Why AI Agents Abandon Complex Checkout Flows

An AI agent has done everything right. It found the product the user wanted, confirmed the price, checked stock availability, and added the item to the cart. Then it hits the checkout flow: six pages of forms, dropdowns, radio buttons, address validation, card entry, and "would you also like to add..." upsells. This is where things fall apart.

The Failure Rate

We ran 500 end-to-end purchase attempts across 50 e-commerce sites using three different AI agent frameworks. The results are stark:

• Product discovery success: 89%
• Add to cart success: 76%
• Checkout completion success: 23%

Agents succeed at finding and carting products far more often than they succeed at buying them. The checkout flow is where three-quarters of successful shopping journeys die.

What Goes Wrong: Step by Step

Step 1: Guest checkout vs. account creation. 34% of sites we tested require account creation before checkout. The agent needs to fill in name, email, password, possibly verify the email (which requires switching to an email inbox), and accept terms. Each field is a potential failure point. Password requirements ("must contain a special character, uppercase letter, number, and be 8-12 characters") trip agents up because the requirements are often communicated as a small text hint below the field or, worse, only as an error message after submission.

Step 2: Address entry. Address forms are deceptively complicated. The agent needs to fill in name, line 1, line 2 (optional?), city, county/state, postcode. Some forms auto-complete addresses from a postcode lookup and require the agent to select from a dropdown of suggestions. Others validate addresses against a database and reject input that doesn't match. International addresses are worse: the form expects different fields for different countries, and the layout might change dynamically based on the country selection.

We watched agents get stuck on address validation loops. The agent enters a valid address, the form rejects it because the postcode format doesn't match the validator's expectations, the agent tries again with slight variations, and eventually gives up after five attempts with a perfectly valid address.

Step 3: Delivery options. Radio buttons or dropdown selectors for delivery speed and type. These are usually manageable, but some sites present delivery options as a map interface ("click to select your nearest collection point") or a calendar widget ("select your preferred delivery date"). Map interactions and calendar date pickers are notoriously difficult for agents to operate.

Step 4: Payment. This is the hardest step. Entering card details into payment forms, especially those hosted in iframes by Stripe, PayPal, or other payment providers, requires interacting with cross-origin iframes that agents often can't access. The agent can see the iframe exists but can't inspect or interact with its contents due to same-origin policy restrictions.

Even when the agent can access the payment form, card entry is fragile. Some forms auto-advance from card number to expiry date. Others expect the expiry in "MM/YY" format while the agent types "03/2026." CVV fields sometimes block paste operations. The 3D Secure verification popup that appears for some cards adds another interaction the agent must handle.

Step 5: Order review and confirmation. After filling in everything, many sites show an order review page with a final "Place Order" button. Some agents miss this step, thinking the previous page was the final one. Others click "Place Order" and get redirected to a 3D Secure challenge or additional verification step.

The Upsell Trap

Between steps, many checkout flows insert upsell offers. "Add gift wrapping for £3?" "Would you like to insure this item?" "Customers also bought..." These interstitials require active dismissal. The agent needs to click "No thanks" or "Continue without" to proceed. Some upsells are designed as modals that overlay the checkout form. Others rearrange the page layout, pushing the "Continue" button to a new position.

We counted an average of 2.7 upsell interruptions per checkout flow. Each one is a chance for the agent to get confused, click the wrong button, or accidentally add an unwanted item.

Sites That Agents Can Handle

The checkout flows with the highest agent success rates share common features.

Single-page checkout. Everything on one page: address, delivery, payment. No multi-step wizard. The agent fills in fields and clicks one "Place Order" button. Amazon's "Buy Now" button is the gold standard here; one click and the order is placed using saved details.

Stored payment methods. Sites that support saved cards, Apple Pay, Google Pay, or PayPal Express let agents skip the manual card entry step entirely. The agent clicks a saved payment method and confirms.

Minimal required fields. Every form field is a failure point. The sites with the best agent success rates ask for the minimum: delivery address, payment method, done. No phone number ("in case of delivery issues"), no company name, no "how did you hear about us" survey.

Consistent form patterns. Building agent-friendly forms means labels associated with inputs via for attributes, error messages linked to specific fields via aria-describedby, clearly marked required fields, and standard HTML form elements instead of custom JavaScript widgets. These patterns help agents (and screen readers) understand what each field expects.

The Business Case

AI agents are increasingly being used for purchasing. We've seen how one retailer doubled agent-driven sales by streamlining this process. Our research suggests that 8% of online purchases in early 2026 involve some form of AI agent assistance, whether that's comparison shopping, finding deals, or completing transactions. A checkout flow that agents can't complete is a checkout flow that's losing a growing segment of buying intent.

Simplifying checkout isn't just about AI agents. The same changes that help agents, fewer steps, fewer fields, clearer forms, also reduce cart abandonment rates for human users. The Baymard Institute's research consistently shows that checkout complexity is the second most common reason humans abandon carts (after unexpected costs). Fixing it for agents fixes it for everyone.