You’ve tried Voice Mode on ChatGPT. You’ve watched the demo where an AI agent calls a restaurant and books a table. You walked away impressed, and you immediately wondered if you could put one on your front desk.
The novelty doesn’t answer that question. It just gets it asked more often.
Voice AI is past the demo problem and well into the production problem, which has a different shape. The work isn’t whether the model can talk. The work is figuring out which calls in your business fit what voice AI handles well right now, and which still need a human. This post is the voice-specific deep cut on a topic we’ve covered more broadly in our SMB AI integration overview.
What changed between the demo and now
A year ago, this was an architecture conversation. Sub-second latency was hard, mid-conversation interruption was experimental, and “can it check a calendar mid-call” was a research question. Real-time bidirectional audio is no longer a differentiator: it’s the floor. Mid-conversation tool use works. Cost per minute has dropped enough that high-volume use cases pencil out for SMBs, not just enterprises.
OpenAI’s Realtime API launched at roughly $0.30 per minute in October 2024. The current generation runs as low as $0.017 to $0.034 per minute for dedicated speech variants, and platform layers like Retell and Vapi advertise base rates of $0.05 to $0.07 per minute. All-in production cost (telephony, speech-to-text, text-to-speech, and model passthrough) typically lands $0.15 to $0.30 per delivered minute. The advertised rate isn’t the delivered cost, and it’s worth pricing the full stack before scoping a project.
The constraint has shifted from “can we build it” to “where does it fit.”
The shape of a call voice AI handles well
A call that voice AI handles well has four properties. Any one missing makes the math shaky. Two missing means hire a human.
- High volume. The build cost amortizes across enough calls per month to matter. A receptionist taking ten calls a day is a payroll question, not an automation one.
- Narrow scope. The set of things the caller might want is small enough to design for. Scheduling, intake, confirmations, and routing all fit. “Anything a customer might call about” doesn’t.
- Low empathy load. The call doesn’t require reading the caller’s emotional state to do the job right. Booking an appointment doesn’t. Handling a billing complaint does.
- Bounded failure cost. When the assistant gets something wrong, the consequence is recoverable in a follow-up call or message. A wrong appointment time is easily fixed. A wrong medication dose recommendation is not a job for voice AI.
If your highest-volume call type doesn’t hit all four, voice AI isn’t the answer yet. That’s a feature of this post, not a hedge.
Where voice AI actually pays off
The use cases below all hit the four properties. Each one includes a signal that a business is ready to deploy.
Scheduling and booking. Healthcare practices, dental offices, home services, salons, fitness studios: anywhere callers want to book or reschedule. Calendar and CRM integrations are the tool calls. The assistant confirms details, books the slot, and sends a confirmation. The market lines up with this. Grand View Research projects the AI voice agents market to grow from $2.54B in 2025 to $35.24B by 2033, and scheduling-heavy industries are the segment seeing the earliest returns. The signal a business is ready: the front desk regularly fields the same fifty scheduling calls a week and can’t grow without adding headcount.
Inbound triage and routing. The call comes in, the assistant identifies who the caller is and what they need, and routes to the right human or workflow. Replaces the “press 1 for X, press 2 for Y” IVR that callers have hated for two decades. Works because the scope is narrow (identify intent), the empathy load is low (you’re not solving the problem, just routing it), and the failure mode is bounded. Worst case, you route to the wrong queue and a human re-routes.
Outbound confirmations and reminders. Appointment confirms, prescription pickup notices, delivery windows, payment reminders. The assistant initiates, confirms the detail, and logs the outcome. The use case has a real cost on the other side: MGMA’s 2025 patient no-show poll found 27% of medical practices saw no-show rates worsen year-over-year. Voice reminders catch the segment of patients who ignore SMS and email.
After-hours and overflow. The most under-discussed use case, and often the cheapest first deployment. Calls that come in when the office is closed, or when the queue is too long. The assistant captures intent, books what it can, escalates what it can’t, and the human team picks up the next morning with a triaged queue instead of a voicemail backlog. For businesses where the alternative is “voicemail nobody listens to,” after-hours overflow is the use case that pays for the whole build.
Order taking in constrained menus. Quick-service restaurants and drive-thrus where the menu is fixed and the order shape is predictable. Works because the catalog is closed (the assistant can’t be asked about something off-menu) and the failure cost is bounded (a wrong order is caught at pickup). Doesn’t generalize to open-ended retail.
Where voice AI still doesn’t pay off
Naming what doesn’t work signals that the rest is honest.
- Complaints, billing disputes, and collections. All three require reading the caller’s frustration, de-escalating, and exercising judgment about exceptions. The model can sound polite without being empathetic in the way the call needs.
- Complex sales conversations. Anything that requires building rapport, qualifying nuance, or improvising around objections. The assistant can take a lead and book a call with a human, but it can’t close the deal.
- Long, context-heavy conversations. Multi-turn calls where the early context shapes later decisions. Context windows degrade across long voice sessions, and so does the assistant’s reliability.
- Anything where the assistant might commit the business to something it can’t verify. Quotes that aren’t fully scoped, promises about delivery times that depend on inventory the assistant can’t see, refunds outside policy. Without strict guardrails, the assistant will agree in the moment to things the business can’t honor.
- Mid-call language switches. The assistant handles one language at a time well. Switching mid-sentence, or handing off between assistants in different languages, still breaks more than it works.
How we build voice AI for the cases that work
When a use case fits the shape, the engineering pattern looks the same across deployments. The voice AI work we ship is built on this stack.
Real-time bidirectional audio sits at the floor. Tool use happens mid-conversation: calendar checks, record lookups, and confirmations all run while the caller is still on the line. The assistant never says “let me look that up and call you back.”
Guardrails define what the assistant can commit to. A booking the assistant can complete is one thing; a discount the assistant can authorize is another. Those limits are scoped at design time and enforced at runtime.
Every deployment includes a fallback to a human. “Let me get you to someone who can help” is a feature, not a failure. Some calls don’t fit, and the assistant’s job is to recognize that quickly and hand off without making the caller repeat themselves.
The operational scaffolding mirrors what every production AI feature needs: transcripts logged, latency tracked, tool-call success monitored, a sampling rate for human review. We’ve covered the production-readiness pattern in detail, and voice AI is one of the features that needs it most, because failures are audible to the caller in real time. The model layer follows the model-agnostic pattern. The streaming side of the architecture is the same shape as the streaming code generator we built for a global user base, with audio replacing text on the response path.
A self-test before you scope a build
Five questions you can answer in under a minute. They translate the four-property frame into a decision.
- How many of your highest-volume call type do you handle per month? If the answer is under a few hundred, payroll is still the right answer.
- Could you write down what the caller wants in five categories or fewer? If yes, the scope is narrow enough.
- What’s the worst thing that happens if the assistant gets it wrong? If the answer involves regulatory exposure, lost revenue beyond the call, or a customer the business can’t afford to lose, voice AI isn’t ready for that call.
- What system has to be live for the call to be useful? Calendar, CRM, inventory, order system: if those don’t have an API the assistant can reach, the build cost is mostly integration, not voice.
- What happens at night and on weekends? If the answer is “voicemail nobody listens to,” after-hours overflow is the cheapest first deployment.
So, do you need a voice AI?
The “can the model talk” question is answered. The question worth asking is which calls in your business fit the shape voice AI handles well right now, and which still need the front desk you already have. If a meaningful share of your call volume hits all four properties, the math has already moved. If it doesn’t, the right move is to wait, not to force it.
If you’re trying to figure out whether your call volume fits, we’re happy to walk through it with you.