After the pilot.

A pilot is, by design, a test on borrowed time. It has a small audience, a forgiving environment, a generous budget per query, and a champion who is paying attention. Most pilots succeed in the narrow sense. They show the thing can be done.

What pilots almost never test is the part that breaks first in production.

In production, the audience is ten times larger and a quarter of them are using the system in ways the pilot did not anticipate. The forgiving environment is replaced by an integration with three legacy systems, one of which is owned by a team that is not part of the AI program. The budget per query is now measured against a real margin. The champion is busy with the next initiative.

The work of moving a pilot into production is not a deployment. It is a renegotiation. The thing that worked in the pilot needs to be re-scoped against constraints that did not exist when the pilot started. The owner of the legacy system needs to be in the room. The unit economics need to be redone with the assumption that volume will be ten times higher and tolerance for error will be ten times lower.

The mistake we see most often is treating the pilot-to-production step as a calendar problem. “We will ship in Q3.” It is not a calendar problem. It is a scoping problem. The pilot proved the model works. Production requires a fresh decision about what the model is for, who owns the integrations it touches, and how the unit economics survive scale.

When we triage an AI portfolio and find a pilot that is genuinely worth shipping, the first deliverable is rarely a deployment plan. It is a one-page rescope. Without that, the deployment plan is a fiction.

Most of the value of the portfolio triage is identifying which pilots are ready for that rescope and which are not. The pilots that are not, almost without exception, should be killed rather than shipped. A pilot that cannot survive a rescope is a research result, not a product.