To build an AI agent for your Shopify store, answer two questions before you touch a tool: do you need an agent at all, and would a plain workflow do the job? After that you need four things (clean data, tools the agent may use, written instructions and limits, a way to evaluate it), you pick one of four routes, and you work out the cost from five factors instead of a price list.
We are noprob agency. If your store is not on Shopify yet, the first thing worth building is a clean base: the Shopify Launch Sprint moves your store to Shopify in 14 days with our Custom AI System. If you are already on Shopify, the AI Accelerator is 30 days one-to-one to find where AI pays in your store and switch on the first solutions. Either way, this guide gives you the method, with the official source for every price it mentions.
What is an AI agent, and how is it different from a workflow?
An AI agent is a model that decides its own next step, calling tools until the job is done. A workflow follows a path you wrote in advance. Anthropic draws the same line in Building effective agents: look for the simplest solution that works and add complexity only when needed, because agentic systems often trade latency and cost for better results.
In a store the difference is easy to see. “Where is my order?” is a workflow: find the order, read the date, answer. A customer who writes that a gift arrived broken, after the return window, and asks for a replacement is where an agent can earn its place. It has to read your policy, check the order, weigh the exception and decide whether a person should take over. A chatbot answers, a workflow acts along a fixed path, an agent chooses the path. The full comparison is in AI agents for Shopify: chatbot, workflow or agent.
Do you actually need an AI agent? Two questions before you build
Ask the objective question first and the personal one second: do the numbers hold, and is it worth it for you? The order matters. If you start from “I want one”, you read every number looking for the ones that agree with you.
Question one: do the numbers hold?
- Volume: how many requests arrive each week, and of what kind (order status, returns, sizing, product questions)?
- Cost of the time today: who answers, and how many hours a week does it take?
- Cost of a mistake: what happens if the answer is wrong, a refund granted by error or a delivery date promised and missed?
- Speed: how fast does the customer expect an answer, and does a late one cost you sales?
- Coverage of a fixed path: how much of the volume could a set of rules handle without any judgement?
Question two: is it worth it for you?
- What kind of store do you want to be: the one that answers in seconds at midnight, or the one where a person reads every complaint?
- Who will read the agent’s failures every week and update its instructions when a policy changes?
- Could you explain how it works to the person who takes over from you?
Two opposite stores show why both questions count. In the first, the owner reads ten messages a week: no agent will pay for itself, and the owner’s reading is part of the service. In the second, support takes a person every day and most messages are order status, returns and sizing: automation pays, and most of it is a workflow, with an agent only for the hard cases.
My rule, as a developer by training: if you have not done the numbers, do them first. If they do not hold, look for a smaller version, and if there is none, stop. If they hold but nobody wants to maintain it, do not start. If both hold, build, and write down what “enough” means before you begin: how many test cases must pass, how long a customer may wait, how much you will spend a month.
When do you not need an AI agent?
You do not need an agent when a fixed path covers the request, when the volume is low, or when nobody will check its work. The usual cases:
- A few requests a week. A person answers better and the cost of building never pays back.
- The answer must be identical every time. Shipping terms and your returns policy belong on a page and in a reply template, not in a model.
- The job is “when X happens, do Y”. That is Shopify Flow, with no model in the loop.
- You need drafts, summaries or reports for your own team. Sidekick or Shopify Magic with a human review is enough: see Shopify Magic and Sidekick.
- Your data is not reliable. An agent repeats wrong stock, wrong sizes and an old returns policy fluently, which is worse than repeating them badly.
- Nobody owns the weekly review. An agent nobody checks drifts.
In each of these cases the useful move is smaller and cheaper than an agent. That is why the AI Accelerator starts by mapping where AI pays in your store, before any tool is chosen.
What do you need to build an AI agent for a Shopify store?
You need four things: clean data, tools the agent may call, written instructions and limits, and a way to evaluate it. Skip the fourth and you will not know whether the other three work.
