AI Shopping Assistants on Shopify: Build, Buy and What to Measure

The AI part of an AI shopping assistant is rarely what goes wrong. A model can describe a linen shirt well enough. Projects fail on the boring plumbing: a price quoted from last week’s catalog, or a cheerful “yes, that’s in stock” for a variant that sold out an hour ago.

So the useful question for a Shopify brand in late 2026 is not “should we add AI chat?” It is which route you take, and how you will prove it pays. You can switch on the free Inbox agent inside Shopify Inbox, pay for a per-resolution app, or build your own assistant on Shopify’s commerce MCP servers. We build all three for clients. This guide covers what each can do today, what Shopify changed underneath developers this year, how to measure the result honestly, and what each option costs at three traffic levels.

What an AI shopping assistant should do on a Shopify store

A good AI shopping assistant does four jobs: it narrows the catalog to the right two or three products, answers pre-purchase questions from your real policies, builds the cart, and hands off to a human when it should. Anything beyond that belongs in phase two.

On the stores we work with, the chats that matter commercially are fit and compatibility questions (“will this fit a 2019 Model 3?”) and delivery questions in the last minutes before checkout. Both need live catalog data far more than clever prose.

What we leave out of version one:

• Refunds, cancellations and address changes. These touch money and fulfilment. Route them to your helpdesk.
• Open-ended discounting. If the bot offers 10% to anyone who hesitates, shoppers learn to hesitate.
• Medical or safety advice about products, even when the shopper pushes. Supplements and baby products are where this bites.

One distinction helps when you brief vendors. A support chatbot is judged on tickets it closes; a shopping assistant is judged on revenue per visitor. Decide which one you are buying before you watch demos. If support deflection is the main goal, our guide to building an AI chatbot for your website covers that side.

Your three options in 2026: Inbox agent, an app, or a custom build

Most Shopify stores should begin with the Inbox agent, move to an app when they need helpdesk integration or more control, and build custom only when volume or catalog complexity justifies owning the stack. Here is how they compare in November 2026.

Shopify Inbox and its built-in Inbox agent

Shopify Inbox now ships with an opt-in AI agent that answers product and policy questions, recommends products and adds them to the cart. The app is listed as free on the Shopify App Store, with 4.5 stars from 5,817 reviews and 21 supported languages when we checked.

Go to Sales channels > Inbox and find the Inbox agent section. There are three modes: All conversations, Only when your staff is unavailable, or Never. Stores that installed Inbox before the agent existed default to Never, so plenty of merchants have it and don’t know.

The details that matter, from Shopify’s Help Center page on assigning the agent:

• It needs new customer accounts (Settings > Customer accounts). Legacy accounts are the usual reason it won’t switch on.
• It answers questions on product details, availability and pricing, and looks up orders only after the customer signs in.
• Its sources are your catalog, policies, pages and Knowledge Base, with web search used only as a secondary source for product-adjacent content.
• Shoppers see a notice that they are talking to AI powered by Shopify. You cannot remove or reword it.
• You, the merchant, are responsible for accuracy, and Shopify asks you to update your privacy policy because chats are processed to improve the agent.

Handoff is decent for a free tool. If staff are online, the conversation moves to them with the AI’s context attached. If not, the agent gives out your sender email address.

Training is lighter than the word suggests. Under Sales channels > Inbox > Training you rate practice scenarios or real conversations, and Shopify proposes an updated persona. That shapes tone, not knowledge. If the agent gets your returns window wrong, fix the policy page.

The limits: it lives only in the online store chat widget, you cannot add your own tools, and reporting is thin.

Shopify AI chatbot apps: what you are really paying for

A third-party Shopify AI chatbot makes sense when you already run a helpdesk and want one agent covering pre-sale and post-purchase questions, or when you need proactive prompts and deeper analytics than the Inbox agent offers.

Pricing models differ more than features do. The comparison Intercom publishes for its own Fin product shows the spread (read it knowing who wrote it): Fin for Ecommerce at $0.99 per outcome, Gorgias at $0.90 to $1.00 per AI resolution on top of a helpdesk plan of $10 to $750 a month, Rep AI from $39 a month on session pricing, and Tidio with a free tier and paid plans from $29.

Per-resolution pricing sounds fair until you read the definition of a resolution. Ask before you sign:

1. What counts as resolved? If a shopper simply leaves, is that billed? Some vendors count it.
2. How does it learn price and stock changes? Product and inventory webhooks are fine. A nightly storefront crawl is not.
3. Does it quote market prices? If you sell in GBP and EUR through Shopify Markets, test it.
4. Can you export transcripts? You need them for quality review and your own holdout analysis.

Our view: if you already run Gorgias or Intercom, use their AI rather than adding a second widget.

Building a custom assistant on Shopify’s MCP servers

A custom AI shopping assistant on Shopify now means your own chat front end and model, talking to Shopify’s Universal Commerce Protocol (UCP) MCP servers at https://{your-store}/api/ucp/mcp for catalog, cart, checkout and policy answers. This changed a lot this year, and many articles ranking for the topic are out of date.

The original Storefront MCP endpoint at /api/mcp is gone. The migration page on shopify.dev opens with a “Removed” notice and maps old tools to new ones:

• Storefront Catalog MCP: search_catalog, lookup_catalog and get_product.
• Cart MCP: create_cart, get_cart, update_cart, cancel_cart.
• Checkout MCP: checkout sessions; requests must be authenticated or signed.
• Ask MCP: ask_business for policy and FAQ questions, still a draft capability.
• Customer Accounts MCP: unchanged, with order tracking also available through Order MCP.

