AI in an ecommerce conversion journey: four bounded scenarios with human decisions

Ecommerce conversion AI applications should begin as narrow, reversible tasks for an England store, not autonomous decisions about customers. No tool was tested for this article, and no evidence reviewed establishes adoption, accuracy, conversion uplift or commercial benefit.

The current boundary is clearer than the outlook. CMA guidance for businesses using AI agents, published on 9 March 2026, says businesses remain responsible when an agent acts unlawfully. It discusses marketing, refunds and customer service, with appropriate human review and prompt correction. This is observed UK guidance, not a forecast of what merchants will deploy by 2027.

What to take away

Four bounded application scenarios

Product-page discrepancy triage. A system could compare a draft page with an approved evidence record and flag missing attributes, warnings or price components for a trained editor.

The human product owner approves or rejects every change. It must not invent claims, infer product safety or publish automatically. For goods, OPSS product-safety advice is a current official boundary, not training data that makes an output compliant.

Worked example. Inputs: an approved evidence record, a draft page and current price data. The model flags a missing warning and a price mismatch. It must not publish or invent claims. The product owner approves each flag.

Event-quality investigation. A model could group missing, late or duplicate journey events for an analyst to inspect.

The source record must preserve product view, basket, checkout, order placed, payment authorised, payment captured, fulfilled, cancelled, returned, refunded and disputed as different states. The analyst owns the correction. The system may not fill refused tracking, assign a customer identity or declare causal impact.

Experiment preflight. An assistant could compare a proposed test record with required fields: eligibility, allocation, actual exposure, primary outcome, outcome window, missingness, contamination, stopping rule and rollback. A statistician approves the design. HM Treasury's QPIE methods guidance, updated 15 May 2026, supports distinguishing stronger counterfactual designs from weak attribution. It supplies no store benchmark.

Support-draft preparation. A system could draft a response from an approved policy and the stated order state. A trained agent verifies identity, facts, rights, accessibility and the final answer. It must not decide whether a customer is legally entitled to a refund, alter money records or close a dispute without the authorised human process.

Controls before any trial

Before any trial, write down who owns the task, the system name and version, and the supplier. List permitted inputs, prohibited data and prohibited actions. Name who receives outputs, who decides, how long records are kept and how the task is monitored. Set the incident route and the manual baseline. Use synthetic or specifically authorised records.

The ICO's AI and data protection risk toolkit covers rights and freedoms but is under review after the Data (Use and Access) Act; recheck with privacy lawyer before use.

Predeclare test cases for false facts, missing evidence, refused tracking, accessibility, prompt manipulation, unauthorised access, price changes and each relevant payment outcome. NCSC's secure AI development guidance covers design, development, deployment, operation and maintenance. It does not certify a model.

Outlook record: these four uses are buyer scenarios dated 6 September 2026, with confidence left blank for performance. The observable indicator is a controlled error or successful evidence trace. The consequence is admit, revise or stop the exact task.

15 January 2027 is this article's scheduled recheck point. Recheck before each model, prompt, data-purpose or supplier change and on 15 January 2027. Rollback means restoring the approved manual route and reconciling affected customer and order records.

Before you act

Common questions

What does the CMA guidance say about responsibility when an AI agent acts unlawfully?

The current boundary section sets out the CMA position on responsibility when an agent acts unlawfully.

What controls should be in place before any trial of AI in a store journey?

The controls section above lists what to name before any trial. It covers the task owner, inputs, prohibited actions and decision authority.

What are the four bounded application scenarios described in the article?

Product-page discrepancy triage, event-quality investigation, experiment preflight and support-draft preparation. Each keeps a human owner for every decision. The system must not invent claims, fill refused tracking, declare causal impact or decide legal refund entitlement without the authorised human process.