US shoppers sent back $706 billion of merchandise, according to Appriss Retail’s February 2026 benchmark report. Behind each of those returns is a customer waiting for their money, and waiting is what they like least: in a January 2026 survey by TodayPay, a refund-payments company, 44% of US shoppers said they would rather visit the dentist than wait seven days for a refund.
While they wait, they write to you. Each of those messages is a support ticket, meaning a customer message your team has to read and answer, and an AI order support agent is software that takes the routine ones off your team by reading the real order, replying to the customer, and, in the right cases, issuing the refund itself.
Not every customer writes to you, though. Javelin Strategy & Research’s 2026 study for Mastercard found that in three out of four card disputes, the customer never contacted the merchant first. A refund that goes quiet can therefore end up with the customer’s bank, which can take the money back from you as a chargeback.
How big is this problem in your own store? You can find out in about ten minutes. Open Shopify, look at your last 100 returns, and count how many messages each customer sent about them. Most stores never do this count, because reports show returns as a single percentage while the messages around them sit scattered across email, chat, and social media. A typical customer asks how to return the item, then asks for a label, then checks whether you have received the parcel, and finally asks where the refund is.
This guide shows how many of those tickets automation can remove, what that is worth in dollars, and how to do it while keeping CSAT, your customer satisfaction score, intact.
What is your returns ticket load actually made of?
Your ticket load is the number of customer messages your team has to answer. For returns, you can work it out in four steps: how many returns you get, how many messages each return creates, how long each message takes, and what an hour of your team's time costs. The table below does this for a hypothetical example store with 5,000 orders a month. Anything that is my assumption is labelled, so you can replace it with your own number.
| Number | Where it comes from | Result | |
|---|---|---|---|
| Orders per month | 5,000, at $80 each | Online store | 5,000 orders |
| Share that comes back | 19.3% | NRF and Happy Returns, October 2025 | 965 returns |
| Messages per return | 2 (can range from 1 to 3) | Industry estimate, 2026 | 1,930 tickets |
| Minutes per message | 6 | My assumption, use your own | 193 hours |
| Cost of one hour of support time | $21.53 | US median pay for customer service reps, 2025 wage data | $4,155 a month, about $49,900 a year |
In this example, returns are costing this store about $4,155 a month in support time. The real-world figure will probably be much higher, because the table counts only the hourly wage and leaves out benefits, software and training. In reality, one industry estimate puts the full cost of a returns message at $3 to $7, while our table works out to $2.15.
Whatever the exact cost, the more useful question is where those 1,930 tickets come from. Mostly they come from customers waiting. After a return, the customer hears nothing until the money arrives, so they write in to ask. You may already know this pattern from "where is my order?" tickets, which we looked at in The WISMO Tax. "Where is my refund?" is the same kind of ticket, just one step later in the customer's journey. Many returns automation ecommerce tools help only at the very start, when the customer asks to return something, so these waiting tickets carry on as before. To remove them, we first need to see exactly which ones they are.
Our take: Most stores undercount return tickets because they sort them by topic, such as "shipping" or "refunds", instead of by return. After the 100-return check, keep counting for a month, because messages per return are the number that tells you whether automation is a nice extra or a real saving.
Which of those tickets can an AI order support agent remove?
A return creates five different kinds of tickets; each has its own cause. The table lists them, and the last column says how automation removes each one. The three words in that column (prevent, resolve, approve) are explained in the next section.
| Customer message | Why they send it | What automation does | How it goes away |
|---|---|---|---|
| "How do I return this?" | The policy is hard to find | Answers from your live return rules and starts the return | Resolve |
| "Can you send me a label?" | Someone has to make it by hand | Checks the order, then makes the label | Resolve |
| "Did you get my parcel?" | Nobody says anything after delivery | Confirms receipt before the customer asks | Prevent |
| "Where is my refund?" | Refund sent, but the bank has not shown it yet | Gives the exact date in the first message | Prevent |
| Exchange or dispute | No exchange was offered, or the refund felt slow | Offers an exchange first, passes disputes to a person | Approve |
The two waiting messages ("did you get my parcel?" and "where is my refund?") are the easiest to remove and the least noticed. They happen because of how refunds work. When you refund a customer, Shopify shows the refund as pending for up to two business days, and then the customer's bank can take up to ten more business days to show it, according to Shopify's Help Center as summarised in a 2026 guide. For up to two weeks, the customer sees nothing on their statement, so they write to you.
