Building Gets Easier Near AGI. Knowing What to Build Does Not.

Sequoia's Sonya Huang and Konstantine Buhler sit down with Alfred Wahlforss about a year after Listen Labs launched. The hosts' intro is a stack of logos: the company now serves about 20% of the Fortune 500, including Microsoft, Anthropic, Sweetgreen, and NBC. Under the hood is an AI agent that talks to customers, drawing on an audience of 30 million people, and can run hundreds of interviews at once. The product thesis comes early. As we get closer to AGI, Wahlforss says, it will be easier to build things. The hard part will be knowing what to build. That is what Listen is building.

The favorite customer story is not a Fortune 500 boardroom. Chubbies, the shorts brand and an early client, uses Listen for almost everything: marketing tests, shirts, what will sell and what will not. One finding: chest hair interfaces poorly with one of their materials. The shirt was uncomfortable to wear. They changed it, and it became radically more comfortable. Small things and big things, Wahlforss says. Manscaped reworked a Super Bowl ad from Listen insights. Skims is a customer too. Buhler, still chewing on the chest-hair detail twenty minutes later, treats it as a parable for all the tiny product truths customers never get a chance to say.

The loop is simple on a slide. Ask a question — how do you improve Cursor's onboarding, in the example — and Listen writes an interview guide, which is just instructions for the agent. It finds the right people in the 30-million pool, from oncologists to software engineers, runs hundreds of conversations, analyzes the data, and returns recommendations. The step they are launching in a couple of months is simulation: after tens of thousands of interviews in the platform, can you predict how customers will answer questions you have not asked yet?

Huang is a survey skeptic, and she says so. People get paid to take surveys, which is already selection bias. What they say they would do is not what they do. She comes from the school that trusts real-world telemetry more than stated preference. Wahlforss does not pretend the gap is closed. Listen has gone back to the same person with the same multiple-choice survey and found the answers radically inconsistent. Do the same thing as a spoken Listen interview, where you have to think and reason through the answer, and people are much more consistent at least with themselves.

They also back-test. When Chubbies tests shirts, Listen looks a couple of months later at actual sales. He agrees that an A/B test is the holy grail, and also that it is hard to get right in practice because you need a very large volume of users. Some input, he argues, is still better than none. The interview itself is not a form. It is essentially a Zoom call with the agent, on video, with emotion detection looking at your eyes and the way you say things — another bridge between what people claim and how they actually think.

Advertising is where the face and voice pay off. A Likert-scale "extremely likely to buy" can score high while the person on camera is flat. When someone also reacts with real enthusiasm, those ads then perform better in performance marketing on Meta and LinkedIn. Traceability is built in so the claim is inspectable: for every data point you can click through to the video or the quote, which is how you know the AI is not hallucinating the finding. Huang calls that awesome. It is also the design constraint that keeps a research product honest.

The origin is a consumer app, BeFake: an early ChatGPT-images-style toy that fine-tuned Stable Diffusion so you could drop an AI avatar of yourself into a scene. It went viral overnight — 20,000 users — and then churned hard. Wahlforss and his co-founder were already experimenting with AI, so they built an interview agent for themselves to ask why people left, how they heard the positioning, which use cases landed. It was useful enough that they turned it into the company.

The old world, they discovered, was not just Qualtrics. There is a multi-billion-dollar services industry that writes the questions, finds the people, and analyzes hundreds of calls, because talking to *prospective* customers, not current ones, gets harder as you scale. Designing those questions is an academic subject. You cannot just ask how much someone is willing to pay; different methodologies predict behavior very differently. Traditional CPG companies, and Microsoft, spend tens of millions of dollars bringing people into rooms for focus groups. Listen's first-order pitch is to do that much faster.

