Promo Advisor
Getting revenue teams to trust an AI that told them their promotions were losing money. An end-to-end platform, live across multiple markets, that cut promo-planning time 30–40% and moved historically money-losing promotions onto data-driven planning.
- Role
- Product lead
- Context
- Heineken · 2021 – 2026
- Domain
- Revenue Management
- Approach
- 3-module ML platform
What Promo Advisor does.
Trade promotions are one of the largest controllable lines a beverage business spends against. Planning ran on spreadsheets, habit, and last-year's-calendar instinct, repeated across every market, every year.
Promo Advisor is the end-to-end platform Revenue Management teams use to design, test and optimize promotions. It combines historical data with prediction through three AI modules: Diagnostics (what worked), the Scenario Simulator (what will work), and the Calendar Optimizer (the best calendar within real constraints).
The result: promotions stop being a leak and become a continuous, optimized driver of growth, with what one market learns shared across the rest, instead of trapped in local spreadsheets.
Nobody likes hearing their promos lose money.
Heineken had the data science to model which mechanics actually paid back ("15% off" versus "buy 3, pay for 2") across customers and markets. Building an accurate forecast was not the hard part.
The hard part was human. The model regularly disagreed with experienced revenue managers, and told them, in numbers, that promotions they had run for years were destroying gross profit. A more accurate forecast that no one trusts changes nothing. The blocker was never the math; it was getting commercial teams to act against their own gut.
Accountable for whether teams actually planned with it.
I was the founding product hire in the analytics group and the only product owner on this system, from discovery to delivery: the problem framing, the module scope, the roadmap, and the adoption strategy into how teams plan.
I was accountable for whether Revenue Management teams actually planned with the tool, not just whether it shipped. I was the standing bridge between the data scientists building the models and the commercial org that had to change how it worked.
Three product bets that earned trust.
Each one traded a better-looking model for something a commercial team would actually act on.
Bet on adoption, not accuracy.
The obvious path, the one the data-science instinct pulls toward, was to keep improving model precision. I argued against it. Teams were already ignoring forecasts that were good enough; more decimal places wouldn't move usage. I redirected scope from squeezing the model to making its output legible and trustable.
Show the model's uncertainty, not one clean number.
The intuitive move is a single confident figure so the tool looks decisive. I chose to show the prediction with its confidence range and a plain-language "why", the drivers: cannibalization, stockpiling, competitor switching. Legibility builds more trust than false precision.
Prove incrementality, not just report sales.
A promo always looks good if you count every unit sold during it. I made the harder call: the model had to separate true uplift from the baseline that would have sold anyway, and credit the promo only for the incremental part. That made every number smaller and every claim defensible.
Show the range, not a false single number.
Credit only what's truly incremental.
Three modules: hindsight, foresight, prescription.
A production platform of three AI modules on the shared React/MUI design system, deployed across multiple markets. To de-risk the rollout, I shipped the Simulator before the automatic Optimizer, so teams learned to argue with the model on their own scenarios before trusting it to build the whole calendar.
Diagnostics
Evaluates past promotions and isolates their true effects, separating real uplift from cannibalization, stockpiling and competitor switching.
Scenario Simulator
Tests a promo before it runs: models trained on thousands of past promotions estimate likely sales and ROI, with a confidence range, for any mechanic.
Calendar Optimizer
Builds the revenue-maximizing calendar within real constraints, testing combinations to keep negative-ROI promotions out of the plan.
From money-losing promos to a data-driven baseline.
The real proof wasn't the accuracy: it was the behaviour change. Revenue teams moved from defending last year's calendar to planning against the model and asking it why. Delivered outcomes are stated as delivered; benchmarks are labelled as such.
The hard part of an AI product is almost never the model. It's the human layer: framing the objective, and making a probabilistic output trustable enough that someone changes what they do.
The execution engine of a three-part system.
Promo Advisor executes inside the budget Allocation AI sets and the timing Smart Flighting schedules.
Sets the annual budget across markets, brands and channels.
Read the case study ›Fills the calendar with the promo mechanics that actually pay back.
Times each budget week by week, so pressure lands where it still converts.
Read the case study ›