मुख्य सामग्री पर जाएँ
संचालन 7 नवंबर 2026 10 मिनट का पठन

Erlang A vs Erlang C: Staffing With Abandonment

Erlang A vs Erlang C explained: how caller patience and abandonment change staffing, which way each model is biased, and when Erlang C is still good enough.

D
7 नवंबर 2026

त्वरित उत्तर

Erlang C assumes no caller ever hangs up while waiting. Erlang A adds average caller patience, so abandonment is modelled. A simulation study found Erlang A usually more accurate but optimistically biased, while Erlang C tends to be pessimistic, and neither dominates in every case. Use Erlang C for quick, safe-side estimates, and Erlang A or a simulation when abandonment is material.

Most call centre staffing starts with Erlang C, which Robbins describes as the most commonly used model, and for good reason: it is simple, widely implemented and needs only three inputs. But anyone who has watched a real queue knows it leaves something out. Callers lose patience and hang up. When they do, the queue shortens, waits for the callers who stay get shorter and the agents you "needed" on paper were not all needed. Erlang A is the standard extension that models this.

This guide explains the one assumption that separates the two, what extra input Erlang A needs, how to estimate it from your own data, and what the research says about which way each model errs. It includes a worked example computed for one illustrative queue, so you can see how much the answer can move.

The one assumption that separates the two models

Both models describe the same basic system: calls arrive at random, wait in a single queue and are answered by a pool of identical agents with random handling times. In the standard version, arrivals follow a Poisson process and handling times are exponential. That much is shared.

The difference is what happens to a caller who waits. In Erlang C, every caller waits for as long as it takes. Garnett, Mandelbaum and Reiman describe the family this belongs to as models that "lack a central prevalent feature, namely that impatient customers might decide to leave (abandon) before their service begins" (Garnett, Mandelbaum and Reiman, 2002, abstract). In Erlang A, each caller has a patience time. If the wait would exceed it, the caller abandons.

In queueing notation, Erlang C is the M/M/N queue. The simplest abandonment model, with exponentially distributed patience and unlimited waiting room, is M/M/N+M, the extra M standing for the patience distribution. That is the model Garnett and colleagues analyse, and they proposed calling it Erlang-A: A for abandonment, and because it interpolates between Erlang C and Erlang B (Garnett, Mandelbaum and Reiman, 2002).

What Erlang A needs that Erlang C does not: average caller patience

Erlang C needs three inputs: arrival rate, average handling time and the number of agents. Erlang A needs a fourth: the average time a caller is willing to wait. In the exponential version this is a single number, for example "callers are willing to wait 120 seconds on average".

That sounds small, but it carries a lot of weight. A short average patience means many callers leave early, the queue stays short and abandonment is high. A long average patience means the queue behaves almost like Erlang C. As patience goes to infinity, Erlang A converges to Erlang C, which is why Erlang C can be seen as the special case in which nobody leaves.

Estimating patience from your own abandon data, and its limits

Your queue reports already contain what you need, but they need careful reading. Pull the wait time of every offered call over a few weeks, and flag whether each call was answered or abandoned.

  1. Do not average only the abandoned waits. Those are the callers whose patience ran out. Callers who were answered after 10 seconds might have waited 10 minutes; you never learned. Robbins notes that patience is in general unobservable for exactly this reason: only callers whose patience expires abandon, so the data is right censored (Robbins, East Carolina University, working paper).
  2. Start with a range. Try patience values around your first estimate, for example 60, 120 and 180 seconds, and see how much the staffing answer moves. If it barely moves, the estimate does not need to be precise.
  3. Back-test. Run the model on a past week using the actual arrivals and staffing. If it predicts 6 percent abandonment and you saw 15, the patience value or the arrival data is wrong.
  4. Check the shape. The exponential assumption means the chance of hanging up is the same at every moment of waiting. Real callers often behave differently: some leave in the first seconds when they hear a queue message, others hang on. Robbins used a Weibull patience distribution in his simulations. He cites empirical hazard-rate graphs that are not constant, with a sharp peak near zero from callers who will not wait at all, and reports that several studies of patience curves concluded patience is best modelled as a Weibull distribution rather than a simple exponential.

Patience also varies by queue type, time of day and what callers expect. A sales line, a support line and a payments line will not share one number. Estimate per queue and revisit it when you change the waiting message or add a callback option.

The same queue under both models

The numbers below are for one illustrative queue, not a benchmark. We computed them ourselves: Erlang C from the standard formula, and Erlang A by solving the M/M/N+M queue with exponential patience. Every figure was recomputed independently, and the abandonment figures were also cross-checked from the average queue length (abandonment equals the patience rate times the average number waiting, divided by the arrival rate).

