Real-Time AI: The Breakthrough Is a Listener, Not a Talker

An episode of Dan's AI Intel

The wait is ending by splitting the model in two — and the real-time AI with measured returns is the one that listens rather than speaks.

Published · By Dan Walter

Executive summary

The waiting is ending, but not because the models got faster. It is ending because the industry quietly split one product into two machines: a cheap, purpose-built talking layer that listens and speaks inside human conversational time, bolted to a slow reasoner that keeps working behind it. NVIDIA's open PersonaPlex switches speakers in 0.07 seconds; a maxed-out reasoning model like Claude Opus 5 takes around 59 seconds to produce its first token on the same independent benchmark. Those are not two settings of one dial. They are two different organs, and 2026 is the year products stopped pretending otherwise — OpenAI's GPT-Live-1 answers you while handing the hard part to GPT-5.5 mid-sentence, and GPT-6 Instant starts talking before it has finished thinking.

The uncomfortable finding is where this pays off. Every deployment with measured, published returns is a listener, not a conversationalist: ambient scribes that sit in a consultation and save a clinician 13 to 14 minutes a day, live sales prompts, meeting assistants that stay silent until asked. Where the AI does the talking — phone support, the flagship consumer case — independent reports put the share of contacts it finishes alone at roughly 5% to 25%, against vendor case studies claiming 70% to 90%. The unit economics agree: speech out costs about four times speech in and runs to roughly 80% of the bill, so the cheapest real-time product listens a lot and speaks rarely.

That makes the near future less "an AI you chat with" and more "an AI that is present." The talking layer is now cheap to build — PersonaPlex is a 7-billion-parameter model trained in six hours on eight A100s — so the constraint has moved to trust: a model that can be interrupted can be hijacked by the interruption. Spoken-interruption attacks published in September 2026 lift jailbreak success on open duplex models to 40–49% purely by timing when the attacker speaks. Speed is solved enough; whether we let the fast layer act is the 2027 question.

Why a few seconds of waiting is an AI-revolution problem

For three years the dominant way humans have met this technology is a text box with a spinner: ask, wait five seconds to several minutes, come back. That shape is why most people's relationship with AI is transactional — you do not converse with a vending machine.

Latency decides which jobs an intelligence can hold. At a few seconds, AI is something you consult; under a quarter-second it can sit in a meeting, a consultation, a car or a phone call and participate — the first time the technology touches the parts of life that are live and unrepeatable. Pew's February 2026 survey of 5,119 US adults found about half now use chatbots and roughly one in four use them daily. The audience is already there. The machine that can answer at conversational speed is what has been missing.

The wait reorganised how we work before anyone measured it

The coping strategy arrived before the fix. People do not watch the spinner; they start something else. Running three to five things in parallel while each one thinks is now the normal rhythm of working with AI — so universal that it has broken the instruments we use to measure the technology's value.

The friction has five parts, and only the first is speed: the wait itself, anywhere from 0.3 seconds to nearly a minute depending which model answers; the switch you make to fill it; the cost of rebuilding your own context when you come back; the verification tax on anything the model did while you were away; and the inability to interrupt or steer an answer already in flight. Duplex models attack the last of those directly. Nothing yet touches the verification tax.

METR — which ran the most-cited controlled study of AI's effect on experienced developers — published an update on 24 February 2026 explaining why it was redesigning its experiment. Developers, it reported, "would often work an unrelated task while waiting for the agent to complete its work," which made self-reported task time meaningless; the authors concluded their "measurements of time-spent on each task are unreliable" for the fraction of developers running multiple agents at once. The study that found AI made developers 19% slower in 2025 could not cleanly measure them in 2026, because the waiting had been absorbed into multitasking.

That is a strange fact. Wait-and-switch never shows up as lost time on a spreadsheet; it shows up as fragmented attention, the one cost nobody bills. Perception follows: in the Pew data three in ten US adults say chatbots help their productivity, 5% say they hurt it, and most say neither — exactly the shrug you would expect from a technology that saves minutes and quietly taxes attention.

Humans hit 200 milliseconds by predicting, not by being quick

Here is the mechanism nobody designing voice interfaces can escape. Across ten unrelated languages, the gap between one speaker finishing and the next beginning is about 200 milliseconds — but planning a spoken utterance takes a human brain well over 600 milliseconds. The arithmetic only works if listeners predict where your sentence is going and start building their reply before you stop talking.

