How to Calculate Wait Time in a Queue

Count the people waiting in front of you, leaving out anyone already being served. Add one. Multiply by the average time it takes to serve one person, then divide by the number of staff serving that line. The answer is the wait until your turn starts, assuming every member of staff is busy and nobody ahead gives up.

Take a post office counter, invented for this example. Two staff are each serving a customer, four more people are waiting, and transactions take 3 minutes on average. None of those numbers comes from a real branch. Someone joining the back now has four people ahead. Four plus one is five, five times 3 minutes is 15, and 15 split between two staff is 7.5 minutes.

That 7.5 is a fair starting point, and it can still miss badly. People giving up, urgent cases and rushes all pull a real wait away from it, and what follows shows which way, and where the sum stops working.

Key Takeaways

  • One customer’s wait is roughly the people waiting ahead, plus one, times the average service time, divided by the staff on that line.
  • That’s the wait until service starts. The whole visit adds the customer’s own service.
  • Waits climb steeply as staff get busier, so a little more demand near capacity can multiply the queue.
  • Very uneven jobs and urgent cases push waits above the simple sum, and rushes break the averages. People giving up pull the other way.

How do you calculate one customer’s wait time in a queue?

Start from the queue in front of them, not from the day’s averages. In research on call centres, the estimate based on queue length is the standard one, commonly used in practice. It needs two inputs: how many people are waiting, and the average gap between one person finishing and the next.

That gives the counter sum, in four steps:

  1. Count the people waiting, leaving out anyone already at the counter.
  2. Add one.
  3. Multiply by the average service time.
  4. Divide by the number of staff serving from that one line.

The average service time divided by the number of staff is that average gap. With two staff each taking 3 minutes, as at the post office above, someone finishes every minute and a half. Without an average service time to hand, timing your own service from the till or a clipboard will give you one.

The “add one” is easy to mistake for the person being served. In a 2016 conference paper on predicting waits, it is the one extra finish needed to free a server. Each person ahead needs a member of staff to come free, then one more must come free for the newcomer. With one person serving, that’s the customer at the counter now. With several, it’s whichever finishes first.

Service time means every minute a member of staff is tied up with one customer, not only the talking. HMRC’s definition of call handling time includes talking, time on hold and the wrap-up after the call. At a counter, wrap-up is the card machine, the receipt, the bag and any note on the file. Where an order passes through several stations, use the time at the station the work piles up behind.

The counter sum leans on the model behind it: all staff busy, one shared line, a fixed staff count, service lengths that vary at random, and nobody leaving. Random lengths matter because they mean time already spent at the counter says nothing about time left. That’s what lets a half-finished customer count as a full service.

Waiting in the queue is not the whole visit

The counter sum gives the wait until service starts, and the customer’s own service comes on top. A transport engineering textbook keeps the two apart. Its average wait time “excludes service time”, while its average total delay is the wait plus the service.

Larson and Odoni’s operations research textbook calls the same pair “the time spent waiting in the queue” and “the system occupancy time”. It writes the first as Wq and the second as W.

Back at the invented post office, the 7.5 minutes is the wait. Add that customer’s own 3 minutes at the window, and they spend about 10.5 minutes between joining the queue and walking away.

The difference is a whole service time: small for a book of stamps, but possibly longer than the wait itself for a phone contract. So when a formula or a screen offers one wait figure, check which it means.

Why does the wait jump when you’re only a little busier?

Because the queue wait doesn’t grow in step with how busy staff are. With one person serving, it grows with busy time divided by idle time, a ratio that climbs faster and faster as spare time runs out. Near full stretch, a few more customers an hour can more than double the wait.

How busy staff are has a technical name, utilisation. A presentation to an NHS analysts’ network on hospital bed occupancy defines it as the mean arrival rate divided by the number of servers times the mean service rate.

ρ = λ/(sμ)

Here ρ (rho) is how busy staff are, and λ (lambda) is customers arriving per hour. The s is the number of staff, and μ (mu) is how many customers one of them serves in an hour.

With one person serving, and arrivals and service lengths both random, the queue wait is a simple multiple of the average service time. This is the busyness sum. The multiple depends only on how busy that one person is:

How busy the one person serving isWait in the queue, in service times
50%1
60%1.5
70%2.3
80%4
90%9
95%19

Source: computed for this article from Kingman’s approximation, as restated in a 2026 preprint by Bartolucci and Vivo, for one server with random arrivals and random service. These are worked-out values, not measurements.

At half busy, idle time equals busy time, so a customer waits about one service time. At nine-tenths busy, the multiplier is nine.

The same paper states the limit: “The queue is stable only if ρ<1”. Staff must be idle at least some of the time, and as they get close to that line, “the waiting time diverges”. Beyond it, Larson and Odoni note, “no steady state is ever reached”. Near the line, the queue gets harder to predict as well as longer.

Take an invented café to see what that does. One person serves. Orders arrive at random, and their lengths vary at random around an average of 2 minutes, so the most they can manage is 30 an hour. At 24 orders an hour they’re four-fifths busy, the multiplier is 4, and the queue wait averages 8 minutes.

At 27 an hour, only an eighth more customers, they’re nine-tenths busy. The multiplier jumps to 9, and the average wait to 18 minutes.

Uneven jobs lengthen the queue

The table assumes the amount of spread that comes with fully random arrivals and service lengths. Real counters can have more spread than that, or less. More lengthens the wait, even at the same busyness, and less shortens it. Kingman’s approximation, as Bartolucci and Vivo restate it, puts busyness, spread and service time into one line:

Wq ≈ ρ/(1−ρ) × (ca² + cs²)/2 × E[S]

Wq is the average wait before service begins, leaving out the customer’s own service. The authors label it G/G/1, shorthand for one server with any distribution of arrival gaps and service lengths. ρ/(1−ρ) is the multiplier from the table. E[S] is the average service time, with S meaning, in the authors’ words, “the amount of calendar time for which the server is occupied by one task”.

