Passer au contenu
Opérations 7 novembre 2026 10 lecture minimale

De combien de canaux SIP ai-je besoin ? Dimensionnement Erlang B

Size SIP trunk channels with Erlang B: busy-hour traffic, blocking targets, transfer legs and predictive-dialing headroom, with worked call centre examples.

D
7 novembre 2026
Dernière mise à jour

Réponse rapide

To size SIP channels, work out busy-hour traffic in erlangs (calls per hour multiplied by average call length in hours), pick a blocking target such as 1%, and read the channel count from the Erlang B formula. Example: 450 calls an hour at 4 minutes is 30 erlangs, which needs 42 channels at 1% blocking. Outbound predictive dialing needs more, because ringing calls hold channels too.

Agents are expensive and obvious, so teams size them carefully. Phone lines are cheap and invisible, so teams guess. The guess shows up later as busy signals at peak, a wave of failed dials at the top of the hour, or a trunk paid for and never used. This guide sizes the lines with the same rigour you apply to the agents.

It completes a pair. Our guide to how many call center agents you need uses Erlang C to size people. This one uses Erlang B to size channels.

Erlang B sizes lines, Erlang C sizes agents

Both formulas turn a workload in erlangs into a headcount, but they answer different questions because the two resources behave differently when they run out.

  • Erlang C assumes a caller who finds every agent busy waits in a queue. The question is how long they wait, so it sizes agents against a service level such as 80% answered in 20 seconds. Use the Erlang C calculator for that.
  • Erlang B assumes a call that finds every line busy is blocked and cleared: the caller gets a busy tone and is gone. The question is what share of attempts you are willing to lose, so it sizes lines against a blocking probability, also called the grade of service.

Lines do not queue. A SIP trunk with all its channels in use rejects the next INVITE, which is why the blocking model fits. The ITU-T has long published guidance on this problem: Recommendation E.520 is titled "Number of circuits to be provided in automatic and/or semiautomatic operation, without overflow facilities", and E.600 defines the terms of traffic engineering.

The inputs

You need four numbers.

  1. Busy-hour calls. The number of call attempts in your busiest hour, not the daily average. Take it from call detail records for a few of your peak days.
  2. Average holding time. How long each call occupies a channel, in seconds, from the moment the channel is seized until it is released. Include ringing and any time before answer, because the channel is in use for all of it.
  3. Traffic in erlangs. One erlang is one channel kept busy for a whole hour. Calls per hour multiplied by holding time in seconds, divided by 3,600.
  4. Blocking target. The share of attempts you will accept losing at peak. 1% is a common planning figure for a business trunk; 2% or 5% is tolerable where a lost attempt can be retried at no cost, and a tighter target costs more lines.

Worked inbound example

Illustrative numbers, not a benchmark. A support line receives 450 calls in its busiest hour and the average call lasts 240 seconds (4 minutes).

Traffic = 450 × 240 ÷ 3,600 = 30 erlangs.

Running the Erlang B formula for 30 erlangs gives these channel counts:

Blocking targetChannels needed for 30 erlangs
5%36
2%39
1%42
0.5%44

Two lessons sit in that table. First, you need more channels than erlangs: 30 erlangs of traffic needs 42 channels for 1% blocking, because arrivals are random and lines must absorb bursts. Second, the cost of a tighter target is gentle at this size. Going from 5% to 1% blocking adds six channels. Compare that with 30 channels, where the offered traffic would be blocked about 13% of the time.

The same 30 erlangs of workload also means you need more than 30 agents on the phones, since the agents carry that workload. The two models together tell you both the people and the pipes.

Larger traffic gets more efficient per erlang. As reference points from the same formula, at 1% blocking: 10 erlangs need 18 channels, 20 need 30, 50 need 64 and 100 need 117.

Outbound: why a predictive dialer needs more channels than agents

An outbound dialer breaks the Erlang B picture in one important way. The traffic is not arrivals you cannot control. The dialer chooses how many calls to place, so it can run into your channel limit by its own pacing. The model still helps, but you size from the dialer's behaviour instead of from customer arrivals.

Each outbound channel is busy for three things: the time the call rings, any time spent classifying the answer, and the conversation. A predictive dialer places more calls than there are free agents, because many will not answer, so at any moment several channels are ringing for each agent who is talking.

Illustrative example, with assumed figures. A team of 40 agents, with 32 on live calls at any moment. The dialer places 3,000 attempts an hour, and an attempt holds its channel for an average of 25 seconds from dial to answer or give-up.

  • Channels occupied by dialling and answer handling = 3,000 × 25 ÷ 3,600 = about 21 erlangs.
  • Channels occupied by conversations = 32.
  • Average in use at peak = about 53 channels, even though only 32 agents are talking.
  • Add around 20% headroom, which is the margin our BYOC setup guide suggests over peak concurrency: about 64 channels.

The dial ratio is what multiplies this. A 3:1 over-dial ratio for 50 agents can open up to 150 concurrent calls, as that same guide notes, so the ratio you let the pacing algorithm reach is the ceiling your trunk must cover. Every point of dial ratio, every second of ring time and every second of answer-handling time moves the number. If your pacing reduces the ratio when abandon rates rise, your lines sit idle at that moment, but you still size for the highest ratio you allow. See the numéroteur prédictif page for how pacing is controlled.

Transfers, conferences and recordings that take extra legs

Count channels, not calls. Anything that adds a call leg through the trunk adds to the total.

