Case Study · ML Product Strategy

Ending a 30-Year Industry Problem with ML

Director of Product Management · Engage3

We built an ML-powered pricing solution that replaced 30 years of manual competitive price collection: 98 to 99% predictive accuracy, an ~83% reduction in manual collection workload, and $150 to 200M in annual budget freed for our largest customer alone. That freed budget became the engine of a land-and-expand motion that made Engage3 the partner of choice for 8 of the top 10 global retailers.

My role

I was the product lead for the Data Integrity product area that this solution grew out of. I owned the what and the why: problem framing, the product design of the sampling-and-validation model, pricing and packaging, adoption and education, and roadmap prioritization for these efforts across the entire product suite. Our data science team and the application engineering teams owned the how, and we iterated on that boundary constantly.

The problem nobody had solved

For three decades, grocers and major retailers answered one question, "what are my competitors charging?", by sending human auditors into competitor stores with pen and paper, across the country, continuously. Our largest strategic customer spent $150 to 200M annually on this single function.

Our own first attempt was a failure worth naming: a "Data Integrity Tool" that asked customers to manually QA pricing data. It created more work instead of eliminating the underlying need. That failure sharpened the real insight.

The reframe

The opportunity wasn't making manual collection incrementally better. It was eliminating the need for 80%+ of it. And the second-order effect mattered more than the first: once we solved price collection, the massive budget it consumed became available for price optimization, our premium offering. Solve the cost center, then sell into the freed budget.

What we built

Before: audit everything

  • Human auditors in competitor stores, pen and paper, nationwide, continuously
  • $150–200M annual spend for our largest customer
  • Slow, error-prone, and impossible to scale
  • Data trusted only because it was observed firsthand

After: predict, then validate

  • ML models trained on in-store audit data plus web-scraped online pricing
  • 98–99% predictive accuracy on competitor shelf prices
  • Lightweight strategic sampling: 10 to 15 products per location
  • Suspect records automatically resampled, no customer QA work

Predictive pricing models. Combined in-store audit data with web-scraped online pricing to train models predicting competitor shelf prices at 98 to 99% accuracy, replacing exhaustive auditing with prediction.

Strategic sampling with continuous validation. Customers shifted from auditing everything to lightweight sampling that validated and continuously improved the models. Suspect records were automatically resampled without customer manual review, the exact mistake our failed first product had made.

Key decisions and tradeoffs

Prediction over verification. Retailers had spent 30 years trusting only observed prices. Convincing them to act on predicted prices was the adoption risk of the whole product, which is why accuracy transparency and the sampling-and-validation loop weren't features. They were the trust mechanism.

Pricing against a $150 to 200M cost center, not against software comps. The value conversation changed from "what does the tool cost" to "what does the freed budget buy," which set up the expansion motion into optimization.

Results

What it taught me

Our failed Data Integrity Tool and the successful ML solution addressed the same problem. The difference was who did the work: the failure asked customers to work for the data; the success made the data work for customers. Most "AI opportunities" I've evaluated since come down to that same test.

← All case studies