Everyone talks about foundation models and use cases. Nobody talks about the integration layer in between — the part where most AI projects actually break.
The foundation model is the interesting problem. Which model? What architecture? Fine-tuned or prompted? These are the questions that attract engineers, generate conference talks, and feature in case studies. They are also, increasingly, the easy part. The hard models are good. The tooling has matured. You can get a capable AI system off the ground in weeks.
The use case is also well-understood. Organisations know what they want to automate or augment. They have a business case, a sponsor, and a target outcome. That part is fine.
The middle layer is where projects die quietly. Data pipelines that don't generalise. API contracts that break when the model version changes. Latency assumptions that made sense in the demo and don't survive contact with production traffic. Escalation paths that were designed without thinking about what the AI system does when it's wrong.