The catalog needs no API key. Every request instead carries a link to your agent profile, a small hosted JSON file declaring which UCP capabilities your agent supports. A trimmed call looks like this (the Storefront Catalog MCP reference has the full schema):

POST https://your-store.com/api/ucp/mcp
{
  "jsonrpc": "2.0", "id": 1, "method": "tools/call",
  "params": {
    "name": "search_catalog",
    "arguments": {
      "catalog": {
        "query": "waterproof hiking boot, wide fit",
        "context": { "address_country": "GB", "currency": "GBP" },
        "pagination": { "limit": 5 }
      },
      "meta": { "ucp-agent": { "profile": "https://example.com/agent.json" } }
    }
  }
}

Three details that save debugging time. Prices come back in minor units, so 1899 means 18.99, and a template that forgets to divide will show a $1,899 sock. Availability is a flag on each variant, so read the variant, not the product. And buyer context changes what comes back, so send it on every call.

The cart has its own trap. Per Shopify’s carts and checkout docs, update_cart replaces the whole cart with what you send. Send only the new item and you have deleted everything else. Checkout usually ends with a continue_url into your normal, prefilled Shopify checkout, and you remain merchant of record.

For data the MCP tools don’t expose, such as compatibility tables in metafields, we add a small service the model calls as one more tool, so the model never queries Shopify with free-form input. If ask_business handles your returns rules poorly, retrieval over your help content works well; see our explainer on retrieval-augmented generation. This pattern is the build our Shopify AI development team does most often.

Shopify Sidekick vs custom: different jobs

Shopify Sidekick is not a shopping assistant. It is the AI assistant for you and your staff inside the Shopify admin, and shoppers never see it, so “Shopify Sidekick vs custom” is usually the wrong comparison. The real one is Inbox agent versus custom.

Sidekick still helps with prep, such as finding products missing sizes or materials. More in our Shopify Sidekick guide.

AI product recommendations that hold up

AI product recommendations in chat are only as good as the attributes in your catalog. If “wide fit” lives inside a paragraph of description, the assistant guesses. If it is a metafield or variant option, it can filter.

Our pre-launch catalog pass:

• Decision attributes as structured data: fit, material, compatibility, capacity, whatever your support tickets ask about most.
• Complementary pairs set by hand for your top 50 products, so “what goes with this?” has a merchant-chosen answer.
• Out-of-stock behaviour decided in advance: nearest in-stock alternative, a back-in-stock alert, or both.

One rule we hold to: offer two or three options with a one-line reason each. That fails more gracefully than a single confident pick when the model misreads the question.

Wrong prices and phantom stock: the risks to design out

The biggest risk with any AI shopping assistant is a confident, wrong statement about price, stock, delivery or discounts that a shopper then screenshots. Shopify’s own Inbox documentation puts responsibility for accuracy on the merchant.

Caching deserves its own line. Shopify’s catalog docs tell agents not to cache search results or reuse images, and catalog queries are rate limited. A nightly vector index is fine for “what do you sell?” questions. It is not fine for any sentence containing a number.

How to measure an AI shopping assistant: conversion, AOV and deflection

Measure an AI shopping assistant with a holdout test: hide it from a random share of visitors and compare conversion rate, AOV and revenue per visitor between groups. Every other number is a proxy.

The Inbox agent’s Overview page, per Shopify’s conversations guide, shows assisted sessions, assisted orders, satisfaction rate and average response time. Assisted orders is the number people quote, and the one that misleads. A shopper asking about sizing is already deep in the funnel. Many would have bought anyway.

A worked example with hypothetical numbers. Your store gets 80,000 sessions a month at 2.0% conversion. You hold out 10%, so 72,000 sessions see the assistant. If that group converts at 2.15% against the holdout’s 2.0%, that is a 7.5% relative lift: about 108 extra orders, or roughly $7,560 a month at a $70 AOV. Run it for at least four full weeks.

What an AI shopping assistant costs

The Inbox agent costs nothing beyond staff time. Apps run from a free tier to a few thousand dollars a month at volume, mostly through per-resolution fees. A custom build has a one-off build cost and a running cost that grows slowly with traffic. Which is cheapest depends on conversation volume.

A hypothetical model, with our assumptions stated:

• 60% of conversations become billable resolutions at $0.99 each (the Fin rate above); helpdesk seats excluded.
• Custom build: 220 hours at a blended $60 an hour, so $13,200 one-off.
• Custom running cost: $0.04 of model usage per conversation (check your provider’s current prices), $100 a month for hosting and logging, and 8 maintenance hours a month at $60.

The custom column ignores the cost of getting version one wrong, which is real. Falling model prices help it over time; model routing and prompt caching keep spend low without touching the price-and-stock rules above. Our Shopify intelligence solutions discovery produces exactly this table with your own numbers.

Frequently asked questions

What we’d do first

This week, switch on the Inbox agent in Only when your staff is unavailable mode and fix whatever blocks it, usually legacy customer accounts. Read the first 50 overnight transcripts. Every vague answer points at a missing attribute or an unclear policy page, and fixing those helps whichever route you take later.

Then decide on evidence. If chat volume stays low, stop there; it cost you nothing. If it climbs into the thousands of conversations a month, or your catalog needs compatibility logic the built-in agent can’t handle, price an app against a custom build with the table above, and plan the holdout before launch.

We have built for brands like Delsey, EcoFlow and Kerbholz. If you want help choosing or building an assistant, our Shopify AI team starts from your transcripts and catalog, not a sales deck. And if you also want to sell through ChatGPT and other outside agents, our guide to agentic commerce on Shopify shows how the same catalog work pays off there.