Real-time conversation, in other words, was never about being fast. It is about overlap: listening and preparing at once. A system that waits for silence, transcribes, thinks, then speaks can only be polite-but-late, however many milliseconds each stage shaves. One that processes while the other party is still talking can sit inside the human gap with unremarkable components.

The stack split in two: a mouth at the front, a mind at the back

Through 2026 every serious player converged on the same architecture, and the clearest statement of it was OpenAI's. On 8 July 2026 it released GPT-Live-1 and GPT-Live-1 mini, replacing the old voice mode, which had chained a speech-to-text model, a language model and a text-to-speech model into one laggy pipeline. The new models are full-duplex: they speak and listen simultaneously, so you can interrupt them. Crucially, they hand queries needing search, reasoning or agentic work to a bigger text model — GPT-5.5 — while the conversation keeps going. ChatGPT's voice lead described 30-to-40-minute walking conversations, and a mode where the model stays quiet for long stretches, absorbing context until spoken to.

That handoff is the whole design. On 7 October 2026 OpenAI shipped GPT-6 Instant, which per its release notes starts answering 44% sooner than its predecessor on questions that need a web search, and can begin replying while it continues to think and use tools. Behind it sit GPT-6 Astra, Sol and Luna — flagship, cost-efficient high end, fast tier — with routing inherited from GPT-5. Apple took the same shape from the other direction: the rebuilt Siri runs on a Gemini backend with a handoff to ChatGPT. Amazon's Alexa+ sends every command to the cloud and sells Echo Frames built around it.

Exhibit — The wait didn't disappear — it moved behind a fast front end. Nothing got 800x faster; the slow work was moved off the critical path of the conversation, which is why the experience changed before the reasoning did. Source: TechCrunch on GPT-Live-1 (8 Jul 2026); OpenAI release notes, GPT-6 Instant (7 Oct 2026); Artificial Analysis provider latency (Oct 2026). Compiled by Dan's AI Intel.

The split is not merely an OpenAI product decision. TypeSafe AI's Jev, launched 15 September 2026 and the subject of our Jev episode, number 61, is a deliberately non-conversational "System One" model: it makes small decisions in a claimed 70 to 500 milliseconds and is built to sit beside a large language model rather than replace it. An independent reproduction of its mechanism on a different open model measured 463 milliseconds per decision against 848 for standard generation — a real gain, roughly half, and a long way from the sub-15-millisecond figures some developer guides repeat without a source.

Only two of the three speed tricks are a real step change

Three techniques get the credit, and they are not equal.

Native duplex audio is the genuine break. Dropping the transcribe-think-speak chain for a single model that consumes and emits audio removes both the hand-offs and the need to wait for silence. PersonaPlex, released open by NVIDIA in January 2026, is the clearest public demonstration: 7 billion parameters, built on Kyutai's Moshi, 0.07-second speaker switches against 1.3 seconds for Gemini Live — inside the human turn gap, where a turn-based assistant is six times outside it — a dialogue-naturalness score of 3.90 against 3.72, and 99.2% smooth speaker changes across 350 customer-service questions in 50 role scenarios. It took six hours on eight A100s to train, and it is English-only.

Other figures fill the middle — a fast non-reasoning text model reaches its first token in about 0.32 seconds, production phone agents at 680 to 850 milliseconds median over 100 concurrent calls — but a speaker switch, a first token and a voice-to-voice round trip are not the same quantity, so read them as orders of magnitude.

Parallel decoding is the second. Generating one token at a time is the floor on conversational speed; everything that breaks that order — speculative decoding, non-autoregressive sampling, diffusion-style language models — lifts it. A trajectory-level speculative decoding method presented at ICML 2026 cut denoising iterations 30–40% and raised tokens per step from 2.6 to 4.3, reporting 7–14x speedups over vanilla diffusion language models with under 1% accuracy change. It shows up on the independent leaderboards: the fastest models measured by Artificial Analysis are Celeris-1 at around 1,500 tokens per second and the diffusion-based Mercury 2.5 at roughly 720–820.

On-device inference is real but small. Qualcomm's Snapdragon AR1+ Gen 1 runs small language models directly on smart glasses without a phone, which makes wake-words and simple commands instant. Multi-turn reasoning still goes to the cloud, and the honest field number for conversational inference on glasses is 800 milliseconds to over 4 seconds — five to fifty times the laboratory figures. A large slice of that is network, not model: geo-steered signalling and relay infrastructure shave round trips no decoder trick can touch, and a buyer cannot inspect any of it.

