An AI agent that answers “where is my order” for a dollar is still charging you a dollar for an email you never sent.
A founder doing $60,000 a month starts the day clearing the inbox before the kids are up. I hear some version of that morning from merchants and agency partners every few weeks. Most of what is in there is the same five questions: where is my order, how do I return this, can I change my address, does it come in another size, and something arrived broken. A vendor demo shows an AI agent answering all five in seconds, and it looks like the answer to the 6 a.m. inbox.
It might be. But the AI agent is the third decision, not the first. On episode 491 of the podcast I put it plainly: “an AI chat bot does not fix unclear policies.” A bot on your site is only as good as what is already on your pages. This guide owns AI in support and service for a Shopify store: what to automate, what it costs, how to hand off and how to measure it. The repeat purchase side of service lives in the customer retention pillar, and AI that sells on chat lives in Conversational Commerce.
An AI customer service agent reads a customer’s question, looks up your policies, products and the customer’s order, and either answers and closes the conversation or hands it to a person with the context attached. That is different from the older chatbot that matched keywords to canned replies, and different again from AI that only drafts a reply for a human to send.
On Shopify there are two kinds worth knowing about in November 2026. The first is free. Shopify Inbox now includes an Inbox agent that answers product questions, recommends products, looks up an order once the customer signs in, and adds items to the cart. It works from your catalog, policies, pages and Knowledge Base, only inside your online store’s chat widget, on the Basic plan or higher, and only with new customer accounts. You choose whether it takes every conversation, only the hours your team is offline, or none. Shopify shows customers a notice that they are chatting with AI, and you cannot remove it. Shopify’s help page also says you are responsible for the accuracy of what the agent tells customers.
The second kind lives in a helpdesk such as Gorgias, Gladly or Fin (the company formerly called Intercom). These agents work across email, chat and social, and the better ones take actions: change a shipping address, cancel an unfulfilled order, start a return. Their AI usage is billed per outcome, on top of the helpdesk subscription, which I cover below. If you are still choosing a helpdesk, Fastlane’s comparison of Gorgias and its alternatives walks through the field.
The gap between the two is the action. Inbox’s agent can tell a customer where an order is. It does not edit the order or process the return. If most of your tickets need something changed, the free agent covers the questions and a person or a helpdesk agent still covers the changes.
Many of the tickets an AI agent would resolve on a Shopify store start with missing information: a shipping email without a tracking link, a returns policy buried in the footer, a size chart that does not answer the question. How many is different for every store, and your own tagged tickets will tell you. Paying an agent to answer those is paying to patch a hole you could close without a new software fee.
Austin Reece of Gladly listed the easy wins on episode 386: “Things like, where’s my order? Can I do a return? Can I do an exchange?” He is right that those are the conversations to automate. I’d argue the first move is to ask why each one exists. A “where is my order” ticket often means the customer never got, or never found, a tracking link, though some come from real carrier delays or a customer who just wants reassurance. A “how do I return this” ticket often means the policy is unclear or hard to reach. When a delay comes from the warehouse, the fix is upstream, and the fulfillment pillar covers where those delays start.
On episode 491 I suggested pulling the questions for your product page FAQs straight from your support inbox and your reviews, because customers have already told you what confuses them. The same inbox tells you which policy page to rewrite. On that episode I also pointed merchants to Shopify’s Knowledge Base app for editing the answers AI shopping assistants give about your store. Shopify has since moved that job: the knowledge base now lives in your admin under Sales channels > Agentic, the app can no longer be installed, and it is being retired on January 15, 2027. The Inbox agent reads that knowledge base too. Its answers do not show on your storefront, so the FAQ still has to live on the page for shoppers. How those outside assistants decide what to say about a store is covered in the generative engine optimization pillar.
There is also a legal reason to fix the source first. In Moffatt v. Air Canada, a British Columbia tribunal ruled in February 2024 that the airline was responsible for wrong refund advice its website chatbot gave, and ordered it to pay $650.88 in damages plus interest and fees. The tribunal called it obvious that a company is responsible for all the information on its website, whether it comes from a static page or a chatbot. An agent that improvises around a vague returns policy is making promises in your name.
The fastest way to decide what an AI agent should touch is to tag one month of tickets by reason and put every reason in one of three piles. Delete is anything a page, email or setting could have answered before the customer asked. Automate is anything with one right answer that lives in your order or product data. Keep human is anything that involves money leaving the business outside a written rule, a judgment call, or a customer who is already upset. A cancellation refund your policy already spells out can be automated; an exception to that policy cannot.
