How to Measure Customer Wait Times
Measuring customer wait times starts with three decisions, and none involves equipment. Name the event that starts the clock. Name the event that stops it. Decide who is inside the count. Only then does the instrument matter.
Get one of the three wrong and you still get a number, just not one anybody can check. Take June 2026 in Northern Ireland. Type 1 emergency departments reported a median of 5 hours 5 minutes for patients sent home (Department of Health, 2026).
For patients admitted to hospital, the median was 13 hours 45 minutes (the same bulletin, 2026). One month, one set of buildings, one written clock rule. Only the counted population moved, and the two figures sit more than eight hours apart.
That transfers to a café without much translation. Time the customers who reach the till and you have one number. Time everyone who joined the queue, including the two who walked off, and you have another. Neither is wrong. They answer different questions, and somebody chose which question to ask. Once you have that number, managing a café queue at peak hours is the next practical step.
Key Takeaways
- A wait time is the gap between two named events. Write both down first.
- Same departments, June 2026, same clock rule, two medians: 5 hours 5 minutes and 13 hours 45 minutes (Department of Health, 2026). Only the population differed.
- Four instruments suit a small venue, and each is blind to something. Your own timestamps miss anyone who never registered.
- How long it felt is a second measurement, taken with one published question.
- An average hides the queue people complain about. Official statistics print a percentile beside it.
- With nothing timestamped, an arrival rate and a headcount still give an estimate.
A wait time is two events and a rule about who counts
A wait time is the gap between a named start event and a named stop event. Until both are written down you have no measurement, just an impression.
The Northern Ireland bulletin writes both rules into one sentence (Department of Health, 2026):
“Time is measured from when a patient arrives at the ED (time of arrival is recorded at registration or triage whichever is earlier (clock starts)) until the patient departs the ED (time of departure is defined as when the patient’s clinical care episode is completed within the ED (clock stops))”
Nobody stands at the door with a stopwatch. So the department takes whichever record lands first, registration or triage. That’s a proxy, chosen on purpose and published.
The stop rule is a definition too. The clock stops when the clinical care episode is complete, not when somebody walks out.
One wait can also hide several. NHS England’s Emergency Care Data Set records at least four timestamps for a single attendance (NHS Data Model and Dictionary, 2025). They bracket one visit: the start, the first assessment, being seen for treatment, and the end.
Every published wait is one of them minus another. Record one timestamp and you can report one gap, no more.
Ambulance services in England and Northern Ireland publish the clearest “close enough” start rule. For the most urgent calls, the clock starts at the earliest of three named events (Government Analysis Function). Those are the call handler coding the incident, the first resource being assigned, or “within 30 seconds of call connect”.
| System | The clock starts | The clock stops |
|---|---|---|
| Type 1 emergency departments, Northern Ireland (the bulletin’s own category) | At registration or triage, whichever is earlier | When the clinical care episode is completed |
| Ambulance incidents, England and Northern Ireland | At the earliest of three named events, the last a timer after call connect: 30 seconds for C1, 240 seconds for C2 to C4 | When the first resource reaches the scene, or when the transporting vehicle arrives if the patient needs one |
| HMRC telephone lines | After the caller asks for an adviser, past the automated menu | When an adviser answers |
Source: the same bulletin, 2026; the ambulance page; GOV.UK, 2021.
Queue shape decides what those two events mean. One shared line puts the start of a wait in one place. Several separate lines put it somewhere else, for the same customer. That is one strand of whether one queue or several moves people faster.
The definition decides the number, and so does the population
Two lawful definitions of the same-sounding wait can run side by side. Each produces its own number. Who you count is a second lever.
In July 2022 the UK statistics regulator wrote about the two twelve-hour measures then in use (Office for Statistics Regulation, 2022). One covered patients who “endure a wait of more than 12 hours from decision to admit to admission”. The other covered a wait “from the point of arrival in A&E to the time they are discharged, admitted or transferred”.
The regulator had concluded the statistics were not misleading. It also said two 12-hour measures “led to the potential to confuse users”.
