Building Herbalife’s AI Center of Excellence

the problem

By 2018, AI work at Herbalife was happening everywhere and nowhere. Data science had a corner of it, operations had another, marketing was experimenting on its own — fragmented efforts with no shared infrastructure and no in-house ML expertise to speak of. The CTO wanted a Center of Excellence to centralize it, but there was no dedicated budget, no team, and no internal model to copy. It had to be built, not requested.

the approach

I started with discovery, not a pitch. I went across every business unit — not to sell AI, but to find where the real pain was and what was actually technically tractable. That gave me a backlog prioritized on effort versus impact instead of whatever was loudest in the room.

Infrastructure came next. I made the case for Microsoft Azure specifically because it leveraged Herbalife’s existing enterprise agreements and let us skip a security approval process that would otherwise have taken six months we didn’t have. Fighting for a new cloud vendor at that stage would have stalled the whole program before it started.

Team was the harder constraint. I couldn’t hire externally at the scale the work needed, so I contracted an external data science agency for technical depth and paired them with internal volunteers who brought domain expertise. Knowledge transfer was written into the engagement from day one — the agency’s mandate wasn’t just to deliver, it was to leave capability behind so the program didn’t stay permanently dependent on outside help.

what shipped

The first win was a sales forecasting model. I worked as the technical-business bridge on it — translating business requirements for the vendor, defining acceptance criteria, and coordinating delivery. It improved forecast accuracy by 10%, and that early, visible win funded the next wave of investment.

From there the CoE became the delivery engine for AI at Herbalife: object detection deployed in distribution centers, and a cloud DAM platform modernization that delivered a 56% operational efficiency gain. Over five years the program shipped 10+ AI use cases across six global regions and grew into a 50-person AI enablement program.

what I learned

The program didn’t survive on technical ambition — it survived on sequencing. Picking the infrastructure that avoided a six-month approval fight, building the team out of contractors plus internal volunteers instead of waiting on headcount, and shipping one visible, defensible win before asking for more investment: none of that is glamorous, but it’s what let a program with zero starting resources become the standard infrastructure for every AI deployment that came after it.