AI agents for ecommerce are systems that choose their next step from what they find, using tools on your store data: a Shopify AI agent might read an order, check your policy and prepare a return, where a chatbot only answers and a workflow follows a path drawn in advance. One question tells the three apart: is the path already written?
We are noprob agency. We work every day with AI agents built on Claude Code: they write and test the code of the stores we build with the Shopify Launch Sprint, and they read store data for us. This page is the map for our series on agents: the criterion to tell the three forms apart, the questions to ask a vendor, what to hand to an agent and what to keep, and the five patterns behind most of what is for sale. Each section points to the article that goes deeper.
What are AI agents, and what do they mean for an ecommerce store?
An AI agent is a system that perceives a situation, decides what to do and acts toward a goal, usually in several steps, adapting each step to what the previous one turned up. The language model inside it is only one part.
Take Claude. There is Claude the model, the engine that turns text into text; Claude the chat app, which adds an interface, some memory and a few tools; and Claude Code, an agent. You give it a goal such as “refactor this checkout script and run the tests”, and it reads files, runs commands, looks at the result and decides what to do next, until it is done or needs your decision. Same family of model, three different systems.
For a store, the lesson is that the word on the box tells you nothing: a vendor can call a chat window with one extra button an agent. Look at what the system does. The model keeps nothing between calls, so whatever the system remembers about an order or a half-finished job lives outside it, in code and data somebody built. An agent on Shopify is mostly engineering around a model: instructions, tools, access to your data, a place to keep state and rules about when to stop.
Keep two directions apart. This page is about agents that work for the store: answering, sorting, checking, preparing. Agents that shop for your customers through ChatGPT or Copilot are in agentic commerce for Shopify, and Shopify documents its side at shopify.dev/docs/agents. For the wider view of what AI does well in a store, read AI for ecommerce. We do not repeat it here.
AI agent vs chatbot vs workflow: how do you tell them apart on a store?
Ask whether the path is already written. A chatbot has one move (answer), a workflow has several paths all drawn in advance, and an AI agent has a goal and tools and chooses the path step by step. Anthropic draws the same line in Building effective agents: workflows run through predefined code paths, agents direct their own process and tool use.
The clearest way to see it is one message handled three ways. A customer writes: “I bought this jacket as a gift, it is the wrong size, can I swap it?”
- Chatbot. It replies from the text it was given: your returns policy and a link to the form. It never looks at the order. Fast, cheap and identical for every customer.
- Workflow. A router labels the message “return”, the system looks up the order in Shopify, sends the return instructions, and anything that does not fit a label goes to a person. Every branch was drawn before the customer wrote. Shopify Flow works this way: triggers, conditions, actions.
- AI agent. It has the goal “resolve this request within policy” and tools to read orders, read the policy and check stock. It finds the order, sees it was a gift and is still inside the return window, checks that the right size is in stock, proposes an exchange, prepares the request and stops at the step that needs a person. The route depended on what it found. Nobody wrote it.
The workflow handles the cases you foresaw and parks the rest. The agent copes with the unforeseen, at a price: every step is a model call, so it is slower and costs more, and a path you did not write is harder to check.
My rule: start with the simplest form that solves the case and climb only when it stops being enough. Anthropic gives the same advice: find the simplest solution and add complexity only when it demonstrably improves the outcome. Most store tasks end up as a workflow with one agent step inside.
Which of the three fits each case on a Shopify store?
For each task, pick the simplest form that still handles the cases you actually meet. This is the matrix we use when we look at a store, eight typical cases:
- “Where is my order?” Workflow with a tool: look up the order and tracking, write the reply. A chatbot with no order access cannot answer it; an agent is more than the job needs.
- Questions on shipping, returns and sizing. Chatbot over your written policies, quoting the page it used. If the policy pages are vague, fix them first.
- A return request. Workflow for the standard case. Agent when gifts, late requests or damaged items make the next step depend on what it finds. The refund stays with a person.
- Product advice (“which one for rain?”). Chatbot or agent that searches your catalog. Quality depends on your product data far more than on the model.
- Abandoned cart. Workflow in Flow or your email platform: trigger, wait, condition, message. An agent adds cost and no value.
- Product descriptions for a large catalog. A fixed pipeline: take the facts from the supplier sheet, write only from them, check that nothing was added, import a sample a person reads.
- Catalog cleanup and migration mapping. Agent, because every source store is different and the next step depends on the last finding. Code reconciles counts and totals.