Clean data
The agent can only be as right as what it reads: product attributes, variants, metafields, policies, order data. Wrong stock or an outdated returns page becomes a confident wrong answer. Cleaning is the unglamorous half of the project, and a migration is the cheapest moment to do it, because every product is touched anyway. The Shopify Launch Sprint runs in 14 days with go-live between day 15 and 30, and the Custom AI System behind it is built for AI agents, with the data of your catalog mapped in detail. Our guide to migrating to Shopify shows the steps.
Measurement belongs to the same discipline. On Cumini, the tracking audit score went from 31 to 92 out of 100 around the migration. The rule holds for an agent too: take the baseline first, because you cannot judge it against a number you never took.
Tools
Tools are what the agent can do: read an order, search the catalog, open a return, tag a ticket, hand over to a person. Each tool is a permission, so start with read-only ones and add write actions one at a time. Shopify’s Universal Commerce Protocol (UCP) is how agents work with a store: its MCP tools cover catalog search, cart and checkout, and the old catalog and cart tools on /api/mcp have been replaced by UCP at /api/ucp/mcp. The agent documentation and the Storefront MCP page, which covers the move, list what exists; Shopify MCP explained covers what it changes. What an agent may touch is a security decision, treated in AI agent security on Shopify.
Instructions and limits
Write down the role, the tone, the policies, what the agent must never do, when it must hand over to a person, the refund amount above which a human decides, and the maximum number of steps per request. These are files in a repository, versioned like code, not a prompt someone pastes into a box.
Evaluation
Collect real past requests with the right answer next to each, starting from a first set of ten and growing it, and run them again after every change. If the score does not move, the change did nothing, whatever it felt like. Track the cost per conversation in the same run. The method is in how to evaluate an AI assistant for your store.
Which route should you take: Flow and Sidekick, an app, a custom agent or the Custom AI System?
Take the smallest route that covers the job: Shopify’s own tools first, then an app, then a custom agent. Each step adds control and adds work.
Shopify Flow and Sidekick
Shopify Flow is a free app available on the Basic, Grow, Advanced and Plus plans; you build workflows from triggers, conditions and actions. The Send HTTP Request action is available on Grow, Advanced and Plus, which means a Flow can call a service of your own. Sidekick is included with your plan (Shopify says features and usage limits vary by plan), can build Flow workflows for you and presents changes for your review before applying them. This route fits internal work and fixed rules. More in Shopify Flow and AI: ready workflows.
An app from the Shopify App Store
For customer-facing support or recommendations, an app is the quickest start. Read which store data it can access, ask what happens when it does not know, and test it on your own evaluation set before you pay for a year. The price is set by the developer on the app page.
A custom agent on Claude or another model
You, or a developer, write the instructions, the tools and the checks, and call a model such as Claude, GPT or Gemini through its API. It is worth it when your process is specific: unusual policies, a back office that is not standard, cases with several steps. We build with AI agents on Claude Code every day, and it shapes how we design them: small steps, written limits, tests before anything touches a customer.
The Custom AI System
The Custom AI System is the store itself, written from scratch for your shop and ready for AI agents, with an automatic GEO layer and data mapped in detail. It does not replace the three routes above: it makes every one of them start from data that is already clean. It comes with the Shopify Launch Sprint.
How much does an AI agent for a Shopify store cost?
An AI agent for a Shopify store costs the sum of five things: your Shopify plan, the model usage, the setup, the apps, and the time to maintain and check it. Only the first two have public prices. Use this as a checklist for any quote you receive:
- Shopify plan. Basic €19, Grow €56 and Advanced €289 a month with yearly billing, Plus from €2,100 a month (Shopify pricing, as shown on 29 October 2026; prices vary by country). Sidekick is included in the plan and Flow is a free app.
- Model usage. Billed per million tokens, input and output separately. Prices below.
- Setup. Developer hours for instructions, tools, tests and the connection to your store. No official price exists: ask for a quote that lists hours per piece.
