Questions about the work
First, does the work repeat? Automation pays back on tasks done many times in much the same way. If every instance is different, you are describing judgment, and judgment is not the place to start.
Second, can you say what a good result looks like? If two of your own people would disagree about whether a given output is correct, an AI employee has no standard to meet. Agree the standard first, in writing, with examples.
Questions about data and tools
Third, is the information the work needs written down somewhere a system can read? Policies in someone's head, prices in a spreadsheet on one laptop and customer history in personal inboxes all have to be fixed before a build, not during it.
Fourth, do the tools involved allow a connection? Most modern business software does. Where a system has no way in, the honest options are to replace it, to keep a person on that step, or to pick a different workflow.
Questions about people and risk
Fifth, who owns the workflow? There must be a named person who decides the rules, reviews the exceptions and can pause it. Shared ownership usually means nobody notices when quality slips.
Sixth, what is the worst thing a mistake could do? If the answer is a mildly wrong internal note, start there. If it is a wrong payment or a wrong message to a customer, the workflow can still be automated, but only with approval gates on those steps.
Use this before you buy or build.
- The workflow repeats often and in much the same way.
- A correct result is defined in writing, with examples.
- The data it needs is stored where a system can read it.
- A named owner can set the rules and pause the workflow.
Score one workflow, not the whole company. If it repeats, has a written standard, readable data and a named owner, it is ready to scope.

