Aaru Uses AI to Simulate Consumer Decisions

Aaru, a two-and-a-half-year-old startup valued at $1 billion, uses AI simulations to predict how groups of people will make decisions, according to its co-founders Cam Fink and Ned Koh in a Goldman Sachs interview. They said businesses use the predictions for consumer research, including product launches and marketing, and for decisions in financial services, utilities and other sectors. Fink said the company combines large language models (AI systems trained to process and generate text) with statistical and mathematical models because language models alone do not provide the behavioral depth businesses need to trust decisions.

Aaru says its simulations are grounded in real-world outcomes and judged by whether clients make better business decisions, such as improving sales, rather than simply reaching a decision faster. Its data comes from public sources such as census information and licensed academic research, purchased sources including credit-card purchase histories, receipts, music streams and social-media data, as well as proprietary data-sharing partnerships. Fink said each client is sandboxed, the company is SOC 2 compliant, and it does not train on customer data.

For consumer brands, Aaru can simulate product development from early ideas through identifying likely buyers, choosing product concepts and flavors, and refining caffeine content and package size. Fink said one client shortened a product innovation cycle from nine months to eight weeks; the product reached store shelves earlier this year and performed better than the client's previous launches. He also said some clients take 12 months or 18 months to move from an idea or market entry to manufacturing readiness.

The company says its simulations can describe groups as small as a few hundred people, including buyers of specialized software, and as broad as U.S. household purchase decision-makers, British residents or people in the European Union. Fink said Aaru does not model individuals or process personally identifiable information. Its approach has shifted from training a large language model on survey results to models built by Aaru, which he said improved the technology's accuracy and extensibility.

Aaru, a two-and-a-half-year-old startup valued at $1 billion, uses AI simulations to predict how groups of people will make decisions, according to its co-founders Cam Fink and Ned Koh in a Goldman Sachs interview. They said businesses use the predictions for consumer research, including product launches and marketing, and for decisions in financial services, utilities and other sectors. Fink said the company combines large language models (AI systems trained to process and generate text) with statistical and mathematical models because language models alone do not provide the behavioral depth businesses need to trust decisions.

Aaru says its simulations are grounded in real-world outcomes and judged by whether clients make better business decisions, such as improving sales, rather than simply reaching a decision faster. Its data comes from public sources such as census information and licensed academic research, purchased sources including credit-card purchase histories, receipts, music streams and social-media data, as well as proprietary data-sharing partnerships. Fink said each client is sandboxed, the company is SOC 2 compliant, and it does not train on customer data.

For consumer brands, Aaru can simulate product development from early ideas through identifying likely buyers, choosing product concepts and flavors, and refining caffeine content and package size. Fink said one client shortened a product innovation cycle from nine months to eight weeks; the product reached store shelves earlier this year and performed better than the client's previous launches. He also said some clients take 12 months or 18 months to move from an idea or market entry to manufacturing readiness.

The company says its simulations can describe groups as small as a few hundred people, including buyers of specialized software, and as broad as U.S. household purchase decision-makers, British residents or people in the European Union. Fink said Aaru does not model individuals or process personally identifiable information. Its approach has shifted from training a large language model on survey results to models built by Aaru, which he said improved the technology's accuracy and extensibility.