Forty emails since yesterday evening. The first is “hi, has my order shipped yet?” The second is a returns question. The fourth asks whether you ship to Ireland, which you do, and which the shipping page says in a place nobody reads. By the time you reach the one that actually needed you, from the customer whose payment failed twice and who is now writing in capitals, it’s an hour later.

Here’s the reframe that matters if you’re drowning in repetitive customer support emails. You don’t have a support volume problem, you have a repeated-question problem, and the two have completely different fixes. Volume problems get solved with more hands. Repetition problems get solved by removing the reason the question gets asked at all.

The fix has an order to it, and the order is not optional: eliminate, deflect, template, automate, delegate. Each step shrinks the input to the next. Run them backwards, which is what buying a chatbot in week one amounts to, and you automate a mess instead of deleting it. I’ve watched founders spend a month tuning an AI agent to answer a question that a one-line change to a product page would have stopped anyone asking.

Before anything else: count your repetitive support emails

You cannot fix repetition you haven’t measured, and your memory of what fills the inbox is wrong. It over-weights the emails that annoyed you and under-weights the boring ones you answer on autopilot.

For fourteen days, tag every inbound. Not a taxonomy workshop, just a flat list that grows as you go. Gmail labels work; so does Shopify Inbox, Gorgias, Help Scout, Zendesk or Front. Twelve to twenty tags is the useful range. If you end up with sixty, you’re tagging root causes instead of questions.

Then build one table: tag, count, your honest average handling time, and the two multiplied. That last column is the only ranking that matters. A tag with 90 tickets at 2 minutes costs three hours a fortnight. A tag with 20 tickets at 11 minutes costs nearly four. The second feels smaller and isn’t. Add a fifth column that’s just yes or no: could this question have been prevented? That’s where the real money is, and almost nobody fills it in.

Step 1, eliminate: the questions that are product defects wearing a costume

A support email is a bug report about your website that the customer was kind enough to file in prose. Four of them show up in nearly every ecommerce and small SaaS inbox.

“Do you ship to [country]?” A shipping page that lists carriers instead of countries, or a selector that only tells the truth at checkout. Fix: name the countries on the product page near the buy button, with a country-specific delivery estimate. If someone has to email to find out whether they can buy from you, most don’t email. They leave.

“What’s your returns policy?” Baymard Institute, whose findings come from over a thousand moderated usability sessions, reports that 20% of ecommerce sites have no direct footer link to returns and shipping information, burying it under a generic “FAQ” or “Help” instead. Their testing also finds 13% of users abandoning a cart over an unsatisfactory return policy. Two costs from one missing link: the tickets you answer and the orders that never happened. Fix: standalone footer links labelled “Return Policy” and “Shipping Info”, plus the return window in plain words on the product page and in the order confirmation email.

“What size am I?” In apparel this drives both tickets and returns, and a size chart that opens as a blurry JPEG in a modal is not a size chart. Fix: real measurements in both units, a garment-versus-body distinction, and fit guidance written by whoever handles the returns.

“I want to change my plan, cancel, or add a seat.” For SaaS that’s a self-service gap, not a documentation gap. Every one of those emails is a customer telling you your billing screen is missing a button.

The test for this category: after you make the change, does the tag go to zero on its own? Not lower. Zero. That’s what elimination means, and it’s the only step with a permanent payoff.

Step 2, deflect: killing “where is my order” on Shopify

Order status is the biggest repeated question in ecommerce, and I’d treat the specific percentages you see quoted with suspicion, because they come from helpdesk vendors selling the cure and I can’t find an independent benchmark behind them. The mechanism doesn’t need a statistic. Every order opens a window of anxiety at checkout that closes at the doorstep, its length is set by your carrier rather than by you, and the customer has one way to resolve it if you haven’t given them another. That’s an information-delivery problem, and Shopify’s own furniture solves most of it.

