The checklist before automating anything

Before automating anything, with or without AI, I run the same five question checklist. No book taught it to me: the processes I automated and should not have did.

The five questions

1. Is the process stable? If the flow changes every month, you will be automating a version that no longer exists by the time you finish. First stabilize the process, then automate it. Automating chaos only produces faster chaos.

2. How much does it hurt manually? The math is simple: frequency times minutes times people. A two minute task done by twenty people every day is a serious problem. A one hour task that happens twice a year is not, however annoying it feels.

3. What happens when it fails? Every automation fails eventually. If the error is reversible, the risk can be managed. If it is not, everything else gets multiplied by caution: in my formula, the risk of an irreversible operation weighs five times more than a reversible one.

4. Who owns it? An automation without an owner is a future ticket with your name on it. Someone has to know it exists, know how to switch it off and find out when the underlying process changes.

5. Does it really need AI? A clear, repeatable rule is solved with a script or iLogic, and that solution is cheaper, more predictable and easier to audit. AI earns its place when there are exceptions to interpret, natural language in the loop or decisions that need context.

The formula that orders the queue

Questions 2 and 3 produce my prioritization formula: frequency times manual minutes times people, divided by effort in days times risk (risk 1 if the operation is reversible, 5 if it is not).

Three real examples from my own Vault backlog: mass property assignment scores 360, the pre-release audit scores 126 and a revision change barely 16. All three can be automated. The formula does not say what is possible: it says what deserves to go first, and what can wait without guilt.

The underlying pain is real: in the field I have seen engineers spend between 3 and 8 hours a week on repetitive PDM operations. That is a field observation, not an instrumented measurement, but it is enough to justify the math above.

Script or agent

My short rule: a script automates the rule, an agent handles the exception. If the whole process fits in a flowchart with no ambiguous branches, a script or an iLogic rule solves it better than any LLM. When the process demands interpreting a request, deciding with context or conversing with the user, the agent comes in. And it comes in with gates: dry-run, validation, verification and hard limits, because question 3 does not go away when you use AI. Quite the opposite.

The checklist also says no

Half the value of this checklist lives in the projects it prevents. Every automation that was not built because the process was unstable, the pain was small or nobody was going to maintain it, is time and trust that did not get burned. Automation is easy to sell and hard to undo: the discipline of the list is what separates a portfolio of solutions from a graveyard of scripts.