The two definitions produce numbers a long way apart. For January 2025 in England, 61,529 people waited 12 hours or more after the decision to admit them (Royal College of Emergency Medicine, 2025). Under the other, 172,515 patients at major Type 1 departments waited 12 hours or more from arrival.
Those are two separately published counts, on two different populations, so no ratio between them is available. The Royal College was presenting figures published by NHS England, not producing them.
A third mechanism is quieter. It works underneath the metric, not on it. In England, median waiting time in A&E and time to treatment “have been calculated using provisional Emergency Care Data Set (ECDS) data since June 2020” (Nuffield Trust, 2026). A series that crosses that date crosses a change in where its figures come from.
The fourth moves out of healthcare. A 2021 HMRC update sets out a telephone measure that starts the clock late (GOV.UK, 2021). Recording runs “from the time a customer requests to speak to an adviser, after going through” the routing system, until an adviser answers. Four minutes in the menu is real waiting, and outside the published figure by definition.
There’s a rule for the operator here. Change when you start the clock and you have changed your numbers. Any before-and-after comparison across that change measures the new definition as much as the queue.
Which instrument should you use?
Four instruments are realistic for a small venue, and each is blind to something different. The choice comes down to what you can afford not to see.
The timestamps you already have
Most venues already record the start event without calling it that. A 2025 paper describes an outpatient clinic in Lebanon’s Bekaa region (Archives of Public Health, 2025). Records went in digitally on arrival. The times were “automatically generated as system timestamps”.
A ticket machine, a booking system or a till does the same job at no extra cost. The blind spot is built in. Anybody who never entered the system is not in the data.
A clipboard and one shift
A 2025 study of an Egyptian private hospital used a “voice of the process observation sheet, which tracked the journeys of 100 patients” through named stages (BMC Nursing, 2025). Its 07:00 to 22:00 window ran daily, from September 2023 to March 2024. Of the four instruments here, only an observer can watch somebody arrive, wait and leave without registering. Only an observer counts the person who eyes the queue and walks on.
It costs one person watching, for a shift. How many shifts? Enough to cover the conditions you worry about, because a Tuesday morning says nothing about Saturday lunchtime.
A timed visit
Send somebody in as a customer to record the two events. A 2025 study in Kisumu County, Kenya, did exactly that (BMC Health Services Research, 2025). Briefed visitors “discreetly recorded the exact time of arrival to the facility, the exact time they were seen by a provider”. Phone notes and a watch are enough.
A briefed visitor who leaves is a dropout in the design, not a customer who gave up. So it cannot count walkaways.
When nothing is recorded at all
Little’s Law relates three averages. John Little set the notation out in 2011, fifty years after his first proof. L is the “average number of items in the system (items)”. The Greek letter lambda is the “average arrival rate of items to the system (items/unit time)”. W is the “average wait in the system (time units/item)” (Operations Research, 2011).
“In the system” includes the person being served. Count only the standing line and W comes out too low. Little states the theorem for a window that is empty at both ends. A trading day does that on its own.
He gives the retail version too. The day’s customer total comes from “tabulating the number of checkouts”. Divide that by “the number of hours the store is open” and the arrival rate follows.
Take an illustrative bakery, with easy numbers rather than real ones. The till records 240 sales across an eight-hour day, so arrivals run at 30 an hour. Ten headcounts average 5 people inside at a time. W is 5 divided by 30: a sixth of an hour, or 10 minutes.
Cameras, Wi-Fi and footfall sensors are a fifth option. They count devices, not people. Apple’s support page says a device uses a different Wi-Fi address for each network, and “may rotate (change) the address periodically” (Apple, 2025). Rotation is a setting, and it runs every two weeks. So an identity holds for a day, not a season.