• Turn on the whole shipment lifecycle. Shopify’s customer notifications include order confirmation, shipping confirmation, shipping update, out for delivery and delivered. Several are optional at the point you fulfil or update tracking, and merchants leave the later ones off because they worry about over-emailing. Wrong worry. A delivery notification removes the exact moment a customer would otherwise write to you.
• Know what the order status page actually does. With supported carriers it shows real-time updates and map-based tracking, with statuses including Confirmed, On its way, Out for delivery, Delivered and Attempted delivery. With unsupported carriers it shows only a tracking number linking out, which is worth knowing before you blame Shopify for a vague page.
• Know the access window. Per Shopify’s documentation, customers can reach their order status page from the confirmation email for three weeks without logging in, on the same browser. After that they need the order number plus the email or phone used at checkout, or a customer account. A slice of your “I can’t see my order” emails is that window quietly expiring on someone who changed device.
• Rewrite the shipping confirmation so it answers the next three questions. Shopify notification templates take Liquid variables and your branding, and most stores ship the default forever. Put tracking at the top, a realistic delivery range rather than a carrier-optimistic one, one sentence on what to do if tracking hasn’t moved in 48 hours, and the returns window. Every sentence there is a ticket you won’t receive.
• Get ahead of exceptions. Customs delays, weather, a pre-order slipping. A proactive email about a delay generates a fraction of the contacts that silence does.

Be honest about the ceiling, though. Gartner surveyed 5,728 customers in December 2023 and found that while 73% use self-service at some point, only 14% of issues were fully resolved there, and 43% of those who failed couldn’t find relevant content. Deflection is a titling and placement discipline, not a content-quantity one. Write the page in the customer’s words: “Where is my order”, not “Fulfilment and logistics information”.

Step 3, template: macros that don’t read like macros

Whatever survives steps one and two gets a saved reply. People hate macros because most are written as fill-in-the-blank corporate skeletons, and customers smell that instantly. Three rules fix it.

Answer in the first line. No “thank you for reaching out regarding your recent enquiry”. Start with “Yes, we ship to Ireland, and it’s usually 4 to 6 working days.” Pleasantries can follow the answer or not appear at all.

Build in one slot that forces a human touch. A bracketed spot near the top that can’t be left generic: their actual situation, their order number with what’s in it, the date their parcel cleared customs. One specific clause makes the whole message read as written by a person, because canned replies give themselves away through the absence of specifics, not through their structure.

End with the next question answered. If someone asks about a return, the next thing they’ll ask is who pays postage. Put it in. A macro that pre-empts the follow-up halves the thread, and thread length is what eats your day.

Keep them in the helpdesk rather than a doc, date them, and re-read the top ten every quarter. A macro describing a policy you changed in March is worse than no macro.

Step 4, automate: what AI support agents genuinely do well

My honest read: AI support agents are very good at retrieval and terrible at policy invention, and the gap between those two is where all the damage happens.

Good at: pulling an order status through an API and stating it; matching a question to the right help article and summarising it; multilingual first replies at 3am; tagging and routing so your morning queue arrives pre-sorted; drafting a reply for you to approve, which is the most underrated mode and where I’d start.

Bad at: anything where the right answer depends on a judgement you never wrote down. Refund edge cases. Whether this customer gets an exception. Anything about security or data handling. And crucially, saying “I don’t know”, because a language model’s default behaviour is to produce a plausible answer rather than stop.

That failure mode has been tested in a tribunal. In Moffatt v. Air Canada (2024 BCCRT 149), the airline’s website chatbot told a customer he could apply retroactively for a bereavement fare, contradicting the policy published elsewhere on the same site. Air Canada argued, among other things, that the chatbot was effectively responsible for its own statements. The British Columbia Civil Resolution Tribunal rejected that, held the company responsible for all information on its website whether it came from a static page or an interactive tool, and awarded roughly CAD $650 plus interest and fees. The money is trivial. The principle isn’t: a policy your bot invents is a policy you may have to honour.

Which leads to the configuration decision that matters more than vendor choice. Scope the agent to retrieval over your own help centre and order data, not to open generation. It answers only from documents you wrote, it links the source, it’s explicitly forbidden from stating refund amounts, discounts, delivery guarantees, roadmap commitments or security claims, and when confidence is low it hands over instead of improvising. Then read a sample of its conversations weekly for the first month, the way you’d review a new hire’s replies. Not a dashboard. The actual transcripts.

Designing the handoff so it doesn’t trap an angry customer

The worst outcome isn’t a wrong answer. It’s a loop: a frustrated customer, a bot cheerfully offering article suggestions, no visible exit. That’s how a refund request becomes a public review. Build the escape hatches deliberately. A visible “talk to a person” option on every turn. Automatic handoff after two consecutive failed attempts at the same question, with the transcript attached so nobody repeats themselves. Instant handoff on trigger words regardless of confidence: cancel, chargeback, lawyer, fraud, broken, allergic, wrong item, plus anything touching data or security. And when no human is available, the bot should say so with a real timeframe, which beats a cheerful non-answer.

