Hershey's $250M AI Bet: Margin Protection Through Physics

TLDR: Hershey's successful AI implementation in manufacturing started with understanding the physics of candy making rather than technology, resulting in significant improvements in product weight variance and data analysis time.

Summary:

Candy making is an art, and the factory operators at Hershey were initially skeptical of new technology. They had rejected an IoT initiative four times before finally allowing sensors on the production floor. These were people who could feel when the Twizzler dough was off, so an algorithm claiming it could do better was met with understandable skepticism.

However, the system proved its worth by predicting weight drift before it happened, not 15 minutes after when an operator finally checked. That difference alone was worth $500,000 per year on a single production line. This demonstrates how a brownfield manufacturer built AI into operations without blowing up in the process.

The implementation achieved impressive results: a 50% reduction in product weight variance on instrumented lines, 75% reduction in data analysis time after consolidating fragmented data, and reduced time for new product innovation cycles from five months to five weeks. The key insight was that most AI strategies start with technology, but Hershey started with the physics.

Every gram of chocolate exceeding the label weight is margin walking out the door. Manufacturers call this "giveaway," and it's invisible until you measure it. The fastest AI wins sit where repetition meets variance. Find the process you run thousands of times daily where small inconsistencies compound into real money.

For architects and teams, this case study shows the importance of starting with domain expertise and physical processes rather than jumping to technology solutions. Understanding the fundamental physics or core business processes before implementing AI can lead to more effective and valuable implementations.

Key takeaways:

Start AI implementations with domain expertise and physics rather than technology
Focus on processes that run thousands of times daily where small inconsistencies compound into real money
Operator buy-in is crucial - demonstrate clear value before expecting adoption
Measure "giveaway" - small inefficiencies that compound into significant losses
Consolidate fragmented data to reduce analysis time and improve insights

Tradeoffs:

Traditional expertise and intuition may be challenged by algorithmic approaches, requiring cultural change management
Initial skepticism from experienced operators can slow implementation timelines

Link: Hershey's $250M AI bet: margin protection through physics