How I think about enterprise AI.
A distillation of opinions from a decade of building. Not a manifesto. Just the working notes I'd hand to a peer.
Run GenAI as a program, not a project.
A one-off prompt is a feature. Production GenAI is infrastructure. The earlier you treat it as infrastructure, the less rework later.
Agents aren't a feature you ship once.
They're a class of system. You don't "deploy an agent" the way you deploy a microservice. You operate it. Plan for evals, regressions, memory, escalation, and replay before plan for launch.
Prompts are technical debt.
Every prompt you write is a contract with a model version. Models change. Have a migration framework before you have a thousand prompts.
The hardest part of enterprise AI is not the model. It's the people.
Capability is cheap now. Adoption, fluency, and trust are not. Most of the value comes from how the organization uses the tools, not which tools it has.
Cost governance is a first-class concern, not an afterthought.
A single team can blow a quarter of the budget in a week if you don't have caching, quotas, and tier-routing on day one. Build for it before you need it.
Multi-vendor model-agnostic abstraction beats vendor lock-in.
Frontier capability swaps between labs every 6 to 12 months. Your platform should make that swap trivial. Your applications shouldn't notice.
Eval before deploy. Trace before scale.
If you can't measure how the system is behaving in production, you can't ship the next version safely. Observability is the line between a demo and a system.
Want to discuss any of these? I'm open to advisory conversations and speaking engagements.