Back to News

AI · Sep 11, 2026

AI voice agents can now handle real phone calls. Should yours?

On September 10, OpenAI made GPT-Live-1 available in its API. It is a voice model that listens and speaks at the same time, handles interruptions, and supports telephony, which means developers can put it on an actual phone line.

For a small business, the phone is still where a lot of revenue starts. A missed call after hours is often a lost job. So the question of whether an AI should answer is a fair one, and it deserves a careful answer.

What changed with this release

Older voice bots worked like a relay race. Speech was converted to text, the text went to a language model, the reply was converted back to speech. Each handoff added delay, and the bot fell apart when a caller interrupted or changed their mind halfway through a sentence.

OpenAI says GPT-Live-1 handles listening and speaking in a single model, deals better with background noise and pauses, and can hand off deeper reasoning and tool use to a separate model in the background while the conversation continues. It also supports phone calls directly.

The company quoted early users. Yelp's CTO said call handling rates improved when its reservation and ordering product used the model. Those are vendor-selected examples, but they point to the right use cases: structured, repetitive calls with a clear goal.

Where a voice agent makes sense

The calls that fit an AI voice agent well share a few traits. The caller wants something specific. The answer lives in a system you already have. And the stakes of a small mistake are low.

  • Booking, confirming, or rescheduling appointments
  • Answering hours, location, and service area questions
  • Taking a message with the right details and routing it to the right person
  • Order or job status checks pulled from your system
  • After-hours intake so a lead gets a callback first thing in the morning

Where it does not, at least yet

Some calls need a person. An upset customer. A quote that depends on judgment. Anything involving payment disputes, legal questions, or medical details. A good voice agent should recognize these and hand off quickly, not try to talk its way through.

The worst outcome is not an AI that says "let me get someone for you." It is an AI that confidently gives the wrong answer about pricing or availability.

How to set one up responsibly

  1. Start with after-hours only. You learn how it behaves without risking your busiest calls.
  2. Connect it to real data. A voice agent that cannot see your calendar or job system will guess. Guessing on the phone is worse than voicemail.
  3. Write the handoff rules first. Decide which topics always go to a person, and test those paths before going live.
  4. Tell callers they are talking to an AI. It builds trust and, depending on where you operate, may be required.
  5. Review transcripts weekly. For the first month, read a sample of calls. You will find gaps fast.

What this means for your business

  • Voice AI is now good enough for routine, structured calls. It is not a replacement for your best people on your hardest calls.
  • The value comes from connecting the voice agent to your scheduling, CRM, or job system. That integration is the real project.
  • Start small, measure missed calls recovered and handoff quality, then expand.

Answering the phone is one of the most common places owners lose time and leads. If you want to explore a voice agent that is wired into your actual systems, that is the kind of automation we help with at Centervert.

Sources