Here's a number that should stop you cold: a market research startup with fewer than 60 employees is now worth $2 billion — roughly 60% of the market cap of NIQ, a global top-five research firm with 38,800 staff. The industry it's attacking does $160 billion in revenue a year. And the startup took about two years to get there.
That startup is Simile, founded by Joon Sung Park, the computer scientist behind the famous "Smallville" experiment in which 25 AI agents spontaneously developed social behavior. Simile's pitch is simple to state and hard to overstate: replace human respondents with AI "digital twin consumers," and compress a four-to-six-week research project into 15 to 30 minutes.
This isn't a funding story. It's a structural one: AI may be the first thing that turns market research — a business that has always scaled by hiring people — into software that scales by running more simulations.
How it works: build the people, then run the experiment
Simile's system works in two steps. First, it "builds" the population. Enterprises feed in authorized interview transcripts, historical surveys, and behavioral data; Simile generates tens of thousands of AI individuals with distinct ages, incomes, preferences, and shopping habits. CVS Health's collaboration, for example, is built on roughly 2.9 million authorized responses from more than 400,000 participants covering 200+ behavioral scenarios — from which Simile produced over 100,000 digital twin consumers.
Second, it runs the experiment. A company drops a new ad, a price change, or a product concept into the simulated environment, and the AI agents choose, talk, and influence each other — propagating reactions the way news spread through Smallville's pixel town. In CVS Health's internal testing, some simulations calibrated against company data matched historical human responses with up to 95% similarity.
Two distinct business models have emerged from this wave. One replaces respondents: Simile and Aaru build agent populations to stand in for real consumers. Aaru once simulated ~5,000 agents to model voters ahead of a New York Democratic primary — its prediction missed the actual result by 371 votes. The other replaces researchers: Listen Labs keeps real human interviewees but automates the questioning, follow-up, and synthesis, completing over a million interviews in nine months.
The framework: human-density arbitrage
Here's the puzzle the old industry could never solve. A $160 billion market, and barely a single company worth more than $10 billion — Gartner is the exception, and its revenue is mostly IT research subscriptions, not consumer research. Why? Because the business model was people-stacked. Every new project meant new questionnaires, new recruiters, new interviewers, new analysts. Revenue and headcount rose together. NIQ's operating costs absorbed 88.6% of revenue in 2023 and still 82.4% in 2025, with the improvement bought mostly through layoffs and efficiency programs.
AI breaks that coupling. The expensive part of Simile is upfront — data acquisition, model training, calibration. Once the system is built, each additional simulation costs little more than electricity and tokens. The old moat — a global network of interviewers and decades of accumulated sample panels — stops being an asset and starts being a cost line.
Call it human-density arbitrage: in any service industry where cost is proportional to headcount, an AI-native company with the same output quality gets to price like software while competitors price like labor.
What comes next
Market research is just the opening act. Consulting, investment research, due diligence, legal discovery — every knowledge service whose billable unit is the person-hour sits in the same crosshairs. The 60-person company at $2 billion is a preview of the repricing that will hit an entire category.
The risks are real, though. People say one thing in surveys and do another at the shelf; simulated agents inherit that gap, and well-calibrated models can also inherit the blind spots of the data they're built on. There's an authorization question — whose data, used how — and a subtler one: if every company simulates the same "average consumer," differentiation may quietly converge toward the mean.
That's why the winning pattern is likely bounded, not total. Use simulation as a fast filter — cheap, broad, iterative — and keep human judgment for the expensive final mile: interpreting why a result happened and deciding whether to act on it.
What to do about it
For research buyers: run a pilot with a tightly scoped question and calibrate the simulation against your own historical data before trusting it. For builders: the defensible asset isn't the model — it's authorized proprietary data and vertical calibration no competitor can rent. For incumbents: stop selling hours. Start selling decision systems.
Simile's $2 billion is less a verdict on one startup than a re-pricing of an entire industry's cost structure. The same math is coming for every business that still sells people by the hour.