Where real-time AI already pays, it says nothing at all

Now the part that gets lost in demos. If you sort real-time deployments by the quality of their evidence rather than the volume of their marketing, the winners are all machines that listen in the moment and deliver afterwards.

Ambient clinical scribes are the most studied case. A 2026 prospective study in JMIR Medical Informatics found high users of Abridge cut documentation time 21% — about 13.6 minutes a day, plus 3.6 minutes off after-hours charting — with no significant change in burnout. Cleveland Clinic reported Ambience saving about 14 minutes a day; Cooper University Health Care reported 4.15 minutes per patient with Microsoft's Dragon Copilot, roughly an hour a day. At Sutter Health, burnout fell from 42.1% to 35.1%, a drop that missed statistical significance. The pattern is modest and consistent: minutes, repeatedly, with the human effect still contested.

Exhibit — The evidence is strongest exactly where the AI stays quiet. The columns run in opposite directions: the more the system talks, the thinner the evidence that it works. Scribes (Abridge, Ambience, Dragon Copilot) have peer-reviewed baselines; phone agents (Sierra, Decagon, Parloa) have case studies; Alexa+ and Siri have neither. Source: JMIR Medical Informatics (2026); AHA Market Scan (14 Apr 2026); vendor case studies and a practitioner review of live deployments; editorial assessment of public evidence, not a measurement. Compiled by Dan's AI Intel.

The same shape holds at work. The meeting tools people keep are listeners: Granola records from the desktop with no bot in the call and improves your rough notes afterwards; Zoom AI Companion and Microsoft Copilot in Teams answer questions mid-call when asked. The genuinely live, prompting-in-your-ear products cluster in sales — Clari Copilot pushing battlecards, Winn.ai prompting and updating the CRM, Otter giving real-time guidance and writing the follow-up — because a sales call has a known script and a measurable outcome, which is what a live suggestion has to be judged against. Nobody has shipped a credible "AI that helps you in a general meeting": no one can say what a good intervention would be.

Customer service is the honest test, and the claims shrink under it

Phone support is where real-time AI meets a number someone has to defend. The headline metric is containment: the share of contacts finished without a human.

Vendors and their customers report the top of the range. Sierra and its case studies cite 70% to 90% for specific implementations. Decagon cites 50% to 80% for mature deployments and more than 50% voice deflection at the mortgage servicer Valon. Parloa's own page reports BarmeniaGothaer cutting switchboard workload by 90%. Then the independent-leaning numbers arrive: a practitioner review of live deployments puts PolyAI at about 25% containment at Simplyhealth and 5% to 21% over a year at Quicken, and argues that many companies remove only a fifth to a quarter of calls from human queues. A statistics roundup of early adopters lands in the middle at 45% to 65%.

Exhibit — Containment falls by two-thirds as the source gets more independent. Parloa's figure is BarmeniaGothaer's switchboard; the two lowest are PolyAI deployments reported by a practitioner review. Containment has no standard definition — the one number buyers compare is the one nobody measures the same way. Source: Vendor case studies (Sierra, Decagon, Parloa); practitioner review of live deployments; 2026 voice-AI statistics roundup. Compiled by Dan's AI Intel.

No audited study measures containment across these vendors — and that absence is the signal: in the one real-time application with millions of daily conversations, the public evidence is almost entirely self-reported.

Listening is cheap, speaking is dear — and the product follows the meter

The economics explain the design better than any roadmap. Realtime audio is priced per token and the asymmetry is brutal: roughly $32 per million audio input tokens against $64 per million output, cached input near $0.40. At about 600 tokens heard and 1,200 spoken per minute, listening costs about 1.9 cents a minute and speaking 7.7. An analysis of around 4,000 measured sessions found audio output alone was about 80% of the bill.

Exhibit — The moment the model opens its mouth, the cost jumps about fourfold. At 1,000 daily users a flagship always-on assistant runs to roughly $1,000 a day; the mini tier cuts it to about $320. The design that survives contact with a budget is one that listens by default and speaks on demand. Source: OpenAI realtime audio token pricing as reported by multiple 2026 implementation guides; analysis of ~4,000 measured sessions. Compiled by Dan's AI Intel.

