# Refund approval workflow: small refunds in seconds, big ones checked by the right person

> How to design a refund approval process: write the policy, let small refunds run in chat, cap what goes out each day, and send big ones to the right approver.

Source: https://pengo.chat/blog/refund-approval-workflow/
Topic: Playbooks
Updated: 2026-10-21

## The short version

- Write the **refund policy down before you automate it**: the window, the reasons you accept, partial versus full, when the customer keeps the item, and the alternatives you'd rather offer. Automation can only follow rules that exist.
- Sort refunds into **tiers**: small ones run on their own up to an amount, bigger ones ask first, and high-risk ones go to managers. A **daily cap** limits how many refunds go out without a person on any one day.
- The amount must come **from the order, never from the chat**. A customer can type any number. The order can't.
- Refund only **verified customers**, note the reason on the order, and tell the customer honestly what is happening while a request waits for a person.
- Read the **Activity tab weekly** and move your limits in small steps, based on what your team actually approved and rejected.

Refund requests are where support speed and support caution pull hardest in opposite directions. The customer wants their money back now. Your finance side wants to know every pound that leaves, and why. Most teams settle the argument by sending every refund through one person, which means a missing pair of socks waits a day for a manager while a much bigger, stranger request gets the same quick glance.

A refund approval workflow fixes that by deciding in advance which refunds need a person and which don't. Small, clear-cut refunds happen in the chat in seconds. Big or unusual ones reach the right person's eyes with everything they need to decide. This playbook walks through designing one, in the order we'd do it, using Kettlewell Outdoor, a walking boots and outdoor clothing store in Leeds, as the example.

## What a refund approval workflow is

> **Refund approval workflow**
>
A refund approval workflow is the set of rules that decides who can approve a refund, up to what amount, for which customers, and what happens while a request waits. It covers the rules (who qualifies for what), the routes (who decides at each amount) and the records (why each refund was given and who approved it).

The same three questions apply to every refund request:

1. **Does it qualify?** Inside the window, for a reason you accept, for an order this customer actually placed.
2. **Who decides?** Nobody, because it's small and clear. A teammate. A manager.
3. **What's recorded?** The amount, the reason, who approved it and when.

Answer them once, in writing, and the rest of this guide is plumbing.

## Step 1: Write the refund policy down first

You can't automate a policy that lives in someone's head. Before you touch a setting, write down what your team already does, including the exceptions they make without thinking. A refund policy worth automating covers:

- **The window.** How long after purchase or delivery a customer can ask. Kettlewell Outdoor takes back unworn items within 60 days.
- **The reasons you accept.** Damaged in transit, lost parcel, wrong item, changed mind. Be specific about which reasons get a refund and which get a return.
- **Partial versus full.** One missing pack in an order of three is a partial refund. A parcel that never arrived is a full one.
- **When the customer keeps the item.** Asking someone to post back a scuffed £8.00 pair of socks costs you more than the socks. Write down the amount below which you don't want things back.
- **The alternatives.** A late parcel that arrived fine may deserve an apology, not a refund. A single-use discount code or account credit often fits better, and gives them a reason to order again.
- **What always goes to a person.** Outside the window, a parcel tracking says was delivered, a second claim in a month, anything involving a chargeback.

For Kettlewell, the written version fits on a page:

```text
Refunds at Kettlewell Outdoor

Window: unworn items within 60 days of delivery. We send a prepaid Royal Mail label.
Damaged in transit: refund the damaged items to the original payment. Customer keeps them.
Lost in transit: full refund once tracking shows no delivery after 10 working days.
Late but fine: say sorry with a single-use discount code instead of a refund.
Changed mind on worn boots: no refund. Offer the size guide and a person to talk to.

Up to £20: no approval needed.
Above £20: a manager approves.
Always a person: outside 60 days, "delivered" per tracking, second claim in 60 days.
```

Then put the customer-facing half of it into your help content. If your returns article says one thing and your team does another, any AI answering from that article will repeat the article. Our guide to [writing help content an AI can answer from](/blog/write-help-content-for-ai/) covers how to structure a page like "Returns and exchanges" so it gets found and quoted correctly.

