What if, before launching that new product or restructuring your company, you could test it on digital versions of your actual customers—AI agents trained to think, shop, and complain just like the real people would?
That's the pitch from Simile, a Stanford-spawned startup that just landed $100 million in Series A funding to build what amounts to a crystal ball for corporate decision-making. Index Ventures wrote the lead check on February 12, joined by Bain Capital Ventures, Hanabi, and A* Capital. The deal also pulled in a who's-who of AI royalty: Fei-Fei Li and Andrej Karpathy took angel stakes, alongside Adam D'Angelo, Guillermo Rauch, and Scott Belsky.
It's an audacious amount of capital for technology that essentially promises to clone human behavior at scale. Whether that promise holds up to scrutiny is another matter entirely.
From Academic Paper to Corporate Oracle
The company's origin story runs through Stanford's computer science labs, where CEO Joon Sung Park published a 2023 research paper that turned heads in academic circles. "Generative Agents" earned a Best Paper award at UIST by creating AI characters that mimicked realistic human behavior inside a virtual town called Smallville—think The Sims, but with agents that could form relationships, hold grudges, and throw impromptu parties.
Park teamed up with Michael S. Bernstein, a Stanford professor now on leave, and Percy Liang, who runs the university's Center for Research on Foundation Models. Rounding out the founding team are Elaina Yallen, previously at Hebbia, and Mihika Kapoor, who cut her teeth in product roles at Figma and Meta.
What began as an academic exercise has morphed into something corporates apparently find irresistible.
Testing Decisions Before They Happen
CVS Health is using the platform to figure out store layouts and inventory decisions. Telstra, Banco Itaú, Suntory Beverage & Food, and Wealthfront have signed on as early customers. The use cases range from mundane product testing to the genuinely high-stakes: modeling litigation outcomes and scenario planning for major strategic pivots.
During a Bloomberg TV appearance, Park made a claim bound to raise eyebrows among skeptics: Simile correctly predicted eight out of ten analyst questions on a simulated earnings call for one (unnamed) client. It's the kind of stat that sounds impressive until you start asking about sample sizes and what "predicted" actually means in practice.
Still, the company has moved quickly to bolster its credibility. A newly announced partnership with Gallup will anchor Simile's simulations in the polling giant's nationally representative panel data—a savvy hedge against accusations that the technology amounts to expensive guesswork.
How They Built the Thing
Training the underlying foundation model took seven months, according to the company. The data diet: hundreds of interviews with real people, transaction logs, and reams of text scraped from behavioral science journals. Shardul Shah, the Index Ventures partner who led the round, described it as building "digital twins orchestrated to answer what real people will do and why."
One independent analysis—though Simile hasn't disclosed who conducted it or under what conditions—found that the company's AI agents replicated 85% of responses from the General Social Survey when compared to humans retaking the same survey two weeks later. That's perhaps more reassuring than it sounds, given that humans themselves aren't perfectly consistent when answering surveys weeks apart.
A Crowded, Contentious Space
Simile isn't alone in chasing this vision. Artificial Societies pulled in roughly $5.35 million in seed funding last year. Dialogue AI and SYMAR are hawking their own versions of synthetic focus groups and digital market research. The space has become a magnet for venture capital, driven by the tantalizing possibility that AI can finally crack the code on predicting human behavior at scale.
Park's stated ambition certainly doesn't lack for scope: he wants to simulate all eight billion people on Earth. It's the kind of moonshot goal that either attracts $100 million rounds or invites pointed questions from scientists about whether the underlying methodology can support such grand claims. In this case, it did both.
Whether Simile can deliver on that vision—or whether its technology represents a genuine leap forward versus an expensive statistical parlor trick—remains to be seen. For now, the company has the backing, the pedigree, and the early customer traction to at least try.
The rest of us will just have to wait and see if these digital twins really know us as well as Simile thinks they do.
