Poisson Process in Contact Centers

The Poisson process is the standard mathematical model for random arrivals in contact centers. It assumes that callers arrive independently, at a constant average rate within each planning interval, and one at a time. Nearly every staffing formula in WFM — Erlang B, Erlang C, Erlang-A — requires Poisson arrivals as an input assumption. Understanding when this assumption holds, and when it breaks, is essential for choosing the right planning tool.

Overview

A contact center's phone rings. Was that call predictable? Not individually — each call is the independent decision of one consumer among millions. But collectively, the pattern is remarkably regular: random arrivals at a statistically stable rate within short time intervals. This is precisely what the Poisson process models.

The Poisson assumption is not merely convenient. It is empirically justified for large-scale inbound consumer contact centers during intervals short enough that the arrival rate is approximately constant (typically 15–30 minutes). Decades of ACD data analysis confirm that inbound call arrivals pass standard Poisson goodness-of-fit tests within individual intervals.

The assumption breaks down in specific, identifiable circumstances — and knowing those exceptions is as important as knowing the baseline.

Mathematical Foundation

The Poisson Process

A Poisson process with rate λ (arrivals per unit time) has three defining properties:

1. Independence. The occurrence of one arrival does not affect the probability of another. Caller A's decision to dial does not influence Caller B.
2. Stationarity (within interval). The probability of an arrival in a small time window dt is λdt, independent of when dt occurs within the interval.
3. No simultaneous arrivals. The probability of two or more arrivals in an infinitesimally small window is negligible.

The Poisson Distribution

The number of arrivals N in an interval of length t follows the Poisson distribution:

$P(N=k)=\frac{(\lambda t{)}^k{e}^{-\lambda t}}{k!}$

Properties:

• Mean: $E[N]=\lambda t$
• Variance: $\text{Var}(N)=\lambda t$
• Mean equals variance. This is the diagnostic signature of a Poisson process.

Inter-Arrival Times

The time between consecutive arrivals follows an exponential distribution:

$P(T>t)=e^{-\lambda t}$

Mean inter-arrival time: $E[T]=1/\lambda$

The exponential distribution is memoryless: the probability of the next arrival in the next dt seconds is the same regardless of how long since the last arrival. This memorylessness property is equivalent to the independence assumption above.

Coefficient of Variation Diagnostic

For a Poisson process, the coefficient of variation (CV) of the number of arrivals across repeated intervals of the same length is:

$CV=\frac{\sigma }{\mu }=\frac{\sqrt{\lambda t}}{\lambda t}=\frac{1}{\sqrt{\lambda t}}$

More practically, for interval counts: variance should equal mean. The index of dispersion (variance/mean ratio) should be approximately 1. If it is significantly greater than 1, arrivals are overdispersed — more variable than Poisson predicts. If significantly less than 1, arrivals are underdispersed — more regular than Poisson.

Variance/Mean Ratio	Interpretation	Likely Cause
≈ 1	Poisson	Random independent arrivals
> 1.5	Overdispersed	Arrival rate varies across days (non-stationarity), or arrivals are correlated (word-of-mouth, marketing bursts)
< 0.7	Underdispersed	Scheduled arrivals, appointment systems, regulated inflow (e.g., IVR throttling)

When Poisson Holds

The Poisson model is appropriate when:

• Large customer base. Each customer has a small, independent probability of calling. The superposition of many independent rare events converges to Poisson (a direct application of the Poisson limit theorem).
• Short intervals. The arrival rate can be treated as constant. 15-minute or 30-minute intervals are standard.
• Inbound, consumer-initiated contact. The timing is driven by individual need, not coordination.
• No queue feedback. Callers' decisions to call are not influenced by current queue conditions. (This fails when IVR announces expected wait time — see below.)

Typical environments where Poisson holds well:

• Large inbound voice queues (consumer banking, telecom support, insurance claims)
• Chat initiation on high-traffic websites
• Email/ticket arrivals (over intervals of hours)

When Poisson Breaks

Scheduled Arrivals

Appointments, callbacks at promised times, and scheduled outbound callbacks produce arrivals that are clustered at specific times rather than random. The inter-arrival times are not exponential; they are concentrated around the schedule. Arrivals are underdispersed.

Impact: Erlang formulas overestimate required staffing because they expect random peaks that scheduled arrivals do not produce.

Callback Queues

Virtual hold / callback systems convert waiting callers into scheduled return calls. The "arrival" of the callback is determined by agent availability, not customer randomness. This creates a feedback loop between supply and demand that violates independence.

Impact: The callback stream is self-regulating — it speeds up when agents are available and slows when they are not. Erlang models are inapplicable to this stream. Simulation is required.

Outbound Campaigns

Predictive dialers generate calls to agents at controlled rates. These are deterministic or algorithmically paced, not Poisson.

Impact: Outbound blending must be modeled separately. The combined inbound + outbound stream is not Poisson even if inbound alone is.

Post-IVR Routing

An IVR system that batches callers (e.g., "press 1 for billing, press 2 for support") can create arrival patterns at downstream queues that are overdispersed relative to Poisson. When IVR navigation takes variable time, the downstream arrivals are the convolution of Poisson input with IVR traversal time — producing a smoothed, non-Poisson pattern.

Impact: Typically small for large queues but can matter for low-volume specialty queues where IVR batching creates arrival bursts.

Announced Wait Times

When IVR announces expected wait time, some callers balk (hang up before queueing) or defer (call back later). This creates a feedback loop: long waits reduce arrivals, which shortens waits, which increases arrivals. The arrival rate becomes endogenous — a function of system state.

Impact: Arrival rate is no longer independent of queue length. Steady-state Erlang assumptions weaken. For heavy-traffic periods where balk rates exceed ~10%, simulation with endogenous arrival modeling is warranted.

