Is There Something Between Real UXR and Just Asking Claude?

Full disclosure before we start: I’ve advised Vectorial for a while since they were in the Berkeley Skydeck program and I’m a small investor, which is exactly why I tried to play this one straight. When I sat down in person with cofounders Taran Singh and Rahul Garg, with their third cofounder Jasmine Kaur joining us later, I wanted to understand two things. The first was what it actually looks like to build a company from scratch in 2026, when every startup is an AI startup and finding a defensible lane is harder than it has ever been. The second was a question I suspect many of you are quietly asking during annual planning season: if you need to build things customers love, is there anything in between a six week UXR study and typing “what would my users think of these ten ideas?” into a chatbot? The founders Taran has spent roughly a decade modeling human behavior, starting in defense research where he predicted how large population segments would act, then moving through a research collaboration with Coursera at Carnegie Mellon on how learners from different backgrounds learn, and later joining a small research group that tried to predict career decisions from enormous volumes of resume data. What stuck with him across all of it was a slightly uncomfortable finding, which is that people are more predictable than we like to believe, both individually and collectively. As a visiting research fellow at Berkeley he helped build a behavior model called Sapiens, which the team says is about 87% accurate at predicting human behavior, and that became the seed of Vectorial. Rahul came at the same problem from the builder’s side. He started in logistics operations managing around 150 people on the ground, then built more than ten products from zero across seven industries, from hyperlocal delivery to travel to real estate marketplaces, and he describes the common thread as an obsession with user psychology that owes a lot to Thinking, Fast and Slow. Jasmine brings deep consumer health experience, most recently leading AI initiatives at Pfizer, where she sat on the buyer’s side of the very decision Vectorial now asks enterprises to make. Why not just ask Claude? This was the first thing I pushed on, because it’s what every skeptical PM will ask. Taran’s answer was the most interesting idea of the conversation: frontier models like Claude, Gemini and ChatGPT are trained and tuned to solve rational problems in code, science, law and medicine, since that’s where the near term money is, while human behavior is fundamentally about the irrational side of people. By his account, models asked to role play a customer have plateaued at roughly 51 to 56% accuracy in predicting behavior over the past two years, and he pointed to academic work from CMU and Berkeley suggesting that tuning for benchmarks may be making them worse at it. Vectorial claims Sapiens is about 40% more accurate because it learns from data annotated specifically for behavior, including the unstated needs and opinions people rarely articulate. I found this framing genuinely clarifying. You wouldn’t hire your CMO out of an MIT math PhD program, and while that might be a fine place to find a CTO, your CPO, your CMO and arguably your CEO need to understand how humans actually work, including the messy motivations and feelings that shape how people use a product and why. If an LLM treats irrational, emotional behavior as noise to minimize on the way to a correct answer, then it is structurally the wrong tool for predicting customers, for whom that noise is the signal. As Taran put it, there are only so many ways to be logical and endless ways to be irrational, and your particular irrationality differs from mine. That said, I’d treat these figures as the company’s own claims rather than settled fact, since “accuracy at predicting behavior” can mean very different things depending on the task, the baseline and who designs the benchmark. If you evaluate any simulation vendor, ask exactly what was predicted, against what ground truth, and on whose data. What simulation actually is The word “simulation” gets used loosely, so it’s worth being precise about what Vectorial means by it. There are two parts, the first being a model of your target customers built at the level of individuals, often thousands of them, and the second being scenarios you run against that modeled audience, whether a landing page, an onboarding flow, a campaign concept or a feature idea. Because each person is modeled individually, you can see how a specific segment reacts and then roll that up into how the collective responds, which is what Taran means when he describes simulation as exploring every plausible scenario rather than returning a single averaged opinion. The data comes from two places. The team starts with public communities like Reddit threads, Facebook groups, Quora and Twitter, where people talk candidly about their lives, and then fills gaps with AI moderated interviews of real people recruited through curated panels (they claim reach into more than 30 million users) to capture the unstated context that public posts miss. The goal is what they call holistic behavioral coverage, a whole person picture spanning life stage, socioeconomic background, personality and other purchasing decisions, rather than narrow answers about your product alone. Jasmine drew the sharpest distinction of the day, one I think every enterprise PM should internalize, which is that behavior simulation is not customer insight. Mining your existing data summarizes what customers have already said and done, whereas modeling a person means predicting how they’ll react to something new. Her healthcare example stayed with me, because pharma companies sit on terabytes of clinical and claims data, yet two clinically similar patients are not behaviorally similar, which helps explain why adherence to many treatments sits below 30%. That gap is also her core argument for buying rather than building in house, since most companies’ data is too narrow to model the whole person. Finding their lane As someone who has been through two founder journeys, what impressed me most was how deliberately the team has narrowed its focus. Other companies sell simulation for broad market research and high level strategy, while Vectorial has chosen to specialize in user research, the execution layer after a bet has been approved, where the questions are about what shape the product should take and how a feature will land. The logic is frequency, since customers reportedly run about 400 simulations a month, roughly 20 times the research volume they managed before, and a tool used daily becomes a habit in a way a quarterly study never does. Early traction is concentrated in healthcare and gaming, an unusual pairing given how differently those industries weigh the consequences of getting it wrong. Two use cases surprised even the founders. PMs love simulated focus groups in which personas debate each other, surfacing disagreements they didn’t know existed, and a growing number of customers want to test AI agents against simulated humans who behave messily, adding the noise and wandering intent that a rational LLM acting as test user or judge would never produce. In one case, acting on those agent simulations reportedly led to a 25% increase in engagement. Where I think it’s also still early Several open questions deserve honest airing, and the first is trust. Vectorial runs a calibration period of about six weeks in which it uses a client’s past studies and live interviews to tune the audience, then runs blind studies and shows the overlap, and it exposes a “proof of work” view tracing each simulated opinion back to anonymized profiles and inferred traits. That transparency is welcome, but the team also argues that matching past human studies isn’t the true north star, because those studies carry their own small sample, incentive and interviewer biases. I agree in principle, yet it creates … This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit firesidepm.substack.com