Open a new AI tool, and the first thing you do, every time, is explain yourself. Who you are, what you're working on, what you already tried, what didn't work. It feels like a minor tax. A few minutes, tops.
It isn't minor. It's one of the best-studied costs in knowledge work, and the numbers are larger than most people assume.
What context switching actually costs, measured
This isn't a new problem invented by AI tools. It's a much older problem in cognitive science and software engineering research that AI tools have simply made more frequent.
Gloria Mark's research at UC Irvine, one of the most cited bodies of work on attention and interruption, found that it takes an average of 23 minutes and 15 seconds to fully regain focus after being pulled away from a task. Her broader research on knowledge workers found people switch tasks roughly every 3 minutes on average, spending only around 11 to 12 minutes in a given working sphere before switching or being interrupted again.
For developers specifically, the picture is just as stark. Parnin and Rugaber analyzed 10,000 real programming sessions from 86 programmers and found that developers routinely spend 15 to 30 minutes reconstructing their working context before they can meaningfully resume an interrupted task. A separate large-scale industry study tracking 44,515 tasks across 23 professional developers over four years found that:
- Developers switch context on 59% of their daily tasks.
- 40% of those switches require genuinely rebuilding context, not just picking up where they left off.
- 29% of interrupted tasks are never resumed at all.
That last number is worth sitting with. Nearly a third of the time, an interruption doesn't just cost you time getting back. It costs you the task entirely.
There's a mechanism behind why this recovery takes so long, not just an observed correlation. Sophie Leroy's research on attention residue found that when you switch tasks, part of your attention stays stuck on the task you just left. You're not actually giving your full focus to the new task. Some of your cognitive bandwidth is still processing the old one, even though you've consciously moved on. For developers working with complex code abstractions specifically, this residue has been observed to linger for 30 to 60 minutes.
And it's not just slower. It's worse. Research by Parnin and DeLine found that interrupted tasks take twice as long and contain twice as many errors as tasks completed without interruption.
Why AI tools make this worse, not better
AI coding assistants and chat tools were supposed to reduce friction, and in one real sense they have. Writing code, drafting text, and answering questions is genuinely faster with them. But they've also introduced a new, very specific version of the same context-switching cost the research above describes.
Every time you move between tools, say from a chat assistant to a coding agent, or from one coding agent to another, you're not just switching windows. You're re-establishing an entire working context that the previous tool already had and the new one has no access to:
- What you're building and why
- What you already tried, and what specifically didn't work
- Your actual codebase conventions, not generic best practices
- Decisions you made three conversations ago that still apply today
None of that carries over automatically. As one 2026 industry analysis of AI coding tools put it plainly: "No single AI coding assistant natively retains full project context across sessions by default." Config files like CLAUDE.md can hold standing instructions, but as the same analysis notes, they "cannot recall a specific decision made three sessions ago, surface the exact command that fixed a build error last week, or carry context from one tool to another."
Even Anthropic's own engineering guidance describes context management as "an active design challenge, not a solved problem." This isn't a minor gap one vendor overlooked. It's an acknowledged, industry-wide limitation.
The compounding effect
Here's what makes this genuinely expensive rather than just mildly annoying. The cost isn't paid once. It's paid every single time you switch.
A developer moving between a chat assistant for planning, a coding agent for implementation, and a separate tool for debugging, a completely normal workflow in 2026, can easily switch contexts a dozen or more times in a single day. If even a fraction of those switches cost the 15 to 30 minutes the research describes for reconstructing working context, the daily total isn't a rounding error. It's hours.
And unlike a human colleague, who can be told "we talked about this yesterday" and actually recall the conversation, most AI tools start from zero. You're not just switching. You're re-teaching, every time, at full cost, with no memory of having taught it before.
What this actually points to
The fix isn't to use fewer AI tools. Different tools are genuinely better at different things, and that's not going away. The fix is closing the actual gap the research identifies: memory that survives the switch, so re-establishing context isn't something a human has to do by hand, on every tool, every single time.
That's a real, measurable cost with decades of research behind it, not a vague productivity vibe. The 23 minute refocus cost, the 15 to 30 minute context-rebuild window, the 29% of tasks that never get resumed. None of that is specific to AI tools. AI tools just made the switching more frequent, which means the same well-documented cost is now happening more often than it ever did before.