> **Write the exceptions as rules**
>
"Use your judgement" is not a rule anyone can follow, human or software. "A second damage claim within 60 days goes to a manager" is. Every exception you can phrase that way is one less refund that depends on who happens to be on shift.

## Step 2: Sort refunds into tiers

Most refund requests fall into three groups, and each deserves a different route.

| Tier | Example | Who decides | Why |
|---|---|---|---|
| **Small and clear** | One pair missing from a 3-pack: £8.00 | Nobody: it runs on its own | The cost of waiting is higher than the cost of the refund |
| **Bigger** | Two sock packs lost from a torn parcel: £48.00 | A manager approves | Worth a ten-second look at the order |
| **Unusual** | Outside the 60-day window, or a parcel tracking shows as delivered | A person handles the conversation | Needs judgement, not a button |

The point of the top tier is speed. The point of the middle tier is a second pair of eyes, and the point of the bottom one is to make sure a person is in the conversation, not just approving a request. In Pengo, write each "always a person" case as escalation guidance in **Intelligence › Behavior**, and Pengo hands those conversations to your team with a summary.

Here is how one illustrative policy routes a refund. Try a few amounts, switch between Pengo refunding in the chat and a teammate refunding from the inbox, and see what happens when the day's cap is used up.

_Figure: An illustrative refund policy: Pengo refunds up to £20 on its own, up to 10 automatic refunds a day, managers approve anything bigger, and teammates refund up to £60 before a second person signs off. Every route checks the customer and takes the amount from the order._

In Pengo, **Refund an order** (Shopify, WooCommerce and BigCommerce) and **Refund a payment** (Stripe) each have a mode: **Auto**, **Ask first**, **Team only** or **Off**. Refunds change something in another tool, so they start on **Ask first**. On Auto, a guardrail lets Pengo refund on its own **up to an amount** and ask above it. Set it in **Intelligence › Actions**; see [Guardrails and limits](/help/actions/guardrails-and-limits/).

## Step 3: Add a daily cap

A per-refund limit answers "how big can one refund be?" A daily cap answers a different question: "how many can go out without a person on a bad day?"

Bad days happen: a courier crushes a van of parcels, or a product page shows the wrong price for an afternoon. On a normal day, every refund under £20 is fine. On one of those days, fifty of them in an hour should make someone look up.

Set the cap from your own numbers. Count how many small refunds Pengo would make on its own on a busy day, and set **Automatic runs a day** comfortably above it, so it never bites on a normal day and always bites on an odd one. Together with the £20 limit, it also caps the money: 10 runs of at most £20 is £200 at most without a person.

When the cap is reached, nothing is refused. Further requests simply ask first until the next day, so customers still get their answer, just from a person. In the router above, switch the cap to "Used up" to see an £8.00 refund that would normally run on its own wait for approval instead.

## Step 4: Take the amount from the order, never from the chat

This is the rule that matters most, and the one that's easiest to get wrong when you build refunds into a chatbot yourself.

A customer can type anything. "I paid £90 for that" costs nothing to say. If your automation reads the amount from the conversation, your auto limit protects nothing. The amount has to come from the order or invoice itself, the one record the customer can't edit.

The same goes for what is refunded. The order says what was bought, at what price, and whether it's already been refunded. Refunding the two packs of socks on order KW-21086 means refunding the line items as the order records them, to the original payment, not to a card number someone pastes into the chat.

> **Watch for split claims**
>
An auto limit per refund can be gamed by asking for several small refunds on the same order. Taking amounts from the order closes most of that gap, because the order shows what's already been refunded. The daily cap closes the rest.

## Step 5: Check who's asking

A refund moves money, so it should only ever go to the person who paid. That means knowing who you're talking to before anything runs.