Stack platform fees on top — Vapi charges about 5 cents a minute before the models, Retell 7 to 31 cents all-in — and the viable product's shape is set. Always-on listening is affordable; always-on talking is a budget line that scales with every user-minute. This is why the assistant in your glasses answers briefly and the scribe in the consultation room says nothing: not taste, arithmetic.

That default carries a politics. A product built to listen constantly and speak rarely is, from outside, indistinguishable from one built to record you — and the installed base is already domestic: about a third of US adults keep a smart speaker, one in five an AI doorbell. Pew found majorities believe AI is advancing too quickly and will put their personal information at risk, younger adults included. The fights that follow are specific: consent and disclosure when the voice on a live call is synthetic, whether an always-on assistant's audio is kept or discarded, and what becomes of contact-centre employment in the economies that bank on it. Duplex systems also destroy the written artefact — speech in, speech out, no transcript — removing the audit trail regulators reach for, just as the labs race these models into glasses, earbuds and cars.

What it still can't do: the interruption it invited

The capability that makes duplex models feel human — being interruptible — is also their new vulnerability, and the evidence arrived in September 2026. A paper titled DuplexJail tested four open full-duplex models, PersonaPlex and PersonaPlex-RL among them, against 720 harmful requests from AdvBench and HarmBench by speaking a fixed jailbreak prompt while the model was still generating. Timing was the attack. At a one-second delay, whole-response attack success rose to 40.3% and 48.7% — up 33.8 and 39.3 percentage points on baseline; on the held-out benchmark a half-second delay added 14.7 and 12.3 points.

Safety alignment was trained on whole turns, and these models no longer have whole turns. The rest of the list is less dramatic and equally real: there is no agreed evaluation protocol for duplex systems, and the ICASSP 2026 HumDial study found timing degrades in multi-speaker rooms, transient noises trigger false activations, and models struggle to resume a task after interruption. Reasoning over voice still lags reasoning over text, and agentic reliability is worse again: the best computer-use agents were scoring around 10% on comprehensive real-world task suites in early 2026.

So the honest division of labour: conversational AI is now good enough to be present, and not yet good enough to be trusted to act. Deep agentic work will stay slow, and that is fine — nobody needs an hour-long refactor to answer in 200 milliseconds.

On the same evidence, what it will do next is narrow and plausible: brief you continuously from glasses and earbuds, both of which Apple is reported to be shipping by 2027; translate a live meeting; and run scripted phone work end to end, which containment numbers show it half-does today. What it will not do on this horizon is hold a long-horizon task inside a conversation, match text-grade reasoning aloud, or earn the right to irreversible actions — Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and the audit-trail problem gets worse, not better, as the interface becomes speech.

Bottom line

Real-time AI stopped being a speed problem in 2026 and became an architecture: a cheap duplex front end that listens continuously and speaks sparingly, with the slow, expensive thinking moved off the conversation's critical path. The breakthrough is presence, not chat. Three hinges decide how far it goes from here. First, action rights: whether the fast layer is ever allowed to execute rather than relay — today it hands off, and the agentic reliability numbers say it should. Second, the cost curve on generated speech, because at four times the price of listening, "always talking" stays a premium feature. Third, safety under interruption: the DuplexJail results are early and on open models, but if alignment cannot be made robust to being spoken over, the most natural interface we have built will stay fenced off from anything consequential.

What would change this read? A duplex model that holds its safety behaviour under timed interruption, an audited containment study in customer service, or audio generation costs falling by an order of magnitude — any one of those moves the fast layer from a mouth to a decision-maker. Watch those three, not the demos.

Sources

Transcript

Alex: Okay, a machine just switched speakers in seventy milliseconds. Your own brain can't close a sentence that fast.

Sam: And the real-time AI product that's actually making money right now doesn't say a single word. So, this week I was reading about a health insurance agency that just moved its AI voice agent out of pilot and into handling real customer calls full time.

Alex: The stage every one of our own AI rollouts at Fernloft gets stuck in for months before Legal will even let it near a real customer.

Sam: That's it exactly. Not a demo anymore, an actual phone line, with their actual customers on it.

Alex: Huh. That's a real line to cross.

Sam: It is. And it's exactly the kind of move my own PR team would fight to keep quiet, because the second it's official, every customer on that line knows they're talking to a machine.

Alex: Low drama, real consequence.

Sam: Right? Anyway, that's been rattling around in my head all week, because it's this whole episode in miniature.

Alex: Which, speaking of. Welcome back to Dan's AI Intel, the show that tries to get underneath the fastest, strangest shift probably any of us will live through.

