Which Process Is Clear Enough to Delegate to an AI Agent?
Assess process readiness through outcomes, inputs, decision boundaries, tools, exceptions, permissions, test data, and recovery.
A process with a written SOP is not automatically clear enough for an AI agent. Many documents describe the happy path while real decisions live in employee experience: which data to trust, which exceptions matter, when to ask again, and who owns an error.
A ready process has a verifiable outcome, minimum inputs, decision boundaries, contracted tools, known exceptions, task-scoped permissions, and a handoff or recovery path. Without these elements, an agent pilot converts business ambiguity into difficult-to-observe failures.
First step
Write the definition of done before the prompt
The agent needs a completion condition and authoritative evidence.
Choose a case unit and describe start and end states. “Process a lead” is too broad. “Verify required fields, classify according to policy, and create a task for the correct owner within ten minutes” is testable. Completion comes from a record or event in the target system, not from the model’s statement.
List required and optional inputs, authoritative sources, and acceptable freshness. If systems conflict, define precedence or route to review. For unstructured input, specify extracted fields, evidence, and corresponding confidence or validation.
Every step should map to a concrete action or output. OpenAI recommends using existing procedures, breaking dense documents into smaller steps, defining clear actions, and capturing edge cases in agent instructions. See the agent guide.
