You're not losing 23 minutes per interruption. Count this.

The "23 minutes and 15 seconds to get back on task" figure is not a research finding. It comes from a magazine interview: Gallup Business Journal, June 8, 2006, where the researcher, Gloria Mark, offered it as the good news and added "which I guess is not so long." The peer-reviewed paper covering the same data reports a different number, with a spread wide enough that no single interruption can be priced at all. If you feel like the bottleneck in your own business, the honest answer is that nobody's study can tell you what that is costing you. Your own week can, in five days, with a pen.

This post does two things. First it takes apart the statistic every article about founder overwhelm quotes, because we read the sources and they do not say what the articles say. Then it gives you the arithmetic that actually applies to your business, which has been sitting in a 1961 theorem the whole time and needs no research at all, only two numbers you can count yourself.

Where does "23 minutes and 15 seconds" actually come from?

From a Q&A. On June 8, 2006, Gallup Business Journal published "Too Many Interruptions at Work?", an interview with Gloria Mark, then an associate professor at the Donald Bren School of Information and Computer Sciences at UC Irvine, conducted by Jennifer Robison. Asked how long it takes to get back to work after an interruption, Mark answered:

"There's good news and bad news. To have a uniform comparison, we looked at all work that was interrupted and resumed on the same day. The good news is that most interrupted work was resumed on the same day — 81.9 percent — and it was resumed, on average, in 23 minutes and 15 seconds, which I guess is not so long."

That is the origin of the most-repeated productivity number in circulation. It is a researcher describing her own results conversationally, in a business magazine, and calling the figure reassuring. Twenty years later the same number is quoted as the cost of your distraction, usually multiplied by the number of times your phone rang.

What does the peer-reviewed paper say instead?

It says 25 minutes and 26 seconds, with a standard deviation of 54 minutes and 48 seconds. The study is Gloria Mark, Victor González and Justin Harris, "No Task Left Behind? Examining the Nature of Fragmented Work", published at CHI 2005 — the year before the interview. Its exact sentence: "When people did resume work on the same day, it took an average length of time of 25 min. 26 sec (sd=54 min. 48 sec.)." The 81.9 percent same-day figure is in the paper too, so the interview and the paper are describing the same body of work.

We searched the text of that paper for the string "23 min". It is not there. We searched the follow-up experiment, "The Cost of Interrupted Work: More Speed and Stress" (Mark, Gudith and Klocke, CHI 2008). Not there either. We are not claiming the number appears nowhere in twenty years of publications; we are reporting that the two papers it is usually attached to do not contain it, and the earliest statement of it we could find is the 2006 interview.

That third row is the one that matters and nobody quotes it. When the spread around an average is more than double the average, there is no typical case. Some interrupted work came back in two minutes and some came back four hours later, and the mean sits in a gap where very little actually happened. Multiplying it by your daily call count produces a number with no meaning.

Who was actually measured, and were they anything like you?

No. Per the methodology section of the 2005 paper, it observed 24 information workers — 7 managers, 9 analysts and 8 developers — across two teams at one company, in two phases over a thirteen-month period. Each person was shadowed for about three and a half days, averaging 25 hours 42 minutes of formal observation each, for more than 700 observation hours in total. A researcher sat in the cubicle with a stopwatch and a notepad and timed events to the second.

It is careful, honest, labour-intensive fieldwork. It is also a study of people who sat at desks in 2004, whose interruptions were email, instant messages and colleagues leaning into a cubicle. Nobody in that sample was quoting a job from a truck, deciding whether to eat a mistake on a customer's kitchen, or being called by a supplier who cannot deliver on Thursday. Applying its averages to a service business with staff in the field is not a small extrapolation. It is a different system.

So what does the research say interruptions really cost?

Not time. This is the finding that survives, and it is the opposite of what the quoting industry claims. In the 2008 lab experiment, Mark, Gudith and Klocke reported that "people completed interrupted tasks in less time with no difference in quality" and concluded that "people compensate for interruptions by working faster, but this comes at a price: experiencing more stress, higher frustration, time pressure and effort."

Read that as an owner. You already knew you get the work done. You get it done at 9pm, you get it done in the truck, you get it done while someone waits. What the interruptions take is not your output. It is your capacity to absorb the next one. That is why the tiredness is not proportional to the hours, and why a week that looks fine on the invoices can feel unsurvivable.

Two honest limits on that study, because it is a lab: 48 subjects, all German university students with a mean age of 26, working a simulated email-answering task with interruptions imposed every two minutes. It measures a mechanism, not your Tuesday.

What number about your business can you actually get?

