What an AI receptionist costs to run
The monthly subscription is rarely the number that decides this. Most of what you pay for a voice agent is consumed per minute of conversation, which means your bill is set by how long your calls last and how many arrive, not by which plan you chose. Two businesses on identical plans can pay very different amounts, and the difference is usually average call length. Here is the shape of the bill and which parts move.
What are you actually paying for?
Five things, usually bundled into one invoice so you cannot see them. A phone number and telephony minutes, charged by your carrier or the platform's carrier. Speech recognition, which turns caller audio into text and is billed per minute of audio. A language model, which decides what to say and is billed per token of input and output. Speech synthesis, which turns the reply into audio and is billed per character or per second. And a platform fee for the orchestration, dashboard, integrations and support.
Bundled per-minute pricing hides which of these dominates, and that matters when you try to reduce the bill. If synthesis is the largest line, a shorter greeting saves real money. If the language model is, the fix is a shorter system prompt and less conversation history resent on every turn.
Ask any vendor for the per-minute rate and what it includes, then ask what happens when a call runs long. Some platforms bill in whole minutes, which makes a thirty-five second call cost the same as a fifty-nine second one, and that rounding is significant for businesses whose calls are mostly short.
One structural question decides a lot: whether you bring your own telephony and model accounts or buy everything through the platform. Bringing your own is usually cheaper per minute and gives you an itemised bill, at the price of owning the configuration and the support conversation when something breaks between two suppliers. Buying the bundle costs more and gives you one number to ring. Neither is wrong, and the choice is far easier to make before migration than after.
Which part of the bill scales with call volume?
Almost all of it. Recognition, generation, synthesis and telephony are consumption costs, and they rise close to linearly with total minutes handled. Only the platform fee and the number rental are fixed, and on most plans they are the smaller half once volume is real.
The important nuance is that cost scales with minutes rather than calls, and minutes are driven by conversation design more than by caller behaviour. An agent that reads a long disclosure, confirms every field twice and asks an open question at the end will double the length of the same call. That is a design choice you control, and it is the single largest lever on the bill.
There is also a superlinear effect in the language model line. If the agent resends the whole conversation on every turn, input tokens grow with the square of the number of turns, so a call with twenty turns costs far more than twice a call with ten. Vendors who cache or summarise context avoid this; ask whether they do.
| Cost component | Billed by | Scales with | How to reduce it |
|---|---|---|---|
| Phone number rental | Month | Nothing | Fewer numbers, no vanity numbers |
| Telephony minutes | Minute | Total talk time | Shorter calls, avoid warm transfer where possible |
| Speech recognition | Minute of audio | Total talk time | End calls promptly, cut dead air |
| Language model | Token in and out | Turns, and prompt size | Shorter prompt, cached or summarised history |
| Speech synthesis | Character or second | Words spoken by the agent | Shorter greetings, no repeated preambles |
| Platform fee | Month or seat | Nothing, until a tier boundary | Right-size the plan after a month of real data |
| Human overflow | Minute or per call | Handover rate | Fix the questions that cause handovers |
Which costs turn up after the first invoice?
Recording storage, if you keep audio. Recordings are large, they accumulate, and a retention policy of keep everything becomes a line item within a year. It is also a liability question rather than only a cost one, which is why retention should be decided deliberately rather than defaulted.
Integration and maintenance are the bigger surprise. Connecting a calendar and a CRM is project work, and it does not stay done: when your booking system changes a field or your CRM enforces a new required property, the agent starts failing silently. Budget for someone owning that, whether internal or on a support agreement.
Then transcript review, which is genuinely a cost even though nobody invoices for it. Someone has to read a sample of calls every week to find the confidently wrong answers, and without that the system quietly degrades. Half an hour a week is realistic; zero is what most teams budget.
How do you model it before you buy?
Three numbers, all of which you can get this week. Calls you want it to answer per month, taken from your phone provider's report of unanswered and out-of-hours calls. Average length of those calls, which for an agent is usually shorter than a human equivalent because there is no small talk. And the platform's all-in per-minute rate.
