Your first AI workflow
should be easy to check.

1. Describe the task before the tool.

Write down what starts the work, what information you receive, the steps you follow and the output you need. “Prepare a reply to this type of inquiry” is easier to test than “automate our customer service.”

Look for repeated work where the source facts are available and a mistake can be caught before the result is used. A checklist, template or ordinary rule may solve the problem without AI.

2. Identify the source of truth.

List the facts the task depends on: approved business information, a product specification, a customer policy or the original document. Mark anything the workflow must not guess.

Begin with made-up or redacted examples. Review your business’s policies and the chosen tool before using customer, staff or confidential information.

3. Define the person’s role.

Agree who checks names, numbers, commitments and missing information. Specify what happens when the source is incomplete, the AI output is uncertain, or the task falls outside its scope.

Keep important decisions with the person authorized to make them. Drafting a reply and deciding on a refund, appointment, quote or financial action are different responsibilities.

4. Establish a baseline and test.

• Time a few ordinary tasks using the current process.
• Choose comparable examples for the proposed process.
• Include an incomplete input and an unusual request.
• Count corrections as well as time spent.
• Record software and usage costs alongside any saved effort.

Expand only after the result is useful and understandable to the person who will use it. A successful demonstration is a starting point; real work also needs ownership, instructions and a way to handle exceptions.

Understand the costs of an agreed project →