Step 5, delegate: when the inbox breaks, and when a person beats more tooling

A shared Gmail or Outlook inbox is genuinely fine for one person. It stops being fine at three points, none of which is volume alone: when two people reply to the same email, when you can’t answer “how many returns questions did we get last month” without counting by hand, and when a customer’s history matters to the reply and you’re reconstructing it from your own inbox. That’s when you move to a real helpdesk.

Then the point everyone gets wrong in the other direction. Automation has a cost curve that turns upward, and past a certain complexity a part-time human is cheaper than the next increment of tooling. Low-volume, high-judgement, emotionally loaded question types should not be automated. They should be staffed.

Rough arithmetic: the US Bureau of Labor Statistics puts the median wage for customer service representatives at $21.53 an hour, or $44,770 a year, as of May 2025, before benefits and tooling. A part-time contractor or a dedicated remote agent in a lower-cost market comes in under that. If you’re clearing ten hours a week of tickets, you’re spending half a working day every day on labour with an established market rate, while the work only you can do waits. Delegate last, not first, because handing a repetitive mess to a person converts your problem into a recurring invoice. But delegate.

A worked example, including where it stops working

A Shopify apparel store, roughly 1,400 orders a month, founder answering everything. Two weeks of tagging produced 310 emails, around 620 a month.

Eliminate. International shipping named on the product page with per-country estimates, and a rewritten size guide with real garment measurements. The shipping tag effectively disappears; sizing drops by more than half over the following month once the guide is linked from the buy button.

Deflect. Out-for-delivery and delivered notifications switched on, the shipping confirmation rewritten to lead with tracking plus a realistic range plus a “hasn’t moved in 48 hours?” line, and a proactive email whenever a shipment stalls. Order-status contacts fall substantially but never to zero, because a real proportion of them are lost parcels that need a human.

Template, then automate narrowly. Six macros covering returns, exchanges, damaged items, address changes, pre-order timing and the residual status reply. Then an AI agent scoped to the help centre and the Shopify order API, answering status and policy questions only, forbidden from stating refund amounts, with a visible handoff on every turn.

What’s left is damaged-and-wrong-item, change-or-cancel, and the long tail. They share one thing: they need a decision, sometimes money, and usually an apology. That residue is roughly twelve hours a month here, and it’s the point at which you hire part-time instead of buying another tool. The queue didn’t get automated away. It got sorted into the part that shouldn’t exist, the part a machine can handle, and the part that was always a job.

One thing worth saying plainly: your ticket mix gets harder as this works. Automation absorbs the easy questions first, so the average difficulty of what reaches a human rises. Staff and train for that mix, not the one you have today.

If the middle of that ladder is the part you’d rather not build yourself, that’s the work AB7 Solutions does: automating repetitive inbox and order workflows, wiring retrieval-scoped AI agents over your own help content rather than open generation, and placing dedicated remote support people for the residue that needs a human. Send us two weeks of your tags and we’ll tell you which of the five steps you’re actually on. Call +1 321 341 7733 or email ab@ab7solutions.com or director@ab7solutions.com.

Questions founders ask next

Should I set up an autoresponder while I’m still answering everything myself? Yes, but make it useful rather than polite. State a response window you can actually hit, link the three highest-volume help articles by name, and tell order-status enquiries where to track. An autoresponder that only says “we’ve received your message” flatters your first-response metric and deflects nothing.

Is Shopify Inbox enough, or do I need Gorgias or Zendesk? Inbox handles chat well and is free, but it isn’t a ticketing system, and you’ll feel the absence of tagging, macros and reporting quickly. If you’re doing the two-week tagging exercise, a proper helpdesk pays for itself on that alone. Choose on tagging and macro quality, not on the AI feature list.

Will customers be annoyed that an AI answered them? Far less than they’ll be annoyed by waiting nine hours, provided the bot doesn’t pretend to be a person and there’s a visible route to one. Disclosure costs you nothing. Trapping someone costs you the customer.

Sources: Shopify Help Center, Understanding order status pages, Setting up customer notifications and Store notifications; Baymard Institute, Have Direct Links to Returns and Shipping Info (1,000+ moderated usability sessions); Gartner, Survey Finds Only 14% of Customer Service Issues Are Fully Resolved in Self-Service (5,728 customers, December 2023); Moffatt v. Air Canada, 2024 BCCRT 149, summarised by the American Bar Association; US Bureau of Labor Statistics, Occupational Outlook Handbook: Customer Service Representatives (May 2025).