Philippe Lehoux of Missive described the risk in the third pile on episode 406: “You can do a demo and it’s extremely impressive, but on the long tail of customer questions and tickets, it often creates a lot of mistakes.” The demo shows the first pile and the second. Your inbox also holds the broken mug, the wrong size sent twice and the customer who has written three times, and those are where a confident wrong answer costs you a customer.
Here is how the sort plays out on a typical mix:
The same split shows up in how small teams staff support. Fastlane’s guide to building an efficient ecommerce team makes the matching call on hiring: automate the questions with one right answer and hire people for the work that needs judgment.
Helpdesk AI agents now charge for outcomes, roughly $1 per resolved conversation at list price as of November 2026. All prices here are in US dollars. On Gorgias’s pricing page, monthly plans bundle the AI Agent at $30 a month for 30 automated interactions on Starter and Basic, $190 for 190 on Pro and $530 for 530 on Advanced, on top of helpdesk fees of $10 to $900 a month by ticket volume; annual billing is lower. Fin lists $0.99 per outcome with a 50 outcome monthly minimum when it runs on another helpdesk. It counts a resolution when no further help is requested after its last answer, and it also bills a configured procedure that ends in a handoff to your team. Shopify’s Inbox agent has no per resolution fee.
Read the fine print on “resolution.” A customer who gives up and closes the chat can count as resolved. That matters for the measurement section below.
Now the worked example. This is a labelled composite, built from the patterns above and not from any one store, and every number is illustrative. A pet supplies brand does about $80,000 a month at a $50 average order, so 1,600 orders and 400 tickets a month. The mix: 160 where is my order, 60 return how-to, 60 product questions, 40 address changes or cancellations, 40 damaged or missing items, 40 complaints and everything else.
The delete pass comes first and needs no new software, only a few hours of someone’s time. Assume a tracking link and order status link in every shipping email, a delay email when a carrier stalls, and a plain returns page linked from every order email cut where is my order tickets in half and return how-to tickets by half. That removes 110 tickets, leaving 290.
Then the automate pile: 80 tracking questions, 30 return questions, 60 product questions and 40 address changes, 210 in all. Product questions go to the free Inbox agent on chat. Assume a helpdesk agent fully resolves 60% of the rest, about 90 conversations. At $1 to $1.50 each that is $90 to $135 a month in AI fees, on top of a helpdesk plan the store already pays for. If each of those took a person five minutes at $22 an hour, the agent saves about 7.5 hours, or roughly $165. On a store that already runs the helpdesk, the net gain is about $30 to $75 a month. On a store buying the helpdesk to get the agent, the plan fee can erase it. The delete pass removed 110 tickets, about 9 hours of work worth roughly $200, with no new fee.
What I would remove: the separate pre-sale chatbot app, if the Inbox agent covers product questions, and any upgrade to a bigger AI bundle until the delete pass has run for a month. What stays human: the 40 damage claims and 40 complaints, plus whatever the agent hands back.
Customers accept an AI agent when they can still reach a person, and they punish it when they cannot. A Gartner survey of 5,728 customers (December 2023, published July 2024) found 64% would prefer companies did not use AI for customer service, and their top concern was that it would get harder to reach a person.
Klarna learned this at scale. Its OpenAI powered assistant was handling about two thirds of service chats, the equivalent of 700 full time agents, and in May 2025 its CEO told Bloomberg the company was hiring humans again because cost had been “a too predominant evaluation factor” and quality suffered, as EMARKETER reported. Klarna is not a Shopify brand, but the lesson travels: an agent judged only on cost per ticket drifts toward closing conversations instead of solving them.
Austin Reece’s framing on episode 386 is the one I’d hold onto. Automate the easy conversations so your team can handle “the things that only agents can do.” In practice that means three settings. Offer a person in the first reply, not after three failed tries. Carry the conversation and the order context to that person so the customer never repeats themselves; Shopify’s Inbox agent does this when staff are available, and shares your store email when they are not. And route repeat contacts and angry language to a person by rule. On the same Missive episode, Philippe described using AI inside the rule engine for exactly this kind of triage, surfacing the urgent ticket instead of answering it.
Fastlane is a media company, not a store, but our own inbox runs on Gorgias, and the AI work that saves me the most time there is not answering anyone. It sorts the inbox, closes the noise and puts the conversations about money or a podcast guest at the top for a person.
Measure an AI agent as a pilot with a fixed end date: 60 to 90 days, cost per resolved ticket, and customer satisfaction on AI handled conversations compared with human ones. On episode 463 I gave voice agents the same test: measure the cost per resolved ticket and your CSAT score, then decide “is it the scale or is it a kill?” The test works for chat and email agents too.
