Production AI in 2026 is mostly plumbing. Sorry.

Executives want a model recommendation. Senior engineers want the data path, access model, rollback plan, eval harness, latency budget, incident process, and cost envelope. The engineers are right.

The production pattern we still trust is boring: start with the smallest valuable workflow, use retrieval only where it beats curated context, keep humans in the loop until quality is measurable, log enough to debug failures, and keep sensitive data out of places you cannot defend later.

That pattern is Bootcamp → Deploy → Run. A working path on real data, then a 6–10 week system, then a monthly operate and eval loop. Plumbing is the product.