The one nobody has to publish for you, because it is a theorem, not a survey. In 1961 John Little proved that in any queue,

L = λW

where L is the average number of items sitting in the system, λ is the average rate at which items arrive, and W is the average time an item spends waiting. Fifty years later Little wrote a review of his own result for Operations Research, volume 59, number 3, May–June 2011, pages 536–549, and the useful part for you is what he added there rather than what he proved in 1961.

He argued that measuring a real queue over a finite window — a week, say — is not a compromise. In his words, "the finite time interval guarantees that the relationship L = λW is numerically exact." He also states plainly that with this method "we are in the measurement business, not the forecasting business." You are not modelling anything. You are counting.

Four properties of that theorem, all stated in the 2011 paper, decide whether it applies to a business owner's week:

• It does not assume things arrive at a steady rate. Little writes that the finite-window result "holds under nonstationary conditions" and that he "did not have to assume that the underlying arrival process was stationary." Your Monday and your Friday can look nothing alike.
• It does not care what order you handle things in. "LL.1 holds independent of queue discipline" — first in first out, last in first out, at random, or by whoever shouted loudest. Which is exactly how an owner's day runs.
• It works when the queue is never empty. His second theorem covers a window that starts and ends with things still waiting, and the corollary reads: "LL.2 holds exactly even if the queuing system is never empty." You have a backlog on Monday morning. Fine.
• It works per category, and the categories add up. The paper's third corollary splits items into classes, gives each class its own L, λ and W, and shows the totals sum. So you can measure "needs my judgement" and "only needs my permission" as two separate queues and they will reconcile.

Little's own explanation of why the law is true is one sentence long: "An item in queue is also waiting." While the thing is sitting there being counted, it is also accumulating minutes. That is the whole mechanism.

How do you run the one-week tally?

Keep one page, on the dashboard or on the desk, from Monday morning to Friday evening. Every time something comes to you — a call you have to take, a text you have to answer, someone walking up, an approval, a price question, a supplier problem, a decision somebody is standing still without — write one line:

1. What time it landed.
2. What it was, in four words.
3. What time you actually dealt with it. If it waited until the evening, write the evening.
4. One letter: J or P. J if it genuinely needed your judgement — something only you know how to weigh. P if it only needed your permission, your presence or a fact that is in your head and nowhere else.

That is it. No app, no categories, no scoring. Anything already waiting on you at Monday morning gets a line too, timed from Monday morning, which is what Little's second theorem allows.

What do you do with the three numbers?

Divide twice. Take the hours you personally worked in that window — call it T. Count the lines on the page — call it N. Average the gap between column 1 and column 3 — that is W, in hours. Then:

λ = N ÷ T   and   L = λ × W

A worked example with plausible figures for a 3–20 staff service business, so you can see the shape before you have your own:

Six things were waiting on you at any given instant, all week. That is the sentence you have been trying to say when you tell your spouse the business needs you constantly. Now split it by the letter in column 4. If 34 of the 90 were J and 56 were P, then the judgement queue is 34 ÷ 50 = 0.68 per hour and the permission queue is 56 ÷ 50 = 1.12 per hour. At the same average wait those come to 2.2 and 3.6, which sum back to 5.8, exactly as the third corollary in Little's 2011 paper says they must.

Roughly two of the six things waiting on you needed you. Roughly four were waiting on your availability. That second number is not a personality problem and it is not solved by working harder. It is the size of the pile that a written rule, a price list, a spending limit or an answered phone would have cleared without you.

If timing every item is too much, there is a lazier version that gives the same L. Five times a day, at random, stop and count how many things are currently sitting on you. Average those counts across the week — that is L directly. Then W = L ÷ λ, and you have the average wait without having timed anything.

Why did the week get worse when the business got bigger?

Because delay does not track workload, it tracks unpredictability — and a bigger business is a more unpredictable one. The cleanest demonstration we could find is a simulation study of general practices published in the International Journal of Environmental Research and Public Health: Grot, Kugai, Degen, Wiemer, Werners and Weltermann, "Small Changes in Patient Arrival and Consultation Times Have Large Effects on Patients' Waiting Times", 2023, volume 20, issue 3, article 1767.

Their first scenario is the one to sit with. They set patient arrivals to exactly 7.6 minutes apart, which is the average German consultation length, so on paper the doctor is precisely keeping up and the average workload is not too high at all. As the paper says, "since the average physician's consultation time equals the patient's inter-arrival times, no waiting times occur when there are no variations in consultation times." Then they let the length of appointments vary, keeping the average at exactly 7.6 minutes. Individual maximum waits rose to more than 200 minutes.

Nothing was added. No extra patients, no longer average appointment. Only the variation. Their second scenario changed when people showed up instead, and unscheduled arrivals alone produced about 20 minutes of average waiting even with completely uniform appointment lengths.

