How to build an AI receptionist from scratch

This page may contain affiliate links from which Blur Studio earns commissions.

The pieces that make up an AI receptionist

An AI receptionist, built from scratch, is really four things working together: a phone number and telephony connection to receive the call, speech-to-text to turn the caller's voice into text, a language model to decide how to respond, and text-to-speech to turn that response back into spoken audio. You can assemble these separately, or use a platform built specifically to bundle them for voice agents.

• Vapi: A developer platform purpose-built for voice agents, bundling telephony, transcription, the language model and voice generation. It charges pass-through costs for each component (transcription around $0.0095 to $0.0099 a minute, the language model $0.0077 to $0.0452 a minute, voice generation $0.0146 to $0.0238 a minute using ElevenLabs as the voice provider) plus its own hosting fee of $0.05 a minute on usage-only plans. A Core plan at $29 a month exists for lighter, ongoing use.
• Retell AI: A similar developer platform, with pre-built agent templates that reduce how much you build from zero. All-in per-minute pricing runs $0.07 to $0.31 depending on which language model and features you select, plus $2 a month per phone number and $8 a month per extra concurrent call beyond the first 20, which are included free. A direct comparison of the platforms is in Vapi vs Retell vs ElevenLabs agents.
• ElevenLabs: Supplies the voice layer specifically, and also offers its own conversational AI agent builder if you want fewer separate pieces to wire together. Its general plans start at $5 a month (Starter).
• n8n: A workflow automation tool, useful for connecting the receptionist's output, a booked appointment, a qualified lead, a flagged emergency, to the rest of your business systems (your CRM, your calendar, a Slack alert), rather than for the voice conversation itself.

The actual steps, in plain words

1. Define exactly what the receptionist should handle: Write down the specific call types it needs to manage, booking, basic questions, triage, before touching any tool. This document is the actual spec, and skipping it is the most common reason a custom build goes over budget.
2. Pick a platform: Vapi or Retell for most custom builds, since both bundle the hard parts (telephony, transcription, voice) rather than requiring you to wire each one up separately.
3. Write the agent's instructions and test calls: This is where most of the real work happens, writing out how the agent should respond to each call type you defined in step one, then calling it yourself repeatedly with real, messy, off-script questions, not just the happy path.
4. Connect it to your actual systems: Wire the output, a booking, a lead, a flagged call, into your calendar, CRM or ticketing system, often using a tool like n8n to handle that connection without custom code for every integration.
5. Set up the human handoff: Define specific triggers (an emergency keyword, a caller asking for a person, a question outside scope) that immediately route to a live person, and test that this actually works before going live.
6. Run it in parallel before fully switching over: Keep your existing answering process running alongside the new AI receptionist for a week or two, comparing outcomes, before relying on it exclusively.

Who should actually attempt a custom build

Be honest about whether this is the right project for your situation before starting. A custom build makes sense if you already have a developer on staff or a clear budget to hire one, your call-handling needs are genuinely more specific than any packaged product supports, and you're prepared for ongoing maintenance, not just a one-time setup. If any of those three aren't true, a packaged product, even an imperfect one, is almost always the better use of your time and money than attempting a custom build that ends up half-finished or unmaintained within a few months. Common mistakes are listed in AI voice agents for small business.

When to use a ready product instead

If you're already running Jobber, Housecall Pro or ServiceTitan, three field service management platforms for scheduling and dispatching jobs, their built-in AI receptionist add-ons (Jobber's Receptionist at $29 a month, Housecall Pro's CSR AI, ServiceTitan's AI Virtual Agent on its Pro tier) already handle most of what a custom build would need to be built from scratch to do, and they book directly into the job system you already use. Building from zero with Vapi or Retell makes sense mainly when your call logic is genuinely unusual, you're not on one of these platforms, or you need the receptionist to do something none of the packaged products support. That route is covered in AI answering service for contractors.

What a real spec document looks like

Before writing a single line of configuration, a workable spec for a custom receptionist is short and concrete: a list of the 4 to 6 call types the business actually gets (booking, basic pricing question, emergency, complaint), the exact information the agent needs to collect for each one, what system each outcome should be sent to, and the specific phrase or condition that triggers a handoff to a person. This fits on one or two pages for most small businesses. Skipping straight to building the agent's prompt without this document is the most common reason a custom build runs long, because the "what should it actually do" questions end up getting answered piecemeal during testing instead of settled up front.

A realistic cost picture

Beyond the per-minute platform costs above, budget real developer time for building, testing and maintaining a custom agent, this is usually the larger cost than the per-minute usage fees for a small or mid-size business's call volume. A simple custom receptionist handling a modest call volume might run $100 to $300 a month in pure platform and model costs on Vapi or Retell, on top of the time to build and keep it working as your business changes. If voice quality matters more to you than deep call-flow customization, check ElevenLabs' current plans before committing to a fully code-first platform.

The ongoing maintenance nobody budgets for upfront

A custom voice agent isn't a build-once-and-forget system. As your business adds services, changes pricing, or your common call patterns shift, the agent's configuration needs updating to match, the same way a website needs occasional updates to stay accurate. Budget for this ongoing maintenance from the start, either your own time or a developer's, rather than treating the initial build as the entire cost. A business that builds a custom agent and never touches it again will find it quietly going stale, confidently giving outdated information, within a few months of real business changes.

How does Blur Studio approach this?

When a client's call handling is specific enough that no packaged product fits, I build the agent on a platform like Vapi or Retell, wire it directly into their existing calendar, CRM and dispatch systems, and set up the handoff logic so calls that need a person reach one fast. You keep the accounts and the code, and I keep it running and tuned as your business and call patterns change.