Sam: I'm Sam.

Alex: I'm Alex. Every week, one question that actually matters, past the hype and past the fear, to what's actually happening, and what it means for the rest of us.

Sam: This week's trigger is almost comically small. OpenAI just shipped a voice model that starts answering you before it's finished thinking.

Alex: Which sounds like a party trick. Except it's the last piece of something that's been building quietly all year.

Sam: And the question I actually want answered isn't "is it faster now." Everyone agrees it's faster. It's: if a machine can really be fully present, listening and talking at once, who benefits from that? Because the honest answer surprised me.

Alex: So we're going right through the architecture, how OpenAI, and nearly everyone else, quietly built two machines wearing one face. Then into where this is already making real money today, versus where the big claims fall apart the second someone independent checks them.

Sam: And we land somewhere I did not expect going in. The exact thing that makes this interface finally feel human is the exact thing that makes it hackable. Stick around for that part.

Alex: Quick thing before we get into it. If you're enjoying this, there's a follow button wherever you're listening. Costs nothing, one tap, and you won't miss the next one.

Sam: Can we start with what actually bugs me about this? For three years our relationship with this stuff has been: type a question, stare at a spinner, wait.

Alex: Pew asked a few thousand adults in February. About half of them now use a chatbot. The audience showed up. I feel the missing half of that myself every time I try to actually talk to our own support bot and give up and just type instead.

Sam: Nobody sat around waiting for it, either. On my team, the second a draft's thinking, people open something else. We're running three or four things in parallel without ever deciding to.

Alex: That's exactly what METR flagged earlier this year. They run the biggest controlled study on whether AI speeds up developers, and they had to redesign it, because engineers kept quietly working an unrelated task while the agent ran.

Sam: So their own stopwatch broke.

Alex: Their own words: the time measurements had become unreliable. The same lab that found AI made developers nineteen percent slower a year ago couldn't cleanly measure them this time, because the waiting had already been absorbed into multitasking.

Sam: Which is such a strange invisible tax. It never shows up as lost time on anyone's spreadsheet.

Alex: It shows up as fragmented attention instead. And people mostly shrug about it, honestly, which is exactly what you'd expect from something that saves you minutes and quietly taxes your focus at the same time.

Sam: "Quietly taxes your focus." That's going on somebody's performance review. Every junior person on my team is doing this now without even noticing.

Alex: And here's the part that actually explains everything else today. Across a pile of totally unrelated languages, the gap between one person finishing a sentence and the next one starting is about two hundred milliseconds.

Sam: That's nothing.

Alex: It's nothing, except planning what you're going to say takes a human brain three times that long. So the arithmetic only works one way. You're not reacting to me finishing. You're predicting where my sentence is going and loading your reply before I stop talking.

Sam: Which means a system that politely waits for silence, transcribes it, thinks, and then answers —

Alex: Can only ever be polite but late. However many milliseconds you shave off each stage.

Sam: So speed was never actually the goal.

Alex: Overlap was the goal. Listening and preparing at the same time. And that's the one thing none of the old voice assistants could do, because they were three separate boxes bolted together.

Sam: Which is what actually changed over the summer, right? OpenAI ripped out the old voice mode entirely.

Alex: Completely new models. The old pipeline was speech to text, then a language model, then text to speech, three hops, each one adding lag. The new one is full duplex. It speaks and listens at the same time, so you can genuinely interrupt it mid-sentence.

Sam: And the clever part isn't even the speed. It's that the fast model hands the actual hard thinking to a bigger one.

Alex: Mid-conversation, without you noticing the handoff. Their own voice lead talked about long walking conversations where the model just stays quiet, absorbing context, until you actually speak to it.

Sam: And then recently they went one further.

Alex: Their newest instant model starts answering something like forty percent sooner on anything that needs a web search, and it can start replying while it's still thinking in the background.

Sam: Okay, that one I want to put on your desk, because Fernloft's own support bot does something a little like this now, doesn't it?

Alex: A little. Nowhere near this clean. We still have one model doing the whole conversation. Nothing's handed off mid-sentence. Which, hearing myself say that out loud, sounds a lot more primitive than I thought it was this morning.

Sam: It's not just OpenAI doing this, though.

Alex: Not at all. Apple did the same split from the other direction. The rebuilt Siri runs on a Gemini backend now, with its own handoff to ChatGPT for the hard stuff. Amazon's doing it too, routing every Alexa+ command straight to the cloud, and selling glasses built around exactly that loop.

