[ Funding
](https://www.unite.ai/series/funding/)
[Add Unite.AI to your preferred sources on Google](https://www.google.com/preferences/source?q=unite.ai)
Simile has raised more than $200 million in Series B funding at a $2 billion post-money valuation, providing the Palo Alto-based startup with substantial capital to develop artificial intelligence models designed to simulate how people make decisions.
Greenoaks led the round, while existing investor Index Ventures increased its backing. Hanabi, Bain Capital Ventures, A*, Factory and CVS Health Ventures (CVS ) also returned, with Definition joining as a new investor. The financing arrives only five months after Simile emerged from stealth with a $100 million Series A led by Index Ventures, taking the company’s disclosed funding beyond $300 million.
Since its public launch, Simile says it has increased revenue fivefold, expanded to more than 50 employees and run tens of millions of simulations for Fortune 100 companies. The company plans to use the new funding to advance its foundation model for human behavior, improve the reliability of its simulations and expand the platform across industries including healthcare, financial services, consumer products and media.
Building AI to Predict Behavior Rather Than Generate Content #
Most of the current generative AI market is focused on creating outputs. Large language models can produce software, marketing copy, images, product concepts and business plans within seconds.
Simile is targeting the decision-making process that comes before those outputs are produced. Its platform is intended to help organizations evaluate what they should build, how different audiences may respond and what second-order effects could emerge after a decision is implemented.
The company describes itself as building a foundation model for human behavior rather than another general-purpose language model. Instead of optimizing primarily for logical answers or fluent text, the model is being trained to represent differences in people’s preferences, experiences, values and decision patterns.
Organizations can use the platform to create comparable scenarios and examine how the same simulated population responds when variables such as pricing, messaging, product features or policy conditions are changed. The aim is not simply to ask synthetic respondents for opinions, but to model how behavior may shift across multiple conditions and over time.
From Stanford Research to an Enterprise Simulation Platform #
Simile’s technology builds on research conducted by co-founder and CEO Joon Sung Park and fellow Stanford researchers Michael Bernstein and Percy Liang.
Park was the lead author of the 2023 “Generative Agents: Interactive Simulacra of Human Behavior ” paper, which introduced a simulated town populated by 25 AI agents. Each agent used a memory stream alongside retrieval, reflection and planning mechanisms to maintain context and determine its next actions.
The agents formed relationships, developed routines and coordinated activities without every interaction being explicitly programmed. The experiment helped establish many of the architectural ideas now used in multi-agent simulations, where autonomous agents interact with one another rather than responding independently to isolated prompts.
Later research tested whether generative agents could reproduce the attitudes and behavior of real individuals. In a study involving 1,052 participants, researchers created agents using qualitative interviews and evaluated how accurately they replicated participants’ responses.
The agents achieved 85% of the accuracy that participants demonstrated when answering the same questions themselves two weeks apart. This is an important distinction: the result does not mean that human behavior was predicted with 85% absolute accuracy. It means the agents approached the consistency level of the people they were designed to represent.
Simile’s website also points to subsequent work involving 2.9 million responses from 210 social science experiments. The company says fine-tuning models on this behavioral data improved alignment on previously unseen studies by 26% compared with the underlying base model.
Confidence Scores Are Central to the Product #
One of Simile’s more consequential technical claims is that it has trained a separate confidence model to estimate the reliability of each simulation.
Predicting human behavior is inherently uncertain. The same person may make different decisions depending on timing, context, available information and factors that are not present in a dataset. A system that produces a single confident prediction without communicating those limitations could encourage organizations to treat simulated behavior as fact.
Simile says its platform compares simulated response distributions with responses collected from real people using Total Variation Distance, a statistical measure of the difference between two probability distributions. Each result receives a confidence score based on the amount and relevance of the supporting human evidence.
When evidence for a particular population or scenario is limited, the confidence score is intended to make that uncertainty visible. The company also says its models are recalibrated weekly to incorporate changes in areas such as prices, policy, economic conditions and consumer behavior.
That validation layer could prove as important as the simulations themselves. For enterprises making expensive or sensitive decisions, knowing when a model is unlikely to be dependable may be more valuable than receiving another plausible prediction.
CVS Health Uses Simulations to Test Decisions Before Launch #
CVS Health is one of Simile’s most prominent early customers and also participated in the funding round through CVS Health Ventures.
The healthcare company has used simulated populations to validate existing customer research, examine drivers of patient satisfaction and test strategies related to medication adherence. It has also explored populations that can be difficult or expensive to reach through traditional surveys, including patients with chronic conditions.
According to Simile, the CVS initiative draws on 2.9 million consented responses from more than 400,000 participants across over 200 behavioral scenarios. Simulations have been used to pre-screen ideas and identify which hypotheses warrant further testing through real-world research or pilots.
This distinction matters. Simulation can reduce the number of ideas that need to be tested directly with customers, but it does not eliminate the need to collect real-world evidence. CVS has described the technology as a way to prioritize research and refine potential interventions before exposing patients or customers to them.
Simile is also working with Gallup on policy research, workplace studies and social trend analysis. Other organizations featured by the company include Wealthfront, Deloitte, Banco Itaú and Suntory. Enterprises are applying the platform to areas such as product testing, market entry, customer experience research, competitive positioning and earnings-call preparation.
Moving From Individual Agents to Simulated Markets #
Many of Simile’s current applications resemble accelerated market research. A company can test several product concepts or messages against a simulated population before deciding which ones deserve a real campaign or pilot.
The longer-term opportunity is considerably more complex. Simile wants to model interactions among customers, competitors, institutions, policies and markets. Rather than predicting how isolated individuals will answer a question, these multi-agent simulations would examine how people influence one another and how decisions produce cascading effects.
Possible applications could include modeling reactions to financial shocks, policy changes, product launches or changes in incentives. A simulation might examine not only whether customers adopt a new service, but how that adoption changes competitor behavior, public perception and demand for related products.
This is where technical limitations will become more significant. Small errors can accumulate when thousands or millions of agents interact, particularly in simulations where early events change the trajectory of everything that follows. Simile will need to demonstrate that its validation methods remain meaningful as simulations move from individual predictions toward complex, dynamic systems.
A New Layer of Decision Infrastructure #
The size and speed of Simile’s latest financing suggest that investors see behavioral simulation as a potential new layer in enterprise decision-making.
Organizations already spend heavily on surveys, consulting, focus groups, product trials and market research. AI simulations could make it cheaper to explore more options, test narrower audience segments and identify weak ideas before committing significant capital.
The technology also introduces difficult questions around consent, representation and accountability. A simulated population is only as useful as the people, behaviors and circumstances represented in its underlying data. Predictions may become less reliable when deployed across unfamiliar cultures, rapidly changing markets or high-stakes situations with limited historical precedent.
For that reason, the strongest near-term role for platforms such as Simile may be to supplement human research rather than replace it. Simulations can help organizations decide which questions to investigate, but real customers and communities remain necessary for verifying whether the results hold outside the model. Simile’s new funding gives it the resources to test whether human behavior simulation can progress from an emerging research field into dependable enterprise infrastructure. Its immediate challenge is not simply running more simulations, but proving that organizations can understand their limits and know when the results are reliable enough to influence consequential decisions.