- Apps. Each app has its own plan on its App Store page.
- Maintenance and review. The weekly reading of failures and the update of instructions when policies or models change.
What does the model cost?
Official prices, checked on 29 October 2026, per million tokens (input / output). On Anthropic’s pricing page: Claude Haiku 5.5 $0.10 / $0.50 (prompts up to 100,000 tokens), Claude Sonnet 5.5 $2 / $10, Claude Fable 5.1 $10 / $50. On OpenAI’s pricing page: gpt-6-luna $0.10 / $0.50, gpt-6-astra $10 / $50. Between a small and a top model of the same maker the input price differs by 100 times at both makers, so “always the most powerful model” is a spending decision, not a safety measure. Prices change often: check the pages before you budget.
Two levers lower the bill without touching the answers. Prompt caching: the instructions and tool descriptions repeat on every call, and a cache read costs a fraction of the base input price at Anthropic (10% on Claude Haiku 5.5, 5% on Claude Sonnet 5.5, 2.5% on Claude Fable 5.1). Batch processing: work nobody waits for, such as nightly test runs, is billed at 50% of the standard price at both Anthropic and OpenAI.
An example you can redo with your own numbers
This is an illustration, not a measurement. Assume a support conversation needs 4 model calls, each with 3,000 input tokens (instructions, tool descriptions, history) and 400 output tokens: 12,000 input and 1,600 output tokens in total. On Claude Haiku 5.5 that is about $0.002 per conversation, on Claude Sonnet 5.5 about $0.04, on Claude Fable 5.1 about $0.20. For 1,000 conversations a month: about $2, $40 and $200. Change the calls and the tokens to match your own requests.
The pattern is the useful part: in this example the model is the small line. The big lines are the setup and the hours of whoever checks the agent, so budget for people’s time before you budget for tokens. If you want your own figures, with the tool and the order of work, the AI Accelerator ends with a priority document and a 90-day roadmap including tool costs; its price is fixed and public on its page.
Where should you start this week?
Start with the numbers and the smallest route, not with a tool. In this order:
- Count your support requests for two weeks and sort them: status, returns, sizing, product questions, exceptions.
- Note the hours spent and the cost of the worst mistake in each group.
- Check which groups a Shopify Flow or a reply template already covers.
- Fix the data an agent would read: attributes, stock, the returns and shipping pages.
- Write a first set of ten real requests with the right answer: the start of your evaluation set.
- Write down what “enough” means: pass rate, waiting time, monthly spend. Then pick a route.
Frequently asked questions
How much does an AI agent cost?
It is the sum of your Shopify plan, model usage, setup, apps and ongoing review. Model usage is billed per million tokens at public prices: in our example about $2, $40 or $200 for 1,000 conversations depending on the model. Setup and maintenance are the larger lines and depend on the quote.
How do I build an AI agent for my store without coding?
Start with Shopify Flow and Sidekick, which need no code, or with an app from the App Store. A custom agent that calls a model and your store data needs a developer. Whatever you pick, write the instructions and a set of test requests yourself, because they define what “right” means.
What are AI agents, in one line?
An AI agent is a model that chooses its next step and uses tools, such as reading an order, until the task is done. A workflow follows a fixed path, and a chatbot only answers. Agents suit tasks with exceptions and several steps.
Is an AI agent free?
Parts of the stack are: Sidekick is included in your Shopify plan and Flow is a free app. A custom agent is not, because the model is billed per use and someone has to set it up and check it. The model is often the smallest cost.
Can I buy a ready-made AI agent for my store?
Yes, as an app from the Shopify App Store or as a service from a developer. Test it on your own real requests before committing, and check which store data it can read and which actions it can take.
Are AI agents worth it for a small store?
It depends on the numbers, not on the size. A store with few requests rarely gains from an agent, while one where support fills a person’s day usually does, mostly through workflows and a few agent steps. Run the two questions in this guide before you spend anything.
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 →
