FOUNDER FOCUSIn this edition, Tech in Asia managing editor Thu Huong Le speaks with Ned Koh, co-founder and president of US-based synthetic research startup Aaru, on the sidelines of SuperAI 2026 in Singapore. Aaru, which reached a billion-dollar valuation within just one year, has a pitch that sounds straight out of Black Mirror. The company says its AI models can generate synthetic populations that predict human behavior more accurately than traditional surveys and focus groups, potentially challenging established market research firms such as Nielsen. This interview has been edited for brevity and clarity.  You're saying that your AI can predict human behavior more accurately than surveys and focus groups, without interviewing a single person. How do you do that? What we do is simulate human behavior. One of the ways that manifests itself is market research. With market research today, you can test pricing and marketing, but it won't be very accurate. That's fine because humans are innately biased. What you can do with Aaru is you can simulate the population of Singapore and ask, "Who buys my shoes today?" Or you could even ask, "Among the people who buy my shoes today, who would buy a net-new product?" You can then filter from a top-down perspective because you can generate many thousands of people. You're not limited to saying, "Oh, I can talk to this person and ask for their opinion." If you're talking about modeling a population, there are two important parts. One is how you generate the actual people and make sure those people are accurate. The other is how you make sure the model is able to take on that persona or digital population. The first is a model-training problem, while the second is a data-filtering and acquisition problem. We don't use off-the-shelf models like ChatGPT or Claude. We build our own models, and they're trained on the behaviors people take. So instead of looking at polls, we look at vote counts. Instead of surveys, we look at click-through rates and sales results. We never look at what people say; we look at the actions people take, and that enables us to be much more accurate. What kind of data do you use to create these digital personas, and where do you get that data from? We might look at anonymized credit card purchase histories, health data, voter files, and more. It depends on the geography. We also use academic studies and other data sources that help us understand who people are, without ever having to ask them directly. A lot of the data we use is completely proprietary. Some of it is public, some of it we buy, and a lot of it comes through partnerships. When you think about training a model to understand behavior and outcomes - things that are free from bias - a lot of that data isn't public. What we'll do with certain businesses is, once we've proven the value and demonstrated how commercially successful our models can be, we'll run simulations in exchange for data. The biggest differentiation for us is that our number one value proposition isn't just saving cost and time. We care first and foremost about making the right decision, and because that's the case, we don't survey people. Almost every other synthetic research company that exists today started by training on survey data and human market research data, which means that, by default, they're biased and inaccurate. What types of clients does Aaru work with, and how do you acquire them, especially in more conservative markets such as those in Asia? We worked with a global consumer packaged goods business, one of the largest in the world. They conducted their traditional segmentation study, and we ran ours in parallel. The audiences we simulated were slightly more expensive to target on Meta and TikTok than the ones identified through their traditional approach. But they were four times more impactful because we were able to better simulate who would actually buy a particular product based on an advertisement. Ernst & Young had run a survey with 53 questions on 3,600 people across 30 countries, so it was a very large study. The company gave us the profiles of the people they spoke to and the questions they asked, but nothing else. No data, no outcomes, no results. We took that information and simulated the population they had spoken to in the real world. We then asked the simulated population the same questions and predicted the outcomes of those people's decisions more accurately than Ernst & Young's own study did. That took them six months. It took us four-and-a-half minutes. We're not just comparing our results against a survey. We're comparing them against actual outcomes. We can say: this is what the survey predicted, this is what our model predicted, and this is what actually happened. We can demonstrate that we're closer to reality using outcomes the client already knows, so it's really about earning trust. We have models that can process all sorts of different languages. Asia Pacific is a very large market for us, especially given how dominant the consumer sector is. We spend a lot of time in Asia, and we're very seriously considering putting people on the ground here. I think that, when it comes to deploying AI technology properly, especially with traditional enterprises and conglomerates, it's very important to have a human interface. You keep talking about eliminating human bias. Do you consider yourself an unbiased person? I try to be as unbiased as possible. I'm a very scientific person. I started this company, and I did drop out of Harvard, so I was an academic. People get pressure. All those biases make things less accurate, and surveys lead to a worse society in a lot of cases. Technology has helped the world in so many ways. Behavior has been a totally underrepresented thing, and that's where our business excels. We started this company because we think that understanding and being able to shape behavior is the most important problem on the globe. |
