A machine will not ask whether the data is complete, or notice that someone quietly changed what a category means. It acts, fast and with confidence.
Train it on a good-news version of your business and it learns the good news. Its answers are fast, consistent and wrong in the same direction every time.
Speed is not the risk. Speed built on incomplete information is the risk.
The data it learns from
Where bad news travels slowly, your systems hold what was reported, not what happened. A system trained on that treats the delay as the truth.
Your disagreements, at scale
If 3 executives understand the plan in 3 different ways, automated decisions will follow all 3. People settle that in conversation. Systems do not.
Working out what went wrong
When an automated decision fails, the board asks what the business believed at the time. Accident investigation is built to answer that question. Most governance rules are not.
The question to settle first.
Before you hand a decision to a machine, check that the information it will act on describes the business you actually run.
This is not a case for waiting until everything is fixed. It is a case for checking one thing first: whether the data the machine will learn from is honest.
That check belongs before the approval, not after something goes wrong.