AI & Automation

Where an AI Workflow Should Wait for a Human

Put human approval before the step that changes something outside your workflow.

An employee who receives a half-written quote can usually tell which blanks still need filling. A customer who receives the same draft as an email cannot. That difference is a useful starting point for designing AI automation: distinguish preparation from the act that creates a commitment. A model can organize an inquiry, highlight missing details and draft a response. The ability to send the result is a separate decision, with different consequences and a different owner.

Find the step that changes the outside world

In its human oversight playbook, n8n describes reviewing proposals in chat and approving selected agent tool calls before they run. The design question is not whether your company needs a person somewhere in the diagram. Ask what action would be hard to take back if the model or workflow were wrong. Sending a statement about price or delivery to a customer is one such action. Extracting the order quantity into an internal field is usually a different kind of risk.

Draw a line between those two steps in the workflow. The intake side may summarize and prepare. The action side may send or write only when the review conditions have been met. If an external service is unavailable, leave the request with a visible status instead of letting it disappear between nodes. Use a different status for a draft that has not yet been reviewed and one that was rejected; they call for different follow-up.

Example: enquiries for custom print

Imagine a small print shop receiving requests for branded products. This is a hypothetical workflow, not a client case. A customer asks for 80 pieces, delivery next Friday and a quote for two artwork variants. The automation can extract those details and identify questions: which product, what size, does the deadline refer to dispatch or arrival? The team supplies approved pricing and production rules. When the inputs are incomplete, the draft should ask for clarification instead of filling a price or promising a date.

The record can contain the original request reference, the extracted details and the draft. Show the reviewer the outgoing recipient, subject, exact text to be sent and where any quoted price or deadline came from. A short list of uncertain details is more helpful than the model's confidence score. If the employee cannot tell from the approval card what will actually happen after clicking, make the card clearer before adding more automation.

Bind the decision to the action

  • Give each request a stable ID and each draft a separate version. Record which version the reviewer saw. If a person or a later step changes the price, recipient or text, do not carry forward the previous approval.
  • Route decisions to authorized staff with a clear reject option and an expiry. An expired review, lost notification or unavailable reviewer must not become an automatic approval. Keep the request visible for a person to pick up.
  • Before sending, check that the action is still pending, the reviewed version is current and an authorized person approved it. Claim the action once in durable state before calling the sender. Use the sender's idempotency feature if one is available; a button alone cannot prevent double sends.
  • If delivery returns an ambiguous error, record that uncertainty and check delivery status before retrying. A timeout is not proof that the provider never accepted the request. Never automatically send again just because an execution failed before it could store a receipt.

Consider where credentials live. The draft generation step does not need direct access to the mail token. The component that sends should have only the permissions it needs for that mailbox and operation. Keep customer data out of notifications where possible; a reference and a link to the appropriate review view are often enough. Treat instructions inside an inbound message as customer text, not as authority to change the workflow, modify credentials or bypass review.

Launch with a small, observable test

Start with one mailbox, one request type and one primary reviewer. Use invented requests in a test-only route with sending disabled. Check that an approval enables exactly the expected action, a rejection never sends, and a delayed approval does not send an outdated version. Try two reviewers responding close together and a repeated approval after the first action. Simulate a sender timeout and check that the team can tell the difference between unsent, sent and uncertain.

Once the team is comfortable with the test, run a limited live trial. Keep the default path manual when a request is unfamiliar or information is missing. Record why drafts were returned for changes, such as unsupported promises, missing quantities or the wrong recipient. If the same issue recurs, improve the form or add a deterministic check before the review rather than making the reviewer inspect the same defect all day.

Decide what to change after the trial

Compare review time and corrected drafts for the same request type before widening the scope. Also check the queue for requests that sat without an owner: a fast draft is not helpful if no one reviews it. Use a small sample to identify friction, not to promise a percentage improvement. A team might choose to automate more of the extraction, leave sending behind approval, or remove an approval for a genuinely reversible internal step. These are separate choices.

Make approval explainable

Give staff a way to see why a request is waiting. Separate an unanswered review from an unsupported price and from a failure at the sending provider. On a small team, one queue can be enough if it shows the responsible person and the next permitted step. Avoid placing entire customer emails in shared channel messages. Record what was decided, when and against which revision, while keeping the full details in the system where the team already handles the request.

For a quote, tailor the review to what is being promised. A message that only asks for the missing product model may need a different check from one that states a deadline. Begin with one conservative rule; split it only after you can show the team understands both paths.

The first practical step is to take one existing workflow and write one sentence describing what must never happen without a person's consent. Put the approval immediately before that action, test the failure paths, and keep a manual way to finish the job. The aim is useful automation without delegating a commitment the business still needs to own.

Read the original source