Traffic Intensity and Server Utilization

Traffic intensity (ρ) is the ratio of offered load to service capacity. It is the single number that determines whether a queueing system is stable, how congested it is, and how service quality degrades as load increases. In workforce management, traffic intensity manifests as occupancy — the fraction of time agents spend handling contacts — and drives the fundamental tradeoff between efficiency and service level.

Overview

Every operations leader has seen the pattern: occupancy is 75% and service level is comfortable. Push occupancy to 85% and service level holds. Push to 90% and service level starts to slip. Push to 95% and wait times explode. The relationship between occupancy and service level is not linear — it is convex and steep near capacity. Traffic intensity is the mathematical framework that explains why.

The concept connects directly to practical WFM decisions: how many agents to schedule, what occupancy target to set, when to trigger overtime, and why small changes in volume near capacity have outsized effects on customer experience.

Mathematical Foundation

Definitions

Traffic intensity for a multi-server system:

$\rho =\frac{\lambda }{\mu \cdot s}=\frac{A}{s}$

where:

• λ = arrival rate
• μ = service rate per agent (= 1/AHT)
• s = number of servers (agents)
• A = offered load in Erlangs (= λ/μ = λ × AHT)

Server utilization is synonymous with traffic intensity in standard queueing theory: it is the average fraction of capacity in use.

Occupancy in WFM practice equals ρ when agents handle a single queue:

$\text{Occupancy}=\frac{A}{s}=\rho$

Occupancy is usually expressed as a percentage: ρ = 0.85 → 85% occupancy.

The Stability Condition

A queueing system is stable (the queue does not grow without bound) if and only if:

$\rho <1$

If ρ ≥ 1, the arrival rate meets or exceeds the total service capacity. The queue grows indefinitely and no steady-state exists. In practice, this means:

• If offered load is 133.3 Erlangs and you have 133 agents, ρ = 133.3/133 = 1.002. The system is unstable. Wait times will grow continuously throughout the interval.
• If you have 134 agents, ρ = 133.3/134 = 0.995. The system is technically stable but at 99.5% occupancy — wait times will be extreme.
• At 155 agents, ρ = 133.3/155 = 0.860. The system operates at 86% occupancy with manageable wait times.

The gap s − A (agents minus offered load in Erlangs) represents the spare capacity available to drain the queue. The smaller this gap, the longer queues persist.

Relationship to Erlang C

The probability that an arriving caller must wait (Erlang C) is a function of ρ and s:

$P(W>0)=C(s,A)=\frac{\frac{A^s}{s!}\cdot \frac{1}{1-\rho }}{\sum _{k=0}^{s-1}\frac{A^k}{k!}+\frac{A^s}{s!}\cdot \frac{1}{1-\rho }}$

As ρ → 1, the term 1/(1−ρ) → ∞, and C(s, A) → 1: every caller waits. The expected wait time:

$E[W]=\frac{C(s,A)}{\mu (s-A)}=\frac{C(s,A)}{\mu \cdot s\cdot (1-\rho )}$

The (1 − ρ) in the denominator is the key: as ρ approaches 1, expected wait time grows hyperbolically — not linearly, not even exponentially, but as 1/(1−ρ). A system at 90% occupancy has 1/(1−0.9) = 10 times the congestion factor of a system at 90% occupancy with a reference of zero congestion. At 95%: factor of 20. At 98%: factor of 50.

The Occupancy-Wait Time Curve

For a given number of agents, the relationship between occupancy and ASA follows a convex curve:

Occupancy (ρ)	Relative ASA	Interpretation
70%	1.0×	Comfortable — short waits, agents have breathing room
80%	1.8×	Standard target — acceptable waits, reasonable efficiency
85%	2.8×	Common target — service level starts to tighten
90%	5.0×	High efficiency — service level at risk
92%	7.5×	Red zone — service level degradation visible
95%	15×	Crisis — extended waits, heavy abandonment
98%	50×	Breakdown — queue spiraling

(Relative values are illustrative for a representative multi-server system. Exact multipliers depend on s and the arrival rate.)

This nonlinearity is the most important insight in WFM capacity planning. It explains:

• Why adding 5 agents at 95% occupancy has a larger SL impact than adding 5 agents at 80%.
• Why small forecast errors near capacity produce disproportionate service failures.
• Why the "last few percent" of occupancy is never worth pursuing.

WFM Applications

Worked Example: Occupancy Target Setting

Scenario: 133.3 Erlangs of offered load, AHT = 240 seconds, target SL = 80/20.

Using Erlang C, the minimum agents for 80/20 SL:

Agents (s)	Occupancy (ρ)	P(wait > 0)	SL (80/20)	ASA (sec)
140	95.2%	0.84	51%	54
145	91.9%	0.64	67%	29
148	90.1%	0.52	75%	21
150	88.9%	0.44	80%	17
155	86.0%	0.29	89%	9
160	83.3%	0.18	94%	5
170	78.4%	0.06	99%	1

Key observations:

• From 155 to 150 agents (removing 5): occupancy rises 3.1 points but SL drops 9 points.
• From 150 to 145 agents (removing 5): occupancy rises 3.0 points but SL drops 13 points.
• From 145 to 140 agents (removing 5): occupancy rises 3.3 points but SL drops 16 points.

Each increment of occupancy costs more service level than the last. This is the convexity at work.

