Generative AI did not arrive as a feature request. It arrived as a platform shift — the kind that rearranges how teams learn, decide, build, and govern. I have led that shift in regulated healthcare, and I study it the same way I study any complex system: by watching what people actually do, not what the roadmap claims they will do.
Adoption is a learning problem
The first mistake organizations make is treating GenAI adoption as a procurement problem. Buy models. Ship chat. Declare transformation.
In my experience, the bottleneck is almost always 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 trained 200+ engineers. We ran guilds. We paired on production deployments. The metric I trusted was not workshop attendance. It was Monday-morning reuse — patterns showing up again without a consultant in the room.
Governance that enables speed
The second mistake is choosing between move fast and stay safe. That is a false trade when governance is designed well.
Tiered risk classification, instrumented controls, shared prompt libraries, and evaluation benchmarks gave teams a credible path — lightweight for experiments, rigorous for member-facing systems. Adoption accelerated when ambiguity dropped, not when mandates increased.
Infrastructure is adoption
The third mistake is assuming AI sits above the stack. Agents that cannot reach governed data, reliable events, and observable pipelines are demos with audit risk.
Enterprise GenAI adoption research, for me, includes the platform layer — Kafka, lakehouse, Go orchestration — because adoption without infrastructure is performance art.
Expert Vision connection
The best adoption programs do not replace experts. They make experts more effective — faster retrieval, better synthesis, less toil on repetitive scaffolding. The failure mode is dependency: juniors who skip judgment formation because the model answers confidently.
I study both outcomes. Amplification scales organizations. Dependency hollows them out.
What I am still asking
How do we measure whether AI made your best people better — not just your average output noisier?
That question should drive adoption strategy more than model leaderboard rank. I think the organizations that figure this out will treat GenAI the way mature engineering cultures treat reliability: boring on the surface, disciplined underneath.
That is the adoption work worth doing.