The middle term is the spread. It uses the coefficient of variation, which is the standard deviation divided by the average. ca² is that measure, squared, for the gaps between arrivals, and cs² is the same for service lengths.

With random arrivals ca² equals one, and random service lengths give one too, so the middle term is one and the formula falls back to the table. Fully random service lengths have a standard deviation equal to their average. A counter whose jobs vary less than that has shorter waits than the table shows. One mixing quick sales with the odd very long job has longer ones. That presentation, from March 2024, calls the whole idea the VUT relationship: variability, times utilisation, times time.

So two counters with the same average service time and the same busyness can have very different queues. A 2024 queueing study of nurse staffing in neonatal intensive care uses the same relationship in its model: more variable arrivals and service meant longer waits.

Booking an occasional slow job into a set slot narrows the range of service lengths on the main line. That shrinks the middle term without adding anyone to the rota.

With more than one person serving

Dividing by the number of staff works only while all of them are busy and serving one shared line. If anyone is free, there’s no wait to work out. With a separate line at each till, each line is its own one-server queue, so its sum divides by one. Choosing one shared line or a line per till decides which version applies.

Each extra member of staff helps, but by less than the one before. The neonatal staffing paper restates this as a standard result of queueing models.

HMRC publishes two of the inputs for its phone lines. In September 2026, advisers handled 1.410 million calls, at an average handling time of 16 minutes 37 seconds.

Multiplied together, as worked out for this article, those calls took about 390,000 hours of adviser time in one month. That’s the workload. A line needs at least that much adviser time to keep up. The table can’t say what slack above that is worth here, because it’s built for one person serving.

HMRC’s quarterly performance data doesn’t say how many advisers were on the phones, so neither their busyness nor the wait a formula would predict can be worked out from outside. A counter is easier: its owner knows how many people are serving.

Where the sum stops working

Four things break the arithmetic: rushes, people giving up, urgent cases served first, and groups.

Rushes break the busyness sum first, because it describes a long-run average. Larson and Odoni describe conditions “changing too rapidly over time for the steady-state expressions” to apply. When arrivals outrun the staff for a stretch, “no steady-state expressions exist” at all.

A staffing paper by Green, Kolesar and Whitt notes that standard queueing theory centres on long-run, steady patterns. Where service times are short and the service target is high, they found plain steady-state analysis needs relatively little change. At a quick counter aiming for short waits, that points to working hour by hour, using each hour’s own arrivals.

The counter sum copes better with a rush, because it counts the people actually in front. Customers arriving behind someone can’t slow that person down. But a rush is when the other assumptions give way, so the count needs redoing often.

People giving up break the counter sum. The 2016 paper’s model assumes no abandonment, which it glosses as “infinite customer patience”. When people ahead leave, everyone behind them reaches the counter sooner, so the sum runs long. Research by Ibrahim and Whitt on overloaded queues where customers abandon works out how the estimate has to change.

Urgent cases served first break the count itself. Larson and Odoni’s section on priority queues describes preemptive systems, where urgent users are never kept waiting for the sake of lower-priority ones. So everyone else waits longer than the count in front of them suggests. The number ahead of an ordinary customer can grow while they stand still.

Groups need their own count. A couple buying one set of tickets is one service, and a friend keeping someone company at the counter is none. Count transactions, not heads.

Which sum to use, and when

It depends on the decision being made. For someone joining the queue right now, use the counter sum. It needs only a count and an average service time, and it holds however busy the day is. The conditions are that all staff are busy on one shared line, nobody leaves, and nobody is served out of turn. It gives an average, and some people in that spot will wait longer, which matters when choosing which figure to give the customer.

For deciding when to add staff, use the busyness sum, one hour at a time. Arrivals per hour against what the staff can serve shows how busy they’ll be. With one person serving, the table turns that into a wait. With more, the table’s numbers don’t apply, but the warning does. An hour close to full stretch is the hour to add someone, or to move slow jobs into booked slots.

Within a rush, when arrivals outrun the staff, the busyness sum has nothing to say. The counter sum still works, but only as well as the last count, so redo it every few minutes.

Frequently Asked Questions

Is there a safe level of busyness to aim for?

No single level is safe everywhere. The bed-occupancy presentation to NHS analysts puts it bluntly: “Occupancy is not a target”. The right level depends on how much variation the system has to absorb. Hospitals show the gap between advice and practice. In no month of 2024/25 did general and acute bed occupancy in English hospitals fall below 91.4%. The Royal College of Emergency Medicine recommends a ceiling of 85%, as the Institute for Government reports.

Does the formula work for appointments?

Partly. Booked arrivals are less random than walk-ins, so the arrival spread in Kingman’s formula is smaller and the predicted queue shorter. Lateness and overruns put spread back in, because an early or late arrival, or a slot that runs long, is unevenness of exactly the kind the formula counts.

Can you use the same sum for a phone queue?

Yes, and a phone system makes the hardest input easy, because it already knows how many callers are holding. The catch is what to divide by. Use the advisers actually answering those calls, not everyone logged in. Someone on a break, in training or on another line isn’t serving that queue, so counting them makes the wait look shorter than it will be.

Sources

About Ali Shahi

Ali Shahi is the founder and director of Upd8All Limited. He is a systems engineer with more than three decades in software, including nearly eight years building public key infrastructure systems. He started Upd8All in 2022 after watching customers crowd a fast-food counter waiting for orders, with no way to know when theirs was ready.