The stack you are actually paying for
A voice agent is not one product, it is four things chained together, and each bills separately. Understanding the chain is what lets you read a quote properly.
First, telephony: the number itself and the per-minute cost of carrying the call. Second, speech-to-text, converting the caller's audio into words. Third, the language model that decides what to say. Fourth, text-to-speech, turning the reply into audio.
Platforms like Vapi, Retell, and Synthflow orchestrate these four and add their own margin and tooling on top. Some bundle everything into a single per-minute rate, others pass through the component costs and charge for the orchestration.
This matters because the components have very different price profiles. A premium, natural-sounding voice can cost several times a basic one, and a large language model costs considerably more per exchange than a small one. Two agents on the same platform can differ severalfold in running cost purely on those two choices.
Why the per-minute rate misleads
The advertised rate is usually the floor: cheapest model, cheapest voice, no extras. The configuration that actually sounds good and behaves well is rarely the cheapest one.
Latency costs money too. The setups that feel natural, where the agent responds without an awkward pause, generally use faster and more expensive inference. You can have cheap or you can have responsive, and callers notice the difference immediately.
Then there is everything that is not talk time: failed calls, testing, and the calls where someone rings, hears a voice, and hangs up. During development you will burn a meaningful amount on your own testing, which is fine but should be expected.
None of this makes the platforms bad value. Per minute, DIY is genuinely cheaper than any human service. The point is that the number on the pricing page is not the number on your invoice.
The cost that is not on any invoice
Building the first working version is a weekend. Building one you would trust with real customers is not.
The prompt is the hard part. Getting an agent to reliably ask the right qualifying questions, refuse to answer things it should not, handle interruptions, and hand over cleanly takes many iterations against real conversations. Each iteration is a change, a test call, and a judgement about whether it got better.
Then the integrations, which is where most DIY builds stall. Reading calendar availability correctly, writing bookings without double-booking, matching returning callers to existing records. I go through what has to work in AI phone agent integrations.
And then it needs maintaining. Models get deprecated, APIs change, your prices and services change, and an agent nobody has looked at in six months is confidently telling callers things that stopped being true in the spring.
When DIY is genuinely the right call
If you enjoy this kind of work, build it. That is not a throwaway line. The people who succeed at DIY voice agents are the ones who find the iteration interesting rather than a tax, because the iteration never really stops.
If you have unusual requirements that no packaged product handles, building gives you control that buying cannot. Odd integrations, unusual call flows, and industry-specific logic are all easier when you own the whole thing.
If you are technical and your call volume is high, the per-minute economics compound in your favour and the build time amortises quickly.
And if you want to understand the technology properly rather than rent it, building one is by far the fastest education available. Even if you end up buying, you will buy far better for having tried.
When buying is the honest answer
If your time is the scarce resource in your business, and it usually is for an owner-operator, then the weeks spent tuning prompts are weeks not spent on paid work. That trade is rarely favourable at typical small-business call volumes.
If the thing has to work correctly from day one because it is answering real customers, the gap between a demo that works and a deployment that holds up is wider than it looks.
If you do not want to be the person who gets paged when a model deprecation breaks the phone line on a Saturday, that is a completely legitimate reason to pay someone else.
There is no virtue in either choice. It is a straightforward question about where your hours are worth most. If you want to hear what a finished one sounds like before deciding, ask me for a live demo, and what is an AI receptionist explains what it should be doing.