Practical insights from building AI products in production — what works, what doesn't, and why.
The Fable 5 shutdown ended on a diagnosis: if access can be revoked, it's a dependency. This piece is the cure — the case for owning your intelligence layer, including a worked example, the honest tradeoffs, and where the hybrid architecture actually lands.
On June 12, 2026, the U.S. government took two of the most capable AI models offline overnight. Most coverage frames this as a political story. It isn't. It's the clearest signal yet that the strategic center of gravity in AI is shifting from capability to control — and every organization building on hosted AI should be asking who controls the off-switch.
Agentic AI is genuinely powerful — and precisely what makes it valuable is what makes it dangerous without governance. Unauthorized actions, data leakage, hallucinations with consequences, compliance violations: the exposures are concrete and the bill arrives all at once.
The model router is about to become standard, non-negotiable infrastructure. Different tasks should go to different models, matched to the difficulty of the work — and the benefits compound fast.
Every era of computing settles into a reference architecture. Enterprise AI is now visibly converging on one. Here's the full stack, layer by layer — and why the model ends up in the middle, not the top.
The reframe that makes enterprise AI governance click: agents are functionally employees. Every governance principle we refined over generations of managing people — permissions, managers, KPIs, monitoring, audit logs — applies directly.
A portfolio approach to model selection is now a core engineering competency. Small language models handle the high-volume, well-defined work. Frontier models handle the genuinely hard problems. Knowing where that boundary falls shapes your entire cost structure.
RAG is a solved engineering problem. The thing that kills production deployments is knowledge governance — ownership, freshness, permissions, auditing, and compliance. Almost nobody is talking about it.
Intelligence is becoming a commodity. What separates serious AI organizations from those stuck in pilot purgatory is the control layer — governance, security, observability, compliance, and monitoring.
There's a reflex in AI procurement to reach for the biggest model. Safe and smart are not the same thing — and confusing them is expensive.
Every impressive agent demo was optimized to show what the agent can do. Enterprise deployment cares about the opposite: defining exactly what the agent is allowed to do.
Fleets of small, specialized models — each doing one thing exceptionally well, orchestrated together — are becoming an infrastructure layer in their own right.
We keep asking whether the models are smart enough. It's the wrong question — and it's quietly steering enterprise AI strategy in the wrong direction.
Consumers forgive AI. Enterprises can't. That single asymmetry explains why so many impressive demos never survive contact with a regulated business.
Most RAG demos work on toy datasets. Production is different — here's how to handle chunking, embedding selection, and retrieval failures at scale.
The wrong choice wastes months and GPU budget. A decision framework for when to prompt, when to fine-tune, and when neither is the right answer.
It's rarely a technical problem. The gap between a working demo and a system that earns ROI is mostly organizational — here's how to close it.