twimco.Let’s talk ↗︎
Menu
Field notes / AI & judgment

Shopify's agent tool now answers policy questions in its own words. Read what it says about your returns

When a shopping agent asks a Shopify store about its return policy, the answer is moving from your policy text to a written reply generated from it. Developers have until January 15 to switch tools. Merchants have a quieter job: finding out what that reply says before a customer quotes it back to them.

Editorial engraving of a bright apparel shop where the owner checks a printed policy page against a short message on his support lead's phone.

On November 11, 2026, Shopify deprecated the Storefront MCP tool that returned a store's policies and FAQs and pointed agent builders to ask_business on its UCP endpoint, which answers a buyer's question in plain written text drawn from the store's knowledge base, policies and product information. The old tool is maintained until January 15, 2027. Developers should migrate and update their parsing. Merchants should ask a set of real buyer questions and fix the policy wording behind any answer that is wrong, vague or too generous.

What Shopify changed

On November 11, Shopify marked a small developer tool as deprecated, with an "Action required" flag. The tool, search_shop_policies_and_faqs on the Storefront MCP server, let an AI agent look up a store's policies and FAQs. Its replacement is ask_business, part of Shopify's Universal Commerce Protocol (UCP) endpoint.

For developers, the change is a migration with a date. The old tool is maintained until January 15, 2027. New calls go to https://{shop-domain}/api/ucp/mcp, every request carries an agent profile that declares the Ask capability, the question goes in ask.query, and optional buyer context such as country or language goes in ask.context.

The important line for everyone else is this one: ask_business "returns a written answer ... not raw knowledge base facts or full policy text."

Decision guide

What an agent got before, and what it gets now

Considerationsearch_shop_policies_and_faqs (deprecated)ask_business (UCP Ask)
EndpointStorefront MCP at /api/mcpUCP MCP at /api/ucp/mcp, with an agent profile
What comes backKnowledge base facts and policy textA written answer in plain text
Reading the answerCode that parsed facts and policy textRead result.structuredContent.answer.plain
Product and buyer contextNot described in Shopify's changelogUp to 10 product, variant, collection or cart IDs; country, region and intent
Supported untilJanuary 15, 2027Current
Source: Shopify's November 11, 2026 changelog and Ask MCP documentation read November 12, 2026; simplified.

Why that matters to a store

Your return policy used to be quoted. Now it's paraphrased, and the paraphrase is the version the customer sees.

Under the old tool, an agent received your policy text and did its own summarizing. Under the new one, Shopify's tool writes the answer from your knowledge base, policies and product information, and the agent passes it on. That's probably better for shoppers: a direct answer to "can I return this if it doesn't fit?" is more useful than four paragraphs of policy.

It also means a sentence you never wrote can now speak for your store. If your policy says "unworn items within 30 days, final sale items excluded," a written answer might say "you can return it within 30 days." True for most products, wrong for the clearance rack.

The same goes for anything with a condition attached. "Free returns" that are free only within the US. "Ships in two days" that excludes made-to-order items. "Lifetime warranty" that covers manufacturing defects and not wear. A human reading your policy page might catch the condition. A one-line answer to a one-line question can drop it without anyone meaning to.

Shopify's docs say Ask answers questions and doesn't change carts, place orders or start returns. So nothing happens automatically on the back of an answer. But a customer who was told "yes" by an agent will expect your support team to say yes too.

Your return policy used to be quoted. Now it's paraphrased, and the paraphrase is the version the customer sees.

Who decides what the answer should be

This is the judgment part. Nobody at the store reviews each answer before it goes out; that's the point of the tool. So the store's influence is upstream, in the wording of the policy pages and knowledge base it draws on.

That puts the decision where it always belonged: with whoever owns the policy. Our note on assistants quoting last year's policy covers the same problem with your own chatbot. The difference here is that the assistant isn't yours, and you can't tune it. You can only make the source text impossible to misread.

It's also worth agreeing, before anything goes wrong, what you'll do when an agent's answer and your policy disagree. Most stores will honor a small mistake for goodwill. Fewer will honor an expensive one. Write down where that line is, so support isn't deciding it case by case in the middle of December, and so the fix to the policy wording happens the same day.

In practice that means:

  • Put exceptions next to the rule, not three paragraphs later.
  • Name the product types and countries a rule applies to.
  • Remove old wording you no longer honor.
  • Decide what you'd be comfortable an agent saying when it isn't sure, and write that down too ("contact us for final sale items").

Ask it what customers ask

The quickest check is to ask the tool real questions and read the answers. That needs a developer, because the endpoint expects an agent profile; Shopify's docs provide an example Ask profile for testing and say to check the store's UCP profile at /.well-known/ucp and use tools/list to confirm the store offers ask_business.

Write ten questions the way your customers actually ask them, with a country attached where it matters. For example: "Can I send back a sale item?", "Do you ship to Canada and who pays duties?", "Is this jacket waterproof or just water-resistant?", "What if it arrives after Christmas?" Include one product-specific question and pass that product's ID, since Ask can use up to 10 product or variant IDs per question.

Each request is one question with no memory of the last, so phrase them as complete questions. Save the exact answers with the date. They're your baseline, and if a customer ever says "the AI told me I could return it," you'll know what the tool was saying that week.

Then have the owner and the support lead mark each answer: right, vague, wrong or too generous. Fix the source text for every one that isn't right, and run them again. Expect the second round to find something new, because tightening one sentence often changes how a neighboring rule gets summarized.

Sequence of work

A policy check before the holidays

  1. Day 1

    Write ten questions real customers ask, in their words, with countries.

  2. Day 2

    Developer runs them through ask_business with a test agent profile and saves the answers.

  3. Day 3

    Owner and support lead mark each answer right, vague, wrong or too generous.

  4. Day 4

    Fix the policy pages and knowledge base behind every bad answer, then rerun.

  5. Monthly

    Rerun the same ten after any policy change.

Illustrative plan for a merchant and developer working together; not Shopify guidance.

For developers migrating

If you've built an agent or app on the old tool, the work before January 15 is concrete. Move calls to the UCP endpoint. Host your own agent profile and declare the Ask capability. Send the question in ask.query (up to 10,000 characters) and only the buyer context you need (the context object is limited to 8,192 bytes). Read the answer from result.structuredContent.answer.plain, and don't depend on Markdown or conversation ids in it. Any code that parsed policy text out of the old response has to change. And because the tools a store returns depend on the capabilities your profile advertises, check with tools/list on each store rather than assuming every store offers Ask the same way.

Keep the handling of the answer honest, too. If your agent goes on to take an action, such as starting a return elsewhere, it shouldn't treat a written answer as authority to act. Our agent checkout authority note is about drawing that line.

Where we help

We build and fix the agent integrations on both sides of this: the store whose policies need to read clearly to a machine, and the app or agent that has to migrate before January. That's our AI applications and agents work. Talk to us.

Related work: Spin:Market

twimco.

We build alongside the people doing the work.
Bring us the problem you’re thinking about.

Related work

Spin:Market ↗︎