Last updated: 31 October 2026.
Little’s Law states that, for a stable system, the average number of items inside it (L) equals the average arrival rate (λ) multiplied by the average time each item spends there (W): L = λW. It is one of the few exact results in queueing theory, proved in 1961 by John D. C. Little, and it holds under very general conditions. That generality is its strength and its limit. The law tells you how three averages must relate, but not why waiting times behave as they do. For that you need the second idea in this article: as utilisation approaches 100%, delays grow non-linearly. Together they explain why systems that look fine at 80% busy are fragile at 95%, why work in progress is a lever for speed, and why a small increase in load can produce a large increase in waiting.
Key takeaways
- L = λW links work in progress, throughput and cycle time. If throughput is fixed, lowering cycle time requires lowering work in progress.
- In the simplest random queue (M/M/1), the time in the system is 1/(1-ρ) service times, where ρ is utilisation. That is 2x at 50%, 5x at 80%, 10x at 90%, 20x at 95% and 100x at 99%.
- Variability, not just average load, drives waiting. Less variable arrivals and service reduce delays at the same utilisation.
- Running a system near full utilisation trades resilience for apparent efficiency; spare capacity is what keeps queues short.
The law and what it needs
Little’s Law applies to any stable system, however complicated inside: a bank queue, a hospital ward, a software pipeline, a factory floor. Stable means that over the long run, the number of items entering equals the number leaving, so the system is not growing without bound. If that holds, then:
L = λ × W
- L: average number of items in the system (work in progress, inventory, people in the queue and being served).
- λ: average arrival rate, which in a stable system equals the throughput.
- W: average time an item spends in the system, from entry to exit (cycle time or lead time).
The law does not assume any particular arrival pattern, service distribution or queue discipline, which is why it works in so many settings. Its practical form is usually a rearrangement: W = L / λ. If a team has 30 tasks in progress and finishes 6 a week, the average time a task takes from start to finish is 5 weeks. To cut that to 2.5 weeks without raising throughput, halve the work in progress.
Care is needed with definitions. The items counted must be the same ones whose time is measured, the boundaries of the system must be consistent, and the averages must cover a period long enough for the system to be near steady state. Applied across a period during which the backlog is growing, the law does not hold.
Why utilisation matters: the hockey stick
Little’s Law relates averages but does not tell you W. To learn how W depends on load, use a queueing model. The simplest is M/M/1: random (Poisson) arrivals, random (exponential) service times and a single server. For this model, if ρ is the fraction of time the server is busy (arrival rate divided by service rate), the average time in the system is:
W = (1/μ) / (1 – ρ)
where 1/μ is the mean service time. In other words, the time in the system is the service time multiplied by 1/(1-ρ). That factor is gentle at low load and then rises sharply.
Figure 1. Time in system as a multiple of the service time in an M/M/1 queue. The curve is flat at low utilisation and rises steeply beyond 80%. Dots mark the values simulated for this article.
| Utilisation ρ | Theory: time in system | Simulation (one run) |
|---|---|---|
| 50% | 2x | 2.02x |
| 80% | 5x | 5.03x |
| 90% | 10x | 10.32x |
| 95% | 20x | 22.63x |
| 99% | 100x | 118.39x |
The simulation is a single seeded run of an M/M/1 queue, so it matches theory closely at low load and drifts at high load, where the system takes very long to settle and estimates are noisy. That noise is itself informative: near saturation, outcomes become hard to predict.
Combine the two results. At 50% utilisation, an item spends twice its service time in the system. Raising load to 90% multiplies the wait by five, and then to 95% by two more. Because L = λW, the number of items in the system also explodes: at 90% there are 9 items on average (ρ/(1-ρ)), at 99% there are 99.
The role of variability
The M/M/1 numbers are a baseline, not a rule. Kingman’s approximation for a single-server queue with general arrivals and service shows that the waiting time scales with three factors: a utilisation term ρ/(1-ρ), a variability term (the squared coefficients of variation of arrivals and service, averaged) and the mean service time. The consequence is that lowering variability gives the same improvement as lowering utilisation. Standardised work, steady arrivals, small batches and predictable service all shorten queues without any added capacity. Conversely, big batches, bursts and rare very long jobs lengthen them.
This is why pooling and batching decisions matter. Fifteen small, uniform tasks flow differently than one large task that blocks everything behind it, even when the total workload is equal.
Applications
Software and operations
Servers, thread pools and job queues sit on this curve. When utilisation rises, response times grow non-linearly, requests time out, clients retry and the retries add load. That feedback can sustain an overload after the trigger has gone, as described in metastable failures, and it can propagate to neighbours, as in cascading failures. Capacity planning that targets 90% average utilisation leaves no room for bursts.