Little put it plainly: “putting RFID on carts keeps track of carts, not customers” (Little, 2011). If a supplier quotes an accuracy figure, count by hand against the sensor for one busy hour.
| Instrument | What it costs | What it cannot see | Published example |
|---|---|---|---|
| Timestamps you already hold | Nothing new | Anybody who never entered the system | Outpatient clinic, Lebanon (Arch Public Health, 2025) |
| A person with a clipboard | One observer, one shift | Anything outside the hours watched | 100 patient journeys, Egypt (BMC Nursing, 2025) |
| A timed visit by a briefed customer | One person’s time per visit | Anybody who gave up | Mystery clients, Kisumu County, Kenya (BMC Health Serv Res, 2025) |
| An arrival rate and a headcount | A till count and a few room counts | Any one person’s wait | The supermarket method (Little, 2011) |
| Cameras, Wi-Fi and footfall sensors | Hardware, plus a step from devices to people | People, until that step is checked | Little on RFID carts, 2011; Apple on rotating Wi-Fi addresses, 2025 |
Counting the people who leave needs an instrument that can see somebody who never registered. Only the clipboard can. The rates themselves are a separate problem: what published walkaway figures actually measure.
How long it felt is a second measurement
You find out by asking, and one question is enough. Felt duration is its own quantity with its own instrument. Published studies print their wording, so nothing has to be invented.
A 2023 study at an optometry clinic in China separates the two in plain terms. Perceived waiting time was taken by asking the patient “how long he felt he had waited this time”. Actual waiting time was “the difference between the time of entering the clinic and the time of registration” (Medicine (Baltimore), 2023). That was 292 valid questionnaires, August 2022.
Four wordings come from current work. Copying one takes five minutes.
| The item, as printed | What it returns | Where it was used |
|---|---|---|
| “How much time did you wait to see a doctor (from registration to the time you saw the doctor)” | An estimate in bands, from under 30 minutes to over 120 | Emergency room, Nairobi, Kenya. 941 surveys, April to August 2023 (PLOS ONE, 2025) |
| “How long did you wait to see the provider?” | An open estimate, in the customer’s own units | Household survey, 744 women, Kisumu County, Kenya (BMC Health Serv Res, 2025) |
| “How did you feel about the amount of time you waited? Would you say it was very short, somewhat short, somewhat long, or very long?” | A labelled scale on how the wait felt, not how long it was | The same household survey, 2025 |
| An 11-point format “where 0 denotes no stress, and 10 is extreme stress” | A rating. This one measures stress, not felt duration | Laboratory elevator study, Japan (PeerJ, 2025) |
Two things self-report will not do. In Nairobi, “waiting times were an approximation from the study participants and could not be fully verified via a queue system” (PLOS ONE, 2025). In Kisumu, “recall bias likely impacts accurate reporting of wait times” (BMC Health Serv Res, 2025). Participants there “may under-report negative experiences, long wait times, and dissatisfaction with long wait times due to social desirability bias”.
The Kisumu work shows what measuring both looks like. A household survey of real users ran alongside briefed visitors timing their own trips. Both ran in the same county over overlapping months in 2022, one in the community, one in facilities. Publication is 2025, so it stands as a method, not as current data.
The two numbers will not match, and the reasons they come apart are their own subject: why a wait feels longer than the clock says.
An average on its own hides the queue people complain about
Add the waits, divide by the number of people, and you have the mean. It cannot see anyone who left. A customer who walked out is in neither the total nor the count. Waiting data is skewed too, so the mean often describes a queue nobody stood in.
A 2024 study of dermatology appointment waits in New York City makes it visible on one dataset. “The mean waiting time for an appointment was 50 days”. “The median waiting time was 19.5 days”. The authors add that “the distribution was considerably skewed” (Archives of Dermatological Research, 2024).
The sample was 344 dermatologists reached by phone, from 486 listed. That is an appointment-booking wait in days, not a queue wait in minutes. The authors name the city’s supply of specialists as their own limit.
Official statistics say why the middle figure gets preferred. The Scottish Government’s earnings glossary calls the median “the preferred measure of average earnings”. Its reason: the median is “less affected by a relatively small number of very high earners than the mean is” (Scottish Government, 2025).
That is about pay, not queues. Yet one four-hour Saturday drags a mean the way one very large salary does. The middle customer’s experience sits where it was.