There are two common levels of proof in ecommerce support:

- **Order number and email.** The customer gives the order number together with the email on the order. Good enough to show them their order status or add a delivery note.
- **A verified customer.** The customer is signed in to your site, and your site vouches for them. This is what refunds, cancellations and billing changes should require.

Order numbers and emails turn up in forwarded emails and shared inboxes, so neither proves the person typing is the person who paid.

Every Pengo Action says what it needs to know about the customer, and Pengo checks it every time, whatever the mode. Refunds are for **verified customers only**: shoppers signed in to your site with [identity verification](/help/developers/identity-verification/). If the customer isn't verified, Pengo doesn't run the refund. It asks them to sign in, or hands the conversation to your team. See [Identity and safety](/help/actions/identity-and-safety/).

## Step 6: Note the reason on the order

Every refund should say why it happened, in the place your team and your accountant will look later: the order itself. "Damaged in transit, two packs of socks" on KW-21086 answers the questions that come up months later. Was it a courier problem? A packing problem? The same customer as last time?

Reasons also make your weekly review possible: if damaged-in-transit refunds jump after you change packaging, you want to see it in a list.

Choose a small, fixed set of reasons (damaged in transit, lost in transit, wrong item, price adjustment, goodwill) and use them consistently. Free-text reasons are fine for detail, but a fixed label is what you can count.

## Step 7: Decide what the customer hears while it waits, and after

A refund that's waiting for approval is the moment customers most often get misled, usually by accident. A chatbot that says "Done! Your refund is on its way" before anyone has approved anything has made a promise your team now has to keep, or break.

The rule is simple: say what's actually true. While a refund waits, the customer hears that it's with the team. When it's decided, they hear what happened.

_Figure: A refund above the £20 auto limit. Gareth is signed in, so Intelligence looks up KW-21086, tells him it has asked the team, and the approval card shows the amount taken from the order and the reason._

Write down the three follow-ups before you go live:

| Outcome | What the customer hears |
|---|---|
| **Approved and done** | A confirmation of what happened: the amount, where it's going, and that there's nothing more to do |
| **Rejected** | That a teammate has reviewed the request and will follow up with them directly |
| **Couldn't be completed** | That it couldn't be done automatically and a teammate will follow up |

Notice what's missing from the rejected case: the reason. A refusal deserves a person, not a template. If your teammate wants to say why, they can add a message when they reject it, or take over the conversation.

While a refund waits, Pengo says it has asked the team. Once it's decided, Pengo follows up in the same conversation, in the customer's language and on their channel. Follow-ups are fixed messages or come from the action's own result, never from the AI guessing, so they can't say more than what actually happened. See [What customers see when Pengo acts](/help/actions/what-customers-see/).

## Make approving fast, wherever the approver is

The middle tier only works if approvals don't sit for hours. A refund that waits until a manager is back at a desk is barely faster than email. Approvers should be able to decide from wherever they are, with the customer, the order and the amount in front of them.

Plan for nobody answering, too. A refund nobody approved should never go out on its own after a while.

Pengo sends approval requests to the people who can decide: as an iPhone or Android notification (on iPhone, Approve needs Face ID or your passcode), on Apple Watch, by email, in Slack, or in the inbox banner. The first decision wins, and everyone else sees who decided and where. Approvers get a reminder after 15 minutes, managers and the owner after an hour. After 4 hours the request expires and the conversation is handed to your team. **Expiry never approves anything.** See [Approve from anywhere](/help/actions/approve-from-anywhere/).

By default, any teammate can approve low- and medium-risk actions, while high-risk ones, refunds included, go to managers and the owner. That matches the policy most shops already have: the person who can say yes to money is the person who's accountable for it. Each approver picks where they get requests in **Settings › Your profile › Notifications**.

## When a teammate refunds from the inbox

Not every refund starts with an AI. Often a teammate is already in the chat and just needs to issue it. The same rules should apply to them, with one addition: a personal limit.

