Under 300 Milliseconds: Is Your Phone Becoming an AI Employee?

For a few weeks now, a number has been circulating through the trade press that sounds technical at first glance but has very practical consequences for your company: under 300 milliseconds. That's how fast an AI voice model is now said to be able to respond - faster than most people even consciously notice a pause in conversation. Behind that number is GPT-Live, OpenAI's third generation of the ChatGPT voice model. And although the actual launch already took place on July 8, 2026, the topic is drawing fresh attention in the business press right now, in late summer 2026 - presumably because more and more phone-assistant providers are turning the underlying technology into real products for the Mittelstand.

For you as a business owner, that raises a very concrete question: can your phone now actually become an AI employee that answers calls so fluently that customers barely notice the difference? And if so - what does that mean technically, economically, and above all legally for you? This article breaks down the facts without uncritically adopting the marketing promises.

GPT-Live: what's behind the full-duplex promise

To understand why GPT-Live is particularly relevant for telephony, it's worth looking at the technical innovation. Previous voice models - including ChatGPT's well-known "Advanced Voice Mode" - worked turn-based: you speak, the model listens, the model answers, you listen. Like a walkie-talkie, just with natural language instead of "over."

GPT-Live breaks this pattern. The model works full-duplex, meaning it listens and speaks at the same time. It can interrupt you, let itself be interrupted, briefly acknowledge ("mhm," "I see") while you're still speaking, or deliberately stay silent to give you room. That's exactly the mechanism people constantly use in real conversations - and it's exactly what previous voicebots lacked, which is why they often felt robotic and sluggish.

Important for context: GPT-Live is now the standard voice mode for ChatGPT Go, Plus, and Pro. For users of the free version, a lighter variant, GPT-Live-1 mini, takes on this role. So this is initially a consumer feature - and that's exactly where the first important clarification for businesses begins.

The 300-millisecond mark: a target value, not a guarantee

The widely cited latency mark of under 300 milliseconds specifically refers to the time to the first audio chunk - that is, until the first fragments of the AI's audio response reach the other party - not the entire response duration. OpenAI itself has not published official raw measurements on this so far, which has sparked some discussion in the trade press.

Independent measurements by the provider Agora arrive at a more realistic median of around 300 to 600 milliseconds - depending on network conditions, load, and other factors. What can be independently confirmed, though: p95 latency, the value that nine out of ten requests fall under, is around 25 percent lower for GPT-Live-1 than for the predecessor model, gpt-realtime-2. That's a noticeable technical advance, even though "under 300 milliseconds" is better understood as a best-case target figure than as a guaranteed average - especially once classic telephony codecs with their own latency come into play, for instance on calls over traditional mobile or landline networks.

In practice, this means: the previous response gap of one to two seconds that made typical voicebots feel unnatural is shrinking significantly with full-duplex models like GPT-Live, or its business counterpart. It was precisely that gap that turned many AI phone projects into a gimmick rather than an everyday-usable tool.

ChatGPT Voice is not your phone assistant - the Realtime API is closer

Here lies the biggest misunderstanding currently haunting many summaries: GPT-Live is first and foremost the voice feature within the ChatGPT app for consumers. As far as trade sources report, there's no separate, dedicated business API for it that could simply be plugged into your phone system. "You're now getting ChatGPT on the phone" would be a misleading oversimplification.

What phone-assistant providers actually integrate technically is the OpenAI Realtime API - specifically the models gpt-realtime-2.1 or the more compact variant gpt-realtime-2.1-mini. These run on the same architectural foundation as GPT-Live and benefit from the same full-duplex capabilities and latency improvements, but they're a standalone product that software vendors embed into their phone-system, CRM, or practice-management software solutions.

For you as a company, that means in practice: as a rule, you won't work directly with the OpenAI API, but with a finished product from a phone-assistant provider that has already integrated this technology and equipped it with a phone number, calendar sync, and a user interface. That's not bad news - it makes getting started significantly easier - but it's worth specifically asking, when choosing a provider, which model is actually running in the background and how current the deployed version is.

What an AI phone assistant really costs

The raw API costs alone look vanishingly small at first glance. With the Realtime API, speech is converted into tokens: one minute of user speech input corresponds to roughly 600 tokens, one minute of model response to roughly 1,200 tokens. For the flagship model gpt-realtime-2.1, audio input costs around $32 per million tokens uncompressed, or $0.40 for already cached content, and audio output around $64 per million tokens. The more compact mini variant runs at about a third to a fifth of these figures.

Extrapolated to a minute of conversation, that works out, according to industry analyses, to roughly $0.06 to $0.11 per minute for the flagship model and $0.02 to $0.05 for the mini variant - assuming prompt caching works as intended. That sounds like pocket change. For you as an SME, though, these raw costs usually aren't directly usable, because that would require your own development capacity. Instead, you pay the markup for a ready-made solution.

And that's significantly higher, but still manageable: according to provider information, ready-made AI phone assistants for the German Mittelstand typically cost between 30 and over 500 euros a month, with a realistic range for most businesses of 80 to 200 euros monthly plus a one-time setup fee of 100 to 500 euros. Concrete entry-level offers start at around 99 euros per month including 250 free minutes. Anyone needing deeper, more individual integration into existing phone or practice-management software should expect setup costs of 2,000 to 15,000 euros and total first-year costs of 10,000 to 50,000 euros according to other sources - a considerable range that depends heavily on integration depth and provider.

