The Fake Five-Second Hold: Voice AI Is Already Negotiating With Truckers

Nathan Labenz opens by calling Olivia Moore and Anish Acharya “AI scouts”: people paid to live on the edge of what exists and actually try it. He wants the meta lesson before the object-level one. Where is the top of the funnel, and how do they know they are onto something? Olivia’s answer is that consumer investing requires being chronically online, and the interesting work is sitting in two rooms at once. Twitter is where founders announce companies, models, and breakthroughs; newsletters and meetups help. Ordinary people showing what they actually do with AI live on Instagram, TikTok, and, more surprisingly, YouTube.

YouTube, she notes, is the number-one mobile app and the number-two website in the world. For many consumer and prosumer companies, it is by far the top social referral, because a separate economy of creators now makes how-to videos about AI tools. An early demand signal, she says frankly, is often teenage girls trying to push ChatGPT into being a therapist, a friend, or a coach. Once that pull is visible, standalone products can probably exist. Anish’s addendum is less glamorous than venture jargon: surprisingly few insiders have actually used Operator, Deep Research, DeepSeek, o1 Pro, or Krea. Hands-on use is how you build intuition.

Consumer, Olivia says, is random and magical. A ten-year pedigree does not save a product that misses timing, insight, or a single onboarding feature. They try to let usage data tell the story, including the off-label ways people yank ChatGPT into jobs nobody designed. Anish dusts off a pre-AI joke: every social app collapsed into a dating app. The new version is that every large language model is being tortured into a therapist. It is a dinner-party line, and also a leading indicator of what people want the models to do.

Anish’s frame for the rest of the hour is almost anthropological. Voice intermediates every human interaction and relationship; the three of them are using it in that moment to get to know each other. It is the original, most important form of human communication, and it has been almost completely unaddressable by technology because the infrastructure did not exist. Other AI substrates sit on decades of prior tooling. Voice is a blank sheet of paper, which is why they are as excited about distribution as about product: a new interface is a new entry point.

Where voice is already working has surprised them. A lot of the startups getting real traction are B2B, Olivia says, because so many businesses already run on call centers, and even small firms pay one, two, or three people to answer the phone all day. Once voice models are anywhere in the realm of human performance, after-hours and voicemail calls are an obvious first slice. Her guess is that plenty of people have already talked to an AI receptionist without quite realizing it. On the consumer side, most contact so far has been through ChatGPT or Grok; Sesame was a later, viral leap, still just a web demo of what is coming once that class of model is open and usable by app builders.

Then there is 1-800-CHATGPT, the OpenAI phone number that is easy to forget because the stack moves so fast. Whether the experiment failed or succeeded, Anish says, it pointed at a real insight: maybe the first way most people in the world experience AI will be by voice, both as consumers and as people calling businesses. Nathan keeps thinking of his mother, now in her early nineties, living alone, sharp but not an early adopter. She already asks Alexa to play music. Alexa Plus, if it can actually converse, might unlock a whole digital life. Anish’s twist is that seniors often need the old technology as much as the new: the TV remote, the friend’s email they cannot find.

One of Nathan’s first GPT-4 red-team tests was tech support for seniors, using the same script he uses with his mother: read everything on the screen from top to bottom. She will dutifully start with the carrier name and the time, and eventually they reach the problem. That prompt worked out of the box in text; multimodal models that can see a screen and talk in real time, including Google’s late-2024 Gemini release and an OpenAI counterpart heading toward API availability, turn the same pattern into around-the-clock help. Anish’s line is that the most important context is the context in physical space. Pointing a phone at the remote beats translating the living room, line by line, to a grandson or to Google.

After a sponsor break, Nathan pushes on a phrase from their presentation: basically solved. He still finds interruption gnarly, especially in a group demo of Advanced Voice Mode, which seems tuned for one-on-one. Olivia softens “solved” to “very close” on basic latency and understandability, the difference between being able to have a conversation and not. In most cases, latency is now less than half a second, which feels human. A year earlier, when a16z put out its first voice report in May, sub-one-second latency was hard to imagine.

