Call Center Staffing (Erlang C)

Erlang C call center staffing calculates the number of agents needed to achieve target service levels by modeling call arrival rates, handle times, and queueing behavior. ## Concept Overview Call centers balance service quality against labor costs. More agents mean shorter wait times but higher costs. Erlang C quantifies this tradeoff—given call volume and handle time, how many agents achieve target service levels? Traffic intensity (Erlangs) is the fundamental measure: call arrival rate × handle time. 120 calls/hour with 5-minute handles = 10 Erlangs. At minimum, 10 agents are needed just to keep up. More are needed for service levels since not all agents are always free. Service level typically targets 'X% of calls answered in Y seconds' (e.g., 80/20). Erlang C calculates the probability of waiting, enabling service level prediction for any agent count. ## Key Variables & Intuition • **Calls per hour** — Inbound call volume (forecast) • **Average handle time** — Talk time + after-call work • **Service level target** — Percentage of calls answered in target time • **Target answer time** — Time threshold for service level (e.g., 20 seconds) • **Shrinkage** — Percentage of scheduled time unavailable (breaks, training, absence) • **Occupancy** — Percentage of time agents are on calls ## Assumptions • Call arrivals follow Poisson distribution (random) • Handle times are exponentially distributed • Callers have infinite patience (don't abandon) • Single queue, homogeneous agents • First-come-first-served routing ## Limitations & Edge Cases • **Call abandonment** — Real callers hang up; Erlang C overestimates wait tolerance • **Variable handle times** — Long calls aren't exponentially distributed • **Multi-skill routing** — Skills and priorities add complexity • **Intraday variation** — Single-hour calculations miss peaks within hours • **Blended channels** — Email/chat handled alongside calls breaks assumptions **Scenario:** Erlang C recommends 15 agents for 80% service level. But actual data shows 20% of callers abandon if wait exceeds 30 seconds. With abandonment, you might achieve 80% SL with 14 agents—those who would have waited are gone. Erlang C overcounted needed staff because it assumed infinite patience. ## Interpretation Guide **Service Level Targets:** - 90/10: Premium support (expensive) - 80/20: Industry standard - 70/30: Cost-conscious operations - 60/60: Minimum acceptable **Occupancy Targets:** - Under 70%: Overstaffed - 70-80%: Efficient - 80-85%: Optimal - Over 85%: Burnout risk - Over 90%: Unsustainable ## Practical Tips • **Forecast accurately** — Garbage in, garbage out; invest in forecasting • **Calculate by interval** — Not just hourly; 15-30 minute granularity • **Account for shrinkage realistically** — Track actual shrinkage data • **Monitor real-time** — Adjust intraday as conditions change • **Balance metrics** — Service level, occupancy, handle time, cost • **Use simulation for complexity** — Multi-skill, omnichannel needs more than Erlang • **Validate against actuals** — Compare predictions to real outcomes • **Plan for peaks** — Size for peaks, flex for troughs • **Cross-train agents** — Skill flexibility improves scheduling efficiency • **Consider customer value** — VIP customers may warrant better service levels ## Common Mistakes • **Using daily averages** — Missing peak intervals that need more staff • **Ignoring shrinkage** — Scheduling without shrinkage buffer • **Targeting 100% SL** — Impossibly expensive; diminishing returns • **Forgetting occupancy** — High SL with high occupancy isn't sustainable • **Static scheduling** — Not adjusting for known variation (Mondays, post-holiday) • **Neglecting handle time** — Reducing handle time can be cheaper than adding staff • **Assuming perfection** — Agents get sick, systems fail; build contingency • **Over-relying on formula** — Erlang C is a starting point, not final answer • **Ignoring agent experience** — New agents have higher handle times • **No feedback loop** — Not learning from forecast vs. actual differences ## When NOT to Use • **Low volume** — Under 10 calls/hour, statistical assumptions break down • **High abandonment** — If many callers hang up, Erlang C overestimates • **Complex routing** — Multi-skill, priority-based routing needs simulation • **Non-phone channels** — Chat/email have different workload patterns

Call center staffing evolved from intuitive scheduling through Erlang formulas to sophisticated workforce management systems that optimize service levels against labor costs in real-time. ## Origins & Why It Emerged A.K. Erlang, a Danish mathematician working for Copenhagen Telephone Company (1909), developed queueing theory to determine how many telephone circuits were needed for traffic loads. His formulas predicted blocking probability given traffic intensity. Call centers emerged in the 1960s-1970s as businesses centralized customer service. Manual scheduling was guesswork—either overstaffed (expensive) or understaffed (poor service). Erlang's mathematics offered scientific staffing. Erlang C specifically modeled queued calls—callers waiting on hold rather than receiving busy signals. This matched call center reality where customers waited in queue for available agents. ## Evolution in Practice Early adopters (1980s) used Erlang C calculations manually or with basic software. Workforce Management (WFM) systems productized the math, adding forecasting, scheduling, and real-time adherence monitoring. Forecasting improved. Historical patterns, marketing calendars, and seasonal factors fed models predicting call volumes by 15-30 minute intervals. Erlang C converted forecasts to staffing requirements. Simulation supplemented analytical models. As call centers became multi-channel (phone, email, chat) with skill-based routing, pure Erlang C became insufficient. Discrete event simulation modeled complex routing and blended operations. ## Modern Usage Today Modern WFM platforms (NICE, Verint, Genesys) embed Erlang C alongside machine learning forecasting. Real-time dashboards show service levels, enabling intraday adjustments. Omnichannel complexity challenges traditional Erlang. Agents handling phone, chat, and email simultaneously need different models. Some platforms use modified Erlang or simulation for blended work. AI is entering workforce optimization. Predictive models adjust staffing recommendations based on factors Erlang can't capture: agent skill levels, customer sentiment, handle time patterns. Human + AI collaboration optimizes better than either alone. ## Common Misconceptions • **Erlang C is exact** — It's a model with assumptions that don't perfectly match reality • **Service level is the only metric** — Occupancy, handle time, and cost matter too • **More agents always improve service** — Diminishing returns; 95% SL costs far more than 80% • **Shrinkage is fixed** — Shrinkage varies by time of day and can be managed • **One calculation covers all day** — Call patterns vary; recalculate by interval