Sam: This isn't just a voice-assistant thing either, is it? There's that other one we did a whole episode on.

Alex: Jev. TypeSafe AI's thing, back a couple of months now. It's deliberately not conversational. A "System One" model that makes small decisions almost instantly and just sits beside a big language model instead of replacing it.

Sam: And somebody actually checked that claim.

Alex: Somebody rebuilt the mechanism independently and measured it at roughly half the time of generating a normal response. Real gain. Not the absurd number some guides repeat with no source behind it at all.

Sam: Which is a great general rule for this entire topic. There's three tricks everyone credits for all this speed. Are they all equally real?

Alex: Not even close. The real one is native duplex audio, one model that just consumes and produces audio directly, no hand-offs. NVIDIA open-sourced the clearest example of it this year, a model called PersonaPlex.

Sam: The one that switches speakers in —

Alex: Seventy milliseconds. That's inside the actual human turn-taking gap we just talked about. A normal turn-based assistant sits six times outside it.

Sam: And it's not even a huge model.

Alex: Seven billion parameters, trained over a single weekend on a handful of chips. English only for now, but it's a real architectural break, not a tuning trick.

Sam: A single weekend. Meanwhile I've got a vendor contract that's been sitting in our procurement queue since before the leaves changed color.

Alex: That tracks.

Sam: That's the part that actually surprises me, though. Every vendor pitch that lands on my desk calls itself an architectural break. What makes you so sure this one actually is, and it's not just this year's version of the same slide deck?

Alex: Because it's open. Anyone can run it, measure it, and try to break it themselves, which is exactly what happened a few weeks later, and that's not something a vendor claim usually survives.

Sam: Fair enough. What's trick number two, the smaller but still genuine one?

Alex: Parallel decoding. Normally a model writes one word at a time, in order, and that's the actual floor on how fast any of this can go. Break that order and you go faster. There was a paper this year that cut the number of generation steps by roughly a third, with barely any accuracy lost.

Sam: And the third, running it on the phone or the glasses themselves?

Alex: Real, but small, and kind of oversold. The newest chips can run a tiny model right on the glasses, so a wake word is instant. The second you need actual multi-turn reasoning, it still goes to the cloud, and the honest number out there is several times slower than the lab figure everyone quotes.

Sam: The gap between a spec sheet and actually using the thing. That's the gap my own vendors keep landing in, too.

Alex: Mostly network, not model, too. There's relay infrastructure in the middle shaving round trips that no clever decoding trick can touch, and you can't inspect any of it from the outside. Actually, this connects to something going on with my dad, if you don't mind me bringing it in here.

Sam: Go for it.

Alex: Tomás is in assisted living in San Antonio, and the facility just proposed putting in a camera system called AUGi, from a company called Inspiren. It watches for unsteady walking, or somebody rising off a chair wrong, and pings a staff member before a fall happens.

Sam: Does it record him, though? That's usually where these things fall apart for me.

Alex: That's the part that actually got me to say yes. It doesn't record or speak at all. What staff see on their end is a blurred stick figure, not actual video of my dad, and he had to consent to it being turned on. He can switch it off any time he wants.

Sam: Okay, that's actually a pretty good design. A lot of these things get built around convenience for the staff, not dignity for the person actually being watched.

Alex: Which is exactly what sold me on it. The company just raised a serious new funding round to build more of exactly that. And it's a perfect example of what we keep bumping into today. It's genuinely real-time, and it never says a word.

Sam: Which is the pattern, pretty much, once you sort these by how good the evidence is instead of how loud the marketing is.

Alex: Every deployment with real, published numbers is a listener, never a talker. Ambient clinical scribes are the best-studied case. One study this year found heavy users of a popular scribe cut their documentation time by about a fifth, roughly thirteen minutes a day, with no clear change either way on burnout.

Sam: Thirteen minutes doesn't sound like much.

Alex: It's not nothing multiplied across every patient, every day, for a year. A couple of other health systems reported similar numbers with different scribes. Modest, but consistent. Minutes, repeatedly, whenever somebody actually measures it.

Sam: Meeting tools are the same shape, aren't they?

Alex: Exactly the same. The ones people actually keep just record quietly in the background and clean up your rough notes afterward. The assistants built into the big meeting platforms only speak up mid-call when you actually ask them something.

Sam: The ones that are genuinely live and talking in your ear cluster somewhere specific, right?

