Generative AI did not arrive at Centene as a polite feature request. It arrived as a platform shift — the kind that rearranges how teams learn, review, build, and govern. The mandate was clear enough: move faster. The constraint was equally clear: this is healthcare. Members are not beta testers.
What made this hard
The bottleneck was rarely the model. It was capability formation — do engineers know how to build governed agents? Do reviewers know what to ask? Do leaders know how to measure value without counting tokens like vanity metrics?
We also could not pretend AI floated above the stack. Agents without governed data, reliable events, and observable pipelines are demos with audit risk. Adoption and infrastructure had to be one conversation.
What we did
I led platform strategy across Databricks and Snowflake — the governed data layer that makes retrieval defensible. We built tiered AI governance: lightweight paths for experiments, rigorous paths for member-facing systems. We trained 200+ engineers, stood up an AI guild, and paired on production deployments until “AI project” stopped meaning “special exception.”
Observability was non-negotiable from the start — tracing, cost attribution, evaluation benchmarks, incident playbooks written for probabilistic systems, not deterministic fantasies.
What changed
Teams stopped waiting for permission and started building inside credible guardrails. Deployment patterns reused across groups. Data accuracy and automation improved in measurable ways — but the human signal I trusted most was Monday-morning reuse: engineers shipping governed agents without a consultant in the room.
That is adoption that sticks. Not a campaign. A capability.