Checkout, But Make It AI: Legal Considerations For Shopping Agents

Artificial intelligence (“AI”) shopping agents are increasingly being used to assist consumers in finding, comparing and purchasing products online. These tools are moving beyond conventional search functionality, with retailers deploying conversational and image-aware assistants that can browse, compare and complete transactions with limited human intervention. The convenience is clear, but as organisations adopt these tools faster than regulators can update governing frameworks, important questions arise about data collection, use and responsibility when things go wrong.

For South African businesses, these developments raise questions that existing law does not expressly address. The Protection of Personal Information Act, 2013 (“POPIA”), Consumer Protection Act, 2008 and Electronic Communications and Transactions Act, 2002 all remain relevant, but none were drafted with AI shopping agents in mind. The same difficulty is apparent globally.

Current AI shopping tools are increasingly designed to take steps in the purchasing process, including comparing products, selecting options and initiating checkout. This shift from AI as a search tool to AI as an executing agent is where the governance questions become acute. In practice, businesses are already having to decide how prominently to disclose AI interaction, when to offer human escalation, how far personalisation should go, and whether existing privacy policies adequately describe the AI processing. Privacy policies drafted before AI shopping agents existed are unlikely to cover new processing purposes, meaning documentation updates are often a pre-launch requirement.

Privacy risks and voluntary disclosure

Conversational interfaces may encourage users to share more detailed, natural-language requests than keyword searches would capture. Chatbot logs may therefore become as sensitive as transactional data, particularly where retained or linked to customer profiles. Organisations should establish clear retention periods for different data categories: raw conversation logs, anonymised data and profile information each warrant distinct treatment.

A distinct challenge arises from voluntary disclosure. Unlike structured e-commerce forms, conversational AI invites open-ended input. Customers may spontaneously disclose medical conditions, religious practices or other sensitive details the organisation never requested. Under POPIA, organisations bear responsibility for personal information in their possession regardless of whether they solicited it. If a customer voluntarily shares dietary requirements revealing a health condition or religious affiliation, that may constitute special personal information under POPIA section 26, triggering additional compliance obligations.

Options to address this include implementing input filters that detect and redact sensitive content before processing, limiting retention to the minimum necessary, and ensuring privacy policies disclose whether provision of personal information is voluntary or mandatory as required by POPIA section 18. Image-based search compounds this exposure, as photograph uploads may inadvertently contain faces, home environments or identifying details, and may reveal health-adjacent information through allergy or dietary queries.

Applying South African law to AI shopping agents

POPIA section 71 protects data subjects from decisions based solely on automated processing where the decision has legal consequences or significantly affects them. This may become relevant where AI shopping agents set personalised pricing or extend credit without meaningful human input. However, where an AI only advises and the customer confirms any action themselves, no decision is made by the AI alone. Organisations should document in their privacy policies which steps are automated and confirm that none constitutes solely automated decision-making.

Cross-border transfer risk is a further consideration. Many AI tools are hosted outside South Africa. Feeding customer data into an external AI platform may constitute a transborder transfer under POPIA, requiring contractual protections with the operator. Where special personal information is being transferred, POPIA section 57 requires prior authorisation from the Information Regulator before processing occurs – a separate pre-launch gate independent of section 72 contractual safeguards.

Under POPIA sections 20 and 21, a written agreement is required with any operator processing personal information. For cloud-based AI services, businesses should verify that the applicable data processing agreement covers the specific project, includes a no-training commitment preventing customer data from being used to improve the provider’s models, and provides for deletion on instruction within defined timeframes.

Key implementation considerations

Customers should be informed when interacting with AI. This disclosure should appear within the chat interface itself, not merely in the privacy policy. A persistent indication that the customer is interacting with AI, supported by content covering what the assistant can and cannot do, how recommendations are generated and how to reach a human representative, reflects emerging best practice.

Where cross-session memory is implemented – allowing the AI to retain information from previous interactions – it should be off by default with customers actively choosing to enable it. Under South African law, consent must be voluntary, meaning customers cannot give valid consent if memory retention is always on and cannot be turned off.

Topics requiring human escalation should be identified. Queries relating to refunds, cancellations, consumer rights or allergens should not be answered by freely generated AI responses. The Air Canada chatbot case, where the airline was required to honour a refund policy invented by its chatbot, illustrates that generic accuracy disclaimers are unlikely to be adequate where customers rely on AI responses when deciding to transact. The AI should only respond to policy questions from approved content sources and hand off to a human where the query falls outside that grounded content.

AI-generated outputs should be grounded in approved sources wherever they concern pricing, returns, refunds or product availability. Output filters that block language creating apparent legal commitments unless retrieved from approved policy content can reduce the risk of the AI fabricating commitments the business cannot honour. Organisations should also consider pricing-error clauses permitting cancellation where the AI has quoted incorrect prices.

Looking ahead

The current legal frameworks were not designed with autonomous purchasing agents in mind. Businesses deploying AI shopping tools are applying existing consumer protection, privacy and cybersecurity obligations to a new commercial reality. The governance questions require documented answers before any customer pilot, covering AI disclosure, human escalation, data retention, cross-border transfers, operator agreements and special personal information handling. For organisations considering these tools, the legal review should begin early in the design, procurement and deployment process. Our team can assist with privacy impact assessments, AI governance reviews, data flow mapping, updates to privacy policies and terms and conditions, operator agreement review, prior authorisation applications to the Information Regulator, and appropriate contractual safeguards.

Your pre-launch checklist

• Update privacy notice to disclose AI processing and the categories of data collected
• Implement persistent AI disclosure within the chat interface
• Establish data retention periods for conversation logs, anonymised data and customer profiles
• Implement input filters to detect and redact sensitive content before processing
• Document which steps in the customer journey are solely automated (POPIA s71)
• Assess whether prior authorisation is required
• Verify operator agreement includes no-training commitment and deletion obligations (POPIA ss20-21)
• Ensure cross-session memory is off by default with opt-in consent
• Identify topics requiring human escalation (refunds, cancellations, consumer rights, allergens)
• Ground AI responses on policy topics in approved content sources only
• Implement output filters to block language creating apparent legal commitments
• Consider pricing-error clause in terms and conditions
• Conduct and document a privacy impact assessment

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.