AI customer service for Shopify brands: automate where-is-my-order, keep refunds with a person

Where ticket time goes

Ticket volume follows orders. Gorgias, a helpdesk built for ecommerce, publishes benchmarks from its own platform data. It finds that electronics brands get about 46 support tickets for every 100 orders and food and beverage brands about 20. The median first response time at the $10 million GMV band is 6.3 hours across industries, from 1.6 hours in hardware to 8.8 hours in apparel. These figures come from a vendor's own customers, so read them as a rough guide.

The biggest single question is where is my order, often shortened to WISMO. Gorgias data puts it at 18% of incoming requests on average. Add delivery problems, address changes, cancellations and return status, and you have a group of questions whose answer already sits in your order and tracking data. Someone still opens several systems to find it, then writes much the same reply again.

Returns add policy work on top. The National Retail Federation and Happy Returns estimated that 19.3% of online sales would be returned in 2025, and found that 9% of all returns are fraudulent. The retailer side of that survey covered US merchants with more than $500 million in revenue, so it describes the industry more than a 60-person brand. For a smaller brand the costly returns are the exceptions, the ones a person has to decide.

What helpdesk AI already does

Before building anything, use what you already pay for. Gorgias offers its AI Agent on every plan and charges only when it resolves a conversation. Shopify's Sidekick can manage orders and edit products from the admin, and it presents changes for your review before applying them. Shopify's help pages add a fair warning: AI outputs can contain errors, so review anything a tool creates or changes.

Adoption is still early. Gorgias says only about 1 in 5 ecommerce brands has deployed AI in customer-facing support. For in-policy questions that a bot can answer from one system, these tools are a sensible first step, and some brands will get enough hours back from them alone.

Check what your plan already includes before you compare anything else. Gorgias prices its helpdesk by ticket volume, from 50 to 5,000 tickets a month, and says it never prices per agent. A brand that adds a custom agent on top keeps paying for the helpdesk, so the build has to earn its fee on the tickets the bot hands off.

The handoff gap

Trouble starts when the bot runs out of information or authority. In Gorgias data from October 2025 to April 2026, 55% of AI-touched tickets ended in a handoff to a person. The median shopper then waited 10 hours, and 33% of handed-off tickets were abandoned without a human reply. Handoffs from email took 32 hours to resolve and handoffs from chat 8 hours. Contact forms were the slowest channel in that data, with handoffs taking 36 hours and 42% abandoned, so if your store's contact form feeds the same queue as email, look there first.

Those handed-off tickets are where an agent built for your systems helps. A carrier exception needs the 3PL's records next to the carrier's scans. A damage claim needs the photos and the policy clause in one place. When the person who picks up the ticket finds a drafted reply with that evidence attached, handling time can drop. How long the shopper waits still depends on staffing and your escalation rules.

What to automate and what a person approves

A practical line: the agent may answer anything it can prove from your systems, and a person approves anything that moves money or bends a policy. Applied to a typical inbox, the work splits like this.

Returns exceptions need a file and a decision

Shopify lets brands set return and cancellation rules that decide when customers can request returns and how return fees apply. Return apps such as Loop go further, with policy rules that can differ by region or by product. Loop describes its fraud model as alerting your team to make the call. Both designs leave the exceptions to a person, and they should.

Fraud is one reason. In Loop's own data, the share of US return value flagged as high risk rose from 3.5% to 7.7% during peak season. Loop sells returns software, so read that as one vendor's view of its own merchants.

The agent's work is the file. A manager deciding a late return from a repeat customer needs the order history, the policy clause, the photos, the carrier evidence and the customer's past returns on one screen, with a suggested outcome. The manager then chooses between a refund, an exchange, store credit or a denial. With the file ready, the decision itself goes quickly.

Controls, including card data

Card numbers come first. On Shopify, card payments run through the checkout and the payment processor, so the agent can work from order references rather than card details. The gap is the support channel. Shoppers sometimes type card numbers into chat or email, and those messages land in your helpdesk. Remove card data from everything the agent reads or logs, including attachments, and test that control before launch. Hiding digits on screen is different from removing them from the data. Have any system that receives the original message assessed for PCI scope, and give the agent no payment access beyond the refund action a person approves.

The PCI Security Standards Council published AI principles in September 2025. They say AI systems should not be trusted with unprotected sensitive data, including unsecured account data, and that a person should stay in the loop and be held responsible for what the AI does, with its actions logged and monitored. The Council's September 2026 information supplement adds that AI should be treated like any other technology when scoping PCI requirements.

The rest are the controls of any approval queue:

• Every draft cites the order data and the policy line it used.
• Anything that moves money goes to an approver until you release a category in writing.
• A lead reviews a sample of sent replies every week and logs what was corrected.
• Any chat bot that talks to shoppers on its own says it is AI at the start of the conversation.

Worked example: a Shopify apparel brand

This is an illustrative example with assumed numbers, not a client result. Take a 60-person apparel brand with 12,500 orders a month. Assume 25 tickets per 100 orders, inside the range Gorgias publishes by vertical. That gives about 3,125 tickets a month, or about 720 a week at 4.33 weeks to a month. At 5 minutes each, the team spends 3,600 minutes, or 60 hours a week.

Now assume 45% of tickets, 324 a week, are routine: order status, delivered but not received, address changes and return status. With the agent drafting each reply from live data, a person spends about 1 minute approving it, 324 minutes in all. The other 396 take the full 5 minutes, 1,980 minutes. Total time falls to 2,304 minutes, or 38.4 hours, which gives back 21.6 staff hours a week.

Exceptions sit inside those 720 tickets, so count only the extra work. Assume 60 of them a week are refund and return exceptions that also need 12 minutes of a manager's research beyond the 5 minutes already counted. That extra work is 12 hours today. With the file prepared, assume it falls to 3 minutes each, or 3 hours. Together the two lines give back 30.6 hours a week.

The routine share is the number to check. Gorgias puts WISMO alone at 18% of requests, so the 45% here assumes the other routine types together add more than WISMO does. Count a real week before you trust any share, and remember that review time counts. If you later release order status replies without review, the hours fall further, but only after a written decision.

Starting small with customer service

Export one ordinary week of tickets from your helpdesk and tag each by type, with a rough handling time. Mark which ones your helpdesk bot already closes and which it hands off. The handed-off group is where a custom build earns its fee.

Our five-day AI Workflow Sprint starts from that export. You get a prototype that drafts replies and exception files from copies of your real tickets, plus a fixed quote for the build based on the staff hours it finds. The sprint costs $2,500 and the fee comes off the build. Full refund if we do not find 10 staff hours a week.

Common questions