The less obvious benefit is that most decisions inside companies are not based on customer input at all, because the friction of talking to customers is too high. Lower the barrier and decisions get smarter. Speed is the magic trick: real-people input within five minutes, hundreds of interviews populating a dashboard. Because the sessions are asynchronous, they are also cheaper — you can pay participants less than for a live interviewer. That is not a theory about AI likability. It is a revealed-preference fact: people accept lower pay to talk to the AI than to a human. They are busy, they can start and stop, the pressure is lower. They are also, Listen has found, more honest. People open up. It can feel therapeutic: a non-judgmental entity that is really interested in you, including for sensitive work like interviewing kids about products, which used to be infamous for parental consent, school, and extracurriculars eating the calendar.

Audience is the bottleneck, and it is where Listen spends 80% of its engineering. Every company, Wahlforss says, is driven by a power law in customer segmentation. Sweetgreen looks like salad for everyone. The right audience is typically urban, high household income, mostly female — and they need to know what seed oils are, a health-world term maybe 1% of the population knows. Some people go every day, and that group is 80% of revenue. Find that segment and the research becomes actionable. Miss it and you are interviewing the wrong country.

The long-term goal is a billion people in the audience, stratified by what each person is actually expert in. It might be sneakers: influencers and early adopters, interviewed first, because those insights are worth more. Every interview thickens a profile, and then you can search. Buhler fills in the network-effect picture. Someone mentions in an unrelated session that they are a sneakerhead; the next time a brand like Nike is launching a line, that person is sitting in the database. The old model could not do this. Separate vendors, email lists, spam. Those lists have an incidence rate that can be 10% — only one in ten people even qualifies — which means you screen someone ten times before they get paid once, and the panel churns.

Why not skip Listen and email the power users you already have? Sweetgreen can, and Listen will connect a CRM and send to known customers. The interesting work is talking to prospective customers who are not yet in the power-law core, then comparing the two groups. CRMs are messy. Google cannot just email Gmail users. Ordinary outbound looks like spam and lands in the filter, which Buhler has watched put companies in "Microsoft purgatory." A third-party panel avoids lighting your own domain on fire.

What happens to the McKinseys? Bain already uses Listen to speed up traditional processes, Wahlforss says. Implementation is still valuable. A lot of margins will drop, and firms will have to unbundle so AI agents can take pieces of the work companies used to send to consultants. Buhler is more optimistic: if you can learn faster, maybe you hire Bain to explore five new expansion areas at the same fee. Wahlforss thinks the ceiling is high. You can learn more about customers and build more things. The joke running under the table is Buhler's chest-hair shirt, and the line Wahlforss offers as a mission: "We hope to live in a world that finally works the way people want."

On pricing compression, he argues the opposite of a race to the bottom. Because you can do it faster, you should be able to charge more. They have run studies where they charge hundreds of thousands of dollars to speak to 20 doctors across eight countries. Over the long term the individual interview gets more affordable, but he expects two orders of magnitude more research. Simulation is how you unlock the 99% of use cases where you would never have time to talk to real people.

Buhler wants that in medicine. Doctors are busy, important, and not going to book a focus group, while companies have not invested in finding the paper cuts. If a physician could complain about the EHR or something in the operating room from the home-screen app, that is a life-saving use of listening. Wahlforss is already being pulled toward closing the loop: a churn interview finds a bug, then another coding agent is told to go fix it. Find the problem from an actual human, then let software ship the patch.

Huang sketches the industry in versions. 1.0 is calling 100 people and collating by hand. 2.0 is AI-native: the model designs the guide, talks to thousands at once, synthesizes. 3.0 is generative agent simulation, and she is both drawn to the self-driving-car analogy and worried that synthetic data is just remixing the pre-training set. Wahlforss's version of simulation starts with one person. Interview Buhler for an hour and you can start to predict his preferences. LLMs are quite good at this. Feed in enough information on one individual and, in some cases, they get to 95% accuracy on how that person will answer certain questions.

