Copilot Checkout: The Configuration Guide

Copilot Checkout demos beautifully. A shopper asks an assistant for a gift under fifty dollars, picks one, says "buy it," and the order confirms without a single page load. Shopify shipped it in the Spring '26 edition on June 17, 2026, and every launch demo looks like that: in-stock product, domestic address, standard shipping, saved payment method. Clean.

Real orders are not clean. The shopper in Quebec buying a preorder item with a discount code and a PO box address is where your configuration either holds or quietly loses the sale. This guide covers the unglamorous half of the install: making tax, shipping, payment methods, and fulfillment behave inside a conversation. For the strategic picture, start with the 2026 agentic commerce guide; this article stays in the plumbing.

What is Copilot Checkout?

Copilot Checkout

Shopify's in-conversation checkout, shipped in the Spring '26 edition on June 17, 2026. It lets a shopper complete a purchase inside an AI conversation — through a Brand Agent or a supported assistant — while the order books into your Shopify store like any other. Currently Shopify Plus only.

The important design fact: Copilot Checkout is not a separate checkout. It reads the store configuration you already have — the same tax settings, shipping profiles, market configuration, and payment setup that power your web checkout. Shopify calls this parity; we call it the reason the install is mostly an audit. You do not build a new checkout. You find out whether the one you have survives being operated by a machine.

It pairs naturally with Brand Agents — the conversation happens there, the transaction completes here — but the two are separate switches. You can run a Brand Agent that hands off to web checkout, and you should, until the configuration work below is done. If you have not read it yet, Brand Agents explained covers that half. And because both are currently Plus-only, Shopify Plus for agentic commerce covers whether the upgrade math works for you.

Why the demo path is easy and the unhappy paths kill orders

At web checkout, a human is a self-healing system. Shipping rate looks wrong? They pick another. Card declined? They reach for a different one. Address rejected? They fix the typo. Web checkout gets away with rough edges because the shopper absorbs them.

In a conversation there is no page of alternatives to scan. The agent hits your configuration, gets what it gets, and relays the result. A missing shipping rate is not an inconvenience — it is the agent saying "I can't complete this order." A tax mismatch is not a small discrepancy — it is a quoted total that turns out to be wrong, which is worse than no quote at all. And the shopper's cost of abandoning is one sentence: "never mind."

This is why the configuration work matters more here than it ever did on web. Your checkout has probably carried small defects for years — a shipping profile with a hole in it, a tax override nobody documented — and humans have been silently absorbing them. The agent will not.

Tax: the quiet order-killer

Tax fails politely at web checkout — the number is just wrong, and most shoppers never check. In conversation, tax fails loudly: the agent quotes a total, the shopper approves that total, and the order books at that total. If your configuration produces a different number at booking than at quoting, you get a failed order or a discrepancy you eat. Audit four things before enabling Copilot Checkout:

• Registrations match your markets. If the assistant surfaces you to shoppers in a region you ship to but never configured tax for, the quote is wrong before the conversation starts.
• Product tax categories are set per SKU. Default categories over-tax some products and under-tax others — apparel, food, and digital goods all carry special rules in many US states.
• Overrides are documented and intentional. Every tax override in the admin is a place where the quoted total can surprise you. If nobody remembers why an override exists, resolve it now.
• Inclusive vs exclusive pricing is consistent per market. VAT-inclusive prices quoted to a US shopper, or the reverse, produce totals that look wrong even when they are technically right.

Shipping: where most in-conversation orders die

Shipping is the stage with the most moving parts and the least forgiveness. The agent needs one resolvable rate, returned fast, on the first pass. Three failure patterns account for most of the dead conversations we see:

• Profile gaps. A product assigned to no shipping profile, or a profile with no rate for the shopper's zone, returns nothing. On web that is a confusing error page; in conversation it is the end of the order.
• Carrier-calculated rate timeouts. Live carrier rates are accurate and slow. An agent waiting on a rate that times out does not retry your carrier — it apologizes to the shopper.
• Address edge cases. PO boxes, military addresses, and territories are exactly the addresses that carrier services reject or misprice. Decide the policy now, not per conversation.

Payment methods: fewer is more reliable

The payment stage is the one place where less configuration is better. Copilot Checkout completes payment with credentials the assistant can actually use — wallets and saved payment methods. Anything that requires a manual step somewhere else does not belong in this channel: cash on delivery, bank deposit, invoice-me options. Leaving them enabled does not add flexibility; it adds paths that stall.

Two things to verify in testing. First, the decline path: when a card is declined mid-conversation, what does the shopper hear, and can they offer another method without starting over? Second, verification friction: 3-D Secure and similar challenges interrupt the conversation with a redirect, which some shoppers will complete and many will not. You cannot remove bank-mandated challenges, but you can make sure your fraud settings are not triggering optional ones on low-risk orders.

Fulfillment edge cases the agent will find for you

Every catalog has products that do not fit the standard flow — preorders, subscriptions, items that ship from a different location, digital goods. On web, these get handled by page copy and shopper patience. In conversation, each one is a promise the agent makes on your behalf. The table below is the edge-case list we work through on every install:

The pattern across all six: the fix is rarely code. It is deciding what the policy actually is, then making the configuration say it explicitly. Ambiguity that humans tolerate becomes agent behavior you did not choose.

Do you stay merchant of record? Yes — and it cuts both ways

Merchant of record

The legal entity responsible for a sale: it charges the customer, owes the tax, handles refunds and chargebacks, and answers for compliance. With Copilot Checkout, that entity is you — not Shopify, and not the assistant the shopper was talking to.

This is the right answer, and it is worth being precise about why. The order books into your Shopify store. The customer record is yours. Payment settles through your payment setup. The assistant is a sales channel, not a reseller — closer to a new storefront surface than a marketplace. That means the customer relationship, the email address, and the repeat purchase all stay with you, which is the entire strategic argument for doing this work instead of waiting.

It also means the liabilities stay with you. A mis-taxed order in conversation is your mis-taxed order. A refund on an in-conversation purchase follows your refund policy and hits your books. You cannot blame the channel — which is exactly why the audit above is not optional. Merchant of record is control and responsibility in the same sentence.

The pre-launch checklist

Here is the sequence we run inside Agentic Commerce Setup. Checkout configuration is week-two work for a reason: it depends on catalog and policy decisions made in week one, and it is the last thing standing between a working agent and live orders.

1. 1

   Mine your order history for edge cases

   Pull the last 90 days of orders and tag the weird ones: unusual addresses, discount stacking, split shipments, preorders. That list is your test suite — your edge cases are already in your data.

2. 2

   Fix tax and shipping against that list

   Close the profile gaps, set fallback rates, verify registrations and tax categories for every market the assistant can surface you in. Do shipping first; it fails most.

3. 3

   Trim payment methods for the channel

   Disable manual methods for in-conversation orders and test the decline path end to end, including recovery with a second payment method.

4. 4

   Run scripted unhappy-path orders

   Test the Quebec-preorder-PO-box orders, not the demo order. Every edge case from step one should either complete or fail with a clear, acceptable message.

5. 5

   Watch the first two weeks

   Log every in-conversation checkout that fails or abandons, and treat each one as a configuration bug report. The failures tell you exactly what to fix next.

Once orders are flowing, the question shifts to whether the channel is paying for itself — completion rate against web checkout, incremental orders, support load. That measurement problem has its own playbook: measuring agentic commerce ROI.