So you want an AI receptionist
Buying, customizing, or building voice AI for your business
Part of the Voice AI series
A year ago, I wouldn’t even have considered allowing an AI agent to answer calls and talk to customers for me. The tech existed, but the audio and answer quality just weren’t ready for production. The field has advanced quickly though, and today’s voice AI sounds natural and fluid, can be interrupted (barge-in detection), and can even detect your inflection, tone, and other audio cues, and respond accordingly.
If you’re ready to add an AI receptionist or voice agent to your business, you have a LOT of options now. There are specialized solutions like Sameday for home services and Slang AI for restaurants. And then there are the developer platforms to build your own, like Vapi, LiveKit, and Retell.
So how do you choose the right one for your business? There’s more to it than just choosing between the specialized services and the build-your-own voice agent dev tools. The other thing to consider is the type of audio processing: a cascaded pipeline (speech-to-text → LLM → text-to-speech), realtime speech-to-speech, or a hybrid approach.
This post will go into the pros and cons of each method, and help you decide on the best solution for your business. We’ll look at the types of services available, the different audio processing approaches, and how to deploy your AI receptionist safely and control cost.
Try it first: Call 657-220-1032 to talk to the GreenFlux AI receptionist. It’ll tell you up front that it’s an AI. You can ask it about voice agents for your business, or have it book a consultation on my calendar.
This post will cover:
- Industry-specific voice AI tools
- Buying, customizing, or building your own
- How voice AI works, and which approach you need
- Deploying safely and controlling cost
Let’s get to it!
Ready-made AI receptionists
There are AI phone products built for specific industries, and for a lot of businesses they’re the right answer. They come preconfigured for the industry’s common calls, and they already connect to the software those businesses use.
Home services. Sameday and Avoca both answer calls 24/7, book jobs, handle texting, and route emergency calls. Both list ServiceTitan as a certified integration, along with tools like Housecall Pro and FieldRoutes. Sameday also connects to lead sources like Thumbtack, Angi, and Google Local Services.
Legal. Smith.ai handles new-client intake for law firms:
- qualifying callers
- booking consults
- collecting consult fees
- sending retainers for e-signature
- running conflict checks
It integrates with Clio, MyCase, Filevine, and a bunch of other practice management tools. It also offers AI, hybrid, and human receptionist tiers, priced per call.
Restaurants and hotels. Slang AI handles restaurant calls and reservations through OpenTable, Resy, and SevenRooms. Canary does the same for hotels, replacing the old “press 1 for reservations” phone menus.
If one of these matches how your business already works, it’s probably the fastest way to stop missing calls. Choosing a platform specific to your industry means you’ll have less customization and setup required.
When a product doesn’t fit
Notice the pattern in that list? Each product works with a specific set of systems for that industry. That’s great if you already run ServiceTitan or Clio. But what if your scheduling lives in a custom database, a FileMaker app, or a Google Sheet that runs the whole company?
Other common reasons a product falls short:
- The integration you need isn’t on their list.
- Your booking workflow doesn’t match theirs.
- Service/sales/emergency calls each need different handling.
- You have custom business rules or approval steps.
- You want more control over the call data, how long it’s kept, or how the agent sounds.
This is where people usually jump to “build vs. buy”.
Buy, customize, or build
There are really three options:
| Approach | What you’re working with | Examples |
|---|---|---|
| Buy | A finished product you configure | Sameday, Avoca, Smith.ai |
| Customize | A managed voice platform, plus your own call logic and integrations | Vapi, Retell, ElevenLabs Agents, Deepgram Voice Agent API |
| Build | Your own orchestration, models, phone lines, and hosting | LiveKit Agents or Pipecat, plus speech providers and Twilio/Telnyx |
The middle option is the one most people don’t know about. The platform handles the hard parts (speech recognition, voice generation, phone lines, interruptions), and you decide what the agent says, what it asks, and what it’s allowed to do in your systems.
I’m not going to claim custom is cheaper. Sometimes it is, sometimes it isn’t, and the pricing models are hard to compare because they don’t line up neatly. Smith.ai charges per call. Most voice platforms charge per minute. The industry products mostly quote you after a demo. The real argument for customizing is fit: the agent works the way your business works, and it talks to your systems.
How voice AI works
So what’s actually happening when an AI answers the phone? This is the part that gets jargon-heavy fast, so I’ll keep it plain.
Speech-to-text (STT) turns spoken words into text. This is how you get call transcripts, summaries, and searchable call logs. Deepgram, AssemblyAI, and OpenAI all offer it.
Text-to-speech (TTS) goes the other way, turning text into a spoken voice. ElevenLabs, Cartesia, and Deepgram are common choices.
Neither one holds a conversation by itself. For that, you need one of these setups:
1. The cascaded pipeline (STT → LLM → TTS). The caller’s voice is transcribed, a language model (LLM) decides what to say and what to do, and the reply is turned back into speech. LiveKit’s docs recommend this as the default for most production agents.
- ✅ You get a live transcript of every call.
- ✅ You can swap any piece for a different provider.
- ✅ Tool use, like booking a meeting, is more reliable here.
- ✅ When something goes wrong, you can see exactly which step failed.
- ❌ Each step adds a little delay.
- ❌ The LLM only sees text, so tone of voice is lost.
2. Realtime speech-to-speech. One model takes in audio and speaks back directly, with no text step in the middle. OpenAI’s Realtime API and Google’s Gemini Live API work this way.
- ✅ Much faster than STT → LLM → TTS.
- ✅ It can pick up on tone and hesitation, and it sounds more expressive.
- ❌ Transcripts come later instead of live, unless you add a separate STT.
- ❌ There’s less control over exactly what gets said.