  • 180 calls arrive per hour.
  • Average handling time is 300 seconds (5 minutes), so the offered load is 15 Erlangs.
  • Target: 80 percent of calls answered within 20 seconds.
  • For Erlang A, two assumed average patience values: 120 seconds and 60 seconds.

One definition matters. The Erlang C column is the usual service level, the share of calls answered within 20 seconds, where every caller is eventually answered. In the Erlang A columns, "answered within 20 s" is the share of all offered calls answered within 20 seconds, so callers who abandon count as not answered. That is the strict reading, and it is not the only one in use.

एजेंटोंErlang C: answered within 20 sErlang A, patience 120 s: answered within 20 sErlang A, patience 120 s: abandonedErlang A, patience 60 s: answered within 20 s
1631.7%75.2%9.4%80.7%
1754.5%81.5%6.8%85.5%
1870.4%86.6%4.8%89.4%
1981.3%90.7%3.3%92.5%
2088.5%93.8%2.2%94.9%

Read across the 80 percent target. Erlang C needs 19 agents. Erlang A with 120-second patience reaches the target at 17. With 60-second patience, 16 agents already reach 80.7 percent. That is two to three agents fewer, out of 19, for the same queue and the same target.

Now read the abandonment column. At 17 agents under 120-second patience, the model predicts about 6.8 percent of callers hang up. Erlang C cannot give you that number at all, because in its world nobody abandons. The saving in agents is not free; it is paid for by callers who left. Whether that is acceptable is a business decision, not a mathematical one.

Which way each model is biased

The worked example shows the models disagree. The more useful question is which one is closer to reality, and in which direction each is wrong. Robbins tested both against a steady-state simulation in which many simplifying assumptions were relaxed. His abstract reports that the findings "support the assertion that the Erlang A model is more accurate", but that "in contrast to the Erlang C model, Erlang A tends to be optimistically biased", and that "neither model clearly dominates the other in all situations" (Robbins, abstract).

In plain words:

  • Erlang C is pessimistic on average. The real system performs better than predicted, because callers who abandon relieve the queue. It is the safer bet, and it tends to overstaff. The study also found it can become optimistic when utilisation is high and arrival rates are uncertain.
  • Erlang A is optimistic, often. The real system tends to perform worse than predicted. The author attributes much of that bias to arrival rate uncertainty, a condition that, somewhat paradoxically, reduces the error of Erlang C.
  • The conclusion is a caution, not a verdict. The paper says care must be taken when using Erlang A to make staffing decisions, in particular where arrival rates are very uncertain and service level requirements are strict, and that staffing from Erlang C is more likely to meet the service level even then.

Two limits on that evidence. The study covers call centres with enough capacity to handle all calls, the regimes where Erlang C has a defined answer, and it looks at long-run steady-state averages, not at an arrival rate that shifts through the day. Treat it as a guide to the direction of the errors, not as exact corrections.

When Erlang C is good enough

Erlang C remains a sensible default in several situations:

  • Abandonment is low. If you answer almost everyone quickly, patience hardly matters and both models agree closely. The simulation study found Erlang C errors shrink as patience increases and are lowest when there are many agents and low utilisation.
  • You want the safe side. If missing the service level is costly, a model that tends to overstate waits is a feature.
  • You do not have patience data. A confident-looking Erlang A answer built on a guessed patience number is not better than an honest Erlang C answer.
  • You need a quick answer. The Erlang C calculator takes seconds, and the Erlang C staffing guide shows how to use it.

If you are sizing trunks rather than agents, a different formula applies; see the guide to SIP channels and Erlang B. For the definition of the unit itself, the glossary has what an Erlang is.

When to use Erlang A or a simulation

  • Abandonment is material. If a noticeable share of callers hang up, ignoring it will misstate both waits and staffing.
  • You run near full utilisation. When the load is at or beyond the agents' capacity, Erlang C cannot give a result at all, while Erlang A still can, because abandonment keeps the queue from growing without limit. Robbins makes this point about the efficiency-driven regime and notes that Erlang A accuracy there needs a separate study.
  • You want to size for an abandon target. If the question is "how many agents keep abandonment under 5 percent", you need a model that has abandonment in it.
  • Your situation is awkward. Uncertain arrivals, skills-based routing, callbacks or non-exponential handling times all strain the simple models. A simulation is the honest tool when the simple models disagree and the stakes are high.

Garnett and colleagues also give practical comfort: for large call centres their asymptotic analysis leads to rules of thumb for design, rather than leaving you with only exact but opaque numbers (Garnett, Mandelbaum and Reiman, 2002, abstract).

