Knowing What Your Customers Want, All the Time: Listen Labs' Alfred Wahlforss
In a Nutshell
Listen Labs is an AI platform that runs thousands of voice interviews simultaneously to replace slow, biased traditional market research with fast, traceable customer insights. It builds detailed audience profiles across 30 million participants, enables targeted research on niche segments like high-value Sweetgreen customers, and validates findings against real sales data. The company's simulation feature lets companies predict customer responses without new interviews, positioning AI-driven strategy as the next bottleneck after execution becomes automated.
These notes were generated by AI and may contain inaccuracies.
Listen Labs' goal is to reach a billion people in their audience and stratify them to identify what each person is an expert on. Even niche topics like sneakers have influencers and early adopters whose insights are significantly more valuable. The platform builds profiles of people across interviews, enabling targeted searches to find the right participants.
Alfred Wahlforss is the founder and CEO of Listen Labs, an AI-first customer research platform that can run thousands of voice interviews simultaneously. The company launched about a year ago and now serves 20% of the Fortune 500, including Microsoft, Anthropic, Sweetgreen, and NBC.
Listen Labs has an AI agent that understands customers better than companies can by talking to them directly. The process works as follows: a user asks a question like "how can you improve course onboarding," and Listen creates an interview guide with instructions for the agent. Using an audience of 30 million participants, the platform can find virtually anyone—from oncologists to software engineers—and conduct hundreds of interviews. The system then analyzes the data and provides recommendations. A simulation feature launching in a couple of months will allow companies to predict how customers will answer future questions after tens of thousands of interviews have been completed.
As we get closer to AGI, it will be easier to build things, but the hard part will be knowing what to build.
Chubbies, an early customer, uses Listen for all marketing testing including shirt testing to understand what products perform well. One insight revealed that chest hair interfaces poorly with one material, making shirts uncomfortable. After changing the shirt, it became significantly more comfortable.
Manscaped changed their Super Bowl ad based on insights from Listen.
Skims is also a customer.
Traditional surveys suffer from selection bias since people get paid to participate, and stated intentions often differ from actual behavior. Multiple choice surveys show high inconsistency—even asking the same person the same question twice produces radically different answers. In contrast, Listen interviews require participants to think and reason through answers, leading to greater consistency.
Listen validates findings against real outcomes. For example, Chubbies tested different shirts through Listen, then checked performance against actual sales data months later. While A/B testing is the gold standard, it requires very large user volumes, making it difficult to execute. Having some customer input is better than none.
Listen interviews function like Zoom calls with an AI agent on video. The system detects emotions by analyzing eye movements and tone of voice, providing insight into what participants actually think and feel beyond their words.
In advertising testing, participants may give high scores on a "likelihood to buy" scale, but enthusiastic emotional reactions correlate with better real-world performance in performance marketing on Meta and LinkedIn.
The platform is built around traceability. For every data point, users can click to view the original video or quote, ensuring AI outputs are not hallucinations and providing source verification.
The founders previously built a consumer app called BeFake that went viral. Users could create AI avatars of themselves by fine-tuning Stable Diffusion. The app reached 20,000 users overnight. While experimenting with AI applications, they built an AI interview tool to understand their own product's churn, positioning, and use cases. The tool proved so useful that it became the foundation for Listen Labs.
Historically, companies relied on old-school survey tools like Qualtrics. For market research involving prospective customers (not current ones), companies turned to a multi-billion-dollar services industry. Creating effective questions is difficult—directly asking "how much are you willing to pay" does not yield accurate behavioral predictions. Finding and sourcing the right participants, then analyzing hundreds of calls, requires significant expertise.
In traditional industries like CPG and even at Microsoft, companies spend tens of millions on focus groups bringing people into rooms for interviews.
Most business decisions are made without customer input due to friction in reaching customers. Lowering barriers to customer conversations leads to smarter decisions. Listen enables input from real people within five minutes. Hundreds of interviews can populate simultaneously in the platform.
The asynchronous nature makes research significantly more affordable. Participants accept lower pay for AI interviews than human interviewers, partly because interviews are asynchronous and lower-pressure. People can start and stop as needed.
Participants are often more honest with AI than humans. The non-judgmental nature creates a therapeutic experience where people open up. Listen can conduct sensitive conversations, including interviewing children about products.