Teams and knowledge work
For a team, W is lead time and L is work in progress. Limiting work in progress reduces lead time without reducing throughput, which is the logic behind Kanban-style limits. It is also the reason that overcommitting a team slows everything: starting more work raises L, which by Little’s Law raises W for each item at the same throughput. See the capacity trap for how this dynamic traps organisations, and backlog as a stock-and-flow problem for managing the queue itself.
Healthcare and services
Hospital beds, emergency departments and call centres behave in the same way. Occupancy above roughly 85% is often associated with delays and blocked patient flow, although the precise threshold depends on variability and the number of servers. With many parallel servers the curve stays flat for longer before turning up, which is why large pooled systems can run at higher utilisation than small ones, but the hockey stick is still present.
Connecting to stocks and flows
Little’s Law is a statement about a stock (L) and its flows. In steady state the inflow equals the outflow, and the stock is their product with the average residence time. In stocks and flows language, the residence time is the stock divided by the outflow rate, and it is how the stock sets the delay in the system. When utilisation is high, the outflow cannot rise to clear a growing stock, which is why queues once formed take long to drain.
Common mistakes
- Using it on an unstable system. If the backlog is growing, averages are not stationary.
- Mixing definitions. Counting queued items but timing only the service step gives wrong results.
- Planning at average load. Averages hide bursts; queues form from variability, not mean demand alone.
- Believing high utilisation is efficiency. It may maximise equipment usage while destroying speed and resilience.
Use the law carefully in live operations
Little’s Law is an accounting relationship for a stable flow system, not a formula that guarantees a particular improvement. Before using it, define the boundary: which requests count as arrivals, which states count as work in progress, and when the clock starts and stops. If work is abandoned, rejected or split into stages, those events must be treated consistently. Mixing a queue’s arrival rate with a different population’s lead time produces a meaningless estimate.
The relation is especially useful for making the cost of excess work visible. If a team completes 12 cases per day and carries 120 active cases on average, the implied average time in the system is 10 days. The calculation does not say which cases wait or why; it gives a coherent starting point for investigating the queue. Compare the result with observed case-level durations and inspect the distribution, since a stable average can still coexist with a long tail.
When throughput changes, avoid assuming that reducing work in progress will instantly raise completions. A team may first need to finish blocked work, resolve dependencies or improve quality. Track arrivals, completions, active work and age together. If arrivals stay above completions, the queue grows; if the flow is unstable, a snapshot of inventory divided by a recent rate may mislead. Pair the law with a time series and explicit checks for rework and seasonal variation.
For managers, the practical intervention is often to limit starts, clarify priorities and remove bottlenecks rather than ask people to multitask faster. A work-in-progress limit creates a signal when capacity is occupied and encourages finishing. Set limits with the people doing the work, then observe whether waiting time, completion rate and quality improve over several cycles.
Frequently asked questions
Does Little’s Law need a specific arrival distribution?
No. It holds for any stable system with well-defined averages, regardless of arrival or service distributions or queue discipline.
Why do queues explode near 100% utilisation?
Because there is almost no spare capacity to clear random bursts. In the M/M/1 model the delay factor 1/(1-ρ) grows without bound as ρ approaches 1.
How can a team shorten lead time without working faster?
Reduce work in progress. With throughput unchanged, Little’s Law says average time in the system falls in proportion.
Is 80% utilisation a safe target?
It is a rule of thumb, not a guarantee. The right level depends on variability, the number of servers and the cost of delay. Highly variable work needs more slack.
Conclusion
Little’s Law gives an exact accounting identity, and the M/M/1 delay curve gives the intuition that load and waiting do not move together proportionally. Together they explain why work in progress limits speed up delivery, why spare capacity is not waste, and why many systems fail suddenly after seeming fine. Measure L, λ and W together, reduce variability where possible and plan for the part of the curve where small changes matter most. The glossary has an entry for Little’s Law.
Sources and further reading
- Little, J. D. C., “A Proof for the Queuing Formula: L = λW,” Operations Research 9(3), 1961, pp. 383–387.
- Little, J. D. C. and Graves, S. C., “Little’s Law”, in Building Intuition: Insights from Basic Operations Management Models and Principles, Springer, 2008.
- Kingman, J. F. C., “The single server queue in heavy traffic,” Mathematical Proceedings of the Cambridge Philosophical Society 57(4), 1961.
- Hopp, W. J. and Spearman, M. L., Factory Physics, 3rd ed., Waveland Press, 2011.
- Little, J. D. C., “A Proof for the Queuing Formula: L = λW,” Operations Research 9(3), 1961, pp. 383–387.
- Kingman, J. F. C., “The single server queue in heavy traffic,” Mathematical Proceedings of the Cambridge Philosophical Society 57(4), 1961.

[…] Organizations Drift Into FailureNov 12Systems Thinking and Complexity: The Complete Advanced GuideNov 10Little’s Law and the Utilization Trap: Why Queues Explode Near CapacityNov 10The Duck Curve: Solar, Ramp Constraints and Grid Flexibility as a SystemNov 14Oscillation in […]