1. It happens often
A task your team does hundreds of times a month gives you room to save real time, and plenty of examples to test against. A task that comes up twice a year rarely justifies the work.
2. The input is too messy for simple rules
AI is useful when the work starts with emails, PDFs or free text, and a person has to read and interpret them. If the data is already structured and the rules are clear, ordinary automation is cheaper and more reliable.
3. Mistakes are easy to catch and fix
AI will sometimes be wrong, so start where an error can be spotted and corrected before it does harm. A drafted reply that someone reviews is a good place to begin. An unreviewed decision that affects someone's health, money or job is not.
4. You know what good looks like
If your team can look at a result and agree whether it's right, you can measure the system. Past examples of finished work are especially valuable, because they become your test set.
5. The data is within reach
The information has to be somewhere a system can read it, and you need to be comfortable with where it goes. Sort out access and privacy early, because either one can stop a project late.
When it isn't ready
If everyone does the process differently, AI will only speed up the confusion, so agree on the process first. If it's just data moving between two systems, an integration is probably all you need. Choosing not to use AI is often the right call.