Every large organization we meet has the same inventory: a dozen AI pilots, two or three that "went well," and none running in production against real systems with real approvals. The pilots were not failures. They were never designed to leave.
A pilot is scoped to prove that a model can do a task. Production is scoped to prove that an organization can run that task with the model in it, every day, under audit. These are different projects with different owners, and the second one is rarely started.
The three decisions a pilot avoids
Who approves. In a pilot, a data scientist looks at the output. In production, a named person in the business signs off on a journal entry, a claim, a customer record. Until that person and that step exist, there is no production.
What system it writes to. A pilot writes to a spreadsheet or a demo database. Production writes to the ERP, the core banking system, the policy administration platform. The integration is most of the work, and the pilot budget never includes it.
What happens when it is wrong. Pilots measure accuracy. Production needs a path for the exceptions: escalation, rollback, and an audit record that a regulator can read. The design of the failure mode is the design of the system.
How to leave the pilot behind
Pick one workflow and scope it to production from the first day. Name the approver. Name the system of record. Write the exception path before writing the prompt.
Build it inside your environment, with your team, so the connectors and the governance you build are yours and they stay.
Then let the second workflow inherit everything the first one learned. That is how one production agent becomes twenty without becoming twenty governance problems.
The demo was never the hard part. Getting it across the line is the job.