Companies Could Soon Use AI Consumer Twins to Predict What You’ll Buy

AI startup Simile is commercialising digital replicas of consumers after raising a total of $300 million (£228 million), while separate research has found substantial discrepancies between synthetic survey respondents and real people.

Pew Finds a 12-Point Gap Between AI and Human Respondents

A focus group is usually a room full of people. Simile's version is a room that can answer before anyone walks in, raising a broader question about whether prediction can eventually replace participation.

Pew Research Center found that AI-generated respondents differed from human results by an average of 12 percentage points across nearly 300 questions. Differences exceeded 15 points on around 28% of questions, while some individual gaps passed 20, 30 or 40 points.

Human respondents were around four times as likely as their AI counterparts to select 'not sure', and nearly half of the questions had at least one answer option selected by no AI-generated respondent. Pew did not test Simile's technology directly, so its findings indicate broader limitations in synthetic polling rather than establishing the accuracy of Simile's commercial simulations.

Pew concluded that 'survey-taking is best left to humans'.

Simile's $300M Funding Bet Moves Into Healthcare Research

Joon Sung Park, Simile's chief executive, told Business Insider that the company's models take 'a description of a person or a group of people' and simulate how they might respond in different contexts.

Park said Simile had raised a total of $300 million (£228 million), with its latest funding round valuing the company at $2 billion (£1.52 billion). The company previously announced a $200 million (£152 million) Series B, five months after a $100 million (£76 million) Series A, TechCrunch reported.

Simile says its synthetic panels can help companies test products, advertising and customer experiences. Park described the capability as 'a superpower for these companies'.

CVS Health is using generative-agent simulations to test healthcare journeys, messaging and digital product changes before real-world rollout. The company says its system was built from 2.9 million consented responses involving more than 400,000 participants across more than 200 behavioural scenarios.

CVS says some research processes were reduced from weeks to hours, while stressing that the simulations do not replace real-world research. Experts remain involved in interpreting the results and deciding whether findings warrant further testing.

Gallup and UK Data Rules Keep the Human Question Open

Gallup is evaluating simulated responses with Simile, but says the method should extend research rather than replace established measurement. Potential uses include refining survey questions, pre-testing research and exploring hard-to-reach populations before direct data collection.

For organisations processing UK residents' personal data, the Information Commissioner's Office says AI developers and users must establish appropriate lawful bases for their processing activities.

Its guidance also addresses when AI-generated inferences about identifiable individuals may constitute personal data, including specially protected categories.

Simile's participant privacy notice describes creating text-based digital twins from people's submissions and says the agents will not synthesise participants' voices. Any wording about sharing original audio or video recordings should be retained only after the current notice and participant agreement have been checked directly.

The replacement question is already appearing in practitioner discussions. One Reddit user described the appropriate framing as 'augmentation, not replacement', while the author of a separate Reddit post questioned whether participants were 'basically helping to train an AI that can replace us?'

For now, Simile's commercial proposition is clear: make research faster by modelling people at scale. The unresolved issue is whether a faster answer remains useful when the people being modelled would have answered differently.