Add two numbers the vendor dashboard will not lead with. The first is repeat contact: the share of “resolved” customers who write back within seven days about the same order. Because vendors can count a quiet customer as resolved, repeat contact is the check on the resolution rate. The second is a human read of the transcripts. Ibby Syed of Cotera described on episode 435 how AI changed one team’s quality review: “instead of scoring just 3% or 5% of interactions, it scores 100%.” A person then reads the low scorers. Whether you use AI to score or read a sample by hand, someone has to read what the agent said.
Set the kill line before you start. If AI handled conversations score below your human ones, or repeat contact climbs, narrow the agent to fewer ticket reasons before you add more. If cost per resolved ticket is above what a person costs for the same ticket, turn it off.
Lower volume stores should run Shopify Inbox’s free tools and fix their pages, and higher volume stores with order edits and returns should add a helpdesk agent priced per resolution. The most common mistake at every stage is buying the next stage’s tool before the current stage’s free fixes are done. These revenue bands are rules of thumb, not cutoffs; ticket volume and how many people answer them matter more than revenue.
If you are starting out, you are the support team, and that is an advantage: every ticket tells you what to fix on the site. Turn on Inbox’s instant answers, which Shopify Magic can draft from your policies, and spend the money you would have spent on an agent on clearer product pages. The product page facts that AI tools read are the same facts customers email you about.
If you are growing, the Inbox agent set to answer only when your team is offline is the cheapest test there is. It covers the evening and weekend shopping questions and leaves the rest for the morning. Set your team’s availability hours first, because Shopify treats staff as always available by default and the agent will not pick up anything in this mode until you do. Check which Shopify plan you are on first, since Inbox needs Basic or higher.
If you are scaling, this is where a helpdesk agent usually pays for itself, because order edits and returns are real work and the volume is high enough that $1 a resolution beats staff time. Keep refunds and damage claims with people. If chat, SMS and social messages are spread across tools, Fastlane’s roundup of AI messaging platforms for Shopify sorts them by channel.
If you are leading, the work moves to quality and coverage: one agent across every channel, every conversation scored, and a voice pilot if your phone queue is long. Keep the handoff promise intact as you add channels. As AI shopping assistants start placing orders on a customer’s behalf, more service questions will arrive from them too; the agentic commerce pillar covers what to prepare for.
Disclosure: names on this page may link through /recommends/ links that can earn Fastlane a commission. Fastlane uses Gorgias for its own inbox. Gladly, Missive and Cotera appeared as podcast guests.
An AI customer service agent is software that answers customer questions using your store’s policies, products and order data, and either closes the conversation or hands it to a person with the context attached. On Shopify, the free Inbox agent answers product questions, recommends products, looks up orders after the customer signs in, and adds items to the cart inside your store’s chat. Helpdesk agents from companies such as Gorgias, Gladly and Fin work across email, chat and social and can take actions such as changing an address or starting a return. Their AI usage is billed per outcome, mostly per resolved conversation, on top of the helpdesk subscription.
An AI customer service agent costs roughly $1 (USD) per resolved conversation at list price as of November 2026, and Shopify’s Inbox agent has no per resolution fee. Gorgias bundles its AI Agent on monthly plans at $30 for 30 automated interactions on Starter and Basic, $190 for 190 on Pro and $530 for 530 on Advanced, on top of helpdesk fees. Fin lists $0.99 per outcome with a 50 outcome monthly minimum on other helpdesks. Compare the fee per resolution with what five minutes of your team’s time costs per ticket before you buy a bigger bundle.
The Shopify Inbox agent is enough for many stores under about $100,000 a month whose questions are mostly about products, sizing and order status. It answers from your catalog, policies, pages and Knowledge Base, works only in your online store’s chat widget, and needs the Basic plan or higher plus new customer accounts. It does not edit orders or process returns, so stores with many address changes, cancellations or returns still need a person or a helpdesk agent for those. Setting it to answer only when your team is offline is a low risk way to test it.
Tickets that involve money leaving the business outside a written rule, a judgment call, or a customer who is already upset should go to a person, not an AI agent. That covers damaged, missing or wrong items, refund exceptions, repeat contacts about the same order, and complaints. A wrong answer here costs more than the ticket. A British Columbia tribunal ruled in 2024 that Air Canada was responsible for wrong refund advice its chatbot gave, so anything the agent promises is a promise from your store. Route these tickets to a person by rule.
Your AI support agent is working if its cost per resolved ticket is lower than a person’s and its customer satisfaction matches your human conversations over a 60 to 90 day pilot. Track repeat contact too: the share of resolved customers who write back within seven days about the same order. Vendors can count a customer who gave up as resolved, so repeat contact catches false resolutions. Read a sample of AI transcripts every week. If satisfaction drops or repeat contact rises, narrow the agent to fewer ticket types before adding more.