Customers do not wait quietly either. Some go straight to their bank instead of your inbox: Javelin Strategy & Research’s 2026 study for Mastercard found that in three out of four card disputes, the customer never contacted the merchant first.
Stopping these waiting tickets is the job of the software we introduced at the start. An agent like that looks at the real order in your store, checks it against your return rules, and then takes the next step by itself. It can send the return label, tell the customer the exact date their refund will arrive, issue the refund, or pass the case to your team with the full story attached. The table compares it with tools you may already use. A returns portal is an app that gives customers a page to request a return, and a helpdesk is the inbox your team answers customers in.
| Tasks | Returns portal | FAQ chatbot | Saved replies in your helpdesk | Agent that acts on the order |
|---|---|---|---|---|
| Looks at the real order | Only for the return request | No | Your team looks it up | Yes |
| Answers "where is my refund?" | Shows a status page | Gives a generic reply | Your team types it | Yes, with the date |
| Issues the refund | Often, by a rule | No | Your team clicks it | Yes, up to a limit you set |
| Knows when to hand over | Follows rules only | No | Your team decides | Passes to a person with the full story |
Ephanti's Order Support Agent is one working example of this. When a customer emails about a duplicate order, it finds the order in Shopify, checks the customer is eligible, issues the refund through your payment processor, makes the return label, and logs the ticket, all without a person. You can see this flow on the E-Commerce solution page.
Prevent, resolve or approve: three ways to take a ticket off the queue
Each of the five tickets above goes away in one of three ways, and it helps to know which is which. Stores that mix them up either hand over too much control or end up paying for what is really a smarter-looking FAQ.
Prevent means the ticket is never created. The refund date goes out in your first message, and a "we received your parcel" note goes out as soon as the carrier scans it, so the customer has no reason to write.
Resolve means the ticket is fully answered in the same chat. The agent checks the order against your return rules, makes the label, and offers an exchange before a refund.
Approve means money is involved. The agent can issue refunds up to a dollar limit you set, and anything above that limit waits for a person.
Some requests should always go to a person: returns outside your time window, customers who ask for refunds again and again, high-value orders, and any order where the customer has already opened a dispute with their bank.
Why set a dollar limit at all? Because the usual problem is not criminals; it is being too generous. Appriss Retail's February 2026 report found that customers bending the rules (policy abuse) caused 12% of returns-related losses, while outright fraud caused only 2%. A dollar limit is the cheapest way to control the bigger risk.
The dispute check matters just as much. One Shopify merchant refunded £284 for a late order, and soon afterwards the customer's bank also raised a chargeback for the same order (a chargeback is when the bank takes the money back from the merchant). Shopify cancelled the refund, and the merchant lost the money twice. The lesson is that any refund automation Shopify setup you can rely on checks for open disputes before it moves money, which is why an AI order support agent needs to see the payment side as well as the order. It also gives you a useful question for any vendor demo: ask which of these three jobs each feature does. That is how returns automation ecommerce buyers tell real automation from a nicer-looking FAQ.
Our take: Start with a low refund limit and raise it when the log shows it is safe, not when a demo makes it look easy. A refund cannot be taken back, so the limit should only move on evidence.
How many tickets will you really cut, and what is it worth?
