Do not start with “we need AI”.
Start with a recurring job that involves the same inbox, spreadsheet or system and broadly the same decisions each time.
That gives us a defined beginning, a defined end and a way to tell whether the process worked.
A suitable first job has clear boundaries
Look for work where:
- the trigger is obvious
- most inputs arrive in recognisable forms
- the routine steps can be explained
- exceptions can be handed to a person
- mistakes can be spotted and recovered
- the result has somewhere specific to go
For example: a request arrives by email, details need checking against two systems, a record needs preparing and a person approves the final change.
That is a process that can be examined and tested. “Improve operations with AI” does not define a workable scope.
Write down the current process
You do not need a process diagram. A rough list is enough:
- This arrives.
- Somebody checks these things.
- If this is missing, they chase it.
- If everything is fine, they update that.
- Anything outside the routine goes to this person.
The gaps will appear quickly. Good. Those gaps exist in the real process too; the agent did not invent them.
Sometimes the right answer is to fix the process first. Sometimes a small script is enough. Sometimes an agent makes sense because the work involves messy language, several tools and a few decisions that cannot be reduced to one rigid rule.
The first goal is not maximum automation. It is one bounded piece of work that no longer consumes the same staff time each day.