Accessing the children's market traditionally requires parental consent, and children have busy schedules with school and extracurricular activities. Finding the right children for specific research is challenging.
Audience quality is critical—Listen spends 80% of engineering resources on audience targeting. Every company follows a power law in customer segmentation. For Sweetgreen, the ideal audience is urban, high household income, mostly female, and aware of seed oils (only about 1% of the population). A small segment visits Sweetgreen daily and generates 80% of revenue. Research on this segment is far more actionable.
Listen aims to reach a billion people and stratify expertise. In categories like sneakers, identifying influencers and early adopters first yields more valuable insights. The platform builds profiles across all interviews, allowing users to search and find the right person for future research.
If someone mentions being a "sneakerhead" in an unrelated interview, that information is stored. When Nike or similar brands need sneaker experts, those profiled individuals can be surfaced.
Traditional approaches used separate entities and manual email list spamming, which recipients find annoying. Low incidence rates (around 10%) mean only one in ten people qualify for interviews, causing significant database churn. Being screened out repeatedly is frustrating for participants.
Companies like Sweetgreen can use their own CRM to reach existing customers. However, the more valuable use case is reaching prospective customers who are not current power users, enabling comparison between the two groups. CRMs are often disorganized, and regulatory issues prevent companies like Google from emailing Gmail users directly. Using an external third party avoids spam risks that could result in being blocked.
Listen works with firms like Bain, which use the platform to accelerate traditional processes. Traditional services retain value in implementing changes, but margins will compress. Consulting firms will need to unbundle services to allow AI agents to handle problems previously requiring full consulting engagements.
With AI tools enabling more customer learning, companies can identify and explore five new expansion areas. The ceiling for customer insight and new product development is very high.
Faster research execution could theoretically support higher pricing. Listen has charged hundreds of thousands of dollars to interview 20 doctors across eight countries. Over time, individual interviews may become more affordable, but companies will conduct two orders of magnitude more research. The simulation feature will unlock the 99% of use cases where real-time interviews were previously impossible.
Many small problems exist in medical systems that companies have not investigated. Doctors are busy and unlikely to schedule long conversations, but asynchronous app-based feedback on EHR systems or operating room experiences could surface critical insights.
Listen is being pulled toward connecting research outputs to other AI agents. A churn interview identifying a bug can be routed to a coding agent to solve the problem automatically.
Market research has evolved from 1.0 (calling 100 people manually) to 2.0 (AI-native platforms interviewing thousands simultaneously). The emerging 3.0 phase involves generative agent simulation.
Simulation works by interviewing a single person for one hour and feeding that information into a model. In some cases, 95% accuracy is achieved in predicting how that individual will answer certain questions. However, chaos theory makes future prediction inherently difficult.
Listen hydrates simulation audiences using all interviews running through the platform. A million interviews have been completed, with exponential growth since that milestone. The network effect enables training audiences on real interview data.
A user could ask how software engineers think about cloud code, and the system could predict answers based on a thousand recent interviews. The challenge is determining what can and cannot be accurately predicted.
The model is explicitly told the domain of knowledge possessed by each audience segment, then tested on how far that domain can be expanded. Backtesting removes one question from the training dataset and measures prediction accuracy. Nonsensical questions (e.g., "what is their dog's name") test whether the model recognizes unanswerable queries.
Message testing is one of the most useful simulation applications. Testing 100 different titles for a conference talk showed the top-performing title was twice as effective as the next option. Even if predictions are imperfect, having guidance for decisions is valuable.
When the same title-testing exercise was run through ChatGPT, it selected the wrong title. Listen's simulation selected the correct title. General models are trained on average-person data, while Listen trains models on specific niches, enabling better performance for targeted audiences.
Forcing general models into specific personas (e.g., "you are a 35-year-old grumpy software engineer") does not replicate the accuracy achieved by bootstrapping from real user interview data. The intuition is that simulation trained only on pre-trained data lacks the specificity needed for niche predictions.
Listen Labs tested multiple data inputs for their models. Credit card spend and purchasing behavior were evaluated, but interviews emerged as the strongest dataset. Interviews allow the model to explore tangents and ask behavioral questions. The design of interview questions matters significantly because models lack clean data on how specific personas think and act.
