Do I need AI for this?

I build AI agents for engineering for a living. VaultPilot inside Vault, a design assistant inside Inventor. I am supposed to be the person with the perfect incentive to tell you yes, AI is right for everything. And yet, half the time someone asks me whether AI fits a specific process, my honest answer is no.

Not because the technology does not work. Because there is a simpler tool that solves the same problem with less risk and less cost. This is the decision tree I use with clients, and the one I use on myself before writing a single line of an agent.

First question: is the rule clear?

If the task is repetitive and can be described by an unambiguous rule, the answer is almost never AI. It is a script, an iLogic rule or a well built Excel. Deterministic code is cheaper to build, more predictable to operate, and it does not hallucinate. Creating styles in bulk from a spreadsheet, renaming to a pattern, exporting a complete library: all of that is clear rule territory, and clear rules get automated with code.

Second question: are there exceptions and judgment?

An agent starts to make sense when the task involves exceptions, decisions with context, and human language in the middle. The sentence I took to my Autodesk University session sums up the split: a script automates the rule, an agent handles the exception. If your process is 95% rule and 5% exception, automate the rule with code and leave the exception to a person. If the exception is the process, that is where agent territory begins.

Third question: is it worth it?

This is where I use a simple prioritization formula: frequency times manual minutes times people affected, divided by build effort in days and risk. Risk is 1 if the operation is reversible and 5 if it is not. It is not exact science, it is a way of comparing apples to apples.

In my own backlog, property assignment scored 360, the pre-release audit 126, and the revision change 16. Same technology, three completely different priorities. The revision change is perfectly automatable, and it is still the last thing I would build: low frequency, high risk, few people affected. Without the formula, I would have started with the flashiest thing instead of the most valuable one.

The right question

The question is never "can I put AI on this?". You almost always can. The question is what problem you are solving and what is the simplest tool that solves it reliably. Sometimes that tool is an agent with nine safety gates. Sometimes it is a fifty line iLogic rule. And sometimes, more often than my industry likes to admit, it is a spreadsheet.

Whoever sells you the answer before asking you the three questions is not selling you a solution. They are selling you their tool.