Lorixen.
EN

AI automations with a human approval step: where to put it

Short answer: Put the human approval step immediately before any action that leaves the building or can't be undone: sending to a customer, paying, deleting, publishing. Let the AI do the reading, sorting and drafting freely, then pause the workflow with the draft shown in full, buttons to approve, edit or decline, and a clear rule for what happens if nobody answers. Log every decision so you can loosen the reins later with evidence rather than hope.

"Human in the loop" sounds abstract until you try to decide where, exactly, the human goes. Put the checkpoint too early and the person approves things that never needed them. Put it too late, or nowhere, and the first odd email goes to a client before anyone sees it. The easiest way to get this right is to walk through one real-looking automation, so here is a fictional example first and the general rules after.

The worked example: a quote-reply automation

Hypothetical business: a small events company in Bristol that hires out marquees, staging and lighting. Enquiries arrive through a website form and by email. The owner wants an AI assistant to draft quote replies so they don't sit unanswered over a busy weekend.

Step 1: Trigger and read (no approval needed)

A new enquiry arrives. The automation pulls the details: date, venue, guest numbers, items mentioned. An AI step summarises the request and flags missing information (no postcode, no date).

Nothing has left the business and nothing has changed. No human needed here.

Step 2: Check availability and price (no approval needed, but guard rails)

The workflow looks up the date in the bookings sheet and pulls prices from the company's own price list. The AI is told to use only listed items and prices, and to mark anything it can't match as "needs a human price" rather than guessing.

Reading internal data is low risk. The guard rail is in the instructions, not a person.

Step 3: Draft the reply (no approval yet)

The AI writes a reply: what's available, the line-item quote, questions about anything missing, and the next step. It also writes a one-line internal note: "Confidence: low, venue access unclear."

Step 4: The approval step

This is where the human sits, because the next action sends something to a customer, with prices, in the company's name.

The workflow pauses and sends the owner a message in Slack (or email, or Teams) containing:

  • The original enquiry, unedited
  • The draft reply in full, not a summary
  • The quote total and any "needs a human price" flags
  • Three buttons: Send as is, Edit then send, Don't send

On Zapier, this maps to the Human in the Loop tool's Request Approval action. As of November 2026, Zapier's help page describes it pausing a Zap so reviewers can approve, decline or edit the submitted data, with notifications by email, Slack or a second Zap, custom button labels, and a choice of whether a decline stops or continues the run. Zapier lists it as a premium app needing a paid account, so check your plan first.

Step 5: Timeout and default

What if the owner is on site setting up a wedding and doesn't look at Slack for six hours? Zapier lets you set a timeout in minutes, hours, days or weeks, and choose whether the run ends or skips ahead when time runs out, plus an optional reminder.

For a customer-facing quote, the safe default is: on timeout, don't send the AI draft. Instead, send a short holding reply ("Thanks, we've got your enquiry and will come back with a quote by tomorrow") that was written and approved once, in advance. The customer hears something; nothing unapproved goes out.

Step 6: Send and log

Once approved, the reply goes out and the workflow writes a row to a log: enquiry ID, draft text, final text, who approved, when, and whether they edited it.

After a couple of months, the owner can read that log and see, for example, that drafts for simple single-item hires were almost never edited, while multi-day festival quotes always were. That's the evidence for loosening the rule on the first type and keeping it on the second.

The rules behind the example

What needs a human, what doesn't

Action Approval? Why
Reading, summarising, tagging No Nothing leaves, nothing changes
Drafting text for later use No The draft is harmless until sent
Sending to a customer or supplier Yes, at first Your name, your promises
Paying, refunding, issuing credit Yes Money is hard to claw back
Deleting or overwriting records Yes Often irreversible
Publishing (site, social, reviews) Yes Public and screenshot-able
Internal notifications No Low cost if wrong

n8n's documentation on human review for AI agent tools draws the same line: it suggests review for tools that send messages, modify records, delete data or make purchases, and lets you require approval on selected tools only rather than gating the whole agent. When a reviewer denies a tool call there, the action is cancelled and the agent is told it was rejected.

Make the approval easy to give properly

An approval message the reviewer can't judge in ten seconds gets approved blindly. Show the full outgoing text and the exact recipient. Show the source the AI worked from. Put risky numbers (totals, dates, discounts) on their own line. Let the reviewer edit in place rather than declining and starting again.

Pick the channel where the person already is

The approval only works if it reaches someone who'll act. For a sole trader that might be a phone notification; for a team, a shared Slack or Teams channel with a named owner per day. n8n notes that the review channel can differ from the one the AI uses with customers, so a chat assistant can route its approvals to one person in Slack.

Decide the default before you need it

Every approval step needs an answer to "what if nobody responds?" For outbound messages and payments, the default should be not to act. For low-stakes internal steps, continuing may be fine.

Batch the boring approvals

If the morning brings twenty near-identical approvals (say, booking confirmations), let the reviewer see them in one list and approve the clean ones together, while anything flagged stays separate. Approval fatigue is how the one bad draft gets through.

Loosen with evidence, not confidence

Start strict. Review the log monthly. Remove approval only for categories where the human almost never changed anything, and keep a spot check on a sample.

FAQ

Doesn't an approval step defeat the point of automating?

No. The slow part is usually reading, checking and drafting, which the AI still does. Approving a ready draft takes seconds.

Can the AI decide for itself when to ask for approval?

It can flag low confidence, which is useful, but don't make that the only gate. Fix approval rules to the type of action, and treat the AI's flag as an extra trigger.

What should the log contain?

The input, the AI's draft, the final version, who approved it, when, and whether it was edited. That's enough to spot patterns and to explain any mistake later.

Is this only for big tools like Zapier or n8n?

No. Any setup that can pause and wait for a click, including a simple email with an approval link, can do it. The design questions are the same.

To work out which steps in your own process should wait for a person's yes, walk us through the workflow you have in mind.

Sources

WhatsAppLet’s talk