Alex: Sales. Live tools pushing battlecards, prompting reps, updating the deal record while the call's still happening. Because a sales call has a script and a measurable outcome, you can actually tell if the live prompt helped. Nobody's shipped a credible "AI that helps in a general meeting," because nobody can say what a good interruption would even look like.

Sam: Can we do a Hype Check? Because customer service is where this gets genuinely ugly.

Alex: Go for it.

Sam: One of the bigger voice-AI vendors says its case studies show up to ninety percent containment. Meaning that share of calls get finished with no human at all.

Alex: Another vendor's claiming the same kind of number for a client cutting its switchboard load.

Sam: My rating, skeptical. Those are vendor numbers, or the vendor's own customer talking about the vendor.

Alex: Mine too, bordering on "show me the independent study." Because here's what happens the second somebody who isn't selling the thing actually measures it. A practitioner review of live deployments put one well-known vendor at roughly a quarter of that, for a real client.

Sam: That's a third of the claim. Maybe a fifth, over a longer stretch.

Alex: An independent roundup of early adopters generally lands somewhere in the middle. But there's no audited study that checks these vendors side by side on the same definition. In the one application with millions of conversations a day, the public evidence is almost entirely self-reported.

Sam: Which is exactly the gap I keep hitting with our own team. We've been circling a vendor for our phone line, so I went and looked at what Chewy's doing, since they're basically our size.

Alex: What'd you find?

Sam: On their last earnings call, their CEO said their new assistant is still in beta, live for a small slice of their traffic, running for less than a month. That's it. No containment number, no victory lap. Just: it exists, it's small, we're watching it.

Alex: Which is the honest version of literally everything we just talked about.

Sam: Right. A big home-improvement chain's doing something similar in stores, and their write-up doesn't give a containment number either. Nobody actually running this at scale is bragging about a number yet. The vendors selling it to us are the only ones with the big round percentages.

Alex: Which tells you something about which number to trust.

Sam: Right. And it's changed what I'm actually asking our vendor for now. Not "what's your containment rate," because that number means whatever they want it to mean. I'm asking for the raw call logs from one of their live clients, and I'm reading them myself before my team signs anything.

Alex: That's a smarter ask than most CTOs make, honestly.

Sam: I learned it the expensive way. I don't want to be the cautionary tale everyone else points to in two years, the team that believed the slide deck over the actual call logs.

Alex: There's a reason the honest deployments all stay quiet, and it's not taste. It's the meter.

Sam: Go on.

Alex: Realtime audio gets billed per token, and the asymmetry is brutal. Audio coming in is cheap. Audio going out costs roughly four times as much.

Sam: Four times the price, the moment it opens its mouth.

Alex: One analysis of real sessions found the spoken audio alone was something like four fifths of the entire bill.

Sam: So who pays for that, at scale? "A few cents a minute" sounds fine until you multiply it by a million users.

Alex: At a modest user base, a flagship always-on assistant runs to real money every single day. Drop to the cheaper tier and it's a third of that. Stack a platform fee on top, and the product that survives contact with an actual budget is the one that listens constantly and talks rarely.

Sam: That's the line I'd actually put in front of my own marketing budget, too. Every tool I've ever signed off that talks back to a customer costs noticeably more than whatever just quietly reads and summarises in the background. I never had a clean number for why until just now. Which is also, and I don't think this is a coincidence, exactly why people get nervous about having it in the house.

Alex: A product built to listen all the time and talk rarely is, from the outside, indistinguishable from one built to just record you.

Sam: A third of US adults already have a smart speaker. Pew found most people already think this stuff is moving too fast and will put their personal information at risk. Younger adults included, which surprised me.

Alex: It's a real fight, and it's specific. Is the audio on a live call kept or thrown away, is the voice on the other end synthetic, what happens to a contact-centre job in a city that depends on them. Duplex also deletes its own paper trail, speech in, speech out, no transcript, right as regulators are trying to get a grip on exactly this.

Sam: Which, honestly, loops straight into something that's not abstract for me at all this week.

Alex: Go for it.

Sam: My mom got a call Tuesday. A voice that sounded exactly like my cousin's kid, panicked, saying he'd been in an accident and needed money sent right then.

Alex: Oh, that's awful. Was it —

Sam: Fake. The FTC and the FBI have both been warning about this all year, that a few seconds of audio off a social post is enough to clone a voice convincingly, even to someone who knows that person well. She almost sent it before she called me to check.

