Work out which part of your process needs AI.

Consider a fictional company routing support requests. Some arrive through a form with a selected category; others are free-text messages. The expected output is a recorded request with an assigned owner and no duplicate entries.

Map the current sequence: receive, validate fields, classify, assign and acknowledge. Record who handles exceptions. If two people disagree about the correct outcome, resolve that disagreement before building the workflow.

For the form, create a lookup table: billing goes to administration, technical issues go to support, and an empty category goes to manual review. If reading a field and applying that table resolves the decision, you can implement this step with explicit rules.

Include required fields, allowed categories and a request identifier for detecting repeats. Rules still need maintenance when the process changes. Writing them down makes it possible to check which version was applied.

A message such as “I was charged twice and cannot sign in” contains two issues. You could test an AI classifier that suggests labels from the text. Give it a closed list of categories, examples and a “needs review” option.

Check the suggestion before routing the request: is the label allowed, is information missing, and are there multiple issues? A model suggestion and a system action are separate steps. In this pilot, a person decides ambiguous cases.

Prepare requests with expected destinations agreed by the team. Include routine cases, incomplete messages and requests covering several issues. Keep some cases for a final check without using them to tune the instructions.

Compare rules plus manual review with a version that adds AI classification. Record correct assignments, incorrect destinations, escalated cases and total review time. Distinguish errors that delay a reply from errors that send information to the wrong team.

Here is an invented example: the original workflow takes 80 minutes for a batch. The new one takes 35 minutes to operate and another 30 to review, saving 15 minutes in that batch. These figures illustrate the calculation; they do not predict your results. Record tool and maintenance costs too.

Keep the chosen process, rule table, test examples, results and named owner in one document. Decide what result would justify continuing and what error would stop the pilot. Set those thresholds according to the consequences of your task.

If review consumes all the time you hoped to save, narrow the task or improve the input. If the pilot works, expand one request type at a time and keep a manual route for recovering work.