How many tickets you cut depends on your store; here are three examples using the 1,930 tickets from earlier. These are ranges, not promises, and 70% is about the highest figure that industry estimates suggest for routine tickets.
| If automation removes tickets | Tickets gone each month | Support time saved each month | Saved each year |
|---|---|---|---|
| 30% | 579 | $1,247 | $14,959 |
| 50% | 965 | $2,078 | $24,932 |
| 70% | 1,351 | $2,909 | $34,904 |
Those savings come with tool costs. To remove half your tickets, a store usually pairs its helpdesk with an AI add-on and a returns portal, and the table below prices that set at public list prices. Vendor names appear only in the source links, and the table leaves out team seats, return shipping labels, payment fees, setup costs, and taxes.
| Tool | List price | Monthly cost at this store |
|---|---|---|
| Helpdesk (priced by ticket volume) | $300 a month, 2,000 tickets included (July 2026) | $300 |
| AI add-on (charged per conversation it closes) | $0.90 each, for 965 conversations | $868.50 |
| Returns portal | $297 a month, 300 returns included, $0.90 for each one after (September 2026), so 665 extra returns | $895.50 |
| Total | $2,064 a month, $24,768 a year |
Added together, the tools cost almost exactly what they save: $2,064 against $2,078 of support time. That is less bad than it sounds, because if you already pay for the helpdesk and the returns portal, only the AI add-on ($868.50) is new money.
The AI add-on has one more catch. The helpdesk in our table is Gorgias, and its own pricing FAQ says every conversation its AI closes also counts as a normal ticket, as reported in 2026, so you pay for the same answer twice. If your store also gets 1,000 other tickets a month, you would reach 2,930 against the 2,000 included, and the 930 extra cost about $335 a month at $36 per 100. A cost that never appears on any invoice is the chargeback we met earlier, which happens when a customer’s bank takes the money back because the customer heard nothing about their refund.
Ephanti charges by the number of customer interactions it handles, with any number of team members and no per-user fees, and its site says typical customers replace four to six separate tools within 60 days. You can run your own figures in the ROI calculator, and the comparison page sets Ephanti against that helpdesk on AI architecture, integration depth, deployment time, and pricing transparency. One honest note: if you get only a few returns a month, Shopify's built-in self-serve returns may be enough, with the one limit that customers cannot request exchanges there.
Whichever refund automation Shopify tool you consider, judge it by the two tables above: what it costs against how many tickets it removes. Saving money is only half the aim, though. The other half is keeping customers happy, which is the next section.
Will CSAT drop as the ticket count falls?
CSAT is your customer satisfaction score, usually collected from a short survey after a chat. Fewer tickets only counts as a win if the customers you still hear from are happy, and four things quietly lower CSAT when returns are automated. Good post-purchase support AI, meaning AI that handles customer questions after the order is placed, avoids all four.
A bot that can only say no. A refusal with no alternative feels like a wall, so the agent should offer an exchange or store credit, and pass to a person if the customer pushes back.
A loop with no way out. Gartner's August 2026 survey of 3,566 customers found that 87% say companies using GenAI for service must let them reach a human.
"Processing" with no date. A status without a date just creates the next message, so every update should say when the money arrives.
A handover that starts the story again. When a person takes over, they should see everything already said, because customers hate repeating themselves.
Trust is a real issue here. Appriss Retail's 2026 survey of over 1,000 consumers found that 71% trust human staff more than AI to approve returns, and that 80% want to know how AI makes return decisions. The upside is just as clear, because TodayPay’s survey from earlier found that 92% of shoppers would buy again from a store that refunds faster.
That repeat purchase is worth protecting. Ephanti's founder has written about why the second purchase is where e-commerce is won, noting that 70 to 77% of customers never come back after their first order. A smooth return is one of the few chances you get to change that.
Our take: Watch CSAT on return tickets more closely than ticket count. If a customer gets stuck arguing with a loop, the ticket has only moved, and it comes back later through the bank.
How do you roll it out without a bad refund?
