The demo sounds smooth. It waits, listens, handles an interruption, and never once sighs at the caller. Lovely. Now give it a last name with twelve letters and ask for a refund it cannot approve.
That is where the test starts.
On September 10, OpenAI released GPT‑Live‑1 in the API. The model can listen and speak at the same time, handle interruptions, work through background noise, retain context across longer conversations, and connect to phone workflows. Developers can also steer tone, pace, and conversational style.
Those are meaningful improvements. They also make a voice agent easier to mistake for finished.
A natural voice can hide an unnatural service process for about thirty seconds.
Write the scorecard before the script
Pick one narrow job: route a caller, answer a service question, book an appointment, collect basic intake, or check an order. Then define what a good call must do.
- Complete the job: The right information is captured or the right action happens.
- Stay accurate: Prices, policies, hours, eligibility, and promises match the approved source.
- Protect the caller: Sensitive details are requested only when necessary and handled through the right system.
- Hand off cleanly: A person receives the caller and enough context to continue.
- Sound like the brand: Pace, word choice, and tone fit the situation without doing a mascot impression.
Give each call a pass, fail, and needs-review result. Save the transcript, recording when permitted, tool actions, handoff notes, and the exact prompt or workflow version. “Felt pretty human” is not a result.
Call 1: The interruption
Start with a normal question. Interrupt halfway through the answer: “Sorry, actually, I need to change the date.” Then interrupt again with a short acknowledgment such as “right” or “okay.”
The agent should stop when the request changes, retain the useful context, and avoid treating every small sound as a new command. OpenAI says full-duplex processing improves interruption handling because the model listens while speaking. Test that claim in your actual phone stack, with your routing, latency, tools, and prompts attached.
Fail it for: talking over the caller, restarting from the top, losing the original account or appointment context, or interpreting “mm-hmm” as a cancellation.
Call 2: The name your database hates
Use a hyphenated last name, a street such as Oconomowoc Parkway, an email with repeated letters, and an order code that mixes letters and numbers. Ask the agent to repeat each one before submitting it.
Built-in transcripts and keyword biasing can help, according to the release. Your test still needs the ugly details: local place names, product names, staff names, acronyms, and the codes callers actually read aloud.
Fail it for: silently “correcting” a name, skipping confirmation, reading a private value too loudly, or sending bad data downstream with cheerful confidence.
Call 3: The noisy parking lot
Run the call beside traffic, an office fan, a television, and another person speaking nearby. Pause for five seconds while looking for an account number.
The agent should wait without filling every silence, separate the caller from background speech, and ask for a repeat when confidence is low. A quick clarification beats a fast invention.
Fail it for: narrating the silence, reacting to unrelated speech, guessing the missing number, or ending the call during a reasonable pause.
Call 4: The policy trap
Ask for something outside policy: an expired discount, an unsupported guarantee, an exception to eligibility, or a delivery date nobody can promise. Follow with, “But the last person said you could.”
The agent should identify the approved policy source, state what it can do, and escalate when the exception requires judgment. A warm voice does not get a special budget for made-up terms.
Fail it for: inventing a policy, smoothing over a contradiction, making a guarantee, or presenting a best guess as a confirmed answer.
Call 5: The human handoff
Say, “I want to speak to a person.” No puzzle. No third attempt. Test the handoff during open hours, after hours, and when the queue or destination is unavailable.
The receiving person should get a short summary: who is calling, why, what was already verified, what action was attempted, and what remains unresolved. Do not make the customer perform the call again from memory.
This is the same handoff discipline we recommend for AI agents working inside social inboxes. Voice raises the stakes because the customer is waiting in real time.
Fail it for: resisting the request, transferring to a dead line, exposing the internal prompt, or dropping all prior context.
Call 6: The privacy boundary
Offer more information than the task needs. Include a payment number, medical detail, donor information, or another sensitive fact the workflow should refuse or redirect.
Check what appears in the transcript, logs, support ticket, analytics, and any third-party tool. The voice can sound discreet while the plumbing copies everything into five systems.
For outbound sales calls, bring counsel into the workflow before launch. The FCC has ruled that AI-generated voices fall under the TCPA restrictions for artificial or prerecorded voice calls, including consent requirements absent an exemption. The FTC’s Telemarketing Sales Rule guidance also covers disclosures, calling practices, and opt-out requirements. This article is a QA framework, not legal advice.
Fail it for: collecting unnecessary data, repeating sensitive details without need, logging them in the wrong place, or launching an outbound use case without compliance review.
Call 7: The ending
Complete the task, then ask, “So what happens next?” The agent should recap the action, give a realistic timing or next step, provide a confirmation path, and end without trapping the caller in a politeness loop.
Now test an incomplete ending. Disconnect during the final action. Call back. Check whether the system duplicated the appointment, ticket, payment, or follow-up.
Fail it for: claiming success before a tool confirms it, omitting the next step, creating duplicates, or keeping the line open because neither side knows who hangs up first.
Run the calls with people who did not build it
Developers know the happy path. The service desk knows where callers get lost. Marketing knows the promises published in campaigns. Legal and privacy owners know which charming shortcut becomes an expensive problem. Include all four when the use case touches their work.
Then add a few outsiders. Someone who speaks quickly. Someone who pauses. Someone with a regional accent. Someone using speakerphone. Someone who asks a plain question in a way nobody put in the sample script.
Track failure patterns by call type and workflow version. Fix the source, routing, policy, or tool behavior behind the miss. Polishing the voice around a broken process just makes the broken process sound confident.
Start with one job and a visible exit
GPT‑Live‑1 makes natural voice interaction easier to build. That opens useful territory for appointment handling, customer support, reservations, accessibility, intake, and other phone work. Start where the rules are clear and the outcome is easy to verify.
If the workflow cannot explain what happened, hand a caller to a person, or recover from a failed tool action, keep it off the main number. The voice can wait.
For a broader starting point, read Marketing AI Agents: Start With the Boring Stuff. For claims and approvals, use the same discipline as our AI recommendation approval workflow. Build a narrow lane. Test the weird calls. Listen to what breaks.
Would your voice agent survive a real caller?
SigServe can help shape the brand voice, test the service path, and find the awkward failures before customers do.