Designing Better Products With Less Data: How and Why We Created Sensory AI Drivers+

In today’s crowded categories, the challenge isn’t a lack of ideas - it’s knowing which sensory changes will genuinely improve consumer appeal, without testing endless products with ever larger samples. Product teams are under pressure to move faster, reduce cost and still make confident decisions, yet traditional approaches to identifying sensory drivers often demand more products and consumers than are practical, or rely on averages that mask what truly matters.

Sensory AI Drivers+ was developed to address this challenge. By combining sensory and consumer data, it allows teams to identify the sensory attributes that matter most to consumers - and crucially, to do so with fewer products and smaller sample sizes than many conventional approaches. The goal isn’t to design for an abstract “average” consumer, but to understand how real individual preferences combine, and how to optimize a single product to appeal to as many consumers as possible.

The development of Sensory AI Drivers+ was a true collaboration between MMR’s Data Analytics and Data Science teams. Together, we paired established sensory science principles with modern AI techniques, ensuring the models were not only statistically robust, but also practical and interpretable for real product decisions. This partnership was essential in creating a tool that balances scientific rigor with commercial usability.

Want to see how it works in practice? ⇨ Explore the Sensory AI Drivers+ factsheet

For each use case, the model is built from scratch, using novel data from real products tested with real consumers, creating a model specific to the client’s category and objectives. In simple terms, the AI analyzes thousands of model variations to find the sensory attributes that consistently drive liking, with the final model giving greater influence to the attributes that repeatedly prove important. For those interested in the technical detail, our Principal Data Scientist, Jonathan Sands, describes the process as three key phases:

• Attribute selection: Thousands of models are fitted on different subsets of the data to identify which sensory attributes consistently predict the full liking distribution. A Partial Proportional Odds (PPO) model with Complementary Pair Stability Selection is used to estimate how reliably each attribute is selected, allowing the AI to distinguish stable signals from noise.

• Model Fitting: These reliability scores guide an adaptively regularized PPO model. Attributes with stronger and more consistent evidence receive less penalization and therefore greater influence in the final model. The result is a data-driven estimate of how each sensory attribute impacts liking across the population, integrating diverse individual preferences into a single robust solution.

• Validation: Techniques such as Synthetic Products and Leave One Out analysis are used to train and validate whether the model can accurately predict a product that it hasn’t seen during training, therefore ensuring that the model is generalizable because the model is not overfit, especially a concern on small data sets.

This is not a replacement for preference mapping across larger product sets, which shows how to win by targeting and delighting different consumer preferences. Sensory AI Drivers+ is best suited in a tighter sensory space, where it tells you what wins and why - guiding the optimization of a single product to appeal to as many consumers as possible. In these situations, moving beyond averages and embracing the full complexity of consumer preferences, teams can design products that resonate more widely with greater confidence, speed, and efficiency.