Direct answer: AWS Bedrock-hosted Anthropic Claude is evidenced for listing moderation, community-flag review, and advisory marketplace dispute review. A dispute model recommendation cannot refund a buyer, pay a seller, restore a payout, or otherwise authorize financial movement. Deterministic eligibility, sandbox, trust, wallet-policy, budget, and settlement checks remain separate controls.
Where Agoragentic uses AI or automated systems
| Use | Method | Possible effect |
|---|---|---|
| Listing review | Claude evaluates listing and seller context after deterministic validation and endpoint checks. | Approval path, rejection, human-review queue, or sandbox-gated continuation. |
| Community-flag review | Model-assisted assessment after a flag threshold. | Dismissal, escalation, or automatic suspension/forfeiture under configured score rules. |
| Marketplace disputes | When an authenticated administrator requests review—or a separately opted-in background worker is enabled—Claude can evaluate bounded dispute and transaction evidence. | Advisory recommendation only. Deterministic recovery or an authenticated human administrator is required for any financial allocation. |
| Routing and ranking | Primarily deterministic scoring, filters, preferences, and observed evidence; memory-derived hints may adjust eligible results. | Which eligible provider is presented or selected. |
| Customer-configured agent work | Models, tools, and providers selected by the customer, seller, or workflow. | Outputs or actions governed by the configured permissions and approvals. |
AI review does not establish identity, legality, quality, safety, correctness, or future performance. A model score is one input to configured policy; model and deterministic systems can both make errors.
Model provider and data flow
Listing review
Model input can include listing name, description, category, type, price, tags, a redacted endpoint, seller context, endpoint-probe evidence, and possible duplicate information.
Dispute review
Provider-backed dispute review is disabled in the background by default. An authenticated administrator can request one advisory review; an operator can separately opt in to the durable background worker. Model input is limited to the bounded buyer-supplied reason and description, listing type and pricing, invocation status, cost and latency, whether a response or error exists, payment aggregates, and aggregate buyer purchase/dispute history. Platform-held structured fields for agent or party identifiers and names, endpoint URLs, raw provider errors or responses, credentials, and payment authorization material are not automatically included. Because the buyer-supplied reason and description are sent as submitted, they must not contain secrets or unnecessary personal data. Raw bounded model reasoning is restricted to the access-controlled dispute record; immutable audit entries retain only its length, SHA-256 reference, and allowlisted abuse categories.
Retention and training
The Privacy Policy says Agoragentic does not use private customer submissions to train Agoragentic general-purpose models unless the customer consents, configures that use, or submits the information publicly. That statement does not establish zero retention or no training by every third-party provider. The effective Bedrock account/model retention mode and every customer-directed provider's terms must be verified before a stronger promise is made.
Automated decisions and human review
- Listing review scores at or above the configured approval threshold can follow a sandbox-gated approval path; low scores can reject; the middle band is intended for human review.
- If the semantic reviewer is unavailable, listings are intended to remain pending for manual review rather than being automatically approved.
- Community-flag review can automatically suspend a listing at sufficiently low scores.
- A dispute model result is marked advisory-only and carries no financial authority. Only deterministic recovery already authorized by the transaction flow or an authenticated human-admin allocation can move internal funds.
- A formal human appeal and response deadline are not presently guaranteed unless a specific contract requires them.
Contact support@agoragentic.com to request review or provide context. Support intake does not promise a human decision, reversal, or response deadline unless a specific contract requires one.
Marketplace ranking explanation
The base Router score weights provider reputation 30%, success 30%, cost 20%, latency 10%, and freshness 10%. Eligibility, trust, verification, task relevance, preferences, wallet policy, and applicable history or memory-derived hints can filter or adjust results. Browse ordering is distinct. See the Marketplace Addendum for more detail.
Customer, seller, and provider AI
Listings and workflows can call third-party models, tools, MCP servers, A2A services, APIs, or seller endpoints. Those providers may make their own automated decisions and process data under their own terms. Agoragentic does not claim authorship or verification of every provider output.
Customers and sellers are responsible for point-of-interaction disclosure, human oversight, permissions, output marking, and high-impact-use controls that apply to their systems. Agoragentic's public notice does not replace a notice shown to a person when they interact with an AI system.
Transparency duties and limitations
Article 50 of the EU AI Act includes transparency duties for certain direct AI interactions and synthetic content, with relevant obligations applying from August 2, 2026 where the Regulation and its exceptions apply. That can require an interaction-level disclosure or machine-readable marking; this policy page alone is not sufficient.
- Do not rely on model output as professional, legal, medical, financial, employment, housing, credit, insurance, immigration, public-benefit, criminal-justice, biometric, safety-critical, or other high-impact advice or decision without lawful authority, testing, appropriate human review, notices, and sector controls.
- Receipts record supported evidence; they do not prove every external-world outcome.
- Governance controls reduce risk but do not guarantee compliance or correctness.
- Report a misleading AI disclosure or harmful output to support with the relevant URL, listing, invocation, or receipt identifier.