A voice agent has no queue. That is half the point of it — and it is also why nobody sees the peak coming.
A contact centre with twelve people on shift takes twelve calls at a time. If sixty come in, forty-eight wait in a queue. The queue is unpleasant for the people standing in it, but it is also a warning: a number grows on a screen, somebody sees it, and somebody has time to act. An agent that answers on the first ring has no such intermediate step. It takes everything — right up until it doesn't, and that happens with no run-up at all.
Peaks rarely arrive evenly spread. A power cut in one district, a product recall, a cancelled departure, an invoice with the wrong amount sent to twelve thousand customers: whatever makes people call reaches all of them inside the same few minutes. That is the one hour your phone line really exists for — and the only hour your capacity is genuinely tested.
Capacity is four ceilings, not one
"How many calls can the agent handle?" has no single answer, because capacity does not sit in one place. It is the lowest of four numbers, and the four are owned by four different parties.
The lines in. Your telecoms provider has set a limit on how many simultaneous calls your number can carry. That number is in the contract, it was usually set from how many people you had when the contract was written, and it has rarely been revisited since.
The platform. The voice agent vendor has its own limit on concurrent sessions — per account, sometimes per agent.
The model. The speech-to-speech model behind the voice has quotas: concurrent streams, requests per second, tokens per minute. This ceiling is the least visible to you, because it sits two links away.
The business system. The system the agent looks things up in will take a certain number of simultaneous queries. That number was sized for a handful of case workers clicking through screens, not for eighty agents asking at once.
Your capacity is the lowest of the four. Everything above it is theory.
Each ceiling breaks with its own sound
What makes this hard to spot is that the four failures look nothing like each other.
When the lines go, the caller hears an engaged tone or a message from the carrier. The call was never yours, so it is not in your statistics either. It is the most serious of the four, and the only one that is completely invisible from the inside.
When the platform goes, the call is rejected or sent on to whatever you have configured as a fallback — if you have configured one.
When the model goes, the call connects. The agent answers, greets the caller, and then falls silent mid-sentence. In the log it looks like a completed call where the customer hung up.
When the business system goes, the agent answers perfectly normally but cannot look anything up. It becomes a talking FAQ: pleasant, fast and unable to say anything about that particular order. At that point it is the knowledge base answering instead of the business system, and that is a different conversation from the one the customer rang for.
Only the second of the four looks like a fault. The other three look like something else entirely: a number that doesn't answer, a customer who hung up, an agent that couldn't help.