A sensible set-up gives each teammate a limit they can refund on their own, and requires a second person above it. It's the four-eyes rule finance teams already use, and it protects your team as much as your money: nobody has to defend a big refund they approved alone.

The other rule for teammates is about identity. A teammate can see more context than any automation, but they can still be fooled by a confident customer with someone else's order number. The tool should make them stop and confirm they've checked who they're talking to before a change for a customer who isn't verified, and record that they did.

In the inbox, the **Actions** button in the composer shows the actions that fit the conversation, ranked by Pengo, with the customer, order and amount already filled in. Check the details, click **Run**, and Pengo drafts the reply. If the customer isn't verified, Pengo warns you and asks you to confirm you've checked who they are; your confirmation is recorded. A change above your own limit needs a second teammate to approve it. See [Run actions from the inbox](/help/actions/run-actions-from-the-inbox/).

If you'd rather refunds were always a person's job, there's a mode for that too. **Team only** keeps the action for your team: when a customer asks for a refund, Pengo says a teammate is joining, hands over, and leaves the refund pre-filled in the Actions button for a one-click run.

## Offer the alternative before the refund

Some requests are better answered with something other than money back. A parcel that arrived three days late but intact is an apology, not a refund. A customer who's disappointed but still keen deserves a reason to order again.

Write down which situations get an alternative instead of a refund, and what that alternative is:

- **A single-use discount code** for a late delivery or a small disappointment. Pengo's **Send a discount code** creates one only this customer can use, valid for 30 days, on Shopify, WooCommerce and BigCommerce.
- **Account credit** for subscription businesses, where the next invoice is the natural place to make up for a problem.

Alternatives go through the same tiers. A discount code is still money, and it starts on Ask first like any other change in your store that moves money.

## Refunds for subscription businesses

Everything above applies to subscriptions, with a few differences worth planning for. Take Tallybook, a bookkeeping app for small agencies with Growth and Scale plans, billed through Stripe.

Subscription refunds are usually about a specific payment: a customer charged twice in a month, or charged after they meant to cancel. So the refund attaches to that payment, in full or in part, back to the card it came from. As with orders, the amount comes from the invoice, not the conversation.

Subscription businesses also have more alternatives than shops do, and they matter more, because the customer asking for a refund is often deciding whether to stay:

- **Account credit** that comes off their next invoices, for a goodwill gesture after an outage.
- **A coupon**, but only one from a list you've approved. The AI should never be able to invent a discount.
- **Skip or pause** instead of cancel, for a customer who needs a break, not a refund.

With Stripe connected, Pengo can **refund a payment** in full or in part back to the card, **add account credit**, **apply a coupon** only from your approved list (Pengo can never make one up), and skip, pause or cancel at the end of the paid period. All of them are for verified customers only, and guardrails work the same way: the amount comes from the invoice itself. See [Stripe actions](/help/actions/stripe-actions/) and the [Stripe integration](/integrations/stripe/).

## Map your policy to the settings

Once the policy is written, each line should land somewhere concrete. If a line has nowhere to go, it's either a job for your help content or a reason to hand the conversation to a person.

_Figure: Each line of an illustrative written refund policy, and where it lives in Pengo. The amounts match the refund router above._

The one line you can't loosen is identity. Refunds are for verified customers only, built into the action, so a busy afternoon can't talk anyone out of it.

## Review the Activity tab weekly and tune your limits

Your first limits are guesses. Good guesses, if you based them on what your team already does, but guesses. The way to improve them is to look at what actually happened, once a week, for fifteen minutes.

  ### List last week's refunds by route
  How many ran on Auto, how many asked first, how many were rejected, and how many failed. In Pengo, the **Activity** tab in **Intelligence › Actions** lists every action Pengo ran or requested and how each one turned out.
  
  ### Look at what your team approved without changes
  If nearly every request just above your auto limit was approved as it was, the limit may be too low. That's time your managers spent saying yes to things they'd always say yes to.
  
