Why is human-in-the-loop essential for AI customer support?

Because the AI speaks for your business, and a wrong answer can be held against you. Air Canada learned that in 2024, when a tribunal made it pay out on a bereavement refund its chatbot had promised and its own policy didn't allow. Checking replies catches those mistakes, and their edits show you which documents or playbooks need fixing.

Can Hay.chat require approval before it issues a refund?

Not through approval mode, which holds replies, not actions. You set refund limits in a playbook instead: a cap, a time window, or a rule that every refund goes to your team. The agent follows those rules and escalates anything outside them. Approval mode can still hold the message that tells the customer.

Is human approval on by default, and which plans include it?

It's off by default and included on every plan. Switch it on per agent with Require Approval, or for the whole workspace with Test Mode Default in Settings, then General. You can try it during the 30-day free trial, and you won't need a credit card to start.

Can I require approval for some agents and not others?

Yes. Message Approval is set per agent: Require Approval, Auto-Send, or Inherit from Organization. A common setup is approval on the agent that handles refunds and complaints, and auto-send on the one answering shipping and sizing questions. Changes apply to new conversations with that agent.

What happens when the AI isn't sure of its answer?

Every answer that states facts gets a confidence score before it's sent. With the defaults, 0.8 or above counts as high. Between 0.5 and 0.8, Automatic Recheck can pull more documents and try again. Below 0.5, with Human Escalation on, the conversation goes to your team instead. Switch escalation off and the reply is sent as it is.

Do test chats need approval?

No. Conversations in the Conversation Playground always send straight away, even on an agent set to Require Approval, so you can try an agent without approving your own messages. Approval only holds replies in real customer conversations, which makes the Playground a quick place to check a new playbook first.

Can my team take over a conversation from the AI?

Yes. Take over a conversation and the AI stops responding while you write to the customer yourself. Your messages are logged in the same thread as the AI's. When you're done, release it back to the AI or to the queue. It's the fix for a draft that an edit won't save.

Is there an audit log of what the AI said?

Every conversation keeps its full thread: customer messages, AI replies and system messages, each timestamped, plus anything your team sent after taking over. Export any conversation as a PDF or CSV. Account activity (logins, password changes, API keys) is under My Profile, and your personal GDPR export includes your audit log, up to the last 1,000 events.

How long does Hay.chat keep conversations?

Until you delete them, unless you set a retention period of 30 to 365 days in your workspace settings, after which closed conversations are anonymised. If you'd rather keep the whole record on your own servers, Hay.chat is source-available and can be self-hosted.

Does cancelling a Shopify order through the AI refund the customer?

Not by itself. In Shopify, cancelling and refunding are separate steps, so the agent can cancel and draft the refund message, but the refund is its own step. Opening a return registers it in Shopify, and the label comes from your carrier or returns app. Refunds go back the way the customer paid, so timing depends on the payment provider.

Human approval for AI customer support: read the reply before your customer does

Human approval for AI customer support in Hay.chat holds an agent's replies until someone on your team approves, edits or blocks them. Refunds and other actions follow your playbook rules instead.

Shopify — The lookups ran on their own. The reply waits for someone on your team to approve, edit or block it.

Published 21 October 2026

Human-in-the-loop

What does human-in-the-loop mean in customer support?

Human-in-the-loop (HITL) means a person sits at a checkpoint in an AI workflow and can stop or change what the AI does before it counts. In support, the checkpoint that matters is the reply, and that's the one Hay.chat holds.

Even a good AI is wrong sometimes, ours included. A refund window that's 14 days in your policy and 30 in the reply gets caught by a teammate in seconds. More in our safe AI patterns guide.

Which AI replies need approval, and how does it look?

Speech bubbles sliding down a forked slide: most ride the long branch straight out, one waits in a tray under a yellow lamp

You pick the agents

Approval is off by default and on every plan. Switch on Require Approval for one agent, or for the whole workspace in Settings, then General.

Human-in-the-loop in customer support: blue speech bubbles circling a looped track, one paused under a navy arch with a small yellow lamp

The reply waits

The agent runs its lookups and writes the reply as usual, but it lands in the conversation as "Pending" instead of going out.

A blank speech bubble waiting on a pedestal with a small yellow lamp, a green chip, a red chip and a pencil laid out in front of it

Your team decides

Approve it, edit it first or block it. Nothing reaches the customer until you do. If a draft is beyond an edit, take over the conversation.

Our take: turn it on for any new agent and move it to Auto-Send once you've stopped editing its drafts. The settings are in the agents guide.

A clipboard on a post beside a lever sets railway points: small parcels roll down the main track while one is switched onto a side track toward an open door

What about actions like refunds?

Approval mode doesn't hold actions, and we'd rather you hear that now than in week one. Refunds and order changes follow your playbooks, so the limit goes there:

"Refund damaged items up to €50 within 30 days of delivery. Anything over €50, pass it to the team."
"For refund requests outside the policy, collect the order number and the reason, then escalate."

One Shopify catch: cancelling an order doesn't refund it on its own. The actions list shows what each integration can do.

A tall card-catalogue cabinet with one drawer pulled out, packed with blue cards, one speech-bubble card lifted out and glowing yellow

What record do you get of what the AI did?

Every conversation keeps its full thread, timestamped, including anything your team sends after taking over. Export it as a PDF or CSV, or self-host to keep it all on your own servers.

Account activity is logged separately under My Profile, from logins and password changes to API keys created or revoked.

Should every AI reply need approval, or should the agent send on its own?

Most teams run both, split by agent. Approval costs your team's time. Autonomy costs you the occasional bad reply.

Swipe the table sideways to see every column.

Three ways to run an AI agent in Hay.chat, reviewing every reply, reviewing by confidence or fully autonomous, compared by setting, your team's time, risk and fit.
	Start here Review every reply	Review by confidence	Fully autonomous
Setting	Require Approval on the agent	Auto-Send, with Human Escalation on	Auto-Send, with Human Escalation off
Your team's time	Someone reads every reply, and chats that come in at night wait for the morning	Only the conversations the agent isn't sure about	None, until something goes wrong
Risk	Lowest. A bad draft only gets out if the reviewer misses it	Medium. A confident wrong answer still goes out	Highest. Low-scoring replies are sent as they are
Fits	New agents, refunds and complaints, your first weeks	Order status, shipping and policy questions, once your documents are solid	Hardly anyone

Time and risk are our judgement from running support teams, not measured data.

Switch on approval for one agent and read what it would have sent

Turn on Require Approval, put your refund limits in a playbook, and read its drafts until you trust them. Free for 30 days, no credit card.

Start Your Free Trial or book a demo

30-day free trial
No credit card required
Set up in under 10 minutes
Hosted in Europe

Frequently asked questions

HITL stands for human-in-the-loop: a person reviews or corrects an AI system's output at a set point instead of letting it run unchecked. In customer support that usually means approving AI replies before they're sent, or taking over conversations the AI isn't sure about. Hay.chat supports both.

Written by Damien Mulhall. Last reviewed 21 October 2026 against Hay.chat's agents and conversations documentation.