Back to legal center

Platform agreements

AI and Automation Terms

Agent permissions, approvals, data, human oversight and control of legal documents.

Download bilingual agreementsRead the legal center checklist
On this page

1. Functions and boundaries

These terms cover site-building, catalog, translation, SEO, support and operational agents. AI produces drafts, suggestions or limited actions using user inputs, approved knowledge and authorized tools; output may be inaccurate, incomplete or similar to others’ output. Neither the Platform nor an agent provides professional legal, tax, medical or investment advice, and this template must not be presented as written or approved by qualified counsel. Enabled models and processing arrangements must be identified in the provider register.

2. Permissions and instruction hierarchy

Agents operate only within expressly granted stores, tools, actions and allowances. Content in web pages, emails, product files and support messages is untrusted input and cannot override Platform safeguards, formal merchant policies or access controls. Authorization must be enforced by the server and be revocable immediately; prompt text such as “refund allowed” cannot bypass separate approval or payment permissions.

3. Drafting, publishing and sensitive actions

Product copy, image descriptions, translations and SEO changes default to previewable drafts. Merchants review material public-content, price or promotion changes before publication. Refunds, payment configuration, settlement accounts, domains, bulk data deletion, expanded permissions and operative legal terms require specific approval. Authorizing routine automated support does not authorize these actions. Approval must identify the particular content, amount or action version and must be renewed if it changes.

4. Refunds and consumer rights

An agent may explain approved refund policy, collect necessary facts, retrieve orders and draft requests, but every executed refund requires approval by an authorized merchant employee. The system rechecks refundable funds and prevents duplicate refunds rather than treating predictions as payment facts. An agent must not finally deny statutory withdrawal, defect remedies, refunds or human review. Imminent deadlines, policy conflicts and rights disputes must escalate promptly with the original request time recorded so a queue does not prejudice rights.

6. Data minimization and providers

Data sent to models must be limited to the current task, using order references, masked fields or aggregates where possible. Full PANs, CVVs, passwords, private keys and unrelated personal information are prohibited. Model providers, countries, logging and training policies require verification and DPA arrangements before launch, with general-purpose model training and cross-tenant retrieval disabled by default. Merchant-supplied models or plugins require the same recipient, permission and transfer review.

7. Output verification and records

Merchants must check product facts, intellectual property, language and cultural context, pricing, stock, delivery and advertising claims. The system records necessary input references, knowledge versions, tool calls, approvals, outcomes and reversals with sensitive fields masked and limited retention. Low-risk changes should be versioned for rollback; financial actions that cannot simply be reversed require dedicated processes rather than a retry used in place of idempotency and reconciliation.

8. Disclosure, human handoff and stopping

Consumer-facing conversations must identify the AI assistant at the start and provide an available human contact path and actual service hours. Merchants may pause agents, revoke tools or take over manually; disabling AI cannot erase complaints or requests already received. Suspected unauthorized action, data leakage, duplicate transactions or material errors require stopping the affected task and notifying the responsible people. Responsibility for AI errors remains based on actual conduct and law; these terms do not excuse either party’s own fault.