Back to Blog
Blog

What should an AI change automatically in a procedure, and what needs human approval?

What should an AI change automatically in a procedure, and what needs human approval?

The simple rule is that AI can handle structure on its own, and anything that changes what an operator physically does needs a person to approve it…

The simple rule is that AI can handle structure on its own, and anything that changes what an operator physically does needs a person to approve it. Splitting a paragraph into steps, flagging a missing warning, catching inconsistent tag numbers or pointing out duplicate content are all things that save a reviewer hours without changing the work itself, since the operator would perform the same actions in the same order either way.

Structural work is a good fit for software because it is tedious, repetitive and easy for a person to check afterward. A reviewer can look at a split step next to the original paragraph and confirm in seconds that nothing was lost, and a flag about an inconsistent tag number is just a pointer to something a person then decides on. Even here the results should be visible to reviewers, so the team can see what was restructured and spot any pattern of mistakes early.

Changing a step’s order, a setpoint, an isolation or a safety critical instruction is different, since the person who approves the procedure is accountable for it and the software is not. These are the changes that can hurt someone if they are wrong, and they often depend on knowledge that is not written anywhere in the document, such as how a particular valve behaves, what happened during the last turnaround or why a step was added after a near miss years ago.

The grey area sits between those two, with things like rewording a step for clarity or merging two similar warnings, which look harmless but can shift meaning in ways that matter. A sensible default is to treat any wording change as needing approval until the team has seen enough of the AI’s suggestions to agree on which kinds of edits are safe to accept in bulk, and to write that agreement down so it does not depend on who happens to be reviewing that week.

We think about it the same way at Interface, where the AI does the structural work and flags what needs attention, and the approval stays with your people. The software makes the review faster and more focused, but the decision about what an operator does in the field belongs to someone who is accountable for it.