Build your own AI receptionist, or buy one

You can build an AI receptionist yourself, and the parts are cheaper than most people expect. Raw vendor bills land at roughly R2 to R3 a minute of talk time. That is not the hard part, and the price is not what should decide this.

The hard part is everything that happens after the call works.

The five parts you have to assemble

A voice agent that answers a phone is five services in a row, and each one is a bill.

1. A number and the telephony. Something that answers a South African number and turns the call into audio your code can read.
2. Speech recognition. Turning the caller's audio into text, fast enough that the pause does not feel wrong.
3. A language model. Deciding what to say next, with your prices and your rules in context.
4. A voice. Turning the reply back into audio that a stranger will listen to without hanging up.
5. Orchestration. Holding the turn, handling interruptions, knowing when the caller has finished talking, and cutting the model off when they start again.

Platforms exist that bundle three, four or all five. That is the market you are shopping in when you look at the developer tools, rather than the receptionist products.

What the parts actually cost

One platform publishes its component pricing openly, which makes the maths easy. Converted at roughly R16 to the dollar:

Add them up two ways. A lean build comes to $0.13 a minute, about R2.08. A build using a premium voice and a frontier model comes to $0.19, about R3.04.

Another platform charges $0.05 a minute for hosting and passes model costs through at cost, or free if you bring your own API keys. A third sells bundled minutes, $99 a month for 1,238 minutes, with extra minutes at $0.08. Read the footnote on that last one: the model and the telephony are billed separately on top, and calls past your concurrency limit bill at double.

Then the number itself. A South African number runs about $4 a month, roughly R64, and closer to $8 in the first month once the once-off setup is in. Inbound is about a cent a minute, so a thousand minutes adds another R170 or so. Published rates from the big international providers vary a lot and are often higher than that, so get a real quote rather than trusting a pricing page.

A local South African SIP provider will sell you the same number for a tenth of that, sometimes R10 or R50 for a block. The catch is that you inherit the SIP trunk, the registration, the codec negotiation and the channel limits, which is real work and the first place a build like this goes wrong.

The worked example

A thousand minutes a month, which is a busy single-location business.

• Lean build. About R2,080 in platform and model bills.
• Premium build. About R3,040.
• Plus the number and inbound minutes. About R235, a bit more in month one.

Call it R2,300 to R3,300 a month in vendor invoices, for a system that answers the phone and nothing else yet.

For comparison, a finished AI receptionist in this market sits between R1,000 and R3,000 a month. The ingredients cost about what the meal costs. That is the number that should give you pause, because every hour you spend is on top of it.

What the per-minute price does not buy

The demo is the easy 20%. The list below is the other 80%, and none of it appears on a pricing page.

• A playbook that survives contact with real callers. Your hours, your prices, your service area, what you do not do, and the forty edge cases you discover in week three.
• Handover to a person. Not just a transfer. Knowing when, warm context so the human is not starting cold, and what happens when nobody picks up.
• What it does when it does not know. The difference between a system that says so and one that guesses is the difference between a useful tool and a customer booked for the wrong service on the wrong day.
• Where the conversation goes. Structured into a diary or a CRM, not a transcript nobody reads.
• Recording, retention and erasure. You are the responsible party under POPIA. That means knowing where recordings live, how long you keep them, and being able to delete one when a customer asks. See our POPIA page for what that involves.
• Somebody to phone at 22:00 when it breaks. If you built it, that person is you, on a Saturday, while the calls go unanswered.

When building it yourself is the right call

Be honest about which of these is true for you.

• You are a developer and this is the product. You are not building a receptionist, you are building a business. Then obviously build it.
• Your call flow is genuinely unusual. Something no product handles, wired deep into systems nobody integrates with.
• You have real volume. At tens of thousands of minutes a month, the margin between wholesale and retail starts to justify an engineer.
• You enjoy it and the phone is not urgent. A legitimate reason, as long as you say it out loud rather than dressing it up as a cost decision.

When it is not

If you run a business that is not a software business, and your phone is already ringing unanswered, building this is the expensive option. Not because the parts cost more, they do not. Because the thing you are short of is evenings, and this will eat them.

The tell is simple. If you are reading this to work out whether you can save R1,500 a month, the answer is that you probably cannot, and the attempt will cost you more in missed calls than the subscription would have.

Where Municate fits

Municate is a South African AI front desk. It answers voice calls, WhatsApp, SMS and web chat on your existing number, follows a playbook you write in plain language, captures the job, and hands over to a person when it should. Customer data stays in South Africa.

R2,495 a month, flat, whatever the volume, with the setup done with you rather than handed to you as a blank config file. The details are on the pricing page.

If you would rather build it, genuinely good luck, and check the POPIA points above before you go live. If you would rather not, book a demo and we will run a test call on your own line.