- ❌ Tool use is less mature than in the cascaded pipeline.
- ❌ You’re more tied to one provider’s model.
3. Full-duplex. This is the newest one. The model listens and speaks at the same time, the way people actually talk, and decides on its own when to take a turn. OpenAI’s GPT-Live works this way. It hands the heavier reasoning and tool calls to a separate backend, so the caller can keep talking while that work runs.
4. Hybrids. You can also mix and match. For example, a realtime voice model can handle the conversation while a separate system handles the business logic, or a separate TTS can read the scripted lines that need to come out word-for-word.
The differences are about control, visibility, and speed, not just how natural it sounds. For me, the live transcript alone is a big reason to start with a pipeline. When a call goes sideways, I want to read exactly what the agent heard.
Which one do you need?
| If you need… | Start with |
|---|---|
| Call transcripts or summaries | STT |
| Announcements or narration | TTS |
| A receptionist that takes messages and books appointments | Cascaded STT → LLM → TTS |
| Open-ended, natural conversation | Realtime speech-to-speech |
| Natural conversation plus a tightly controlled backend workflow | Full-duplex or hybrid |
These are starting points, not rules. How complex your workflow is and how natural the conversation feels are two separate questions. A complicated booking flow might still be best on a plain cascaded pipeline, where you can see every step. A simple Q&A line might sound better on a realtime model.
For the GreenFlux demo, I went with the cascaded pipeline on Vapi. Why? Booking a meeting is a structured job. I want to confirm the caller’s name, email, and time slot before anything gets written to my calendar, and I want a transcript of every step.
Who makes the building blocks?
These categories overlap a lot. ElevenLabs, for example, sells voices and a full agent platform. But here’s roughly how the pieces fit together:
| Layer | Providers |
|---|---|
| Speech recognition (STT) | Deepgram, AssemblyAI, OpenAI |
| Voice generation (TTS) | ElevenLabs, Cartesia, Deepgram |
| Realtime voice models | OpenAI, Google Gemini Live |
| Managed agent platforms | Vapi, Retell, ElevenLabs Agents, Deepgram Voice Agent API |
| Developer frameworks | LiveKit Agents, Pipecat |
| Phone lines | Twilio, Telnyx, Plivo |
If you’re a developer and want to go the full build route, Cole Medin’s LiveKit voice agent tutorial is a great place to start. And Kwindla Hultman Kramer’s advice on building voice AI is worth reading before you choose an architecture. He’s at Daily, the company behind Pipecat.
I’m intentionally not naming specific model versions here. They’re changing too fast. One of Google’s Gemini Live voice models is already scheduled for retirement on December 13, 2026. 🫠
Rules to know before you go live
I’m NOT a lawyer, and this isn’t legal advice. But a few rules come up for anyone putting AI on a phone line in the US:
- Call recording. About a dozen states require consent from everyone on the call before it’s recorded, including California, Florida, Pennsylvania, and Washington. Your callers can be anywhere, so the simple approach is to announce recording at the start of every call.
- Outbound calls. In February 2024, the FCC ruled that AI-generated voices count as “artificial” voices under the TCPA, the federal robocall law. That’s aimed at outbound calls, not at answering your own phone. But if your agent will call people back or send texts, plan for consent first.
- Saying it’s an AI. California and Utah both have laws about telling people when they’re talking to a bot. California’s new AB 1609 (signed September 2026) also requires a way to reach a human, but it only applies to companies with more than $500 million in revenue.
The easy way to cover all of this is to have your agent say it’s an AI, and that the call is recorded, in its very first sentence, and to always offer a way to leave a message for a human.
Keeping costs under control
Voice AI is usually billed by usage, often by several providers at once: the voice platform, the speech and language models, and the phone line. That makes it easy for a bug, a looping conversation, or a spam caller to run up a bill. Before you go live, set billing alerts and spending limits with every provider in the stack, not just the main platform. Then cap the maximum call length, limit how many calls can run at the same time, and set a silence timeout so the agent hangs up on dead air and voicemail instead of waiting.
After launch, check your call logs and usage dashboards every week for the first month. Look for unusually long calls, repeat callers, and calls that end without a result. Keep the system prompt short and focused, since it gets sent to the model on every turn, and only give the agent the tools it actually needs. Small settings like these keep costs predictable as call volume grows.
Voice agents that actually DO things
Here’s where it gets interesting. A voice agent that only answers questions is a fancy FAQ. The value comes when it can take action, like:
- checking calendar availability
- booking an appointment
- creating a lead in your CRM
- transferring an emergency call to the on-call tech
- logging the call details for follow-up
Each of those is just an API call to a system you already use. That’s the part I build: connecting a voice agent to the systems your business already runs on, with rules for what it’s allowed to do and confirmation before it does anything important.
Call the demo
Ok, enough reading. Time to test it out!
Call 657-220-1032 to talk to the GreenFlux AI receptionist. It can:
- answer questions about building a voice agent for your business
- check my calendar and book a consultation
- take a message if you’d rather not book
It’ll tell you it’s an AI, calls are recorded, and calls are capped at 5 minutes. It only confirms a booking after the calendar actually accepts it.
Prefer not to talk to a robot? Totally fair. Here’s a regular booking link.
Conclusion
AI receptionists are no longer just for big call centers. There are solid ready-made products for common industries. And when those don’t fit, managed voice platforms make it possible to build a custom agent that works with your existing systems, without building the voice technology yourself.
What’s Next?
In the next post, I’ll show you how I built this demo with Vapi and Google Calendar, including the setup steps and the settings that matter. After that, I’d like to try branching call logic for service vs. sales calls. Let me know what kind of calls your business would want to automate!