Multiply calls by minutes by rate, add the fixed monthly fee, and then add a margin for the calls the agent will handle that a human would not have: agents get called by people who would previously have hung up, so volume tends to rise once the phone starts being answered.
Then do the comparison that actually matters, which is not against a human receptionist's salary. Compare the monthly figure against the gross margin of one average job. For most businesses with a meaningful job value the whole question is decided by whether the system recovers one or two jobs a month, and if it does, minute-level cost optimisation is a distraction.
When does the arithmetic stop working?
At high volume with low value per call. If you take thousands of calls a month and each one is worth very little, per-minute consumption pricing works against you, and the honest answer may be a phone tree, a better website, or fewer reasons to call. Voice is an expensive channel per interaction and it does not get cheaper with scale in the way software usually does.
It also stops working when average handle time creeps up. A pilot with ninety-second calls that becomes a production system with four-minute calls has more than doubled its running cost without anyone deciding to. Watch average handle time as a cost metric, not only a quality one.
And it stops working if the handover rate is high and your humans are paid to take those calls anyway. You are then paying twice for the same conversation. That is a design problem rather than a pricing one, but it shows up on the invoice first.
The last trap is a plan bought on the wrong unit. Allowances quoted in calls look generous until your average call runs to four minutes, and allowances quoted in minutes look generous until you find that ringing and transfer legs are counted. Establish which unit the allowance uses, then convert it into your own numbers before comparing two vendors. Quotes that appear to be within a few pounds of each other frequently are not.
What should you ask a vendor about pricing?
Five questions. What is the all-in per-minute rate, and which of the five components does it include? Do you bill in whole minutes or per second? Are minutes counted for ringing, hold and transfer legs, or only for conversation? What happens when I exceed the bundled allowance, and is the overage rate the same? And is there a charge for calls the agent answers and immediately transfers?
The last one catches people out. A call that connects, plays a greeting and transfers still consumes telephony, recognition and synthesis, and on some plans it also counts against your allowance as a full call. If your handover rate is high, you are paying for a lot of very short calls.
Get the answers in writing before signing anything, then check the first invoice against your own model. A gap of more than about twenty per cent between your estimate and the bill means one of the five components is behaving differently to how it was described, and it is worth finding out which before month two.
Common questions
- How is an AI receptionist priced?
- Mostly per minute of conversation, wrapped in a monthly fee. Underneath the bundle there are five components: telephony minutes, speech recognition per minute of audio, a language model billed per token, speech synthesis billed per character or second, and a platform fee. Only the platform fee and number rental are fixed, so your bill follows total talk time rather than the plan you picked.
- What makes an AI receptionist expensive to run?
- Long calls, mostly. Every component except the platform fee is consumed per minute, so conversation design sets the bill: long greetings, repeated confirmations and open-ended closing questions all add minutes. Resending the whole conversation to the language model on every turn compounds it, because input tokens then grow faster than the number of turns.
- Does an AI receptionist get cheaper at high volume?
- Not much. Fixed platform fees spread out, but the per-minute components dominate at volume and they scale close to linearly. Voice stays an expensive channel per interaction. If you handle thousands of low-value calls a month, consumption pricing works against you and a phone tree or better self-service may be the honest answer.
- What hidden costs should we budget for?
- Recording storage if you keep audio, integration maintenance when your calendar or CRM changes a field, and transcript review time. The last one is never invoiced and is the one most teams set to zero. Half an hour a week reading a sample of calls is what keeps confidently wrong answers from accumulating unnoticed.
- How do we estimate the monthly cost before buying?
- Take three numbers: calls per month you want answered, from your provider's unanswered and out-of-hours report; the average length of those calls; and the all-in per-minute rate. Multiply, add the fixed fee, then allow for volume rising once the phone is actually being answered. Compare the total against the gross margin on one average job, not against a salary.