People who have tried the Sesame demo, she says, realize the leap is beyond latency. Humans fill speech with pauses, ums, and inflections that a model can treat as errors. Put those back in, as Sesame did, and the voice stops sounding like a better Alexa or Siri and starts being mistakable for a person. Two hard problems remain. One is emotionality: founders building voice agents want happy news in a happy voice and sad news slower, lower, and quieter, and models such as Hume AI’s Octave are pushing that way. The other is interruptability. Humans have not solved it either; two people start at once and still say “no, you go ahead.” Voice AI needs a cleverer answer than people have.

Anish splits the stack further. Sesame showed that conversation models may be different from, or an extension of, voice models. The three of them are constantly coordinating who talks next with nonverbal cues the models were not trained on. Native speech-to-speech is still the minority among companies they see; most still do speech-to-text, language model, text-to-speech. Humans often do not know the next sentence when they start talking. The AI does, which can feel creepy and keep the product in the uncanny valley. Nathan cites a recent Meta brain-reading paper: phrase-level formation about two seconds before you speak, then the next syllable a fraction of a second out. He wonders whether a diffusion structure, coarse to fine, would fit that better than one token at a time.

Why are builders still on the old pipeline? Olivia says voice-to-voice is more expensive and earlier. Among the native options, Gemini Flash is maybe the best of the bunch, but interruptability is not there yet. Her weekly mantra: the models are the worst they will ever be. Anish adds a primitive people under-appreciate because it shares an interface with language models: reasoning models. Friendship, rapport, and the weather of a negotiation benefit from probabilistic language. A price that is definitively better for one party does not. Orchestrating the two on different stretches of a call beats betting that one model is smart enough. Marc Andreessen’s question, which Anish repeats, is whether the bar is as good as humans or perfect. If it is perfect, these technologies are not there and maybe never will be. If it is to outperform humans, they are there in many ways.

Conversation quality, Anish says, predicts business outcomes. a16z is an investor in Happy Robot, voice AI for freight brokers, and the team’s text-to-speech and conversational feel are simply better than off-the-shelf alternatives because the company did more work under the hood. Commodity voice gets you into a call. Specialized voice earns permission for persuasion, negotiation, and disagreement. Nathan asks the obvious question: is this company actually negotiating? Anish’s answer is that it negotiates, befriends, and disagrees, and that the demo is an aha moment. They offer to send it.

Olivia unpacks the design that makes the demo work. An agent that always answers “this is my best price” does not feel like a negotiation. Users will try to circumvent it and demand a human. Happy Robot introduces extra latency on purpose. The agent says hold on, let me talk to my supervisor, puts the trucker on hold for about five seconds, then comes back with a slightly better price. The system already knows its max. What changes is the human’s sense of having given it their best and extracted a concession. Acceptance of that final offer, she says, is much higher. Nathan admits he does not know how to feel about it: genius, and a little far out.

The agent discloses that it is AI. The surprise is that the truckers, driving big rigs across the country, not Stanford technology enthusiasts, do not mind at all. Anish’s explanation is reptilian: once you are in the conversation, intellect knowing it is an AI does not stop you from falling into the rhythm, expectations, and cadence of a human call. Olivia quotes a line Anish uses a lot: in the best cases, AI can be more human than the humans. You get the same friendly, patient counterpart every time, willing to hear about your day, rather than a grumpy person on a bad shift. Superhuman patience, Anish adds, is not that hard to achieve and is quite valuable. Near-zero wait times are a huge driver of value too.

Nathan turns to Apple’s headline that Siri will not get a major update until 2027, which from where he sits feels like the other side of the singularity. He cites Ethan Mollick’s reminder that the comparison is the best hirable human, not imagined perfection. Is that what Apple is getting wrong? Anish has many thoughts. For anyone who uses both AI products and Siri, it is a stick in the eye five times a day on the most basic commands, then juxtaposed with Apple Intelligence advertising. That degrades consumer trust. AI, he argues, does best when it explores the messy surface of human interaction, and large incumbents are designed to take the humanity out of every technology product. The more they try to neuter the AI, the more dissatisfied consumers become. Genmoji is a valiant effort that, in his view, just looks terrible.

