Define your first governed AI workflow.

Start with a person who can decide whether the result is useful and who has authority over the process. An agent is not the accountable owner. Write a one-sentence objective before discussing tools.

List the inputs, permitted sources, actions and completion conditions. Separate preparing a recommendation from executing it. Leave irreversible actions out of an initial experiment unless review and recovery are explicitly defined.

Identify the precise point where a person must approve a response or action. Give that reviewer the source information, proposed change and uncertainty. Define what happens if approval is declined or unavailable.

Use representative, permitted examples to understand the current process. Evaluate factual accuracy, correction effort, delay and cost. Set stop conditions for unsupported answers, missing permissions and unacceptable errors.

Who owns the workflow? Which source is authoritative? What is the most consequential action? What evidence would justify moving forward? Use non-sensitive examples in an initial conversation.