When the Agent Abandons the Cart: Task Breaks, Autonomous Baskets and Repair

What happens when an agent abandons the cart?

When an agent abandons a cart there is no hesitation involved; there is a break. When a person abandons a cart the reason is usually doubt: an unexpected shipping fee, a forced registration, a long form. An AI agent does not change its mind, it gets stuck. The task was "add this product to the basket and go to checkout"; when the agent meets a modal that will not close, a button with no accessible name or stock buried in JavaScript, it cannot finish and tells the user "I could not complete the transaction on this site". The outcome looks the same, an incomplete order, but the cause and the fix are completely different.

This article extends the March 2017 cart abandonment guide into the agent era. The human-side sections stay: reasons for abandonment, email structure, timing, channels and measurement. We added a second layer on top. Our thesis at Webtures: in the agent era cart abandonment work splits into two jobs, winning the person back and repairing the step where the agent broke; the second is solved with code, not email.

Why do people abandon carts?

The classic reasons have not changed and their order has been stable for years. Unexpected cost and friction at the payment step lead:

• Unexpected extra charges. Shipping, tax and service fees appearing at the last step.
• Forced registration. Long sign-up flows with no guest checkout.
• Errors at the payment step. Flows that stall on card verification with unclear error messages.
• Lack of trust. Return terms, delivery time and contact details not visible.
• Comparison behaviour. Users treating the basket as a wish list while comparing other sites.

The remedies have not changed either: show shipping and tax early in the basket, open guest checkout, write payment errors clearly, move the return and delivery promise into the checkout step. A cart abandonment email does not replace those fixes; it sits on top of them.

At which step does an agent break?

Break points on the agent's task trace are technical and repeatable. The ones we see most often:

You can read this table like a session recording: run a defined task with an agent, log every step, and repair the element at the step where the curve drops. We described this in detail in agent trace maps, and the button-scale detail in our CTA article.

How do you build cart abandonment work in the agent era?

1. Separate the two funnels. Human abandonment and agent breaks do not belong in one report. Tag agent-sourced sessions and track the success rate of "add to basket" and "proceed to checkout" separately for them.
2. Run agent task simulations. Once a month, test with real agents on three defined tasks (find the product and add it, change the variant, reach checkout) and record the break step.
3. Keep the classic flow on the human side. First email within an hour, second the next day, third on day three; reminder and trust information in the first two, a discount only at the end.
4. Move charges earlier. Show shipping and tax in the basket, on the product page and in the schema. That reduces human abandonment and lets the agent meet the delivery constraint.
5. Measure the autonomous basket rate. The order conversion rate of baskets an agent built on its own is a separate metric from human baskets. It is the first indicator of agent-driven revenue.
6. Review your pop-up policy. An exit-intent pop-up may raise human conversion and stop an agent at step one. At minimum it must close by keyboard and not obstruct the DOM.
7. Measure the repair. Rerun the same task simulation after the fix and check whether the curve changed. You cannot claim the effect of a repair you did not measure.

Can an agent-sourced cart be recovered?

Partly. When the agent drops the basket and returns to the user you may not have an email address: the user may never have signed in, permission may not exist, contact details may sit with the assistant. So "recovery" on the agent side is mostly repair rather than marketing: fix the element that caused the break, make the product findable again, and get onto the list on the next query. Where you can build the customer relationship, when the order completes on your own site, the classic email flow applies; we covered its structure in loyalty when an agent shops.

A caution on measurement: part of what an agent recommends but does not complete never appears in analytics. Clicks without a referrer dissolve into "Direct", and so do sessions where the agent stalled on your site without creating an order. We described that gap in dark AI traffic. Task simulation does not replace analytics data, but it shows what analytics cannot see.

What are the most common mistakes?

• Reading an agent break as "low conversion". Raising campaign budget does not fix a modal that will not close.
• One funnel report. When human and agent sessions sit in the same chart neither is read correctly.
• A discount in the first email. It burns margin early; a reminder is usually enough.
• Showing charges at the last step. The place where both people and agents stall most often.
• Aggressive pop-ups. They annoy people and stop agents.
• Repair without simulation. A team that cannot show what improved reproduces the same fault in the next release.

How does the Webtures approach work?

We build cart abandonment work on two tracks. On the human side, the classic structure: fee transparency, guest checkout, a three-stage email flow, channel mix and measurement with a control group. On the agent side, task simulation: defined tasks run with real agents, step-by-step logging, diagnosis of the break point and re-measurement after the repair. The two tracks produce one table: at which step is whom losing, and why?

You can see whether agents can operate your site with the Agentic Commerce Readiness scan and set up task simulation and the repair plan with our Agent Experience team.