The spiritual problem, he says, is committees, lawyers, and the posture of large incumbents. Olivia points at the backlash to AI-generated notification summaries on iPhones, which she guesses spooked Apple. A new Apple AI product has to be production-ready on hundreds of millions of phones, for every age and use case, and it has to feel natural *and* be correct. A startup’s seekers know they are on a beta. She gives Google Labs more credit: NotebookLM and video models such as Veo 2 came out of a waitlist-and-beta pattern that lets early testers go first. Even then, once a Google product hits the public, the pace slows. Her jab: it is like ten VPs for every engineer on NotebookLM right now.

Deep Research is the missed-opportunity exhibit. It was originally a Gemini product, something Google should be the best in the world at, and they never commercialized it in the right way. ChatGPT now owns the association. Olivia’s summary is blunt: it is one missed opportunity after another with incumbents.

Nathan asks about tool use, including calls into an obscure freight system that might even need fine-tuning. Anish’s broad observation is that voice capability alone is insufficient. Integrations, workflows, and a traditional mode still have to be built; the model gets you into the conversation and is not enough to get you to the other side. Olivia agrees, especially because the buyers are traditional enterprises, not Apple and Google. For them, launching a horizontal agent once is a miracle, let alone keeping it current as models improve and as the system of record breaks. That is why customers are excited about vertically focused platforms that are fine-tuned for the actual conversations *and* have done the long tail of integrations and dialogue types.

Nathan quotes Tyler Cowen: context is that which is scarce, a line that predates the AI era and has never been more true. Base models are versatile and still do not know the way *your* freight business runs. That last mile trips a lot of people. Olivia has watched enterprise AI budgets explode, and six months earlier some buyers were shopping because they did not spend their days on the latest models. ChatGPT is the consumer rhyme: fastest product to 100 million users, then roughly a year of flat usage until more models and more obvious jobs-to-be-done appeared. If a product takes more than a few steps to get running, the paying funnel shrinks. Vertical companies are winning because they collapse those steps.

Where those companies land, Anish says, growth is explosive, on both agents and scribes. For a business that can embrace the substitution, phone-calling is a conceptually straightforward swap, with scorecards and negotiation guardrails still required. In the cases where it is working, it is really working: some of the fastest-growing B2B startups they have seen in ten years. Real-time coaching is working too, especially where the job has a heavy in-real-life component. Salespeople and HVAC technicians can win or lose a $10,000 upsell on the nuance of one question. Hundreds of dollars a month for an individual AI coach is cheap against that. Call-center coaching is in huge demand now, even if agents later take more of those seats.

Has the 90 percent call-center headcount cut arrived? Anish says no, not yet. Nobody’s job is only initial screens. Recruiting still includes deeper interviews, salary negotiation, and taking a new hire to a baseball game a month later. AI takes the annoying screen; headcount has not dropped because the rest of the work still matters and is not trusted to the model. Olivia notes that basic call-center seats can be 300 percent turnover, thankless work, and that screening agents free people for better jobs, including a recruiter spending twenty extra hours a week on the five candidates they actually want. A new job is already growing fast: contributing training data and doing online tasks, similarly paid, much better lifestyle. Nathan presses the timeline. They put a 90 percent cut as a 2026-plus phenomenon, and even then Anish is not sure the labor market will move the way the thought experiment imagines.

For small businesses, Olivia’s highlight is that vertical solutions already exist by trade: restaurants, spas, home services. SMBs typically have one or two people doing nothing but answering phones, which is expensive. When those customers switch to a voice agent, they are usually not laying that person off. The human, often core to the business, spends time on customer experience and growth instead. On creators, she points to ElevenLabs for cloning and for describing a voice into existence, and to platforms such as Delphi for a digital clone an audience can talk to. She has not seen AI replace full podcast episodes yet; that is probably a couple of years out. Nathan once redid a guest’s bad audio with a HeyGen avatar on a HeyGen episode, which felt fitting if not quite as good as the original.