UK practice pairs the two rather than choosing. For June 2026, Northern Ireland reports a median of 1 hour 38 minutes from triage to the start of treatment (Department of Health, 2026). The same line adds: “with 95 percent of patients receiving treatment within 8 hours 59 minutes of being triaged”. The middle is under two hours. The top is nearly nine.
The number you quote to a waiting customer is a different sum again. A March 2026 simulation names the methods used in practice. Three of the four, by their inputs, need no software.
| Method, as the paper names it | What it needs | What it does |
|---|---|---|
| Rolling average: “the average wait time of patients who commenced service within the past three hours” | Recent completed waits | Smooths noise, lags a change |
| Last to enter service: “the actual wait time of the most recent patient who completed their wait and entered service” | The last wait, and nothing else | The simplest. The paper calls it robust in congested systems |
| 95th percentile of recent wait times: “the 95th percentile of actual wait times” | Recent waits from the past three hours, ranked | Quotes near the top of the distribution. In use in “the 18 public hospitals in Hong Kong” |
| Exponential smoothing: “a weighted average wait time by assigning exponentially decreasing weights to older observations” | Recent waits, plus a weighting rule | Reacts faster than a flat average |
Source: JAMIA, 2026. A simulation calibrated on visit records from three emergency departments in Hong Kong. Its authors warn that “inaccurate predictions can actively misdirect patients”.
Before anyone buys a system
Two decisions come before any purchase, and neither of them costs anything. Which two events bracket your wait? And who is inside the count?
Go back to the two medians at the top. Those departments recorded a timestamp for every patient and published their clock rule in writing. They still produced 5 hours 5 minutes and 13 hours 45 minutes for June 2026. The timestamps were never the problem. The population was, and no equipment decides that for you.
Everything after the measurement, including how to make a wait feel shorter, rests on the two lines you write down here. So write the two events on a piece of paper. Show them to whoever reads the number in three months. If they cannot tell you who is in the count and who is outside it, no system will rescue that. A star rating that mentions waiting is not a measurement, and whether a bad review is really about the wait is a separate question.
Frequently Asked Questions
Should you use the average or the median wait time?
Report the median, and put something from the top of the distribution beside it. A mean gets dragged by a handful of very long waits. In one 2024 dataset of dermatology appointment waits in New York City, the mean was 50 days and the median 19.5 days (Arch Dermatol Res, 2024). The authors call that distribution considerably skewed. Official statistics prefer medians for the same reason (Scottish Government, 2025).
How do you measure customer wait times without a ticket system?
Pick by what you can spare. If you can free somebody for a shift, put them on the floor with a sheet of paper. One 2025 study tracked 100 journeys that way (BMC Nursing, 2025). Only an observer sees the customers your systems never record. If you cannot spare the shift, send one briefed visitor to time a single trip (BMC Health Serv Res, 2025). One visit tells you about that visit only.
What is Little’s Law, and can a small business use it?
Yes, with one limit worth knowing. Little’s Law turns an arrival rate and a headcount into an average wait, so a till count and a few counts of the room are enough (Little, 2011). That makes it useful for staffing: it tells you what an average Saturday looks like. It cannot tell the person at the counter how long they will wait, because the answer it gives is an average across a window.