A frequently cited example calculation from the provider world: a dental practice with around 300 calls a month and an average call duration of two minutes lands on a plan of about 199 euros a month (166 euros with annual payment). A comparable part-time phone staffer, by contrast, costs upward of around 600 euros a month - without covering peak times, evenings, or weekends. The stated savings of over 400 euros a month come from provider marketing material and are not an independently verified study, but they reflect a plausible order of magnitude.

An AI voice that answers in under 300 milliseconds is, for many callers, barely distinguishable from a human anymore - which is precisely why it must be made unmistakably clear in the first seconds of the call that a machine is speaking on the other end.

What you need to clarify legally before you start

That exact point leads to the most important prerequisite before you put an AI phone assistant into operation at all: the legal situation. Since August 2, 2026, the transparency obligations under Article 50 of the EU AI Act have applied bindingly. Anyone who communicates directly with an AI voice system - chatbot, voicebot, or virtual assistant - must be informed by the provider or operator that they're interacting with an AI. Exceptions apply only if this is obvious anyway to an informed person, which is hardly likely to be the case with a deceptively human-sounding full-duplex voice.

Violations of this labeling requirement are no trivial matter: under Article 99 of the EU AI Act, fines of up to 15 million euros or 3 percent of worldwide annual revenue are possible - whichever amount is higher. For a mid-sized company, that number alone should be reason enough to treat the announcement at the start of the call not as an annoying formality, but as a fixed requirement of the project.

That's not the end of it, though. In Germany, a second, often overlooked legal layer applies: the recording or processing of the call, which is technically necessary for AI telephony in most cases, is a criminal offense without the caller's express consent under Section 201 of the German Criminal Code - violation of the confidentiality of the spoken word - punishable by up to three years' imprisonment or a fine. A mere announcement at the start of the call creates transparency in the sense of the AI Act, but doesn't automatically substitute for effective consent. What matters is that the caller actually has the option to continue the call even without a recording being made - a plain "if you keep talking, you agree" formulation is not legally reliable enough.

For outbound calls, meaning when your AI assistant actively calls customers, an even stricter requirement applies: German law requires the express prior consent of the person being called. An existing business relationship alone is not sufficient. So if you're considering automated callbacks or reminder calls, you should clarify this point with the provider in advance, and legally where necessary.

Is it worth it for your business?

Despite all the legal due diligence: AI phone assistants are no longer a vision of the future in 2026, but according to market observers are already in use at over 120,000 German companies - a figure that comes from a single industry source and should therefore be read with appropriate caution, but it does reflect the overall market dynamic well.

Typical use cases according to industry sources are availability outside business hours, appointment scheduling and management with calendar integration, initial qualification of inquiries, and taking callback requests. Providers put the average call duration at 60 to 120 seconds, and the resolution rate - the share of calls that reach the desired outcome without human intervention - at 70 to 90 percent. Both figures come from provider marketing material and are not independently verified industry benchmarks, but they provide a point of reference.

As a rough rule of thumb from practice: an AI phone assistant usually pays off economically from about 15 to 20 inbound calls a day with a high share of recurring standard inquiries, or from around 300 minutes of conversation a month, compared with the cost of an external call service. The real added value often lies less in pure cost savings than in better coverage of evenings, weekends, and staff absences - times when, until now, simply nobody picked up the phone and inquiries were lost.

For more complex or emotionally charged matters, however, the technology still hits its limits - here, routing to a human still makes sense. Anyone working in healthcare, for example, should also keep in mind that appointment inquiries over the phone can indirectly reveal information about treatments. For this particular use case, a clean data processing agreement with the provider and a GDPR-compliant processing location in the EU are especially important - though beyond the general requirements, no additional, specifically verified requirements were found here, so a targeted legal review is advisable when in doubt.

The technical development around GPT-Live and the Realtime API is real and noticeable: full-duplex conversations with significantly reduced latency are increasingly turning what used to be demo gimmicks into everyday-usable tools for phone-heavy industries like trades, medical practices, retail, or services. At the same time, it's worth reading the marketing messages with appropriate distance: "under 300 milliseconds" is a target, not a guaranteed average, "ChatGPT on the phone" is a simplification of the actual technical architecture, and every cost figure from the provider world should be understood as a range, not a fixed commitment.

Anyone wanting to get started now should pursue three things in parallel: the technical provider choice based on your actual economic call volume, clean legal safeguarding under the EU AI Act and Section 201 of the Criminal Code, and a realistic set of expectations for what the AI can take over - and what should remain with your human team. Because even the fastest and most natural voice doesn't replace the tact needed for an upset customer or a complicated special case that doesn't fit into a standard form.

In practice, for rollout, that means: don't start with your entire call volume, but with a clearly defined subset - for example exclusively with calls outside business hours or with pure appointment scheduling. That way, you gather solid experience within a few weeks on resolution rate, actual latency on your specific phone network, and acceptance among your customers, before extending the assistant to further use cases. At the same time, agree in writing with your provider which model is running in the background, where the data is processed, and how consent for call recording is technically implemented - not only once the first complaint or inquiry from a regulator arrives, but as a fixed part of the project from the start.

In the end, the bottom line is sober: the technology is far enough along to hold up in everyday use - no longer just in demos. But whether the step is worth it for your business isn't decided by the latency mark, but by your actual call volume, clean legal preparation, and your willingness to approach the project step by step with realistic expectations.