For as long as businesses have been sold automated phone systems, the same objection has held: customers can tell, and they hate it. That objection is weakening — and the businesses about to act on it are at risk of buying the wrong half of the solution.
01 — What happened
On 10 September 2026, OpenAI released GPT-Live-1 in its API: a full-duplex voice model. Full-duplex means it listens and speaks at the same time, rather than waiting for the caller to finish before it begins working out a reply. Earlier voice agents were assembled by chaining three components — speech to text, a language model, then text to speech. GPT-Live-1 collapses that chain into a single model that reasons over incoming and outgoing audio together.
Four details matter more than the announcement itself.
It is priced by the second. Voice sessions cost $0.05 per minute, billed per second rather than rounded up. Backend model usage and tool calls are charged separately.
It is built for phone lines. OpenAI’s Live API documents telephony and SIP connection paths directly. Twilio published a working tutorial the same day showing an agent answering a real phone number, in roughly fifty lines of Python, with no custom WebSocket code.
It does not think for itself. GPT-Live-1 handles the conversation and delegates reasoning and tool calls to a separate backend model, chosen independently. The voice layer is deliberately thin.
It is measurably better at turn-taking. Yelp’s chief technology officer, Alex Levy, reported improved turn-taking and accuracy over the company’s previous voice architecture, and better call handling rates on reservation and order calls. Speak, a language-learning app, reported that interruptions during learners’ thinking pauses fell by almost 80% against earlier turn-based systems.
One more line from the documentation is worth holding on to: interrupting the agent’s speech does not automatically cancel the work happening behind it. The application has to handle that. It is a small technical note that tells you exactly what kind of project this is.
02 — Why it matters
The pause was the tell. That half-second of dead air while a system transcribed, thought and re-synthesised was what made callers say “representative” three times. Closing that gap does not make an agent intelligent, but it removes the reason most people refused to deal with one.
So the honest objection has to change. It is no longer “it sounds wrong.” It is “can it actually do anything?”
That is a much harder question, and it is not a question about AI at all.
Here is the trap. Fifty lines of code now produce something that sounds genuinely good on a phone call. Over the next year a great many businesses will be shown a demo of an AI receptionist, be impressed — correctly — by how natural it sounds, and buy it. The quality of the voice now conceals the absence of everything behind it. A system can sound superb and still be unable to tell a caller whether Tuesday at three is free.
03 — The economics
Take a service business with, say, 600 inbound calls a month averaging three minutes. That is 1,800 minutes. At five cents a minute, the voice layer costs about $90. Add the backend model, tool calls and telephony minutes and the real figure is higher — but the order of magnitude holds. The same 1,800 minutes of a person’s attention is thirty hours, and that understates it, because those thirty hours arrive as interruptions to other work rather than as a clean block.
The saving, though, is the least interesting number here. Replacing a person on calls you already answer is a cost-reduction argument, and cost reduction is capped by definition. The uncapped number is on the other side of the ledger: the calls you never answer.
Assume the same business misses one call in six — engaged, out of hours, everyone on site. That is 100 calls a month. Assume one in five of those is a prospective customer who does not ring back. Twenty lost opportunities a month, at whatever an average job is worth. That revenue would have arrived at close to full margin, because the marketing that generated the call has already been paid for.
This is the calculation the announcement changes. Staffing the phone outside business hours to catch those calls could rarely be justified against the value of the ones that came in. At cents per minute, with concurrent session limits starting at 25 and rising to 500 at higher usage tiers, it can be — and capacity no longer collapses at your busiest hour.
Then there is a third number that vendors will not put in the deck: what a wrong answer costs. A missed call is neutral. It annoys someone who may ring a competitor. An agent that confidently books an appointment into a slot already taken, or tells a customer an order shipped when it did not, is worse than neutral — it consumes staff time to unwind and costs trust that was not previously at risk. That asymmetry is the whole argument for spending on the integration rather than the voice.
04 — What businesses should do
Count and classify before evaluating anything. How many calls, at what times, how many unanswered, and about what. Most phone systems export this. If more than a third of your calls are three or four structured requests — booking, availability, order status, opening hours — there is a case here. If most calls are existing customers with complicated, unpredictable problems, there is not.
Apply the look-it-up test. For each of your top request types, ask two questions: does a system hold the answer, and does that system have an API? If availability lives in a diary on the counter, or stock is “whatever Dave remembers,” no voice model repairs that. The agent can only be as accurate as the system it can read.
Start where a mistake is cheap. Out-of-hours enquiry capture and qualification, with a human callback next morning. Overflow when every line is busy. Both fail safely. Taking payment, amending orders and giving out account details belong much later, if ever.
Instrument it from the first call. Record and transcribe, review a sample every week, and track two numbers: how many calls are resolved without a person, and how many escalate. Without this you will discover the agent is failing when a customer complains, which is the most expensive way to find out.
Say what it is. Identify the assistant as automated in the opening seconds and give a clear route to a person. Businesses that hide it will get caught, and the trust cost outweighs any benefit from the deception.
05 — Where technology fits
For a large number of businesses, the honest answer today is: not a voice agent.
If your diary is paper, or a shared calendar nobody keeps current, the agent has nothing to read and will either guess or stall. The sequence is integration first, voice second — never the reverse. That is the same pattern we described in the real cost of disconnected business systems: the tool is rarely the missing piece, the connection between tools usually is.
If your call volume is low and you answer nearly all of it, the economics barely move. Buy nothing. Spend the attention on follow-up after the call instead, which is where most businesses leak more value than they do on the phone.
If you are genuinely missing calls, or losing out-of-hours enquiries, and the answers live in systems that can be queried, then this is a real candidate. What the project actually consists of is three parts: telephony, the model, and the connective tissue between the agent and your systems of record. Only the third part is bespoke, and it is nearly all of the work — which is why it belongs to systems thinking rather than to software purchasing.
It is also worth being clear about what the model does not know. It has no knowledge of your prices, your stock, your staff or your customers unless you connect it to them. A voice agent is an interface. It is not a member of staff, and buying one does not create institutional knowledge that did not already exist somewhere queryable.
06 — RevenueStack’s take
Our view: the significant development is not that AI can hold a natural phone conversation. It is that the conversation has become a commodity input, priced by the second, which shifts the entire advantage to whatever sits behind it. Anyone can now buy a voice that sounds good. Very few businesses have operational data clean and connected enough for that voice to say anything true.
We expect a wave of purchases over the next year, and a meaningful share of buyers discovering they have acquired an articulate, well-spoken way of saying “let me take a message.” That is not a failure of the model. It is a failure to notice that the model was never the part that was missing.
The businesses that get real value from this will not be the ones that move fastest. They will be the ones whose booking system, customer records and order data were already in a state where something could ask them a question — which, as usual, is the unglamorous work that determines whether the exciting work pays.
Sources
- Build more natural voice experiences with GPT-Live-1 in the API — OpenAI, 2026-09-10
- GPT-Live 1 Model — OpenAI, 2026-09-10
- Getting started with GPT-Live — OpenAI, 2026-09-10
- How to Build a Voice AI Agent with OpenAI's GPT-Live, Twilio Agent Connect, and Python — Twilio, 2026-09-10
- OpenAI's GPT-Live-1 Arrives in the API at $0.05 Per Minute — Unite.AI, 2026-09-10
