Square-Root Staffing

356. Square-Root Staffing

A simple rule of thumb for sizing call-center / service-system capacity. Halfin-Whitt asymptotic regime (1981).

356.1. The rule

Given offered load $𝑎 = \frac{𝜆}{𝜇}$ Erlangs, optimal number of servers is approximately:

𝑐 \approx 𝑎 + 𝛽 \sqrt{𝑎}

— a base of $𝑎$ servers (deterministic capacity) plus a safety margin $𝛽 \sqrt{𝑎}$ that grows as $\sqrt{𝑎}$ .

356.2. Three quality regimes (parameterized by $𝛽$ )

Efficiency-driven ( $𝛽 < 0$ ): servers saturated, customers wait often. Wait time scales as $\frac{1}{\sqrt{𝑎}}$ .
Quality-driven ( $𝛽 > 2$ or so): servers idle often, customers rarely wait. Wait time scales as $𝑎^{- \frac{1}{2}}$ with very small constant.
Quality-and-Efficiency-Driven (QED, $𝛽 \approx 0.5 - 1.5$ ): balanced — wait probability and idle probability both moderate.

For typical call centers targeting “80% of calls answered within 20 sec”, $𝛽 \approx 0.7 - 1.0$ .

356.3. Why $\sqrt{𝑎}$ ?

In the heavy-traffic limit, queueing behavior is governed by central limit-style fluctuations. The natural fluctuation scale of an offered load $𝑎$ is $\sqrt{𝑎}$ — same as the standard deviation of a Poisson( $𝑎$ ) distribution.

If you provision exactly $𝑎$ servers, fluctuations of size $\sqrt{𝑎}$ overwhelm capacity routinely. If you provision $𝑎 + \sqrt{𝑎}$ , you absorb typical fluctuations. Bigger $𝛽$ → more safety.

356.4. Example

Call center: 50 calls/min, average 3 min per call → $𝑎 = 150$ Erlangs.

Strategy	$𝛽$	$𝑐 = 𝑎 + 𝛽 \sqrt{𝑎}$	Behavior
No safety	0	150	50% chance customer waits
Modest	0.5	156	30% wait probability
Quality	1.5	168	10% wait probability
Premium	2.5	181	< 3% wait probability

(Wait probabilities approximate, depend on abandonment, service distribution, etc.)

356.5. Practical use

Estimate offered load $𝑎 = \frac{𝜆}{𝜇}$
Choose target service level (e.g., 80% answered in 20 sec)
Look up corresponding $𝛽$ (or simulate to calibrate)
Staff $𝑐 \approx 𝑎 + 𝛽 \sqrt{𝑎}$

Compare to Erlang C / Erlang A exact calculations — typically very close.

356.6. Why it’s the right scale

For a fixed fraction-of-time-waiting target (constant probability of waiting), $𝛽$ stays constant as offered load grows. So if you double traffic ( $𝑎$ doubles), safety margin only grows as $\sqrt{2} \approx 1.4$ x — capacity grows by $\sqrt{𝑎}$ less than linearly.

This is the economies of scale of pooling: big call centers need proportionally less slack capacity than small ones.

356.7. See also

Erlang C — exact wait probability
Erlang B — no queue / blocking
Erlang A — with abandonment
Sakasegawa — G/G/c approximation