The pilot is over. What AI in production actually looks like.

The pilot era ended around Q3 of last year.

By which we mean: the budget for novelty is gone. The companies that ran proof-of-concepts in 2024 and 2025 are now either operating at scale or they have quietly buried the project. There is very little middle ground left. The procurement question has shifted from can this work to can you operate this without breaking us.

What's shipping

• Vertical agents with narrow scope and explicit handoffs to humans. The successful deployments we see are deeply boring — claims triage, supplier reconciliation, customer onboarding.
• Retrieval-grounded assistants in regulated industries. Slow to build, easy to defend, and increasingly the only configuration auditors will accept.
• Internal tooling that quietly compounds. Eighteen months in, the cost reduction shows up in headcount plans, not in headlines.

What's getting deprecated

• Open-ended chat as a primary surface. Users want a button, not a prompt.
• Prompt libraries as a strategic asset. They aged out faster than anyone predicted.
• RAG-only architectures without a structured data layer underneath. They work in demos and fail in production.

The architectural decisions that mattered.

Twelve months on, the projects that are still live share two boring properties: they have a structured data layer that predates the AI, and they have an explicit human-in-the-loop checkpoint at the cost-of-error boundary. Everything else is interchangeable.

Production AI is what's left after you stop calling it AI.

If your pilot is still a pilot in Q3, it is no longer a pilot. It is a project that didn't ship. Treat it accordingly.