The Problem
Almost every AI adoption effort I've seen measures the wrong thing: who clicked "complete" on a training module. That number can hit 100% while actual usage sits at zero, because completing a module and changing how you work are two different events.
It's the same trap I'd seen years earlier in a completely different context: a marketing program optimized toward a mid-funnel metric that looked healthy while the business underneath it wasn't. Trust the behavior data, not the completion data. That instinct is the throughline of everything below.
What I Built
At Robinhood, the lifecycle marketing team was using AI sporadically: a few people, a few prompts, no system underneath it. I built the system instead of another slide deck:
The agent
An autonomous Claude Code agent that runs lifecycle marketing campaigns end-to-end, plus a library of 40 reusable skills and sub-agents spanning planning, targeting, content, QA, and performance tracking.
The enablement
A 2-day AI bootcamp I pitched to leadership and then ran myself, workflow-level training rather than a slide deck, built to convert skeptics into daily users.
The measurement
An adoption framework tracking real workflow behavior against defined baselines, the same discipline that catches a vanity metric before it costs you the business.
The roadmap
A transformation roadmap sequencing rollout across the org by team readiness and tool maturity, so adoption compounded instead of stalling on the teams least ready for it.
How I Think About It
Enablement and building are the same job to me, pointed in two directions. You can't design training that actually changes behavior unless you've built the thing closely enough to know where it breaks. And you can't build something worth adopting unless you understand the team you're asking to change.