Start the conversation
AI automation22 Nov 2026

Let the agent build the basket: customers choose how to buy

An agent-ready buying journey doesn't have to end with an agent spending money. It can end with an accurate basket the customer opens, reviews and pays for. Support levels of delegation, and keep the handoff intact.

The research is wasted if the customer rebuilds the basket.

In Nigel Poole's experience, some sites make it easier than others to prepare a cart for a customer and hand it back in a browser. Where the handoff works, the research an AI assistant has done survives: the right product, the right variant, the right quantity, ready to review. Where it doesn't, the customer starts again from a blank page.

That is the argument of this article. An agent-ready buying journey doesn't have to end with an agent spending money. It can end with an accurate basket that the customer opens, checks and pays for themselves. A completed human handoff is a useful outcome, not a failure of automation.

It follows Build for customers and their agents and Make your store easy for customers' agents to understand. Those cover what to fix first. This one covers what happens at the moment of buying.

Customers are comfortable in different places.

Australians are shopping online in large numbers. The Australia Post eCommerce Report 2026 found they spent $82.6 billion online in 2025, up 14%, with 9.8 million households shopping online. [1] [2] That is market context. It isn't a measure of agent sales.

The same report asked about agentic commerce. 16% of consumers like or love the idea, 61% dislike or hate it, and 23% are unsure. Among businesses, 44% are advocates. [1] These are attitudes, not adoption, and the report states no sample size for them.

Read them as a design brief. Many customers don't want software to spend their money, and some do. A journey that only works at full delegation fails the majority, and one that only works by hand wastes what an agent can do.

There is also a gap we won't fill. We know of no reliable Australian figure for the share of orders completed by agents. We haven't substituted global forecasts or AI-referral statistics, and we make no claim about conversion uplift.

One purchase, four ways to delegate.

These four modes are Matter's way of describing the choice. They are not an industry classification. To show them, take an illustrative purchase: restocking a customer's usual pet food.

1. Research for me. The customer asks an assistant to compare options. It reads product pages, delivery terms and prices, and reports back. No basket exists yet. This mode depends on clear, honest product information.

2. Prepare it for me. This is the centre of gravity. The assistant picks the right product and pack size for the customer's dog, sets the quantity, builds the basket and hands it over. The customer opens it in a browser, sees what the assistant chose, changes anything they like and pays in their usual way. The customer keeps full control of the money, and the assistant has still done most of the work.

This is where the pet food order is won or lost. If the link opens an empty cart, or a different pack size, or a price that has moved, the customer has to redo the work and trust drops. If it opens exactly what was agreed, the customer can review it and finish. A site that does this well is easy to buy from through any assistant.

Instinct is a personal AI assistant that researches products and websites on its user's behalf and prepares purchases, but only checks out with the user's approval. It is a clear example of preparing the purchase while the customer keeps the decision to pay.

3. Ask me before buying. The assistant is allowed to go as far as the final step, then stops for an approval. The customer sees the full order, including the total, and says yes or no. It needs a cart and an approval screen the customer can understand at a glance.

4. Buy within rules I set. The customer sets limits, such as a maximum spend on a named product from a named store, and the agent buys within them. It is optional, tightly scoped and should never be the default. Offer it only after the other three work.

Build in this order. Each mode rests on the one before it.

Saved baskets, cart links and the browser handoff.

What is documented today varies by platform. The examples below show that discovery and payment need not share an interface. They are not universal support.

On 24 March 2026, OpenAI described how merchants can use their own checkout. For Shopify, purchases complete on the merchant's store through an in-app browser. [3] The assistant finds and prepares, and the merchant's own checkout takes the order.

