Human Approval in AI Workflows: What to Check Before Buying

Short answer: an approval step only protects you if it does four things:

• it pauses before the action;
• it knows who approved;
• it has a defined path for decline and silence;
• it is included in the plan you are buying.

The three platforms differ most on those points:

• Zapier has a built-in Human in the Loop step on paid plans. On Professional you can only send approval requests to yourself, so approval by colleagues needs Team or Enterprise.
• n8n builds approvals into its messaging nodes ("send and wait") and into AI agent tools. Its default email-style link buttons cannot tell who clicked, so verified approver identity needs a channel that checks it, such as Slack approvals with Capture Who Responded turned on.
• Make offers a Human in the Loop app only on Enterprise, as a closed beta. On other plans you build the approval yourself from webhooks and stored state, and instructions to an AI agent are not an enforced gate.

Ask every vendor to demonstrate the decline, timeout and wrong-person cases before you buy, not only the happy path.

Approval before the action vs review after it

These are different controls, and buyers often get one when they meant to buy the other.

• Approval before the action pauses the run. Nothing is sent, paid, deleted or published until a named person says yes. Use it for irreversible or customer-facing actions.
• Review after the fact lets the run finish and gives someone a log, a report or a sample to check later. Use it for low-risk, high-volume work where mistakes are cheap to correct.

A run history, an audit log or a weekly report is review after the fact. Only a pause that the platform enforces is approval before the action. A prompt that tells an AI to "check with a human first" is neither, because the model can ignore it.

The approval-state matrix

Every approval step can end in more states than "approved". Decide in advance what each state does, then check that the platform can express it.

The decline row is the one most often missed. On Zapier, a decline that continues the run will still reach later steps unless you add a branch. On n8n, the Gmail approval node shows only an Approve button unless you change the type.

Who can approve, and on which plan

Zapier Team starts at $69 a month billed annually ($103.50 monthly) for 2,000 tasks and up to 25 users. n8n Cloud starts at €20 a month billed annually for 2,500 executions. Make Core starts at $12 a month billed annually ($16 monthly) for 10,000 credits, but its built-in approval app is Enterprise-only. See Zapier pricing, n8n pricing and Make pricing.

Worked example: approving a customer credit note

Everything below is invented. "Example Co" lets an automation draft credit notes, but any credit above $500 needs approval from the finance lead, Dana, within one business day.

1. Draft. The workflow creates a draft credit note in the billing system and stores its ID. Nothing is issued yet.
2. Request. It sends Dana an approval request showing the customer, the amount, the reason and a link to the draft.
3. Wait. The run pauses until the end of the next business day (set as a fixed timeout or a computed date), with a reminder halfway.
4. Decide. Each outcome has its own path:

• Approved: the workflow issues the credit note and logs Dana's name and the time.
• Approved with a lower amount: it updates the draft to the edited amount, then issues it.
• Declined: it deletes the draft, records Dana's note, and tells the account manager.
• No answer: it leaves the draft unissued and alerts the deputy approver.
• Someone else clicks: the response is refused and the request stays pending.

5. Act once. Before issuing, the workflow checks that the draft still exists and has not already been issued, so a replay cannot issue it twice.

How this maps to each platform:

• Zapier: use Team so Dana can be a named reviewer. Set Action if reviewer declines to Stop run, or branch on the decision. Set On timeout to End run and alert on ended runs.
• n8n: send the request through Slack's Send and Wait for Response with Capture Who Responded on and Dana as the only approver, and wrap the wait with Limit Wait Time. Gmail's Send and Wait (set to Approve and Disapprove) cannot refuse a click from someone else, so use it only if a forwarded link is an acceptable risk.
• Make: on Enterprise, ask for access to the Human in the Loop beta. Otherwise build two scenarios (request and decision), a data-store record with an expiry, and a scheduled clean-up.

Questions to ask in a demo

• Show a decline. What runs next, and can a later step still act?
• Show a timeout. Where does the item go, and who is told?
• Have someone who is not the approver open the request. Is the response refused?
• Approve, then make the next step fail. What does a replay re-run, and could it act twice?
• Where is the approver's name and decision stored, and for how long?
• Which plan includes each of these, and what does each approval cost in tasks, executions or credits?

Who should not rely on built-in approval steps

• Processes with legal sign-off requirements, such as contract signature or regulated payment release. Use the system of record's own approval controls and keep the automation to drafting.
• Teams without a named approver and deputy. A pending queue nobody watches turns every timeout into an outage.
• Very high volumes. If people approve hundreds of items a day, approvals become rubber stamps. Narrow the gate to high-risk items and review the rest after the fact.

For approval patterns specific to outbound email, see AI email triage. For escalation in support queues, see AI support ticket routing and escalation.