- “Why did sales drop on Tuesday?” Agent: it looks at orders, ads and tracking and decides where to look next. Only worth trusting when the tracking is right.
The pattern is easy to remember: fixed jobs are workflows, investigative jobs are agents, plain questions are a chatbot with the right data.
Which five questions should you ask a vendor who sells an “AI agent”?
Five questions separate an agent from a chatbot with a new label, and they work on any vendor.
- “Show me the path of one real request, step by step.” A workflow shows a diagram. An agent shows a trace of what it did and why. If they can show neither, you are buying a chatbot.
- “Which tools does it call on my store, and with which permissions?” Reading orders is not editing them. Ask for the list, and start with read-only access.
- “What can it do alone, and what waits for a person?” Refunds, discounts, price changes and promises to customers should wait. The vendor should name the approval step, not describe good intentions.
- “What happens when it does not know?” The right answer is a handoff to a person with the context attached, not a confident guess.
- “How do I test it before and after every change?” A fixed set of real cases with a pass mark. Our guide to evaluating an AI assistant for your store gives you the cases.
Questions two and three are also your security review. OWASP calls the failure excessive agency: too many tools, too many permissions, too much autonomy. The remedy is minimal functions, narrow permissions and a person approving high-impact actions. More in AI agent security for Shopify.
What does it make sense to give an AI agent on Shopify today?
Give an agent work that has several steps, can be checked and can be undone. Customer care, catalog, migration and analysis fit that description; money and promises do not.
Customer care
In customer care, an agent can look up the order, read your policy, draft the answer and prepare the case, while a person approves anything that costs money. Start with drafts that a person sends, and widen the agent’s freedom only as your test cases pass. The full design is in an AI assistant for orders and returns on Shopify.
Catalog
In the catalog, AI works first as a checked pipeline and only second as an agent. An agent earns its place when supplier sheets are messy and each needs a different decision. In my own projects I use Shopify’s built-in AI with clean data and a person reviewing the output. The built-in tools are in Shopify Magic and Sidekick, and the fixed jobs you can run without code are in Shopify Flow and AI.
Migration
Migration is where agents fit best, because the work is wide, repetitive and verifiable. Agents are good at the checkable parts: reading the source store, mapping fields, drafting the redirect map, comparing counts and totals. A person decides the exceptions, and in the Sprint we reconcile counts and totals of products, customers and orders before anything is switched on. The method is in migrating to Shopify with AI agents and the 301 redirects guide; the whole process is in the guide to migrating to Shopify. Integratori DTC Nazzi moved from WooCommerce and Cumini from a proprietary platform tied to its ERP.
Analysis
In analysis, an agent can watch the numbers you would not open every day and tell you what moved, provided the tracking is right. An agent reading a broken dashboard gives you a confident wrong answer, only faster. On Cumini the tracking audit score went from 31 to 92 out of 100 around the migration: an agent reading numbers is only as good as that baseline. Working with Claude on a store is in Claude and Claude Code on Shopify and Claude Duet for a Shopify ecommerce.
What should you not hand to an AI agent?
Keep with people anything that moves money, makes a promise to a customer or cannot be undone. The agent prepares, a person decides.
- Refunds and exceptions. The agent opens the case with the order, the reason and the policy attached. A person approves.
- Price, discount and shipping-rule changes. The agent proposes the change with a preview of what it touches.
- Promises. Delivery dates, remedies the customer or the law is entitled to choose, anything that commits the store.
- Claims in regulated categories. Health, safety and legal statements are checked by someone who answers for them.
- Anything irreversible. Deleting products or customers, or sending to the whole mailing list.
A practical test: could you undo it in a minute? If yes, an agent can do it and you review later. If not, it waits for a click. Begin with read-only access and add write tools one at a time, each with its own approval rule.
Which patterns sit behind most AI agents, and what do they look like on a Shopify store?
Five patterns cover most of what is sold as an AI agent: routing, tool use, reflection, planning and multiple agents. Each solves one problem, so add one only when you have that problem. Anthropic’s guide uses a slightly different vocabulary (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer), but the ideas overlap.
Routing
Routing sorts an input and sends it down the right path. On a store, each customer message is labeled order, return, product question, complaint or other, and each label has its own handling. Keep a fallback route for everything else and read what lands there every week: it shows the route you have not built yet.