The intuition is that models do not have clean data on how a specific persona acts and how they think. Having enough interviews with people who follow a certain group allows the model to generalize and complete the picture of what everybody in that category is thinking.
Listen Labs offers a product where companies can run active interviews through Listen Labs, accumulate thousands of interviews, and then access a special-purpose Listen bot for instantaneous ad hoc questions. This is called augmented responses.
These responses can also integrate into coding agents and other agents. The vision includes a human API where agents can call user preferences to determine what to build, how to do it, who to invest in, or how to help customers best.
The system uses post-training with typical RAG along with proprietary techniques to combine interview conversations with the models.
Listen Labs models one person extremely well and then scales that model up to a thousand people to create a representative sample. This approach is multi-agent in nature. The system does not have those thousand people debate each other.
Competitors are exploring debate-based approaches. The concern is that chaos theory suggests when things compound, it becomes difficult to predict interactions. The AI council approach of sending queries to multiple LLMs and having one synthesize responses produces slightly better average outcomes, but skepticism remains about scaling this to simulation.
Generative simulation opens many new use cases for customer input. Large decisions like a Super Bowl ad still require real interviews. Smaller decisions such as billboard taglines can use simulation.
Human input will still be needed, but the range of decisions informed by customer input will expand significantly.
Listen Labs maintains an eval measuring whether the AI follows instructions and avoids repetitive questions. With GPT-4, the same question was sometimes asked 100 times. The initial eval score was 20 percent and improved to 85 percent. A new, more advanced eval was created measuring screen recording behavior and skipping irrelevant questions, resetting the score to 20 percent.
Vertical AI companies gain advantage through proprietary evals that they continuously improve. The pattern is to push forward better data, harder problems, better data, and repeat.
The core company loop is figure out what to build, then build it. The build component is advancing rapidly and exponentially while figuring out what to build is the bottleneck being addressed. The broader loop is strategy and execution. AI accelerates execution, making it faster, cheaper, and better. Listen Labs positions itself to improve the strategy component, from what to build to what to say.
When a one-person billion-dollar company exists, Listen will be part of that autonomous loop alongside coding agents.
In software, an agent could identify a bug from customer feedback, create a PR, close it, and ship it. The same principle applies to consumer packaged goods and other industries. Procter & Gamble spends tens of millions of dollars launching in new markets, making accurate customer understanding critical before launch.
Procter & Gamble is the archetype of best market research organization. They function as marketing companies identifying niches people care about and building specific brands to solve those problems. Tide Pods were developed after interviews revealed that washing liquid was uncomfortable to use; customers wanted something easier, leading to the pod format.
Mars conducted one of the first market research studies in the 1950s. M&M's were originally designed for military use as a sweet treat that does not melt in pockets. Market research identified young kids as another segment. The advertising strategy pivoted because the product does not melt and ruin furniture.
Human input will remain essential because even with AGI, humans are irrational and chaotic. New TikTok trends can emerge and require complete marketing strategy changes.
Uncertainty exists around the level of role simulation will ultimately play and how accurately models can predict human behavior. As AI improves at understanding human wants, the remaining delta—what exists in a human's mind that is not in the AI's mind—becomes more important.
Listen Labs has network effects on the panel through supply and demand dynamics. Data network effects exist because more interviews improve simulation quality. The product is sticky because accumulated interviews enable longitudinal tracking.
Founders want to build complex products, but customers want something stupid simple that works without configuration or custom workflow building.
Creating interview guides is difficult and is an academic subject. Service firms exist because they know methodology for understanding pricing, brand perception, and other topics. Leading the witness must be avoided.
Initially, customers created interviews with vanilla LLM models and received unusable data. Listen Labs trained the agent to follow best practices so good data is consistently produced. This is a vertical AI advantage.
Market research currently invites judgment on ideas. Hallucinations can function as a feature in generative AI. The question is whether user research will evolve into live product ideation where AI invents solutions during interviews and helps visualize them.
Today, customers manually use AI to generate images of concepts and feed them into interviews. Listen Labs now offers an MCP so users can feed data into Claude, run Listen in a loop, generate marketing ideas or concepts, and iterate. The possibility of AI live brainstorming solutions during interviews, including image generation, is being considered.
Keep Sequoia Capital in your library
Save the videos and channels worth coming back to, and find them again in one place.
