Summary
PepsiCo just published a peer-reviewed paper on how it uses AI for pricing and promotions, and the results are strong. Gustavo Mendonça breaks down what the paper gets right, where it hits real limits, and what companies without PepsiCo’s resources can learn from it.
PepsiCo’s data science team recently published a detailed paper on how they use AI for pricing and promotions. It’s not a press release or a vendor pitch. It’s a peer-reviewed paper in the INFORMS Journal on Applied Analytics (Llenas et al., 2026), and it lays out the actual methods, the system design, and, helpfully, where things get difficult.
The two systems, PromoAI and PricingAI, support promotional planning across dozens of markets and base pricing decisions in the US and Mexico. About 85% of the promotional recommendations are accepted and used by the business. Planning that used to take weeks now takes minutes in many cases. It’s a strong result, and the team behind it deserves credit for both the work and for being open about how it was built.
Reading it, a few things stood out to us at Revenue Management Labs.
The payoff comes from the decision, not the analysis
One line in the paper sums up their own view well: the demand models on their own would just be descriptive analytics. Useful, but not something you can act on directly. It’s the optimization step, turning those estimates into an actual price or promotion plan, that creates the value.
This matches what we see with a lot of companies. Many have good data and good analysis, but the output stops at a report or a dashboard. Nobody has built the step that turns the analysis into a decision. PepsiCo’s paper is a useful reminder, from a company with serious resources, that this last step is where the value actually sits.
The hardest part of pricing, and how they work around it
The part of the paper we found most useful is where they describe their biggest challenge: prices in consumer goods don’t change very often. Maybe once or twice a year per product. That means there isn’t much real history to learn from, but you still need to estimate how demand responds to price across hundreds of products, sizes, and retailers.
Their solution is reasonable. They assume that products with things in common, like the same brand or the same pack size, respond to price changes in similar ways. That lets them fill in the gaps where there isn’t much data for a specific product.
It’s a sensible workaround, but it comes with a trade-off the paper itself points out. Those shared patterns are assumptions. When the data is thin, there’s no easy way to fully check whether the assumption holds for every product. So a price recommendation that looks very precise is, in part, built on an educated guess rather than direct evidence.
This isn’t a criticism of PepsiCo’s approach. It’s a problem every company runs into when it tries to price a large product range, and there’s no way around it without more data or a different way of getting it.
Where we come in
This is the kind of problem we work on regularly. A few things we typically do differently:
We build in the uncertainty rather than hiding it. Instead of pointing to one “best” price, we look for the price moves that still make sense across a realistic range of outcomes, so the recommendation holds up even if the estimate is a bit off.
We go looking for real evidence instead of relying only on old data. Where there isn’t enough price history to work with, we can design and run conjoint studies that are fast and don’t cost a fortune, contrary to what most people assume. Where it’s possible to execute, we can also design small, structured tests, by region or by product, to generate the evidence a model actually needs.
We check results against what actually happened, not just how well the model fits on paper. A model that looks correct isn’t the same as a decision that worked.
It’s also worth saying: PepsiCo built this with its own dedicated data science team over several years. Most companies don’t have that, and don’t need it to get similar value. The tools and algorithms involved are increasingly available to everyone. What’s harder to come by is the judgment: knowing which models to use and how to use them, which assumptions are safe to make, which business rules actually matter, and how to build something your commercial team will trust enough to use day to day. That’s usually where an expert outside partner helps most.
In short
PepsiCo’s paper is an outstanding look at what’s possible when pricing is treated as a serious, data-driven discipline. Most companies don’t need to build what PepsiCo built to get real value from better pricing decisions. They need the right approach for their own data and their own team, with the right capabilities to make it work.
If you’re making pricing decisions with limited data, and/or you need AI-powered tools, whatever your industry, this is exactly the kind of problem we’d like to talk through with you. Get in touch.
Reference: Llenas, A., et al. (2026). “PepsiCo Deploys AI-Driven Pricing and Promotion Optimization at Scale.” INFORMS Journal on Applied Analytics.
