Rules, AI and people solve different parts of a workflow. A reliable system can use all three without pretending they are interchangeable. Start by describing the decision a step must make, the evidence it receives and the consequence of being wrong. Then choose the smallest mechanism that can do the job.

Before you start

Take a real workflow and divide it into individual steps. Avoid assigning one label to the entire process. A customer request may need rules for duplicate detection, AI for extracting intent and a person for an exception. Write the input and expected output for each step before selecting a tool.

The workflow

  1. Use rules for exact conditions you can state clearly: a required field exists, an invoice status equals paid, or the same event ID was already processed. Keep these checks visible and testable.
  2. Use AI for variable language or unstructured material when a grounded draft or classification is useful. Supply the source material and require uncertain cases to remain uncertain rather than inventing missing facts.
  3. Use people for judgment, authority and commitments. A refund exception, sensitive complaint or material change to an agreement needs someone who understands the context and has permission to decide.
  4. Connect the types with explicit handoffs. An AI suggestion should carry its source and proposed next action. A rule should validate required fields before an approved action reaches an external system.
  5. Build a fallback for each step. A missing record, conflicting source or unclear request should route somewhere useful. “The model will figure it out” is not an exception-handling plan.

Worked example

Illustrative example: an incoming message asks for a refund and mentions a damaged parcel. Rules first locate the order and check whether the request already exists. AI drafts a concise summary using the customer’s message and available order facts. A support owner reviews the evidence and chooses the response. The workflow records the decision, then an authorized action uses that approved outcome. No model confidence score substitutes for refund authority.

Your worksheet

  • Step and decision: ___
  • Inputs and source of truth: ___
  • Can the rule be stated exactly? ___
  • Does the task involve variable language? ___
  • Who has authority to commit? ___
  • Selected type and fallback: ___
  • Evidence passed to the next step: ___

Test it

  1. Try a straightforward request and inspect the full path.
  2. Provide conflicting information and verify the workflow asks for review.
  3. Remove a required field and confirm no external action proceeds on an invented value.

What to measure

Track corrections and escalations by step type. A high escalation rate can indicate unclear inputs, not a need for a more powerful model. Improve the decision boundary before adding more automation.