What is voice AI concurrency and why size for the peak?

Concurrency is the number of phone calls your AI voice agents handle at the same instant, not over a day. It is the single number that decides how much telephony and model capacity you must provision, because everything in a voice stack is parallel: ten simultaneous calls need ten live audio channels, ten model streams, and the carrier capacity to ring ten numbers at once. Sizing for the average is the classic mistake. A dental agency might run two calls most of the hour and then take fifteen at 9am Monday when every office opens. If you provisioned for the average you drop or busy-signal the rush, which is exactly the revenue you sold the client on capturing. Size for the realistic peak plus headroom, then you are never the bottleneck when volume spikes.

How many telephony channels do I need per concurrent call?

One channel per concurrent call, plus headroom. A channel, sometimes called a trunk session or a concurrent call path, carries exactly one live conversation. If your peak is ten simultaneous calls you need at least ten channels, but provisioning exactly ten means the eleventh caller in a burst gets a busy signal. This calculator adds a thirty percent buffer by default so short spikes above your stated peak still connect. On most providers, including the carriers Hermes runs on, channels are elastic and you are billed per connected minute rather than per reserved channel, so the buffer costs nothing until it is used. The practical rule: take your worst realistic minute, add a third, and request that as your concurrent-call ceiling so a surge never turns into dropped calls.

What LLM rate limits do I need for concurrent voice calls?

Voice is far more rate-limit hungry than chat because every active call streams tokens continuously. A single call resends its system prompt and growing transcript several times a minute, so a planning figure of a few thousand tokens per minute per call is reasonable. Multiply that by your peak concurrency and you get the tokens-per-minute, or TPM, your provider tier must clear. Ten concurrent calls can easily demand tens of thousands of TPM and a matching requests-per-minute, or RPM, ceiling. New API accounts sit in the lowest tier and will throttle the moment a few calls overlap, which shows up as long pauses or dropped turns mid-conversation. Check the required TPM and RPM this tool returns against your current tier, and request an increase before you launch, not after a client complains about lag.

How many phone numbers do I need for outbound calling?

For inbound, a single number can receive many simultaneous calls, so one number is usually enough on the receiving side. Outbound is different. Carriers and analytics networks flag a number that places too many calls per hour as likely spam, which tanks your answer rate and can get the number labeled Scam Likely. The fix is caller-ID rotation: spread outbound volume across a pool of numbers so no single one trips the filters. A common planning ratio is roughly one number per eight concurrent outbound calls, adjusted for your dialing pace and local-presence strategy. This tool recommends a starting pool size from your peak. It is a floor, not a guarantee, because deliverability also depends on consent, call duration, and proper A2P and STIR/SHAKEN registration on every number.

What is CPS and why does it matter for AI calling?

CPS is calls per second, the rate at which your carrier lets you originate new calls. It is separate from concurrency: concurrency is how many calls run at once, CPS is how fast you can start them. A burst of fifteen outbound calls fired in one second needs roughly fifteen CPS of capacity, even if each call is short. Most carrier accounts start at one CPS by default, which quietly queues or fails a fast burst and stretches a campaign over minutes instead of seconds. For predictable inbound this rarely bites, but outbound campaigns and reactivation blasts hit the CPS ceiling first. This calculator suggests a CPS to request from your peak concurrency so a planned burst dials cleanly. Request the increase ahead of time, since carriers review and approve CPS raises rather than granting them instantly.

How does Hermes handle concurrency and scaling for my agency?

Hermes is the platform agencies run their voice agents on, white-labeled end to end, so concurrency is provisioned and managed for you instead of being five separate vendor tickets. You do not file a CPS increase with one carrier, a rate-limit raise with a model provider, and a number-pool request with another, then glue them together. You build the agent, connect numbers, and launch inbound and outbound campaigns under your own brand, and the telephony channels, model throughput, and number rotation scale behind one transparent per-minute rate. Pricing is straightforward: Starter is $149 a month with three white-label workspaces and 300 included minutes, Business is $399 with 1,000 minutes, and Agency is $699 with 2,000 minutes, all at $0.24 per overage minute. Want the full breakdown for your agency? Book 20 minutes and we will size your peak concurrency and your first launch with you live.

free tool

Voice AI concurrency calculator.

