Many people have had this experience: you call a customer-service hotline, an AI voice system answers, and you have to wait until it finishes an entire sentence before you can speak. Interrupt halfway and it either doesn't hear you or answers the wrong question, until you end up repeating "agent, agent." The problem is not that the AI isn't smart enough; it is that the conversational model of traditional voice systems is too rigid. On 10 September 2026, OpenAI released GPT-Live-1 in its developer API, using full-duplex technology so the AI can listen and speak at the same time. This article explains what the technology changes, what a business needs in place to build AI customer service that actually works, how to calculate the cost, and what to watch out for when rolling it out.
What is full-duplex?
"Duplex" is a communications term. A walkie-talkie is half-duplex: only one side can talk at a time, and the other side can respond only after the first lets go of the button. A telephone is full-duplex: both sides can speak and listen at the same time. Most AI voice systems until now have effectively worked like walkie-talkies. They are usually built as a chain of three steps: convert the caller's speech to text, pass it to a language model to compose a reply, then turn the text back into speech. Each step adds latency and strips out information about rhythm, pauses and tone; the system may mistake a brief pause for the end of a sentence, or have already "decided" what to say by the time the caller interrupts. A full-duplex model keeps incoming and outgoing audio flowing simultaneously, so the AI can continuously decide whether to keep listening, keep speaking, pause, or respond to an interruption.
What does GPT-Live-1 change?
According to OpenAI, GPT-Live-1 cuts turn-taking latency from about 1.4 seconds on the previous GPT-Realtime-2.1 to about 0.8 seconds; its score on a full-duplex interactivity test rises from 45.4% to 80.1%; and accuracy when calling tools mid-conversation, such as querying a system or taking an action, rises from about 60% to about 87%. It handles interruptions, backchannel cues like "mm-hmm" and "right," and callers changing direction halfway through. Its handling of silence and background noise is also improved, reducing false interruptions caused by normal pauses or ambient sound. Especially useful for customer service, it supports keyword biasing and stronger recognition of letters and numbers, which helps with names, order numbers, reference codes, addresses and product models, and it supports telephony integration. Testing by language-learning platform Speak found that switching to GPT-Live-1 cut the AI's interruptions during learners' thinking pauses by almost 80%.
A key design: splitting the voice layer from the backend
Another design point businesses should note is that GPT-Live-1 does not have to do all the thinking itself. It keeps the live conversation responsive and natural, while handing work that needs more compute, data access or longer reasoning to other models or tools in the backend. For example, a simple booking enquiry can go to a lower-cost model, while a complex account dispute can go to a stronger model such as GPT-6 Astra, and third-party models can be plugged in as well. This means a call does not need to run the most expensive model for its entire duration: the AI keeps talking with the customer while the backend searches databases, reads account data and runs workflows, so the customer never hears an awkward "waiting for the machine" silence.
Five essentials for building AI customer service
Technology is only the foundation; whether AI customer service succeeds depends on the overall design. First, the conversation must feel natural: use the full-duplex capability so callers can interrupt and correct at any time, and use system instructions to set tone, pace and speaking style in line with your brand. Second, it must get things done, not just answer questions: connect the AI to your order system, booking system, membership database and ticketing system, so it can look up orders, change bookings and open cases during the call and actually resolve problems. Third, it needs a solid knowledge base: organise product information, pricing, FAQs, return policies and terms of service into clear, up-to-date content so the AI answers accurately and the risk of made-up answers falls. Fourth, design rules for handing over to a human: complaints, refunds and anything involving money or sensitive personal data should have clear escalation conditions, so the AI handles the high volume of everyday enquiries while staff focus on complex cases. Fifth, keep records and check quality: GPT-Live-1 outputs a transcript of the caller's speech alongside the AI's response text, so you can spot-check calls, analyse common issues and keep improving the knowledge base and processes.
How should you calculate the cost?
GPT-Live-1's voice layer is priced at USD 0.05 per minute, billed per second. On that basis, 100,000 minutes of calls a month costs about USD 5,000 for the voice layer; with an average call of five minutes, that is about USD 0.25 per call. Note, however, that this is only the voice-layer price: backend model reasoning, tool calls, data retrieval and the surrounding infrastructure are all charged separately. So when comparing costs, the per-minute price alone is not enough; a more meaningful metric is cost per resolved case. A system that costs slightly more per minute can still be cheaper overall if it finishes conversations faster and reduces escalations and repeat calls. OpenAI also cited an early customer that replaced its multi-stage voice stack with GPT-Live-1 and cut the relevant code by about 80%, or roughly 23,000 lines. That is a single case, but it shows that lower system complexity also affects total cost.
Best-fit use cases
Full-duplex AI customer service suits phone work that requires talking and completing actions at the same time, including general customer enquiries and after-sales support, table and appointment bookings for restaurants or clinics, membership and order-status enquiries, billing reminders and collections, initial screening of sales leads, and internal IT or HR help desks. For businesses with high call volumes or enquiries outside office hours, AI customer service can answer 24 hours a day, so customers no longer have to wait on hold or leave a message.
Rollout steps and cautions
We recommend a phased approach. Step one: choose a single hotline with high call volume and fairly repetitive questions, such as a booking line, or start with after-hours calls as a pilot. Step two: organise the knowledge base, connect the necessary systems and define the conditions for transferring to a human. Step three: test thoroughly. OpenAI's figures are launch-time vendor results, and real-world performance must be verified in your own environment, including different accents, noisy settings, stability over long calls and tool-call success rates. For Hong Kong businesses in particular, test how well the system handles Cantonese and mixed Chinese-English conversation before deciding the scope of service. Step four: ensure compliance and transparency, telling customers clearly that they are speaking with an AI and handling call recordings and personal data in line with privacy rules. Finally, keep tracking resolution rate, escalation rate and customer satisfaction, and expand the AI's scope step by step.
Conclusion: success depends on design
Full-duplex technology finally lets AI customer service converse as naturally as a person, but what really decides success is still the knowledge base, system connections, escalation rules and continuous improvement behind it. The technology is ready; the question is whether a business can design it into a service that genuinely solves customers' problems. To learn how to build AI customer service for your company and deploy AI staff into real business workflows, visit ai.ud.hk to explore UD's AI Staff solutions.
懂AI，更懂你｜UD相伴，AI不冷