The useful trick is message testing: the tagline on the billboard, the title of a talk. He had to name a conference session, wrote 100 titles, dropped them into a simulation of Listen's own customer base, and the winner was twice as good as the next one. He does not know if that is "correct," but it felt correct, and it was nicer than deciding alone. Then the control: he fed ChatGPT and the simulation a talk that had worked and one that had not. ChatGPT picked the wrong one. The simulation picked the right one. General models are trained on the average person. You want to build for a specific niche. Prompting ChatGPT to be a grumpy 35-year-old terminal-using engineer helps a little. Credit-card spend and purchase behavior helped a little. Interviews helped a lot, because they go off on tangents, they let you ask behavioral questions, and the way you design those questions matters. Models do not have clean data on how a specific persona acts and thinks.

That package already exists as a product: augmented responses, a special-purpose bot you can query after enough real interviews. The cooler version lives inside coding agents — a "human API" so other agents can call user preferences when they decide what to build. The stack is post-training plus typical RAG plus proprietary techniques. Simulation itself is model one person well, then scale to a thousand as a representative sample, which is already multi-agent. Huang asks about letting those thousand debate, the council-of-models pattern. Wahlforss is skeptical. Chaos theory says that once things compound, interactions get hard to predict. Competitors are trying it. He wants to explore it and does not trust it yet.

Human input does not go away. For a Super Bowl ad you still run real interviews. For the tagline on a small billboard, simulation can answer. They have also shipped MCP, so you can drop Listen into Claude and tell it to run Listen in a loop: generate a pile of marketing concepts, then test them. Huang wonders whether the interview itself becomes live ideation — the AI inventing a Tide Pod in the session, visualizing solutions while someone complains. Today customers do that manually, generating concept images and feeding them into interviews. She is basically on the product team, Wahlforss says.

The original product idea was just to make the interview less annoying. They had an eval for repetitive questions and instruction-following. With GPT-4 the agent sometimes asked the same question 100 times. That eval started around 20%. They climbed it to 85%. Then they wrote a harder one: can the agent understand a screen recording of what you are doing, and skip questions that are no longer relevant? They are back at about 20%. That, he says, is the value of a vertical AI company: you own a proprietary eval, and you climb it. Better data, harder problems, better data, repeat.

Huang's synthesis is an infinite loop. A company is figure out what to build, build it, figure out what to build. Write code and talk to users, Wahlforss adds. Execution is getting exponentially faster. Listen is pushing the other half — and not only in product, in strategy: what to build and what to say. The end-state in his head is a one-person company in the loop: a coding agent and Listen, running autonomously. For a CPG giant the stakes are even higher, because once you ship atoms you cannot patch them. When Procter & Gamble launches in a new market, that can be tens of millions of dollars. It has to be right on the day. That is one reason they are a Listen customer. P&G is his archetype of a marketing company that finds a niche people care about and builds a brand for it. Tide Pods came from interviews that said liquid detergent was uncomfortable and people wanted something easier. Mars, in a story from the Acquired podcast, learned in the 1950s that M&M's — designed not to melt, originally for the army — had a bigger market in kids, and pivoted the advertising so the candy would not ruin the furniture.

What is he sure of? Future systems will still need human input, because even a perfectly rational AGI sits next to irrational humans who suddenly obsess over a TikTok trend and force the entire marketing strategy to turn. What he is unsure of is how far simulation goes — confident it works for certain questions, waiting to see how good models get at predicting behavior. Buhler, the self-described humanist, adds that if companies exist to serve people, and intelligence approaches what humans want, the remaining delta — what is in a person's mind and not in the model's — only gets more valuable.

On moats he lists the clean ones: network effects on the panel, supply and demand; a data moat, because more interviews make simulation better; stickiness, because the interviews live on the platform and you want to track over time. Then the unfashionable one. Sequoia partner Bryan Schreier told him that founders want to build something complex and customers want something stupid-simple that just works. They do not want to configure a workflow. Interview-guide design is hard, which is why services firms exist, and you do not want to lead the witness. Early Listen used a vanilla LLM. Customers ran studies, got garbage, and came back furious. Listen took the blame and trained the agent on best practices so the data comes out usable. That, as much as the billion-person panel, is the vertical-AI job: train the agent on the craft so the customer never has to become a methodologist.