AI can make processes cheaper and faster, and most of the risk sits in one question: can this change alter what the process produces? Changes that provably cannot are applied automatically. Anything that touches what the AI reads gets tested against old runs and approved by a human. Every version is saved forever.
Pointing an AI at its own prompts sounds like free improvement. In practice the rewrite happens blind, nothing proves the output stayed the same, the old wording is gone, and somebody babysits the loop forever. Even DSPy, the open source tooling for this, requires you to supply the metric yourself.
Improvement gains rarely survive on their own. Kaizen research finds 70 percent of improvements decline within 12 months, and undocumented fixes start reverting in about six. The cause is neither effort nor talent. Gains fade when nobody maintains the current best way of doing the work, and the standard quietly dies.
Panopto research found knowledge workers waste 5.3 hours every week waiting for information or recreating work that already exists. Static SOPs feed that waste. A procedure that lives in a document starts aging the day someone writes it, because a document cannot watch the work change underneath it.