Enter your peak concurrent calls and get the exact capacity to provision: telephony channels, calls-per-second, outbound number pool, and the LLM rate limits your tier must clear. Built for operators sizing a stack before launch. No email, no signup.

/ your peak load

Peak concurrent callsCalls live at the same instant in your worst minuteAverage call length (minutes)Used for sustained-load contextCall directionOutbound drives the caller-ID number poolModel typeChanges how rate limits are framed

/ voice channels

13

peak 10 + 30% headroom

/ llm tokens/min

70,000

TPM your tier must clear

/ outbound numbers

2

caller-ID rotation pool

/ what to provision before launch

Concurrent voice channelsPeak + 30% headroom so a short spike still connects

13

Calls per second (CPS) to requestCarrier default is often 1 CPS, request a raise before launch

1

Outbound caller-ID numbersRotate volume so no number trips spam filters (~1 per 8 calls)

2

LLM throughput (tokens/min)Your provider tier must clear this TPM at peak

70,000

LLM requests/min (RPM)~3 model requests per minute per active call

30

Channel-minutes/hour at peakSustained load if every channel runs for a full hour

600

Calls/hour completed at peakThroughput if channels stay saturated at your average call length

150

Planning estimates from peak concurrency. Deliverability also depends on consent, call length, and A2P / STIR-SHAKEN registration. Verify against a real load test.

Book a 20-min walkthrough to cut this number.

We will size your channels, carrier CPS, number pool, and model tier, then provision them behind one transparent per-minute rate. One platform, your brand, from $149/month.

Book your walkthrough →

Your inputs are encoded in the URL. Send it to a partner and they see the same numbers.

how it works

Concurrency is the number that breaks your stack first.

Most operators size a voice stack by monthly minutes, then get surprised when calls drop during a rush. The reason is that a voice stack is parallel, not sequential. Ten calls at the same instant need ten live audio channels, ten model streams, and the carrier capacity to handle ten conversations at once. None of that is captured by a monthly total. The metric that actually decides your provisioning is peak concurrency: the worst minute, not the average one.

This calculator takes your peak concurrent calls and turns it into the four numbers you have to request before launch. It sizes telephony channels with a headroom buffer so a short spike above your peak still connects. It suggests a calls-per-second ceiling so an outbound burst dials cleanly instead of queuing behind a default one-CPS limit. It recommends an outbound caller-ID pool so volume is spread across numbers and your answer rate does not collapse under spam filtering. And it estimates the tokens-per-minute and requests-per-minute your model provider tier must clear, because voice is far more rate-limit hungry than chat.

Treat the outputs as a planning floor, not a guarantee. Real deliverability also depends on consent, call length, and proper A2P and STIR/SHAKEN registration, and model tiers change. Use the numbers to file the right increases ahead of time, then validate against a real load test. When you would rather not manage five vendors and five rate-limit tickets, that is where Hermes comes in: channels, throughput, and number rotation provisioned for you behind one transparent per-minute rate starting at $149 a month, under your own brand.

frequently asked

Common questions.

next step

Know your peak? Run the whole stack on Hermes.

One platform. Your brand. From $149/month, with telephony channels, model throughput, and number rotation provisioned for you.

Book a 20-min walkthrough Apply for the Founders' Beta

Want the full breakdown for your agency? Book 20 minutes and we will size your peak concurrency, the carrier and model tiers to request, and your first agent launch with you live.

Voice AI concurrency calculator.

Concurrency is the number that breaks your stack first.

Common questions.

Know your peak? Run the whole stack on Hermes.

/ Keep reading