Keeping customers happy also means getting every refund right, and once software is involved, that becomes a legal question as well as a technical one. In May 2026, the Higher Regional Court of Hamm in Germany ruled that what a company's chatbot says counts as the company's own statement, even though the bot had been trained on correct information, and the ruling is not yet final. A 2026 PYMNTS roundup applies the same idea to the earlier Air Canada case. The short version is that if your bot approves a refund, you own it.
So whichever AI order support agent you shortlist, ask for these five protections:
It answers from your live return rules, not an old FAQ.
Refunds stop at a dollar limit you set.
It checks for open disputes before moving money.
When it hands over, the person sees the full chat and why it was passed on.
Every decision and override is logged, so you can review it later.
Ephanti is built along these lines. It passes a case to a person when a refund or dispute goes above a dollar limit you set, as described on the Omnichannel page, and the Trust Center says every agent decision, human override, and handover is recorded and searchable. The thinking behind this is in our write-up on accountable agent architecture.
Once those protections are in place, start small. Ephanti's site describes three steps: connect your systems in the first week, run one channel or queue alongside your current tools, and widen it as your confidence grows. Most E-Commerce teams go from contract to live within 30 days, and the E-Commerce Playbook covers the plays and common mistakes. When you pilot post-purchase support AI, judge it on these six numbers.
| What to track | What good looks like | Question for the vendor |
|---|---|---|
| Messages per return | Lower than your count from the 100-return check | Can I see it per return, not per topic? |
| Days to refund | The customer knows the date every time | Does every message state a date? |
| Repeat messages | The same issue is not raised again within 7 days | How do you spot repeats? |
| Handed to a person | A steady share, with reasons | Is the reason logged? |
| CSAT on return tickets | Same or better than before | Can I filter CSAT by return? |
| Chargebacks on returned orders | Not rising | Do you check for open disputes first? |
Our take: Judge the first 30 days on plain numbers: messages per return, refund dates kept, and no surprise refunds. If a vendor cannot show you the log behind each refund, you are taking on risk, not getting automation.
Run the audit again in 30 days
Go back to the 100 returns you counted at the start of this guide and count the messages again, the same way. If automation is working, that number drops without you hiring anyone. An AI order support agent is how you get there, but the proof is your own recount, not a vendor deck.
Ephanti's manifesto puts the idea this way: "A customer message doesn't have to end with a response. It can end with the order placed, the refund issued..." If you would like to see that on your own store, pick the request your team dreads most, the one that eats an afternoon and still ends in a follow-up. Ephanti's site invites you to "name the job nobody wants to do," and you can book a demo to start there.
FAQ
Can AI issue Shopify refunds without a human?
Yes, within limits you set. Ephanti's Order Support Agent finds the order in Shopify, checks eligibility, issues the refund through your payment processor, makes the return label and logs the ticket, and it passes anything outside its scope to a person with the full conversation attached.
Which return requests should never be automated?
Returns outside your time window, repeat refund requests, high-value orders, anything with an open dispute, and customers who are clearly getting upset should always go to a person.
Returns automation ecommerce: is a returns portal enough?
A portal handles the request step well, but it leaves the waiting messages ("did you get it?" and "where is my refund?") untouched, and those make up much of your ticket load.
Can refund automation Shopify merchants set up reduce chargebacks?
It can, when refunds go out quickly, and every message states a date, because customers who hear nothing often go to their bank. It also has to check for open disputes first; otherwise, a chargeback can cancel the refund.
Will post-purchase support AI hurt our CSAT?
Not if it keeps a clear route to a human and hands over the full story. Ephanti's agents can be overridden at any point, and a person can take over a conversation midway, so someone is always in control.
Do I need a returns app if I am adding an AI order support agent?
Not always. A store with few returns may be fine with Shopify's built-in self-serve returns, while a busier store usually benefits from an agent that also covers the waiting messages.