with a straight face.
AI-assisted and AI-executed work doesn't behave that way. The same ticket might take an agent four minutes or four hours depending on factors that have nothing to do with the complexity a human would perceive: how well-trodden the code path is, how much relevant context already exists, whether the task decomposes cleanly or hides a nasty ambiguity three files deep. Two tickets that look identically sized to a human triaging the backlog can have wildly different execution profiles once an agent gets involved.
This breaks the quiet assumption sprint planning was built on: that effort is a reasonably smooth function of scope. When it isn't, story points stop being a forecasting tool and start being theater, a ritual that produces a number without producing the certainty the number is supposed to represent.
What's replacing it, in teams that have actually adapted, looks less like estimation and more like triage. Instead of asking 'how many points is this,' the useful question becomes 'how confident are we that this can be executed cleanly, and what's actually driving the risk.' That's a different kind of number: not effort, but a probability of clean execution, backed by reasoning about blast radius, ambiguity, and how much relevant precedent exists.
Planning around confidence bands instead of point totals changes how a sprint gets built. High-confidence work gets pulled forward and often executed same-day. Low-confidence work gets flagged for a human to scope down before anyone commits capacity to it. The backlog stops being a flat list sorted by guesswork and starts being sorted by something a team can actually act on.
Estimation isn't dying, but the version built for a world where all execution happens at human pace is running out of runway. Teams that replace it with something that accounts for how differently AI-assisted work actually behaves will plan sprints that hold up. Teams that keep pointing everything the old way will keep being surprised by their own backlog.
Fleet Work Team
Writing on autonomous delivery
Parallel AI tasks look like free capacity, until you notice the real limit was never compute. It was how much a person can hold in their head at once.
Fleet Work Team
Teams have always leaked knowledge as people move on and memory fades. That used to be unavoidable. With AI systems that persist what they learn, it no longer has to be.
Fleet Work Team
AI can now own real chunks of the development process. The hard part was never the technology, it's the letting go.
Fleet Work Team