  ### Look at what was rejected, and why
  Rejections are your policy's blind spots. A cluster of rejected "lost parcel" claims that tracking says arrived is a rule waiting to be written.
  
  ### Check the failures
  A refund that was approved but didn't go through usually means a disconnected integration or a missing permission in your store. See [A Pengo Action failed or didn't run](/help/troubleshooting/action-failed/).
  
  ### Move one limit, one step
  Raise the auto limit or the cap a little, or lower it if something went wrong. Then leave it for at least a week before you judge.
  

Treat the cap and the limit as separate dials. If small refunds are piling up for approval on busy days, the cap is too low. If managers are approving lots of mid-sized refunds without a second thought, the limit is.

## Mistakes we see most often

- **Automating before writing the policy.** The automation then encodes whatever the first person to set it up happened to believe.
- **Reading the amount from the chat.** The single fastest way to turn an auto limit into an invitation.
- **"Done!" before it's done.** A customer told their refund is on its way will hold you to it, whether or not anyone approved it.
- **One approver for everything.** If every refund needs the owner, refunds wait for the owner's holiday.
- **Never moving the limits.** A limit set on day one and never reviewed is either too tight or too loose, and nobody has checked which.

## Your refund workflow checklist

**Refund approval workflow checklist**

- [ ] Write the refund policy down
- [ ] Update your returns and refunds article to match it
- [ ] Set the auto limit low to start
- [ ] Cap automatic refunds a day
- [ ] Choose who approves above the limit
- [ ] Give teammates their own limit, with a second approver above it
- [ ] Make sure amounts come from the order or invoice, never the chat
- [ ] Require verified customers for refunds
- [ ] Agree a fixed list of refund reasons
- [ ] Check the waiting message and the three follow-ups
- [ ] Set where each approver gets requests
- [ ] Decide the alternatives
- [ ] Book a 15-minute weekly review of the Activity tab

A refund workflow built this way doesn't make your team more generous or more strict. It makes them consistent, and it gives their time back for the refunds that really need a decision. If you run a store, the companion playbook on [Shopify customer service automation](/blog/shopify-customer-service-automation/) covers the rest of the inbox, from "where's my order?" to address changes and cancellations, and [live chat for ecommerce](/blog/live-chat-for-ecommerce/) covers the chats that happen before anyone buys. To see every action and its default mode, start on the [Pengo Actions page](/actions/).

## Questions people ask

### What is a refund approval workflow?

It's the set of rules that decides who can approve a refund, up to what amount, and what happens while it waits. A good one lets clear-cut small refunds happen immediately and sends larger or unusual ones to a person with the authority to decide, with every decision recorded.

### Should customer service refunds be automated?

The small, clear-cut ones, yes. A missing £8.00 pair of socks doesn't need a manager, and making the customer wait a day for it costs more goodwill than the refund itself. Keep a person in the loop above an amount you're comfortable with, cap the daily total, and only refund customers whose identity is confirmed.

### What refund amount should I let an AI approve on its own?

Start low: around the price of one of your cheaper items is a sensible first limit. Then look at a few weeks of refunds your team approved. If nearly every request below a higher amount was approved without changes, raise the limit one step and watch for another few weeks.

### How do I stop customers abusing automatic refunds?

Take the amount from the order rather than from what the customer types, refund only to the original payment, require a verified customer, and cap how many refunds run on their own each day so a bad day can't drain your account. Review refunds weekly and look for repeat requests from the same customer.

### Who should approve a refund?

Low amounts can run on their own. Above that, someone with authority over money, usually a manager or the owner, should decide. When a teammate issues a refund above their own limit, a second person should approve it, the same four-eyes rule finance teams use.

### What should the customer see while a refund is waiting for approval?

An honest message that the request is with the team, never a claim that it's done. Once it's decided, they should get a follow-up in the same conversation: a confirmation if it went through, or a note that a teammate will be in touch if it was rejected or couldn't be completed.
