contact-center-math

Erlang C in JavaScript: agents, service level and ASA

Erlang C answers the core staffing question for a queue: how many agents do you need on the phones so that a target share of contacts is answered within a target time?

The formulas

A    = λ × h ÷ I                       traffic intensity (Erlangs)
P_W  = N·B ÷ (N − A·(1 − B))            probability a contact waits (B = Erlang B)
SL   = 1 − P_W · e^(−(N − A) · T ÷ h)   share answered within T seconds
ASA  = P_W · h ÷ (N − A)                average speed of answer
sched = ⌈N ÷ (1 − shrinkage)⌉           agents to schedule

λ is contacts per interval, h is average handle time, I is the interval length and N is agents on the phones. The library computes Erlang B with the recursion B = A·B ÷ (j + A·B) instead of AN ÷ N!, so it stays numerically stable for thousands of agents. The tests check it against the factorial form and the published A = 10 table.

Worked example

100 contacts in a 30-minute interval, a 3-minute AHT, a target of 80% answered in 20 seconds and 30% shrinkage.

  • Traffic: 100 × 180 ÷ 1800 = 10 Erlangs.
  • 13 agents reach 79.56%, just short of the target. Rounding traffic up "plus a few" is how teams end up one agent short.
  • 14 agents on the phones give 88.8% in 20 s, an ASA of 7.8 s and 71.4% occupancy.
  • After 30% shrinkage, schedule 20 agents.
AgentsP(wait)Service levelASAOccupancy
1244.9%64.0%40.4 s83.3%
1328.5%79.6%17.1 s76.9%
1417.4%88.8%7.8 s71.4%
1510.2%94.1%3.7 s66.7%
165.7%97.1%1.7 s62.5%

Try it

Usage

import { solveStaffing, staffingTable } from 'contact-center-math/erlang-c'

const r = solveStaffing({
  contacts: 100,      // per interval
  intervalSec: 1800,  // 30 minutes
  ahtSec: 180,
  targetSL: 0.8,      // 80% …
  targetSec: 20,      // … in 20 seconds
  shrinkage: 0.3,
  maxOccupancy: 0.85, // optional cap
  aiContainment: 0,   // share of contacts an AI agent fully handles first
})
// r.agents, r.scheduled, r.serviceLevel, r.asaSec, r.occupancy, r.feasible

Things Erlang C gets wrong

  • No abandonment. Erlang C assumes every caller waits forever. Real queues lose impatient callers, so it slightly overstaffs busy intervals. Erlang A models abandonment if you have patience data.
  • Steady arrivals inside the interval. Use 15- or 30-minute intervals, not daily totals. A daily average hides the peaks that drive staffing.
  • Chat concurrency. An agent working three chats is not three agents. Model chat separately.
  • AI containment changes the mix. If an AI agent resolves the easy contacts, the remaining contacts usually have a longer AHT. Re-measure AHT after rollout instead of keeping the old figure.

Further reading