Nathan’s kids now say “Daddy, ask AI” when they are stuck in Mario 64. He wants interactive education and does not want to trust just anyone to implement it. Anish is more excited about behavioral and social-emotional products than about homework. His son loves Minecraft and meets toxic teenagers online; why is there not a companion that plays with him and models decent behavior? A multimodal model that watches classroom interaction and feeds back to parents and teachers would help the many kids who do not have two adults in the room, one academic and one social-emotional. Assignment and quiz generation will happen. Pairing it with the emotional layer is where they get most excited.

Olivia sketches the spectrum. Synthesis, ELLO, and Super Teacher are the all-day reading and math tutor that understands how a child learns. On the other end, Curio-style toys are a friend, mentor, or coach that is simply present, tracks progress, and can be a sympathetic ear. Companionship still feels like the first inning despite Character and a cluster of apps in the top fifty and top hundred. Anish wants the unsympathetic version too: a companion that challenges, pushes, and disagrees. They joke about “East Coast mode,” a little more terse. It does not exist. He thinks those products show up in the next two years.

On the far edge, romance and NSFW, Anish had assumed the users would be frisky young dudes. That has not been true. A lot of the audience caters more to women and feels more like interactive fiction than pornography. Definitions of romance are doing a lot of work in the critiques. An AI that trains you at conversation and flirting, or that vents frustration you would otherwise dump on a partner, may make people more capable rather than substituting for traditional romance. Olivia says that whenever they circulate the top fifty or hundred AI apps, the team is always startled by how many companion products are on it. The mix is more AI boyfriend than AI girlfriend, more fanfiction than anything else. Sexuality is part of the human experience. Pretend it does not exist, and you end up as Apple, unable to release a product for five years. a16z will not invest in the fringe products; those are almost the least interesting to talk about, because they have always been there.

Nathan has done two episodes with Replika founder Eugenia Kuyda. The later one reviewed Stanford research: Replika reduced suicidal ideation in a substantial way for people who arrived with that problem, and users more often than not said it encouraged them to get out into the world rather than trapping them. Some of the data is self-reported; he still found it striking. Anish’s conclusion is that the exact shape of what you build matters more here than in most software. You could absolutely make a predatory, addictive, exploitative romantic AI. There is already existence proof for a version that is predominantly positive. Product design, not the underlying model, decides which one you get.

Nathan had opened the show with a red-team update. A year after he reported flagrant holes to two well-known AI calling platforms, both still let him quickly clone Donald Trump’s voice and call anyone at scale with no meaningful controls. He thinks new rules are needed sooner rather than later, to protect the public and to protect the industry from its most careless developers. One proposed rule is Yuval Noah Harari’s: AI must disclose that it is AI. Nathan’s own idea is a do-not-clone registry, a modern do-not-call list where you register your likeness and voice and platforms are barred from copying you.

Olivia’s twist, which Nathan called brilliant in the introduction, is to make the directory a permission registry: default protection, plus the right to license your identity for uses you actually want. ElevenLabs already runs a version of this. Alongside celebrity voice collections, it has been a boon to voiceover artists who could never get a Hollywood job and now pick up work on the platform. The same logic could hit the creator economy. An influencer with 5,000 followers will struggle to get a big brand to reply; an AI avatar that is more available and more extensible might land the deal. Extending yourself with the tools, she argues, is the more interesting story than the everyday non-public person being deepfaked, which she has seen less of so far.

Anish is wary of the talking-heads reflex toward paternalism. He has been more surprised by frustration at large model companies refusing to let people do things than by average consumers being cloned. Every technology can be misused; he is not being glib. He still thinks a generous view of the audience is warranted. People spent thirty or forty years learning that a book is not automatically true, then that the internet is not, then that social media is not. There is no reason this layer will be different. Whatever rules get written, he wants them to assume consumers are smart enough to use the products with an appropriate level of caution.

Asked for the medium-term picture, Olivia sees voice unlocked as a modality switch on every product, interaction, and device: AirPods, glasses, the computer. Plenty of moments are not two-way conversation; you want transcription one way or speech the other. Five years from now, today’s compelling demos will look incredibly limited. Anish reaches for Steve Jobs: the computer was a bicycle for the mind, extending people intellectually. The next extension is emotional, through companionship and much else, and voice will be the primary catalyst and interface. He thinks that impact has been underestimated. That, he suggests, is a subject for the next conversation.