Carbuki Insights
Full-Duplex Voice AI Has Landed. Here Is What It Actually Fixes on a Dealership Phone
Callers put on hold waited an average of more than eight minutes and rated the dealership 19 points below the service average. Callers whose phone rang unanswered rated it 32 points below. Source: CDK Global survey of 2,000+ service customers, February 2025.
A voice model that stops waiting its turn
On July 8, OpenAI released GPT-Live, a new generation of voice models that now power ChatGPT Voice. The headline is architectural rather than cosmetic: GPT-Live is full-duplex. It processes incoming audio while it is generating outgoing audio, and OpenAI says the model decides many times per second whether to speak, keep listening, pause, interrupt, or call a tool.
That reads like an engineering footnote. In a dealership phone room it is closer to the whole ball game.
Ask a BDC manager who has piloted a voice bot what went wrong and the answer is rarely "the answers were wrong." It is that the bot talked over a customer reading a VIN. That it treated a two-second pause as the end of a sentence. That a service customer said "it makes a noise when I brake, kind of a..." and the bot jumped in before the word arrived. The script was fine. The timing was not.
Myth: Customers do not want to deal with AI when they call a dealership.
Data: In a CDK Global survey of more than 2,000 service customers, 31% said they would prefer to book with an AI voice assistant - rising to 44% of millennials and 51% of Gen Z. What customers punish is not automation. It is being ignored: when the phone rang and no one answered, Net Promoter Score for that dealership came in at 26.7, against a service average of 59.
Three generations of voice AI, in plain terms
OpenAI's launch notes are unusually candid about what came before, and the taxonomy is a useful buying framework even if you never touch an OpenAI product.
| Generation | How it works | What the caller hears |
|---|---|---|
| Cascaded (speech-to-text, then a language model, then text-to-speech) | Three models run in sequence for every turn | Long gaps, stilted delivery, information lost between models |
| Turn-based single model | One model handles audio in and out, but in discrete turns; end of turn is inferred from silence | Faster and smoother, but rigid. A pause or background noise can be misread as "you are finished" |
| Full-duplex | Input is processed continuously while output is generated; the model chooses to speak, listen, pause or interrupt many times per second | Interruptions land, pauses are tolerated, and the model can signal it is following along |
Source: OpenAI's description of cascaded, turn-based and full-duplex approaches, July 8, 2026.
Most dealership phone AI in market today lives in row two. That is not a scandal - row two was the state of the art eighteen months ago, and a well-tuned turn-based agent still beats a phone that rings out. But it explains the specific texture of the complaints dealers report, because row two has a structural problem that no amount of prompt engineering solves.
Human conversation runs on a tenth-of-a-second clock
The best evidence for why turn-taking dominates the experience does not come from a vendor deck. It comes from linguistics.
In a study of ten languages across five continents, published in the Proceedings of the National Academy of Sciences, researchers measured the gap between one speaker finishing and the next beginning. The cross-linguistic median was about +100 milliseconds, with individual language medians landing between 0 and +300 milliseconds (Stivers et al., PNAS, 2009). Human conversation is timed at roughly the speed of a blink, and people notice deviations from it without being able to say why.
A silence-based system cannot meet that standard by construction. It has to wait long enough to be confident the caller has stopped talking before it starts. Tune the threshold short and it clips people mid-sentence. Tune it long and every exchange sags, which callers read as a bad connection or a dumb machine. OpenAI names the same trap directly: because turn detection is based on silence, "even a brief pause or background noise could be mistaken for the end of turn."
Now consider who is on the other end of a dealership call. Someone reading a 17-character VIN off an insurance card. Someone standing next to a running engine trying to describe a noise. Someone in a car on speaker with a kid in the back. These callers pause, restart, and correct themselves constantly. They are precisely the population that silence-based endpointing handles worst.
Full-duplex does not repeal the tradeoff so much as move it. Because the model keeps listening while it talks, it can stop when the caller starts, and it can hold the floor without demanding silence to prove the caller is done.
Where duplex earns its keep on a dealership phone
- Number readback. VINs, stock numbers, phone numbers and policy numbers are read in bursts with pauses in the middle. This is the single most common place turn-based agents fail in a way the caller remembers.
- The mid-sentence correction. "Book me for Thursday - no, wait, Friday morning." A caller who cannot interrupt has to start the appointment over.