That is your business. You did not take on more work than you can do. You took on work that arrives and finishes less predictably than it used to, and the queue at your phone is the result. This is why "I just need to be more efficient" fails: efficiency moves the average, and the average was never the problem.

What is the cheapest thing you can do about it?

Claw a few minutes back immediately after a day goes wrong, rather than trying to make the whole week smoother. The same 2023 paper models a morning where two appointments unexpectedly run 15 minutes long, which wrecks everything behind them: in the baseline, patients after the 21st are waiting more than 60 minutes and total patient waiting for the session reaches about 2,790 minutes against about 380 minutes of actual doctor time.

Then they test two repairs. Cutting three minutes from each of the next five appointments — giving back 15 minutes in total — removed 405 minutes of waiting. Holding every appointment to a flat 6 minutes instead of an average of 7.6, which saved the doctor about 40 minutes, removed 834 minutes. Dividing their published figures gives returns of roughly 27 to 1 and 21 to 1; that division is ours, not theirs.

Be careful about what transfers. This is a simulation of a German doctor's surgery, not a measurement of a contracting business, and the exact ratios belong to their model. What transfers is the shape: a delay early in a queue compounds through everything behind it, and a small correction applied early is worth many times its own size. In practice that is the twenty minutes you spend at 10am resetting the day after a job blows up, which is exactly the twenty minutes you normally skip because you are behind.

What this changes about "just delegate more"

Page one for this problem is essays and lists of signs. They are not wrong, they are just unfalsifiable: you cannot check whether you have stopped being the bottleneck, because nothing was measured. The tally gives you a before number and a number to beat.

It also reorders the work. The advice says start with the simplest tasks. The tally says start with the largest P pile, because that is the queue that only exists because a rule is unwritten, and a written rule costs one evening and never has to learn anything. Look at your P lines and you will usually find three or four sentences would have removed most of them: what we charge for the standard job, how much someone may spend without asking, when to say yes to a Saturday, who calls the customer when we are running late.

The J pile is different and should not be shrunk out of pride. Judgement calls are what you are for. The question there is only whether they reach you in a batch you handle at a set time, or one at a time all day, which is the variability finding above applied to your own attention.

Before you do anything else

Run the week. Then, if you want a second pair of eyes on what the page says, our free AI Work Audit is a short questionnaire about how work actually reaches you, ending in an optional 15-minute call. If the honest answer is "go write four rules down and run the tally again," that is what you will hear. Related reading on the same queue, from the customer's side rather than yours: what a missed call actually costs.

What we could not verify

Said plainly, because a post about a misquoted statistic has no business being loose with its own. We did not find any study of interruption load among owners of small service businesses; every figure above about interruptions comes from office and laboratory settings and is presented as such. We are not asserting that "23 minutes 15 seconds" never appears in any of Gloria Mark's publications, only that it is absent from the two papers it is normally attached to and that the earliest source we could read for it is the 2006 interview. The 27-to-1 and 21-to-1 ratios are our own division of the 2023 paper's published minute totals and are not claims the authors make. And no number in this post describes your business. That is the point of the tally.

Frequently asked

Is the 23-minute figure wrong, then? Not wrong, misused. It is a real researcher's real summary of real data, given in an interview. What is wrong is quoting it as a per-interruption cost for people who were never studied, and dropping the fact that the paper's own version of it carries a standard deviation more than twice the size of the mean.

Do I have to do the tally for a whole week? A week is the shortest window that contains both your quiet day and your bad day, and the bad day is where the queue forms. Little's finite-window result is exact over whatever window you pick, so three days is valid arithmetic, but it will flatter you if those three days were calm.

What counts as one item? One thing that could not move until you touched it. A single conversation that settles four questions is four items if four things were stuck, one item if one was. Do not agonise; consistency across the week matters more than the rule.

My L came out under 2. Is that good? It means you are not the constraint and something else is, which is worth knowing before you reorganise anything around yourself. Run the same tally on whoever the work stacks up behind instead.

Where exactly do these sources live? Every one is free to read, so check us. Gallup Business Journal, "Too Many Interruptions at Work?", June 8, 2006, interview by Jennifer Robison. Mark, González and Harris, "No Task Left Behind? Examining the Nature of Fragmented Work", CHI 2005, read as the PDF hosted on the author's UC Irvine page. Mark, Gudith and Klocke, "The Cost of Interrupted Work: More Speed and Stress", CHI 2008, same source. Little, "Little's Law as Viewed on Its 50th Anniversary", Operations Research 59(3), 2011, pages 536–549, doi 10.1287/opre.1110.0940, read in full. Grot and colleagues, International Journal of Environmental Research and Public Health 2023, 20(3):1767, doi 10.3390/ijerph20031767, open access: full text, catalogue record.