A new research paper published last week may signal a seismic shift in how companies study consumer behavior. The study, led by researchers at PyMC Labs and Colgate-Palmolive, outlines a method that allows large language models (LLMs) to simulate human purchase decisions with remarkable accuracy—without needing actual survey respondents.
For professionals in the credit and collection industry experimenting with artificial intelligence to understand borrower intent or consumer sentiment, this represents a major step forward: it shows that AI can learn not only what consumers say, but also how they think.
Traditional market research costs companies billions each year but often produces inconsistent results due to bias and panel fatigue. LLMs have long been viewed as potential replacements for survey panels, but they’ve struggled with one basic task — providing realistic numerical ratings. When asked to rate a product from 1 to 5, models often cluster around the middle, generating data that looks nothing like human surveys, according to the study.
That’s the problem a new method, called Semantic Similarity Rating (SSR), aims to solve. Instead of asking the model for a number, researchers prompt it for an open-ended response (“I’d definitely buy this” or “Maybe if it were cheaper”). The model’s answer is then converted into a numerical rating by comparing its text embedding to five reference statements representing each Likert scale value. The closer the AI’s response is semantically to “I would definitely buy this,” the higher its score.
The research team tested SSR using 57 real-world product surveys and 9,300 human responses from the personal care industry. The synthetic respondents achieved 90% of the test/retest reliability of human surveys and produced distributions that were statistically almost identical to human data, according to the paper.
More importantly, the simulated consumers didn’t just produce scores; they offered qualitative insights explaining why they liked or disliked a product. Compared with short, generic human survey responses (“It’s good”), AI-generated comments were richer, more nuanced, and less biased toward positivity.
While the study focused on personal care products, the implications reach far beyond consumer goods. Credit and collection professionals could use similar LLM-based frameworks to simulate borrower behavior or test how consumers might respond to new contact strategies, repayment offers, or digital self-service tools, all without needing to collect personal data.
This approach could help agencies, banks, and fintechs identify trends, evaluate messaging, and predict outcomes in controlled, privacy-safe environments. Imagine being able to model how different types of consumers would react to a proposed payment plan without needing to actually send out the proposed plan and wait for a response.
