A 21-year-old cofounder’s sales pitch to clients begs them to question the company’s results: ‘Do not trust us. Do not trust our model’

Ned Koh, the founder and president of Aaru, said the AI startup’s ability to predict outcomes better than human-led survey data speaks for itself.

Nearly 2 million New Yorkers voted in the city’s primary mayoral election. 

Aaru, a two-year-old AI startup, simulated all of those voters—agents built from the kind of data the company collects, like credit card purchasing history, food-delivery orders, and demographic records—and tried to predict how the votes would be cast. The simulation came within 2,000 votes of the final count.

Aaru’s 21-year-old founder and president Ned Koh doesn’t care if you believe that. He said at Fortune’s Brainstorm Tech conference recently that while he lacks years of experience in enterprise sales and market research, he’s let the company speak for itself.

“Our entire sales methodology is to go to a business and say, ‘Do not trust us. Do not trust our model,’” Koh said. “In fact, I hope you’re skeptical, because it means you’ll be a better customer in the long run.”

Koh cofounded Aaru in March 2024, at 19, with Cameron Fink, then 18, and John Kessler, then 15 (and too young to sit on the board). The company organizes thousands of AI agents into statistically representative populations—each agent assigned an age, an income, a zip code, a gender, among other things—and surveys them in place of human respondents.

The company begs a somewhat uncomfortable question: Are humans so pliable to conditioning, that you can predict their behavior using second-order data? The company builds AI to train and benchmark on what it calls “outcomes”: election results, purchase data, the things people do rather than the things they say.

The outcomes are pretty spot on, though not always in the way clients anticipate. Aaru counts on clients’ incredulousness for the success of this framework. Prospects hand over an old survey—the people they polled, the questions they asked—and Aaru reruns it blind against the known result.

The proof he cited: Ernst & Young, a partner of Aaru, which spent six months surveying 3,600 people across 30 countries for its global wealth study. One question asked whether respondents would keep their parents’ wealth manager after they passed away, and 82% said yes. However, real world retention is actually closer to 20% to 30%, Koh said. Aaru reran the entire survey blind—no humans—and its simulated respondents said roughly 40%. So the simulated people were closer to the truth than the real ones.

Eliminating people bias

Aaru’s success hinges upon people’s own bias toward themselves, Koh argued. Intentions are often different from behaviors, so when surveys ask people what actions they plan to take, it often deviates from what they actually do. Scientists call this the “intention-behavior gap,” and have used the concept to explain why individuals don’t go to the gym despite their New Year’s resolutions, or why people make impulsive purchases even after planning their monthly budget.

Koh puts it more simply. “People lie,” he said. “If you ask someone how much they drink alcohol, they’re never going to be honest.”

Koh has seen this play out in the top demographic Aaru is asked to survey (though he did not disclose which clients requested this information, or what results would be used for): GLP-1 users, who he claims tend to lie by omission. 

“They’re not even telling their family they’re on these medicines,” he said. “They’re not going to tell your study.”

Investors and clients are buying into Aaru’s methodology, including EY, Accenture, Interpublic Group, McDonald’s, Boston Beer, A24, and Bayer. The startup recently worked with Spindrift on product innovation, which led to the sparkling water company launching a still tea drink.

“That’s a totally separate category,” Koh said. “They’ve never done that before, but we can be confident in our predictions of that space, because we can be confident that our models are good at thinking through the profiles.”

In December, Aaru raised a Series A led by Redpoint Ventures at a $1 billion headline valuation—on annual recurring revenue still under $10 million, according to TechCrunch. The round used a multi-tier structure: some investors paid the $1 billion price, and others paid less, putting the blended valuation below the headline number. Total raised to date is roughly $88 million, per PitchBook. 

There is, of course, risk associated with relying on AI to mitigate the impact of human bias in surveys: AI respondents inherit AI biases. A Cornell University-led study published last year found thousands of bot respondents across four large-language models produced “flattened” representations of marginalized groups, portraying them as outsiders would see them, not as they see themselves. This is largely a result of AI being trained on human data—which contains biases.

Koh sees things differently, arguing that competitors who train their agents on survey data import the bias wholesale. Aaru, he said, trains on behavior, which he said is more objective. Koh did not disclose the startup’s architecture or the mechanisms behind how it is able to reduce partisanship.

There are also things Aaru concedes it can’t do. It tried to simulate behaviors of figures like President Donald Trump and Jerome Powell, but found it was unable to accurately predict outcomes, even with decades of public statements from the individuals. There are certain outcomes that are just too complicated to forecast.

“There’s just so much variance,” Koh said. “It doesn’t matter how much data you have on somebody.”

Sasha Rogelberg contributed reporting to this story.

This story was originally featured on Fortune.com