The Occupancy-Service Level Tradeoff

Operations leaders face a fundamental tradeoff:

• Higher occupancy → lower labor cost per contact, higher agent burnout, worse service level
• Lower occupancy → higher labor cost per contact, more agent idle time, better service level

The "right" occupancy depends on the operation's value equation. Common targets:

Context	Typical Occupancy Target	Rationale
High-value voice (retention, sales)	78–82%	Service quality directly affects revenue
Standard inbound voice	83–87%	Balance of efficiency and SL
Asynchronous digital (email, ticket)	90–95%	No real-time queue; processing delays acceptable
Chat (concurrent sessions)	Effective ρ varies	Concurrency changes the math; 2–3 concurrent chats at 85% is different from single-thread
Blended inbound/outbound	88–92%	Outbound fills idle time, raising effective occupancy

Why 85% Is Common

There is no mathematical theorem that makes 85% optimal. It is a heuristic that balances:

• Agent well-being: sustained occupancy above 88–90% correlates with increased burnout and attrition in most studies.
• Service level: for typical contact center sizes (50–200 agents) and traffic intensities, 85% occupancy usually lands near 80/20 SL.
• Financial efficiency: each percent of occupancy below 85% adds roughly 1–2% to labor cost for marginal SL improvement.

The specific optimum depends on the operation's cost of a bad service experience vs. cost of an incremental agent-hour. Probabilistic Planning provides the framework for making this tradeoff explicit.

Occupancy vs Utilization vs Productivity

These terms are often conflated. Precision matters:

Term	Definition	Denominator
Occupancy	Time handling contacts / time available for contacts	Logged-in time minus shrinkage
Utilization	Time handling contacts / total paid time	Total paid time (including breaks, meetings, training)
Productivity	Output (contacts handled, revenue) / input (hours, cost)	Varies by business definition

Occupancy of 85% with 30% Shrinkage yields utilization of 85% × 70% = 59.5%. A planner targeting 85% occupancy is targeting 85% of available time, not 85% of paid time.

Common Misconceptions

1. "Occupancy and utilization are the same thing."

See table above. Conflating them leads to either overstating efficiency (if occupancy is called utilization without adjusting for shrinkage) or understating it (if utilization is used where occupancy is needed for Erlang calculations).

2. "Higher occupancy is always better."

Higher occupancy is more efficient in a narrow cost-per-contact sense. But the convex relationship between occupancy and wait times means each additional percent of occupancy costs disproportionately more in service quality. There is always a point beyond which the next percent of occupancy destroys more value (customer experience, agent retention) than it saves (labor cost).

3. "If occupancy is 85%, 15% of agent time is wasted."

The 15% idle time is not waste — it is the capacity buffer that enables acceptable wait times. Eliminating it would push ρ toward 1 and cause service collapse. The idle time between calls is the price of the service level guarantee. See Power of One.

4. "Occupancy targets should be the same for all interval sizes."

For the same offered load and agent count, shorter intervals exhibit more variance. A 15-minute interval with 85% planned occupancy will experience more SL volatility than a 30-minute interval at the same occupancy, because the Poisson variance is larger relative to the mean for shorter periods. Occupancy targets may need to be slightly lower for operations planned at 15-minute granularity.

5. "We can run at 95% occupancy if we use Erlang-A."

Erlang-A accounts for abandonment, which does relieve queue pressure. But high abandonment at 95% occupancy is not a staffing success — it is demand destruction. Erlang-A provides a more accurate prediction of what will happen at high occupancy; it does not make high occupancy acceptable.

Maturity Model Position

• Level 1 — Initial. Occupancy not tracked or conflated with utilization. Staffing targets set by headcount rules, not capacity math.
• Level 2 — Foundational. Occupancy calculated and monitored. 85% used as default target. Practitioners understand that higher occupancy means longer waits but may not appreciate the nonlinearity.
• Level 3 — Progressive. Occupancy target varies by queue, channel, and business value of the interaction. The convex occupancy-SL tradeoff is understood and used to justify differential staffing investments. Erlang C outputs validated against observed occupancy.
• Level 4 — Advanced. Occupancy modeled as a distribution across intervals, not a point target. The cost curve (labor cost vs. experience cost as a function of ρ) is estimated and used for optimization. Simulation used where Erlang models are insufficient. Intraday occupancy management triggers real-time adjustments.
• Level 5 — Pioneering. Dynamic occupancy targets adjusted in real time based on current queue state, agent fatigue models, and economic value of marginal service improvement. Closed-loop optimization balances ρ across all queues simultaneously.

See Also

• Erlang C
• Erlang-A
• Erlang B
• Occupancy
• Service Level
• Little's Law Applied to WFM
• Offered Load vs Carried Load
• Queueing Theory Fundamentals
• Pooling Theory
• Power of One
• Shrinkage

References

• Gross, D. & Harris, C.M. (2008). Fundamentals of Queueing Theory, 4th ed. Wiley. Chapters 2–3.
• Gans, N., Koole, G. & Mandelbaum, A. (2003). "Telephone Call Centers: Tutorial, Review, and Research Prospects." Manufacturing & Service Operations Management 5(2), 79–141.
• Koole, G. (2013). Call Center Optimization. MG Books. Chapter 4: Staffing.
• Green, L., Kolesar, P. & Whitt, W. (2007). "Coping with Time-Varying Demand When Setting Staffing Requirements for a Service System." Production and Operations Management 16(1), 13–39.
• Whitt, W. (1992). "Understanding the Efficiency of Multi-Server Service Systems." Management Science 38(5), 708–723.