Small Populations

When the calling population is small (< 200 potential callers) and the call rate per person is non-trivial, the arrival process is better modeled as Engset rather than Poisson. Each call in progress removes a potential caller from the population, reducing the arrival rate.

Impact: Erlang B and C overestimate blocking/waiting because they assume an infinite population. The Engset model corrects for finite populations.

Marketing Events and Outage Surges

A TV advertisement, a service outage, or a product recall triggers a spike in arrivals that is neither Poisson nor stationary. Arrivals are correlated (many people respond to the same stimulus simultaneously) and the rate changes rapidly.

Impact: No steady-state formula applies during the transient. Use intraday reforecasting and real-time management. After the surge, if arrivals return to a stable rate, Poisson resumes.

WFM Applications

Validating the Poisson Assumption

Step 1: Extract interval-level call counts for the same interval across 20+ days (e.g., Monday 10:00–10:30 for 20 Mondays).

Step 2: Compute mean and variance of the counts.

Step 3: Calculate the index of dispersion (variance/mean). If it is between 0.8 and 1.2, Poisson is a reasonable assumption for that interval. If consistently > 1.5, the arrival rate varies across days more than Poisson allows — consider using a mixed-Poisson model or simulation.

Worked example:

Monday	Calls 10:00–10:30
Week 1	102
Week 2	95
Week 3	108
Week 4	99
Week 5	104

Mean = 101.6, Variance = 24.3. Index of dispersion = 24.3 / 101.6 = 0.24.

This is well below 1 — underdispersed. Investigate: is arrival pacing or throttling occurring? In this case, the variation is suspiciously low, suggesting a systematic constraint (e.g., IVR port limit) is capping arrivals.

Choosing the Planning Tool

Arrival Pattern	Poisson?	Appropriate Model
Inbound consumer voice, large base	Yes	Erlang B / Erlang C / Erlang-A
Chat on high-traffic site	Approximately	Erlang with adjustment for concurrency
Scheduled callbacks	No	Simulation
Outbound predictive	No	Dialer-specific pacing model
Post-outage surge	No (transient)	Real-time management, intraday reforecast
Small internal helpdesk (<100 callers)	No (finite population)	Engset model

Common Misconceptions

1. "Arrivals are Poisson over the whole day."

False. The arrival rate changes throughout the day — the intraday pattern shows peaks and troughs. Within each short interval, arrivals are Poisson with a rate specific to that interval. The overall daily process is a non-homogeneous Poisson process. Erlang formulas apply interval by interval, not to daily aggregates.

2. "If arrivals aren't exactly Poisson, Erlang formulas are useless."

Erlang formulas are reasonably robust to moderate departures from Poisson. For the call center context, small amounts of overdispersion (index of dispersion 1.0–1.3) cause only minor errors in staffing calculations. The formulas break down significantly only when arrivals are strongly non-Poisson (scheduled, feedback-driven, or bursty).

3. "Poisson means arrivals are evenly spaced."

The opposite. Poisson means arrivals are randomly spaced, producing natural clustering and gaps. Evenly spaced arrivals (deterministic) are the antithesis of Poisson. The randomness of Poisson is exactly what creates the need for staffing headroom above the offered load.

4. "We can test for Poisson using a single day's data."

A single day provides one observation per interval — impossible to compute variance. Testing requires multiple days for the same interval. Cross-interval analysis conflates rate changes with distributional properties.

Maturity Model Position

• Level 1 — Initial. Poisson assumption not explicitly considered. WFM tool applies Erlang formulas to whatever volume is entered.
• Level 2 — Foundational. Practitioners understand that Erlang formulas assume Poisson arrivals and that arrivals are modeled per interval. No explicit validation performed.
• Level 3 — Progressive. Index of dispersion checked for key queues. Non-Poisson streams (callbacks, outbound) staffed separately or modeled with simulation. Finite-population queues identified and treated with Engset where appropriate.
• Level 4 — Advanced. Arrival process validated statistically as part of forecast quality assurance. Mixed-Poisson models used where day-to-day rate variation is significant. Simulation used for complex multi-stream environments. Real-time detection of non-Poisson events triggers intraday reforecast.
• Level 5 — Pioneering. Arrival process modeling is granular — customer-segment-specific, channel-specific, and updated in near-real-time. Bayesian updating of arrival rates within the day. Non-stationary Poisson models used explicitly rather than piecewise-constant approximation.

See Also

• Erlang B
• Erlang C
• Erlang-A
• Queueing Theory Fundamentals
• Palm's Theorem and PASTA
• Traffic Intensity and Server Utilization
• Probabilistic Planning
• Demand calculation

References

• Ross, S.M. (2014). Introduction to Probability Models, 11th ed. Academic Press. Chapter 5: The Poisson Process.
• Gross, D. & Harris, C.M. (2008). Fundamentals of Queueing Theory, 4th ed. Wiley. Section 1.3.
• Gans, N., Koole, G. & Mandelbaum, A. (2003). "Telephone Call Centers: Tutorial, Review, and Research Prospects." Manufacturing & Service Operations Management 5(2), 79–141.
• Avramidis, A.N., Deslauriers, A. & L'Ecuyer, P. (2004). "Modeling Daily Arrivals to a Telephone Call Center." Management Science 50(7), 896–908.
• Brown, L., Gans, N., Mandelbaum, A., Sakov, A., Shen, H., Zeltyn, S. & Zhao, L. (2005). "Statistical Analysis of a Telephone Call Center: A Queueing-Science Perspective." Journal of the American Statistical Association 100(469), 36–50.
• Jongbloed, G. & Koole, G. (2001). "Managing Uncertainty in Call Centres Using Poisson Mixtures." Applied Stochastic Models in Business and Industry 17, 307–318.