Sources
- Apple, “Use private Wi-Fi addresses on Apple devices”, page dated 9 December 2025, retrieved 2026-08-29, https://support.apple.com/en-gb/102509
- Archives of Dermatological Research, “Wait times for scheduling appointments with hospital affiliated dermatologists in New York City”, published 17 August 2024, retrieved 2026-08-29, https://pmc.ncbi.nlm.nih.gov/articles/PMC11330380/
- Archives of Public Health, “Reducing patient waiting times in humanitarian settings: a data-driven approach to improving healthcare access for vulnerable populations”, published 11 November 2025, retrieved 2026-08-29, https://pmc.ncbi.nlm.nih.gov/articles/PMC12606882/
- BMC Health Services Research, “Comparative perceptions of wait times for family planning services among contraceptive users and mystery clients in Kisumu, Kenya: a mixed methods analysis”, published 30 August 2025, retrieved 2026-08-29, https://pmc.ncbi.nlm.nih.gov/articles/PMC12398003/
- BMC Nursing, “Streamlining emergency nursing care post-pandemic: A lean approach for reducing wait times and improving patient and staff satisfaction in the hospital”, published 22 April 2025, retrieved 2026-08-29, https://pmc.ncbi.nlm.nih.gov/articles/PMC12016415/
- Department of Health (Northern Ireland), “Emergency care waiting time statistics (April – June 2026)”, published 12 August 2026, retrieved 2026-08-29, https://www.health-ni.gov.uk/news/emergency-care-waiting-time-statistics-april-june-2026
- Government Analysis Function, “Summary of ambulance response time data in the UK”, no publication or update date printed on the page, content postdates December 2025 by internal evidence, retrieved 2026-08-29, https://analysisfunction.civilservice.gov.uk/government-statistical-service-and-statistician-group/user-facing-pages/theme-based-support-for-users/health-and-care-statistics/summary-of-ambulance-response-time-data-in-the-uk/
- HM Revenue and Customs, “HMRC monthly performance update: January 2021”, published 11 March 2021, retrieved 2026-08-29, https://www.gov.uk/government/publications/hmrc-monthly-performance-report-january-2021/hmrc-monthly-performance-update-january-2021
- Journal of the American Medical Informatics Association, “Impact of announced wait time information on emergency department overcrowding mitigation: a simulation study”, published 30 March 2026, retrieved 2026-08-29, https://pmc.ncbi.nlm.nih.gov/articles/PMC13197183/
- Medicine (Baltimore), “Effect of waiting time on patient satisfaction in outpatient: An empirical investigation”, 102(40):e35184, published 6 October 2023, retrieved 2026-08-29, https://pmc.ncbi.nlm.nih.gov/articles/PMC10553012/
- NHS England, NHS Data Model and Dictionary, “Emergency Care Data Set Version 4”, August 2025 release, retrieved 2026-08-29, https://archive.datadictionary.nhs.uk/DD%20Release%20August%202025/data_sets/clinical_data_sets/ecds_v4/emergency_care_data_set_version_4.html
- Nuffield Trust, “A&E waiting times”, last updated 25 June 2026, retrieved 2026-08-29, https://www.nuffieldtrust.org.uk/resource/a-e-waiting-times
- Office for Statistics Regulation, “Ed Humpherson to Chris Roebuck and Mark Svenson: Statistics on patients waiting more than 12 hours”, published 8 July 2022, retrieved 2026-08-29, https://osr.statisticsauthority.gov.uk/correspondence/ed-humpherson-to-chris-roebuck-and-mark-svenson-statistics-on-patients-waiting-more-than-12-hours
- Operations Research, Little, “Reprint: Little’s Law as Viewed on Its 50th Anniversary”, volume 59 issue 3, retrieved 2026-08-29, https://projectproduction.org/journal/reprint-littles-law-as-viewed-on-its-50th-anniversary/
- PeerJ, “How long is too long? Examining waiting times and stress in human-elevator interaction”, published 18 August 2025, retrieved 2026-08-29, https://pmc.ncbi.nlm.nih.gov/articles/PMC12369606/
- PLOS ONE, “The Perception of Waiting Times on Patient Satisfaction and Patient Care: A Cross-Sectional Study at a Tertiary Health Care Institution in Kenya”, published 2 May 2025, retrieved 2026-08-29, https://pmc.ncbi.nlm.nih.gov/articles/PMC12047815/
- Royal College of Emergency Medicine, “RCEM describes record 12-hour ‘trolley waits’ as ‘a catastrophe'”, published 13 February 2025, retrieved 2026-08-29, https://rcem.ac.uk/press-release/rcem-describes-record-12-hour-trolley-waits-as-a-catastrophe/
- Scottish Government, “Annual Survey of Hours and Earnings 2025”, glossary page, first published 27 October 2025, retrieved 2026-08-29, https://www.gov.scot/publications/annual-survey-of-hours-and-earnings-2025/pages/glossary-copy/