Alex: Did she have a way to verify it was fake?

Sam: Not really, which is the problem. The advice everyone gives is to agree on a family code word in advance. We didn't have one. We do now.

Alex: That's the exact same vulnerability we're talking about today, just aimed at a family instead of a company. What makes these models interruptible is the same quality that makes them feel human, and it's now brand new attack surface. A paper this September tested several open duplex models by literally speaking a jailbreak attempt while the model was still mid-sentence.

Sam: And timing alone did that?

Alex: Timing alone. Wait about a second before you interrupt, and attack success nearly doubles over just asking normally.

Sam: Because the safety training assumed a whole turn.

Alex: Exactly. These models don't really have whole turns anymore, so the training that's supposed to stop them doing something dangerous was built for a world that doesn't exist on this architecture. And it's not the only rough edge. There's no agreed way to even evaluate a duplex system yet, timing falls apart with more than one speaker in the room, and the best computer-use agents were still failing most realistic tasks earlier this year.

Sam: So it can be present. It can't yet be trusted to actually do anything.

Alex: That's the honest split, and it's a fine place to be. Nobody actually needs an hour of deep reasoning to answer you in two hundred milliseconds. So, where do you actually land on this? Because I came in today thinking this is basically solved and we should all be building on it right now.

Sam: Of course you did. You've been calling things "basically solved" since the dot-com crash, and my husband still brings up the time you told him streaming video was a dead end.

Alex: Okay, that one's fair, I deserved that. But I still think I'm right this time.

Sam: And I still don't think that, even after all of this.

Alex: Why not? The economics work. The latency's inside the human gap. The architecture question's basically answered.

Sam: Because "the architecture's answered" and "it's safe to let it act" are two completely different sentences, and today's episode is full of reasons the second one isn't true yet. My mom almost wired money to a voice. That's not a hypothetical risk model. That's Tuesday.

Alex: That's fair. I think I'm still someone who'd ship the listening layer right now. The scribes, the sales prompts, the quiet stuff. Low risk, real minutes saved, today.

Sam: I'll actually meet you there. Ship the version that only listens. I'm nowhere near ready to hand it the microphone for anything it can act on, not the phone line, not a payment, nothing, until somebody shows me a duplex model that holds its safety behaviour when someone talks over it on purpose.

Alex: So: present, yes. Trusted to act, not yet. Probably not for a while.

Sam: What would actually change your mind on the acting part?

Alex: Three things, really. Whether anyone ever lets the fast layer execute instead of just relaying to a human. Whether the cost of generated speech drops enough that always-on talking stops being a luxury feature. And whether anybody builds a model that holds its safety behaviour when you talk over it on purpose.

Sam: Watch those three. Not the next flashy demo video, whichever lab puts one out this month. Look at that. We actually agree on something.

Alex: Put that on the record.

Sam: So, in one breath. The industry didn't make AI smarter this year, it split one product into two. A cheap, fast mouth up front, and a slow mind working behind it that you never see.

Alex: And the surprise is where that pays off. Every real, measured win is a machine that listens and says nothing. The scribes, the sales prompts, the fall sensor in my dad's building. The second it starts talking, the bill jumps and the evidence gets thinner, not thicker.

Sam: And the thing that finally makes it feel human, being interruptible, is the exact same thing that makes it hackable. That's not solved. That's next year's actual problem. And look, thank you so much for listening to this one. I hope you came away seeing a little more clearly where all of this is actually heading, because it moves fast enough that staying close to it is the only way not to get surprised by it.

Alex: One honest note on how the show itself gets made. It's AI-generated. A custom stack of AI tools researches it, checks it, and puts it together, built mainly so the person behind it can learn this stuff himself, and published for anyone who wants to follow along. AI-assisted, fact-checked, worth a second look.

Sam: Before we go, one genuinely useful thing you can do. Follow the show. Whatever app you're in right now, there's a follow or a plus button, it's one tap, it's free, and it does two real things. You get the next episode the second it's out, and for a small independent show like this one, a follow is honestly the single biggest lever for helping it reach other people trying to make sense of all this.

Alex: So if any of the last twenty-five minutes was worth your time, go hit follow.

Sam: And one last ask before you actually go. If there's someone in your life who keeps asking you where AI is actually heading, send them this one. It's genuinely the kindest move you can make, for them and for the show. It's still small and independent, and every single share does more than you'd think.

Alex: See you next time.