Tool use
Tool use lets the model ask the system to do something, and the system does it under your rules. The model proposes “look up the order for this email”, your code runs it with the permissions you granted. Begin with read-only tools. Shopify exposes the buyer side the same way: its Order MCP fetches the current status of an order on demand. See Shopify MCP explained.
Reflection
Reflection adds a second reader: a checker that reviews the draft against written criteria before it goes out. “Read it again” is not a criterion. Useful ones can be checked: every fact comes from the order or the product page, no promise belongs to a person, an unhappy customer is acknowledged. Give the checker its own context and switch it on only where a mistake costs something, such as refunds, dates and promises. Anthropic calls the pattern evaluator-optimizer.
Planning
Planning is for requests with several steps, where the next step depends on the last result. A gift arrives damaged, the return window has closed and the customer wants a different size: nobody can script that in advance. Plan only the next step, look at the result, plan again, and write the limits in code: a maximum number of steps and a way out to a person.
Multiple agents
Several agents make sense when work has to be kept apart, and rarely otherwise. Anthropic’s guidance on building multi-agent systems names three situations where they beat one agent: context that gets polluted, tasks that can run in parallel, and real specialization. Outside those, it says, coordination costs typically exceed the benefits, and it advises dividing along context boundaries, not job titles.
On a store that looks like this: one read-only agent gathers order and policy facts, while a second one with write access never reads text written by strangers, such as reviews. Decisions stay in one place. For most stores a workflow with one agent step is enough; what building an agent costs, and when you do not need one, is in building an AI agent for your Shopify store.
Where do Sidekick, Flow and agentic commerce fit in this map?
Shopify already ships pieces of each form, so check them before you build or buy. Shopify Flow is the workflow form: the app is free on the Basic, Grow, Advanced and Plus plans, and Send HTTP Request is available on Grow, Advanced and Plus. Sidekick and the built-in AI are an assistant for the merchant inside the admin, covered in Shopify Magic and Sidekick.
AI also reaches your store from outside. Agents that shop for customers are in agentic commerce for Shopify and ChatGPT shopping for Shopify; getting cited in AI answers is in GEO for Shopify and GEO vs SEO.
How do the AI Accelerator, the Sprint and the Custom AI System fit in?
If your store is live on Shopify, the [AI Accelerator](/ai-accelerator) is the route: 30 days one-to-one, up to two quick wins switched on with you, and three deliverables that stay yours. If you are changing platform, the [Shopify Launch Sprint](/shopify-migration-sprint) rebuilds your store on Shopify in 14 days with the Custom AI System.
The Custom AI System is code written from scratch for one store and designed for the AI agents that work on it, which also makes the store ready for the agents that visit it. In the Sprint, go-live falls between day 15 and 30, followed by 90 days of support. In the AI Accelerator we help you decide which of the cases above deserve a workflow, an agent or nothing yet.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers; an AI agent pursues a goal in several steps and chooses each step from what it finds. The test is whether the path is already written. A chatbot follows one move, a workflow follows paths drawn in advance, and an agent builds the path as it goes using tools on your data.
What are AI agents in ecommerce?
In ecommerce, AI agents are systems that act on store data to reach a goal, such as resolving a return request, cleaning a catalog or explaining a drop in sales. Money, promises and irreversible actions should still wait for a person.
Do I need an AI agent for my Shopify store?
Probably not as a first step. Many store tasks need a workflow or a chatbot with the right data, which cost less and are easier to check. An agent is worth it when the next step genuinely depends on what the last one found.
Which AI agents are best for ecommerce?
There is no best agent, only the right form for each job. Use Shopify’s built-in tools such as Flow and Sidekick for what they cover, consider a custom agent only for work that is multi-step, checkable and reversible, and judge any vendor with the five questions above.
What is an AI assistant, and is it different from an agent?
“AI assistant” is a loose label for software you talk to. It can be a chatbot that only answers or an agent that uses tools, so judge it by what it does.
The Sprint has a public price by revenue: €5,800 under €2M, €8,800 from €2M to €20M, €12,800 above €20M, VAT excluded, in 4 installments.
The Shopify Launch Sprint moves your store to Shopify in 14 days with our Custom AI System: a fixed calendar, a public price and 4 written guarantees.
Antonio Manitta
Founder & eCommerce Manager at NoProb Agency
For over 7 years I have helped fashion, supplements, and DTC brands scale their eCommerce. I work in long-term partnership with three brands at a time, building stable systems that compound.
About NoProb →
