AI-Assisted Operations

Summary in one sentence

EBLIT may use artificial-intelligence tools (currently Anthropic's Claude) to help our staff work faster -- but a human reviews and makes or approves EVERY consequential decision, and no consequential action is taken against you, your account, your transactions, or your funds by AI alone.

What a consequential decision means

A consequential decision is any action that affects you or your money -- for example: suspending, restricting, or closing an account; issuing a strike or ban; blocking, cancelling, or reversing a transaction; deciding a dispute; releasing, withholding, or refunding funds held pending confirmation; or removing a listing. EBLIT commits that NO CONSEQUENTIAL DECISION IS EVER MADE OR EXECUTED BY AN AUTOMATED SYSTEM WITHOUT A QUALIFIED EBLIT STAFF MEMBER REVIEWING AND APPROVING IT FIRST. Human-in-the-loop applies at all times for consequential outcomes.

Automated flagging vs. automated deciding

• Flagging may be automated. EBLIT's systems (including AI tools) may automatically surface or flag items for a human to review -- for example, drafting a suggested support reply, drafting a suggested dispute-evidence summary, or flagging a listing or an identifier pattern as potentially requiring moderation.
• Deciding is never automated. A flag is only a prompt for human review. The decision and any user-facing action are made by a person. A flag, score, or AI suggestion, standing alone, never suspends an account, blocks a transaction, moves money, or decides a dispute.

Some of these capabilities may be limited, staged, or not yet active at any given time; this disclosure commits the safeguards above WHENEVER such tools are used, whether or not a given capability is currently live.

Where EBLIT uses AI

• Dispute resolution drafting — AI drafts a suggested response or evidence summary for the EBLIT trust + safety reviewer; the reviewer edits, approves, or rejects before anything reaches Stripe or the parties.
• Customer support response drafting — AI suggests a response template for the EBLIT support agent; the agent edits and sends.
• Listing quality assessment (forward-tracked) — AI flags potentially incomplete or misleading listings for human moderation before publication.
• Identifier validation cross-reference (forward-tracked) — AI checks seller-provided IMEIs/VINs against known patterns for human moderator review.

What human-in-loop means

Every AI suggestion enters an internal queue (Claude Actions queue). An EBLIT staff member must explicitly approve, edit, or reject the suggestion before any user-facing action happens. This is enforced at the system level (Convention 19 of EBLIT's engineering discipline) and is auditable in the platform's admin audit log.

• Approve — staff agrees with AI suggestion; downstream apply executes.
• Modify — staff edits AI suggestion before applying.
• Reject — AI suggestion discarded; no user-facing action.

If the staff queue exceeds its expiry window without review (typically 4-14 days depending on action class), the suggestion auto-expires and is never applied.

What the AI does NOT do

• The AI does NOT make automated decisions about you, your account, your transactions, or your dispute outcomes. GDPR Article 22 alignment.
• The AI does NOT send messages to you or other users. Staff review and send.
• The AI does NOT issue refunds, transfer funds, or take any payment action.
• The AI does NOT close, suspend, or ban accounts.
• The AI does NOT see your password, your government ID images, your photographs of items, or your biometric data. The AI only receives the minimum text context needed for the specific suggestion (e.g., dispute description text + EBLIT policy context for dispute drafting).

Data handled by AI

When AI is invoked for a specific operation, the model receives:

• The relevant text content of the operation (e.g., the dispute description, the support ticket text, the listing description).
• The EBLIT policy context relevant to that operation (e.g., refund policy summary, listing requirements).
• NO password, payment card, biometric, or government-ID data.
• NO data from unrelated users or transactions.

Data sent to the AI is processed by Anthropic (the AI provider). Anthropic operates as a sub-processor to EBLIT under a Data Processing Agreement; EBLIT's other sub-processors include Stripe (payments and identity verification), EasyPost (shipping), and Amazon Web Services (infrastructure, including Amazon Chime for LiveVerify). Anthropic's own retention and use policies apply; their relevant terms are at anthropic.com.

Your rights

You have the same rights with respect to AI-assisted decisions as you do for any other EBLIT operation:

• Right to be informed (this page).
• Right to access (request a copy of any AI-assisted action recorded against your account; see /account/data-export).
• Right to object to processing (contact support).
• Right to human review — ALREADY structurally guaranteed by Convention 19; every AI suggestion is human-reviewed before user-facing action.
• Right to lodge a complaint with your supervisory authority.

Model and provider

EBLIT's AI tooling is currently configured for Anthropic's Claude models (Claude Haiku 4.5 for support drafting and Claude Sonnet 4.6 for dispute-evidence drafting) accessed via AWS Bedrock. AI-assisted drafting is presently disabled in production; EBLIT updates this page when the production configuration changes materially. Model selection may change as the model landscape evolves; the current configuration is sourced from backend/src/config/aiCosts.js. EBLIT will update this page when the production model changes materially.

Questions

If you have questions about how AI was used in a specific interaction with EBLIT, contact support through the platform with the transaction or dispute ID. Our staff can show you the AI suggestion, the staff edit (if any), and the final action taken — the full audit trail is logged.