What one interruption really costs, now that your AI tools ping you too

An interruption costs more than the seconds it takes. In the most-cited field study of fragmented office work — 24 information workers shadowed across more than 700 hours of observation, published at CHI 2005 — people took an average of 25 minutes and 26 seconds to get back to an interrupted task, and worked on 2.26 other things before they did.

Everything load-bearing below comes from three peer-reviewed studies and three vendor help pages, each read in full on 20 August 2026. The field observations are Gloria Mark, Victor González and Justin Harris, "No Task Left Behind? Examining the Nature of Fragmented Work" (CHI 2005); the stress measurements Mark, Daniela Gudith and Ulrich Klocke (CHI 2008); the email findings Mark, Iqbal, Czerwinski, Johns and Sano (CHI 2016). Attention residue is Sophie Leroy's concept, per her faculty page at the University of Washington Bothell. What the AI tools do comes from OpenAI's, Google's and Slack's own documentation. This is information, not medical or mental-health advice.

How long does it really take to get back on task?

An average of 25 minutes and 26 seconds, per the CHI 2005 paper — with a standard deviation of 54 minutes and 48 seconds, the more useful half of the finding. Some resumptions were near-instant. Some took the rest of the afternoon.

The same study found people spent an average of 11 minutes and 4 seconds on a single "working sphere" — one coherent chunk of work with its own documents, people and goal — before switching. Some 57.1 per cent of observed segments were interrupted.

Two caveats the paper carries and the internet usually drops: this was observation of 24 people at one firm, not a controlled experiment, and the enormous standard deviation means "25 minutes" is not a per-interruption tax you can subtract from your calendar.

Is the "23 minutes and 15 seconds" figure real?

Not in the paper it is usually credited to. The number published in the CHI 2005 study is 25 minutes and 26 seconds; the widely circulated "23 minutes and 15 seconds" does not appear in it. If you are quoting the research rather than a blog post about the research, quote the published figure.

Does an interruption actually make you slower?

No — and that is the uncomfortable part. In the CHI 2008 lab experiment with 48 participants, interrupted tasks were finished faster than uninterrupted ones: about 20.3 to 20.6 minutes against a 22.8-minute baseline. People compensated. What it cost them showed up somewhere else entirely.

On the study's 20-point workload scales, the same participants reported markedly worse experiences when interrupted. Stress rose from 6.92 at baseline to between 9.13 and 9.46. Frustration went from 4.73 to roughly 6.5. Time pressure climbed from 11.02 to as high as 12.69, and effort from 9.50 to about 11.5.

The authors' reading is that people absorb interruptions by working faster and writing shorter, and pay in mental demand rather than in minutes. If your day feels expensive but your output looks fine, that pattern is documented — in a lab simulation of email work, which is not your job.

What is attention residue?

Attention residue is a carry-over effect in which part of your attention stays fixed on a previous task after you have already switched to the next one. Leroy introduced it in a 2009 paper in Organizational Behavior and Human Decision Processes, and her university profile describes the research programme as work on when residue occurs, what it does to performance, and how to prevent it.

It is the mechanism that makes the resumption numbers legible: the cost of a switch is not only the time spent away, but the degraded quality of the attention you bring back. Her later work, listed on the same page, extends this to task transitions and interruptions more broadly.

Do AI assistants add a new kind of interruption?

They add a scheduled one, which is genuinely different from a colleague's ping: it arrives on a cadence you set, and it keeps arriving. Both major consumer assistants ship background scheduling that ends in a notification, and both document the limits and the off switch.

Per OpenAI's Scheduled Tasks documentation, ChatGPT can run one-off or recurring tasks, including monitoring tasks that "check for changes and notify" you when something moves. Per Google's Gemini Apps help page, scheduled actions run daily, weekly or monthly and are prepared in the background so the result is ready at the appointed time.

One overlooked detail in Google's documentation: results are pre-generated ahead of delivery — several hours ahead on free accounts, within the hour for subscribers — and the page warns that "rapidly changing data, like stock prices, won't be the latest when you receive them." A notification that interrupts you with stale numbers is the worst of both trades.

Does batching your checks fix it?

Partly, and not the part you were promised. In the CHI 2016 study, 40 information workers were logged for about 12 workdays each while wearing heart-rate sensors. They averaged 1 hour 23 minutes a day in email across 77 checks, and longer daily email time tracked with both lower perceived productivity and higher measured stress.

Batchers did rate their productivity higher than mixed-strategy users on heavy email days. But the authors are blunt about the popular claim: "Despite widespread claims, we found no evidence that batching email leads to lower stress."

One finding does transfer neatly to AI notifications. People who checked mainly by self-interruption — going to look — reported higher productivity at high volumes than people who waited to be notified. Being pulled is not the same as choosing to go.

How do you turn the volume down using only documented settings?

1. Set the cadence at creation. ChatGPT tasks cannot run more often than hourly by design; Gemini actions run at most daily. Choosing weekly where weekly will do removes six interruptions.
2. Choose a channel you visit, not one that visits you. Gemini's desktop behaviour is to mark the chat unread; ChatGPT's notification channels switch individually in Notifications settings.
3. Treat the ceiling as a ceiling. Ten active Gemini actions or fifteen ChatGPT tasks is a limit, not a target. Both platforms document pause, edit and delete for exactly this.
4. Put a floor under the day elsewhere. Slack's work-hours guidance suggests setting a daily Do Not Disturb schedule as the inverse of your hours, and notes that workspace owners can set defaults for everyone.

What these sources do not establish

None of this research studied AI assistants. The observations date from 2005, 2008 and 2016, and none of the three papers measured a scheduled machine notification of the kind ChatGPT and Gemini now send. In these sources, there is no resumption cost specific to an AI-generated ping.

The vendor pages describe features, not outcomes: they document limits and controls, and say nothing about attention. The CHI 2005 figures are field observation, so they cannot tell you that fewer interruptions would have produced better work.

So what should you actually do?

If you use an assistant occasionally, this is not your problem yet — one weekly digest is not what the interruption literature is about, and the setup will cost you more than the pings do.

If you run one daily, the documented controls are worth an hour of your attention precisely once. Set the cadence deliberately, keep the count under the ceiling, and prefer a channel you check over one that checks you. The research cannot promise that this buys back 25 minutes. It can tell you the cost of interruption showed up in stress rather than on the clock — the part you feel at six in the evening, and the part no dashboard measures.

For a related business news perspective, read Is AI actually taking entry-level jobs? Here's what the data says.