Anyone can pick a model
For twenty years I've watched the same pattern: the hard part of a system is rarely the part in the demo. AI has made that pattern louder.
Every enterprise can now reach strong models, and soon the gap between them will matter less than people think. So the model is not the moat. The architecture around it is.
Security isn't a firewall around the system. Compliance isn't paperwork after deployment. Cost control isn't just a cheaper model. Human control isn't an override added later. These are architectural decisions, and where you place them decides everything else.
Get the placement right and performance, cost, accuracy, scale and control improve together. Get it wrong and you have a demo that never becomes a durable product.
In practice, that means three rules I build by:
- One gateway. Nothing calls a model provider directly. Policy, personal-data redaction, routing, caching and audit live in one place.
- Capability, not brand. Agents ask for what they need (reasoning depth, latency, cost ceiling, jurisdiction), and the platform picks the model. Swapping models becomes a configuration change.
- Evals before agents. Capability you can't measure isn't capability. It's risk.
Anyone can pick a model. Architecture decides when, where and how it's allowed to think. That is the work, and that is the moat.