Enterprise AI doesn't have a model problem
May 2, 2026 · 5 min read
Every enterprise AI conversation eventually turns into a debate about which model to use. Which vendor, which context window, which benchmark score. It's a satisfying conversation to have because it feels technical and it has a clean answer. It's also almost never the reason the product isn't working.
Running a large-scale conversational AI and search portfolio across provider, patient, and employee workflows taught me a less comfortable lesson: the model is rarely the constraint. The constraint is the data no one cleaned up, the ownership nobody claimed when the answer is wrong, and the fact that "AI product" quietly became the label for anything nobody wanted to own the roadmap for. Swap the model and the same conversation returns three weeks later with a different vendor's name attached to it.
The teams that ship durable AI products treat the model as the least interesting decision in the stack. The interesting decisions are about failure modes: what happens when the system is confidently wrong in front of a patient, who gets paged, how fast the answer can be corrected, and whether the product was designed assuming the model would sometimes fail — because it will. Most AI incidents I've seen trace back to a product decision made months earlier that assumed the model would just be right, not a defect in the model itself.
If your AI roadmap is organized around model upgrades, you're optimizing the part of the system that's already improving on its own, for free, every few months. The actual product work — the data pipeline, the escalation path, the trust the interface has to earn — doesn't get better just because a new model shipped. That's where the real leverage sits, and it's usually the least glamorous line in the roadmap.