Abandonment is not free: fewer agents can mean lost callers

It is tempting to read the table as "Erlang A says we can cut agents". Be careful. Every abandoned caller is a lost sale, a lost payment or an unresolved problem, and some will call again, adding to the next interval's load. A staffing level that meets an "answered within 20 seconds" target by letting 7 percent of callers walk away is a different promise from one that answers nearly everyone.

So use the model to see the trade, not to justify the lower number. Set the abandon rate you can live with, check it against the cost of a lost call, and then choose the staffing level. The परित्याग दर और service level glossary pages define the two measures. And remember that scheduled agents are not the same as available agents: apply shrinkage for breaks, training and absence before you turn a model answer into a roster. The workforce management guide covers that whole planning chain.

Using the Erlang C calculator and queue reports as inputs

You can combine the tools you already have into a sound routine.

  1. Collect inputs per half hour. Offered calls, average handling time, abandoned calls and wait times. DialerBee's रिपोर्टों include Queue Performance and Abandoned Calls reports, and the Insights centre has a Service Level tab, all of which export for further analysis.
  2. Run the Erlang C calculator first. It gives a conservative starting figure.
  3. Estimate patience from the same period, as a range, and see what Erlang A or a simulation says about waits and abandonment.
  4. Compare with last week's actuals. If real abandonment is lower than the model implies, patience is longer than you assumed; if real waits are longer, the arrival forecast or the shrinkage may be the cause.
  5. Staff to the safer end of the range for strict service targets, and to the leaner end only where abandonment is acceptable and monitored.

अक्सर पूछे जाने वाले प्रश्नों

What is the difference between Erlang A and Erlang C?

Erlang C assumes every caller waits as long as it takes to be answered. Erlang A adds one more input, caller patience, so some callers give up and leave the queue. That single assumption changes how many callers wait, how long the queue gets and how many agents you need for a service level target.

Is Erlang A more accurate than Erlang C?

Usually, yes, but not always. In a simulation study of both models, Robbins found Erlang A made smaller errors over a wide range of conditions, yet neither model clearly dominated the other in all situations. Erlang C tended to be pessimistic, Erlang A tended to be optimistic, and the best choice depends on the call centre and what the model is for.

Does Erlang C overstaff?

On average it tends to, because it ignores abandonment and so predicts longer waits than the real system produces. That is the pessimistic bias reported in the simulation study. The same study notes it can flip to optimistic when utilisation is high and arrival rates are uncertain, so it is not a guaranteed safety margin.

How do I estimate average caller patience?

Start from your queue data: how long callers waited before hanging up, and how long answered callers waited. Abandon times alone understate patience, because callers who were answered quickly never revealed how long they would have waited. Treat the first estimate as rough, test several values and check the model against a week of real results.

Do I need Erlang A for inbound only?

The models describe callers waiting in a queue for agents, so they fit inbound queues. For purely outbound calling there is no queue of waiting callers, so the usual planning tools are different.

Is there an Erlang A calculator on this site?

No. The calculator on this site is the Erlang C calculator. This guide shows how to read its answer next to what changes when abandonment is modelled, using worked numbers computed for one illustrative queue.

To get the Erlang C answer for a full day of intervals, use the call center staffing planner.

सूत्रों का कहना है

  1. Garnett, O., Mandelbaum, A., and Reiman, M., "Designing a Call Center with Impatient Customers", Technion, 2002 (published in Manufacturing and Service Operations Management 4(3), 208-227, 2002). aviman.technion.ac.il, abandon02.pdf (7 नवंबर 2026 को जाँचा गया)
  2. Robbins, T. R., "Evaluating the Performance of the Erlang Models for Call Centers", East Carolina University. myweb.ecu.edu, working paper (7 नवंबर 2026 को जाँचा गया)
  3. Worked example: computed by DialerBee for this article from the standard Erlang C formula and the M/M/N+M queue, illustrative only. Not a benchmark.

संबंधित पठन

DialerBee को काम करते हुए देखने के लिए तैयार हैं?

15 मिनट के लाइव डेमो का अनुरोध करें, या 14 दिन का फ़्री ट्रायल शुरू करें — न स्लाइड, न कोई प्रतिबद्धता। कॉलिंग तब शुरू होती है जब आपका अपना SIP ट्रंक या PBX जुड़ जाए और मंज़ूर हो जाए।

14 दिन का फ़्री ट्रायल · मुफ़्त AI क्रेडिट · कोई कार्ड नहीं · 11 भाषाएँ · BYOC · अनुपालन में सहायक नियंत्रण

पूरी साइट अंग्रेज़ी में देखें →