AI Shopping Agents: Why Purchases Still Fail at Checkout

The gap between AI shopping hype and completed orders

An agent may be able to search for an item or fill in a form, but a real transaction can involve multiple steps and decisions: choosing the correct product, applying an offer, confirming delivery details, selecting payment, and authorizing the charge. A failure or unexpected choice at any point can leave the order incomplete - or require the shopper to take over.

That distinction matters because shopping is more than finding the lowest price. Consumers may care about a preferred brand, a loyalty benefit, delivery timing, or a familiar ordering experience. An agent has to interpret those preferences and act within the merchant’s checkout rules. It also needs to make clear what it plans to buy and what it will charge before the user commits.

For businesses, early order counts are a reminder that AI-initiated purchases are not yet a substitute for conventional ecommerce. The channel’s potential depends on whether agents can finish transactions consistently, not just whether they can produce convincing recommendations.

Why shoppers hesitate to give AI agents access

To act on a shopper’s behalf, an AI assistant may need access to an account, saved payment details, or other personal information. Sharing that access raises practical questions: What can the agent see? What actions can it take? Can the user review a purchase before it is placed? And how can access be revoked?

These concerns are especially important when an agent can do more than browse. A mistaken product selection is inconvenient; an unintended purchase or exposure of account information can have financial and privacy consequences. Users may also be unsure whether an agent is acting within the limits they set, particularly when a task involves recurring preferences or stored credentials.

Bank of America has warned customers that they remain responsible for activity resulting from access they grant. That puts a clear burden on the user: permission to an assistant is not the same as assurance that every action will be correct. Shoppers need understandable controls, while providers need to communicate what access means and keep users involved in consequential decisions.

Useful safeguards include requiring confirmation before payment, showing the final item and total, limiting an agent’s permissions to the task, and providing a straightforward way to stop or revoke access. These controls may add a step, but they can make AI-initiated purchases more predictable and easier to trust.

How checkout details can derail an AI-initiated purchase

Even when a user approves a purchase, merchant checkout flows can be difficult for an agent to complete. Websites may present information in changing layouts, require account-specific steps, or apply rules that depend on timing and eligibility. People often recognize what to do from context; an agent may not know whether a discount applies or what to do when an expected option is missing.

These are ordinary parts of ecommerce, not exotic technical edge cases. They also show why a checkout that works for a person may not work reliably for software. Clearer messages, consistent page structures, and explicit rules for discounts and availability can help agents understand what is happening - and help customers see why an order cannot be completed.

What merchants need to change - and what they may risk

Merchants are working to make websites and ordering flows easier for agents to navigate. Chipotle is one example of a business considering how its experience can support agent interactions. That does not mean every merchant should immediately make every transaction open to automated purchasing. It means businesses need to decide what an agent may access, what actions require customer approval, and how a transaction should be presented and completed.

A useful starting point is to review checkout from the agent’s perspective: Are product options and prices clearly labeled? Can it identify availability, fees, discounts, and ordering limits? Are errors explained in a way that supports recovery rather than leaving the process stuck? Merchants should also consider how to preserve account security and ensure that customers can verify the final order.

Executives have also pointed to factors beyond price: product quality, customer service, and brand trust can continue to influence purchases. Merchants can make information accessible to agents while still emphasizing those strengths. The goal is not simply to let software complete a checkout as fast as possible; it is to support a transaction that reflects the customer’s intent and the merchant’s service standards.

Trustworthy agents need reliable ways to complete transactions

AI shopping agents are an emerging sales channel, not yet a routine replacement for human checkout. Their adoption depends on three things working together: users must feel safe granting access, agents must handle real-world checkout rules, and merchants must provide reliable ways for those agents to interact with their systems.

For shoppers, the practical test is whether an agent makes its choices and payment request clear before acting. For merchants, it is whether checkout can communicate product, price, eligibility, and errors in a consistent way without compromising service or security. As agentic commerce develops, those fundamentals will matter more than a polished demonstration.

Explore how agentic commerce is changing online shopping, and what businesses can do to prepare their checkout experiences.

Step-by-Step Guide

1. 1

   Limit agent permissions

   Grant access only to the account information and actions needed for the shopping task, and make sure you know how to revoke that access.

2. 2

   Review the proposed order

   Before authorizing payment, check the selected product, quantity, delivery details, discounts, fees, and final total.

3. 3

   Map checkout requirements

   As a merchant, identify account-specific steps, loyalty conditions, ordering windows, availability rules, and other requirements that an agent must handle.

4. 4

   Make checkout information explicit

   Label products, prices, fees, discounts, and availability consistently, and explain errors in a way that helps the agent or customer recover.

5. 5

   Keep approval and recovery controls

   Require customer approval for consequential actions, provide clear confirmation of the completed order, and offer a straightforward way to stop or correct a transaction.

Key Statistics

• DoorDash said it had received only several hundred orders from AI agents.The article attributes this reported figure to DoorDash and gives no timeframe, so it should be treated as an indication of early activity rather than a rate or market-wide total.
• Uber described agent-initiated orders as very early and limited.This is Uber's characterization as reported in the article; no exact order count or measurement period is provided.
• Chipotle examples involved difficulties with loyalty discounts and orders placed outside normal ordering windows.These reported checkout issues illustrate how account-specific eligibility and timing rules can interrupt agent-initiated purchases.
• Bank of America warned customers that they remain responsible for activity resulting from access they grant.The article cites this warning to emphasize that granting an assistant access does not guarantee every action will be correct.

Frequently Asked Questions