AI & Human Oversight Notice

How an answer is produced

Huck retrieves relevant material from a customer’s connected knowledge and supplies permitted context to an AI model. Prompts shape the response; gates and server-side authorization determine permitted context. The resulting text is generated, not a verbatim guarantee from a system of record. Citations identify supporting source material, which can itself be incomplete, wrong or outdated.

What checks cannot prove

AI can misunderstand instructions, overlook evidence, invent details or cite material that does not support a claim. Models, source data, policies and upstream services change. Goldens and judges compare selected cases with defined expectations; judges can also make mistakes. A pass is evidence for that test configuration and time, not proof that all possible questions, users, data or actions will behave correctly.

Where a person must stay involved

Verify consequential answers against authoritative sources. Use qualified professionals for professional advice and keep a person responsible for important decisions. For actions that move money, change access, delete information, alter orders or otherwise affect people, validate arguments and authority in your own service and use appropriate confirmations. Do not depend on a chat response as proof that an external operation completed; check its confirmed status.

Responsibilities when you embed an assistant

Explain that people are interacting with AI, identify your organization and link your privacy notice. Limit the assistant to suitable topics, keep its sources current, test representative permitted and forbidden cases, and provide a route to human support. Review failures and rerun evaluations when prompts, knowledge, access rules, models or integrations change. You can customize the assistant’s name and appearance; these duties still apply.

Report a problematic Huck response with its request identifier and a concise explanation through support. Avoid sending passwords or unrelated private records. We can investigate a defect, but cannot promise that every AI error is preventable or that an action already performed by an external system can be reversed.