Agentic commerce: let AI help customers buy without losing control

What is agentic commerce, and what changes for your business?

A shopping assistant might recommend a product and leave the customer to buy it. An AI shopping agent can go further: it may compare options, prepare a basket or place an order within the buyer’s instructions. These are different levels of permission. An agent allowed to search should not automatically be allowed to spend.

For retailers and distributors, the work extends beyond appearing in an AI recommendation. The store must recognise the customer, show the right price, confirm available stock and pass the accepted order to the people who will deliver it. ERP means enterprise resource planning: the system a business uses to manage stock, orders and accounts. An AI sales channel needs the same reliable ecommerce and ERP integration as any other channel.

If orders already go missing between your store and back office, fix that first. The ecommerce ERP integration guide covers that underlying connection. This article focuses on the extra decisions needed when an agent acts for the buyer.

What should an AI shopping agent be allowed to do?

Write the buying rules in terms a customer, finance manager and support team can understand. Specify the supplier, products, quantities, delivery address and total spending limit, including tax and delivery. Decide how long permission lasts and how the buyer can withdraw it. The connected systems must enforce these rules; a sentence in a chat is not enough.

Check the rules again just before placing the order. If two purchases are running at once, each must not treat the same money as available. A refund should not automatically give the agent permission to spend again unless that is part of the agreed rules.

For B2B customers, check which company account the buyer may use before showing private prices, credit terms or saved addresses. Permission to pay by card does not mean permission to buy on account. Product descriptions and messages from outside the business must never be allowed to change these buying rules.

How would an agent purchase work from request to delivery?

Consider an illustrative purchase, not a client result or a tested vendor feature. A maintenance buyer asks an agent to reorder the same approved filter pack from the usual supplier, within an agreed budget, for delivery to the usual warehouse. The buyer does not permit substitutes.

Suppose the agent finds individual filters instead of the approved pack. The purchase should stop, even if it is within budget. A spending limit cannot catch the wrong unit. If the product is correct but delivery costs have risen above the limit, the agent must ask for approval rather than silently changing the order.

Stock also needs a fresh check. Goods on a shelf may already be promised to another customer. At checkout, the seller should set aside the goods, offer a clearly approved backorder or say they are unavailable. An old product feed is not a delivery promise.

Keep the accepted price, quantities and delivery terms with the order. If the store accepts it before the ERP has checked it, explain what that confirmation means and who contacts the buyer if operations cannot fulfil it. A message saying “request received” is not the same as an accepted order.

Do ACP, UCP or Ramp solve the same problem?

No. Commerce protocols are shared rules that let systems communicate. The Agentic Commerce Protocol (ACP) describes exchanges between agents and merchants. OpenAI’s commerce overview also covers structured product information and discovery. The Universal Commerce Protocol (UCP) describes checkout steps and when the buyer needs to take over. They are separate protocols; support for one does not mean a connection to every AI shopping channel.

Ramp’s finance AI and procurement tools address the buying business’s controls: purchase requests, policy checks, approvals, purchase orders and matching invoices. That is different from a retailer making its products available to customers’ shopping agents. Ramp is a useful example of keeping routine work within rules and sending uncertain decisions to people, not evidence that any agent can buy from any store.

Access matters as much as a feature list. Shopify says Agentic Storefronts are automatically enabled for eligible stores, with orders reaching the existing admin. Google’s January 2026 UCP announcement described an initial checkout rollout for eligible US retailers. It does not establish current access for every country or account. Ask the supplier to confirm your store’s eligibility and the exact buying steps it supports.

Neither a protocol nor a procurement tool repairs incorrect product data or connects refunds to your accounts automatically. Ask what your existing systems must supply and what the provider actually handles.

How do agentic payments protect the buyer?

Agentic payments let an agent arrange payment within agreed limits. The agent should use protected payment details provided for that purpose, rather than see or store the customer’s full card details. Keep payment secrets out of chat histories and ordinary activity logs.

For example, Stripe’s shared payment tokens give limited access to a customer’s payment method, with usage limits, expiry and the option to withdraw access. Stripe labels the feature public preview, with country restrictions. Confirm eligibility for the actual account. A payment limit does not check whether the filter pack or warehouse is correct; those checks still belong in the purchase rules.

A payment authorisation usually means the provider has approved a hold on funds. Capture is the later step that collects the payment, although both can happen together. Neither tells the warehouse which goods to send. Your team needs to see the payment result alongside the accepted order, not assume that having protected payment details means the customer has paid.

What happens if the purchase goes wrong?

Return to the filter purchase. The seller creates the order, but the confirmation never reaches the agent. The agent sees a timeout. It must check whether that order exists before trying again. Otherwise, a temporary connection problem can become two orders and two payment attempts. Keep one purchase reference so a repeated request can return the original result.

Give each unresolved problem a named owner and a time for follow-up. Staff should be able to distinguish “waiting for approval”, “accepted”, “partly shipped” and “refund pending”. A single “success” message hides too much.

Who handles cancellations, returns and refunds?

The purchase must remain manageable after the conversation ends. Customers and staff need the seller’s order number, the goods ordered and a way to request help. Decide when an order can be cancelled and what happens if only some goods have left the warehouse.

Stopping the agent or cancelling an unfinished checkout does not necessarily cancel an accepted order. Crossmint’s checkout documentation makes this distinction and says the merchant handles returns and refunds. It also notes that merchants or human checks can block automated checkout. Do not promise that an agent can buy successfully from every website.

Approving a return, receiving the goods, issuing a credit note and sending money back are separate steps. The UCP order specification also separates fulfilment from changes such as refunds, credits and disputes. Keep each step tied to the original goods and quantities. Do not tell the customer “refunded” just because staff approved the request.

Finance should be able to trace the order through payment, invoice, refund if needed and the bank payout. A credit note changes the accounts; it does not by itself send money to the customer. The ecommerce payment reconciliation guide explains why sales, refunds, fees and payouts need separate checks. Check for missing orders and unpaid refunds even when the connection reports that all messages were sent.

Where should you start, and what should a supplier demonstrate?

Start with one low-risk repeat purchase: approved products, one supplier, a known delivery address and a clear spending limit. Keep a person’s approval where the cost of a wrong decision is high. Letting the agent prepare a basket for someone to check may be enough for the first stage.

• Ask the supplier to show the wrong pack size being rejected and the correct customer price being used.
• Withdraw permission before checkout and try two purchases against the same remaining budget. Both must respect the buying rules.
• Interrupt the confirmation after the seller accepts an order. Ask the team to find that order without creating a second one.
• Follow an accepted order into the ERP, then show a partial shipment, a cancellation request and a failed refund with a named owner.
• Turn off new agent purchases and check that staff can still finish deliveries, returns and refunds for existing orders.

These are suggested checks for a demonstration, not claims that the named providers have passed them. Ask for the resulting order and payment records, using the account and payment route your business would actually use. A polished conversation alone is not proof.

Budget for cleaning product data, connecting systems, provider setup, security checks and staff time fixing failed purchases. Include ongoing support and provider changes. A charge for an attempted checkout may still apply when no order is completed, so check the supplier’s terms rather than assuming every attempt becomes a sale.

Wait if you cannot get reliable prices or stock, cannot find an order after a lost confirmation, or have nobody responsible for refunds. For a practical starting point, bring one typical order and its approval rules to Harmatiq’s integration and automation team. Our commerce and B2B services cover the selling process around those connections. Agree what the business can safely promise before adding another buying channel.