- Symptom descriptions. Service intake is halting by nature. The customer is thinking and talking at once.
- Noise. OpenAI says GPT-Live is better at focusing on the caller's voice when there is traffic or conversation in the background. Anyone who has taken a call from a service drive knows why that matters.
- Bilingual households. Listening and speaking at once is the same capability that makes live translation practical, which is a real staffing lever in markets where a third of your calls are not in English. We wrote about that market gap in bilingual AI phone agents.
What full-duplex does not fix
This is where a measured read matters more than an enthusiastic one.
You cannot buy it yet. GPT-Live launched inside ChatGPT. OpenAI says it plans to bring the models to the API "soon," and is collecting developer sign-ups. Until that happens, no dealership phone product is running on GPT-Live. If a vendor implies otherwise this month, that is a useful signal about the vendor.
Naturalness is not accuracy. A model that sounds human and books a 7:30 a.m. appointment into a shop that opens at 8 has not helped you. The value is created in the connective tissue - scheduler, DMS, CRM, parts availability, advisor capacity - and none of that is a voice problem. We have argued before that dealers should judge phone AI on outcomes, not on demos.
Compliance does not move. Consent, disclosure and calling rules are unchanged by architecture. A better-sounding agent may actually raise the stakes on disclosure. Our TCPA and AI calling primer still applies line for line.
Escalation matters more, not less. The more human an agent sounds, the more insulting a dead end feels. The exit to a person has to be fast, obvious and reliable.
The number that should decide it
Two data points from CDK's 2026 Friction Points study, released in January, frame the opportunity honestly. Dealer AI adoption climbed to 39% from 28% the year before. Over the same stretch, the Net Promoter Score for buyers at the dealership fell from +48 to +29, and online-only shoppers dropped to zero.
Adoption is up. Satisfaction is down. Whatever dealers are buying, it is not automatically producing a better customer experience - which is exactly why the phone, the most measurable customer touchpoint in the store, is the right place to start and the right place to hold a vendor accountable.
| What to measure | Why it matters | Where most stores are |
|---|---|---|
| Connect rate on inbound calls | A call that never reaches a human or an agent cannot be sold or booked | 40% of service shoppers reported hold, transfer, callback or no answer (CDK) |
| NPS after a hold or no-answer | Phone failure contaminates the whole visit | 39.7 after a hold, 26.7 after no answer, vs 59 average (CDK) |
| Booked appointments per 100 answered calls | The only number that becomes revenue | Rarely tracked by store, almost never by hour |
Five questions worth asking any voice vendor this quarter
- When a caller pauses for two seconds in the middle of a VIN, what does your agent do? Ask for a call recording, not an answer.
- Can the caller interrupt the agent mid-sentence and be understood, or does interrupting simply stop the audio?
- What is your median response latency, and what happens to it in a noisy environment?
- On live traffic - not demos - what is the connect rate, the booked-appointment rate, and the escalation rate?
- When the agent does not know something, what exactly happens next?
A vendor that answers all five with numbers is telling you something. So is one that answers with adjectives.
The bottom line
Full-duplex is a real advance, and the phone room is one of the few places in retail where the advance maps directly onto lost revenue. It is also not yet purchasable as a dealership product, and it does not touch the integration and compliance work that determines whether a voice agent is worth anything. The right posture for a dealer this quarter is neither to wait nor to rush: fix the measurement, run a real pilot on a real call bank, and hold whatever you buy to connect rate and booked appointments rather than to how pleasant the demo sounded.
If you want to see how your store's inbound calls actually perform - answered, missed, booked, escalated - that is the conversation we like having at carbuki.com.
Sources
- OpenAI, Introducing GPT-Live, July 8, 2026.
- TechCrunch, OpenAI releases new voice models for more natural live conversations, July 8, 2026.
- Stivers, T. et al., Universals and cultural variation in turn-taking in conversation, PNAS, 2009.
- CDK Global, How Unanswered Calls Hurt Service Department Satisfaction, February 2025 (survey of 2,000+ service customers).
- CDK Global, CDK Releases the 2026 Friction Points Study, January 2026.
Carbuki builds AI voice agents for retail automotive — answering sales and service calls, following up on leads, and booking appointments 24/7 in multiple languages.
See how it works →