  • Attended transfers. While the first agent talks to the third party, the customer and the third party are on separate legs. Briefly three parties hold channels for one customer.
  • Conferences. A third party dialled out through the trunk is one more channel for the duration.
  • Calls to a mobile number from a desk or softphone line. An agent who is reached through the carrier, instead of through a browser, uses a channel of their own.
  • Recording and monitoring. Whether these use a trunk channel depends on where media is captured. When recording happens on the platform's media path it does not add a carrier leg, so confirm that with your platform rather than assuming.

Multiply your expected share of transfers and conferences by the extra legs and add them to the peak. A team that transfers 10% of connected calls at peak, with 32 conversations, adds about three legs: small, but worth counting on a trunk that is already tight.

Headroom, failover and two media nodes

The formula gives the minimum for a stated blocking target at a stated load. Real operations add three margins.

  1. Growth. Size for the busiest hour you expect next quarter, not the busiest hour you had.
  2. Failover. If one trunk fails and another takes the load, the survivor needs the combined peak. Split the channels carefully: two trunks each sized to carry 100% give full failover, while two sized at 60% each only protect partly.
  3. Two media nodes. If calls are spread over two media nodes, check that the trunk limit applies to the total and that each node can absorb the whole load if the other is down.

The BYOC architecture lets you configure a primary and a secondary carrier with a failover gateway, as described in the page de présentation BYOC, so the sizing rule is to give each path enough channels for the calls you would move onto it.

Channels versus calls per second

Channels and calls per second (CPS) are two limits on the same trunk, and either can stop you. Channels cap how many calls exist at once. CPS caps how fast you may start new ones.

They connect through simple arithmetic: concurrent calls equal the call rate times how long each call holds the channel. At 5 new calls a second with a 25-second hold during ringing you already have about 125 channels of dialling traffic. Conversely, a trunk with plenty of channels still rejects a burst when the dialer starts a campaign faster than the carrier's CPS limit. Always ask your carrier for both numbers, and set the dialer's limit at or below the carrier's. The BYOC setup guide covers where the CPS limit is set and why bursts happen at campaign start.

Comment DialerBee s'intègre

With bring-your-own-carrier (BYOC), the trunk stays yours: your carrier, your rates, your channel contract. The Explication BYOC covers the commercial side. The gateway form takes a Max Concurrent Calls value (100 by default) and a CPS limit (20 by default), so set both from the sizing above and the limits your carrier gave you.

On cost, tarification se fait par poste : $69 par poste et par mois pour Starter, $129 pour Professional, et sur mesure pour Enterprise. Il n'y a pas de frais de plateforme à la minute sur les postes agents (l'usage de l'IA est inclus pour l'instant), et le nombre de canaux de votre trunk ne change pas le prix de la plateforme. Ce que le nombre de canaux change, c'est la facture de votre opérateur ; c'est pourquoi il vaut la peine de le dimensionner avec une formule plutôt qu'au jugé.

Channel sizing checklist

  1. Pull busy-hour attempts and average holding time from real records.
  2. Calculate traffic in erlangs and pick a blocking target.
  3. Read the channel count from Erlang B for inbound lines.
  4. For outbound, calculate dialling traffic from attempts per hour and time to answer or give-up, add conversations, then add headroom.
  5. Add legs for transfers and conferences.
  6. Size each failover path to carry the load it would take over.
  7. Get the channel and CPS limits from the carrier in writing and set the gateway to match.
  8. Review after the first busy week and re-run the numbers with real data.

Foire aux questions

How many SIP channels do I need for 10 agents?

It depends on call volume and length, not on agent count alone. As an inbound example, 10 erlangs of traffic needs 18 channels at 1% blocking. For outbound predictive dialing, include the ringing calls: with a dial ratio above 1, channels in use exceed the number of agents talking.

What is the difference between Erlang B and Erlang C?

Erlang B sizes lines where a call that finds no free line is blocked and lost. Erlang C sizes agents where a caller who finds none free waits in a queue. Use B for trunk channels and C for staffing.

What blocking target should I use?

1% is a common planning target for business trunks. Choose a looser target if a blocked attempt is cheap to retry, as in many outbound campaigns, and a tighter one if every blocked call is lost revenue.

De combien de canaux un numéroteur prédictif a-t-il besoin ?

Enough for the calls that are ringing or being classified plus the calls in conversation, plus headroom. As an illustration, 40 agents with 3,000 attempts an hour at 25 seconds each need roughly 53 channels on average at peak and about 64 with 20% headroom.

Is a channel the same as a call per second?

No. A channel is one simultaneous call. Calls per second is the rate at which new calls may start. A trunk can fail on either limit, so size and configure both.

Do I pay more for DialerBee if I need more channels?

Non. DialerBee facture par poste, sans frais de plateforme à la minute sur les postes : le nombre de canaux ne change donc pas le prix de la plateforme. Votre opérateur facture les canaux et les minutes selon votre propre contrat. L'usage de l'IA est inclus dans le prix du poste pour l'instant : pas de frais supplémentaires au lancement.

Sources

The channel counts above come from the standard Erlang B recursion, calculated for this article. All traffic figures labelled illustrative are examples, not measurements.

Lectures complémentaires

Prêt à voir DialerBee en action ?

Demandez une démo en direct de 15 minutes, ou démarrez un essai gratuit de 14 jours — sans diaporama, sans engagement. Les appels démarrent dès que votre propre trunk SIP ou PBX est connecté et approuvé.

Essai gratuit de 14 jours · crédit IA gratuit · sans carte · 11 langues · BYOC · contrôles qui contribuent à la conformité

Voir le site complet en anglais →