What an AI voice agent can take off your team
- After-hours and overflow calls. Every call is answered, the details are captured, and nothing waits until morning.
- Intake and qualification. The agent asks the questions your staff ask, in the same order, and writes clean notes to your system.
- Appointment booking and rescheduling. It checks real availability and confirms by text or email.
- Follow-up calls to people who asked to hear from you. Reminders, missed-appointment calls and callbacks to inquiries, with the rules for timing and opt-outs built in.
How a voice agent works
- Phone line. A number from a carrier such as Twilio or Telnyx, or your existing number connected over SIP.
- Speech-to-text, streaming. The caller's words are transcribed as they speak (for example with Deepgram or AssemblyAI).
- The brain. A language model follows your script and rules, and calls tools: check the calendar, look up an order, write to the CRM, transfer to a human.
- Text-to-speech. A natural voice reads the reply back (for example ElevenLabs, Cartesia or Deepgram Aura).
The hard part is not any single piece. It is turn-taking: knowing when the caller has finished, stopping instantly when they interrupt, and replying fast enough that the pause feels human. That is where most of our engineering time goes, along with testing the agent on realistic calls before it talks to a customer.
Platform or custom build?
| Voice platform (Retell, Vapi, Bland) | Custom build on your stack | |
|---|---|---|
| Time to first call | Hours to days | About four weeks |
| Cost per minute | Platform fee plus providers (Retell lists from $0.07 per minute) | Providers only (Deepgram's voice agent API lists $0.075 per minute) |
| Your own rules and systems | Within what the platform supports | Anything with an API |
| Where it runs | The vendor's cloud | Your cloud account, if you need it |
| Code ownership | Configuration only | All of it |
If a platform does what you need, start there; we will tell you so on the first call. A custom build pays off when you have volume, strict rules, sensitive data or systems a platform cannot reach. We often start on a platform and move the agent to a custom stack once the script is proven.
What an AI voice agent costs
Build: from $8,000 for a focused agent (one job, such as intake or booking, connected to one or two systems), scoped as our AI Assistant package. A voice product with its own users, billing and dashboard is an AI MVP, from $12,000.
Running: roughly $0.08 to $0.15 per minute all-in for speech, model, voice and phone line, plus about $1 to $2 a month per phone number. As a worked example, 1,000 calls a month averaging two minutes at $0.12 a minute is about $240 a month. Run your own numbers in the cost calculator, or read how much an AI voice agent costs.
The rules for AI phone calls
In the US, the FCC treats AI-generated voices as artificial voices under the TCPA, so marketing calls to consumers with an AI voice need prior express written consent (see this summary of the 2024 ruling). Inbound calls, and calls to people who asked you to call, are the safe starting point. Every agent we build says it is an AI, honors "don't call me again" immediately, respects calling hours and keeps a transcript of every call. This is not legal advice; check your use case with counsel.
What we have built
We built a voice-first clinical platform where doctors talk with patients and the consultation notes write themselves from the conversation, using streaming speech-to-text with Claude, Groq and Gemini for extraction and summaries. Kortex Labs is founder-led by Saif Ur Rehman, a senior Python engineer with 5+ years and 20+ products shipped, and the person you talk to is the person who builds it.
Where voice agents fit best
- Law firms: intake calls that come in after hours.
- Clinics: booking, rescheduling and reminders.
- Real estate: instant callbacks to listing inquiries.
- Home services and trades: answering the phone while the team is on a job.
- AI receptionist for small business: every call answered and booked.
If the agent mostly needs to answer from documents, see RAG chatbot development; the same knowledge can power both chat and voice.
Frequently asked questions
How do you develop an AI voice agent?
Four parts are joined in real time: a phone line, streaming speech-to-text, a language model that follows your rules and can take actions (book, look up, update the CRM), and text-to-speech. Most of the engineering goes into turn-taking, interruptions and keeping replies fast, then testing on realistic calls before going live.
How much does an AI voice agent cost?
A custom build starts at $8,000, fixed price, scoped as our AI Assistant package; a voice product with its own users and billing is scoped as an AI MVP from $12,000. Running costs are per minute, typically around $0.08 to $0.15 all-in depending on the providers, plus about $1 to $2 a month per phone number.
Which AI model is best for voice agents?
Speed matters more than size. A fast model keeps the conversation natural; a larger model can be called only for the hard steps. We test two or three options on recordings of your real calls and pick by accuracy, latency and cost.
Can I build my own AI voice agent?
Yes. Platforms such as Retell and Vapi let you assemble a simple agent yourself. Teams hire us when the agent must follow strict rules, connect to their own systems, run at lower cost per minute, or live on their own infrastructure.
Will callers know they are talking to an AI?
They should. Our agents say they are an AI at the start of the call, and we build opt-out and do-not-call handling in. Outbound marketing calls with an AI voice to consumers need prior written consent in the US.
Who owns the code?
You do. 100% of the source and IP transfers on delivery, with documentation and a deployment runbook. No license and no lock-in.
Related services: RAG chatbot development · AI development · AI MVP development.