On Shopify, there are two ways to get a customer into a prepared basket, and they behave differently. Use "a saved basket or a link that builds one" deliberately:

  • A saved basket. Shopify's Cart API holds a cart and returns a checkoutUrl that goes into web checkout. [4] The basket exists before the customer arrives.
  • A link that builds one. Shopify's /buy permalinks build a cart from a link in a browser. Each use makes a new cart, with no lock on stock or price. [5] [6]

That difference matters in practice. A link that builds a new cart each time doesn't hold the stock or price the assistant saw, so what the customer opens can differ from what was prepared. Your journey has to flag that change instead of hiding it.

When there is no buy button.

Many B2B and services businesses don't sell through a cart at all. The equivalent of the basket is an approved enquiry or quote request that carries the customer's brief forward, so nobody has to explain it twice.

What follows are Matter's suggestions, not findings:

  • Build buyer-intent pages that state who the service is for and what a good brief includes.
  • Start with a simple human enquiry route that accepts a full brief from a person or from their assistant.
  • Add a structured endpoint only when there is demonstrated demand and the safeguards are in place: disclosure the customer has approved, rate limits, a reference or receipt, and a route to a person.
  • Measure handoffs, not visits. A qualified enquiry that arrives with the brief intact is the conversion.

Login, permission to act and permission to pay are different things.

Signing in proves who the customer is. It doesn't give an agent permission to act for them, and it doesn't give permission to pay. Treat the three separately.

Shopify documents how checkout can recognise a signed-in customer. [7] For anything an agent does on a customer's behalf, our recommendations are explicit account linking, the least access that does the job, an expiry, a way to revoke it, and a guest route for customers who don't want an account.

Two protocol designs show where approval and delegated payment are heading. Google's AP2 (September 2025) describes signed intent, cart approval and constrained payment tokens. [8] The Agentic Commerce Protocol (ACP) documents checkout and payment references of its own. [9] [10] These are designs, not proof of adoption in Australia. They aren't interchangeable, and none is mandatory.

Build first.

Matter's recommendations, in the order we would tackle them:

  1. Preserve the work across the handoff. What the assistant chose should arrive in the basket: product, variant, quantity and any notes.
  2. Make the cart link safe for the customer. Test it signed out, as a new visitor would meet it. Never put cart IDs, tokens or other secrets in a link someone might share.
  3. Recheck before commitment. Confirm stock, variants, price and the final total, and say so clearly when anything has changed.
  4. Make approval understandable. Show the order in plain terms, with edit, cancel and a way to contact a person.
  5. Show the final price. The ACCC expects the final payable amount, including taxes and unavoidable or pre-selected fees, to be clear before the customer commits. [11] This is practical guidance, not legal advice.
  6. Confirm clearly. Give a confirmation, an order reference, and no duplicate orders when a customer or agent retries.
  7. Measure the handoff. Track basket accuracy, successful handoffs, completed orders and duplicates, not only visits.

Controlled handoffs, built to be checked.

Matter builds for this kind of handoff. Matter Content Studio publishes structured information for people and agents. Our conversational contact experience carries a customer's context into human follow-up. It is not a working retail cart. Matter Data Studio's Operate mode prepares work for a team to approve before it goes ahead. Our Agentic Commerce work covers the rest.

If you'd like to start, find one place where a customer has to start again, and fix that first. Start the conversation and bring it with you.

Sources

Checked 22 November 2026 (Sydney time).

  1. Australia Post: eCommerce Report 2026.
  2. Australia Post: eCommerce trends.
  3. OpenAI: Powering product discovery in ChatGPT, 24 March 2026.
  4. Shopify: Manage a cart with the Storefront API.
  5. Shopify: Cart and checkout permalinks.
  6. Shopify: Cart and checkout permalinks now support UCP at /buy.
  7. Shopify: Checkout authentication with the Customer Account API.
  8. Google Cloud: Announcing Agents to Payments (AP2) Protocol, September 2025.
  9. Agentic Commerce Protocol: Checkout reference.
  10. Agentic Commerce Protocol: Payments reference.
  11. Australian Competition and Consumer Commission: Price displays.

More news.

View all news