An AI tool call is not consent: building visitor-approved question forwarding

Описание и применение

В Tasda реализован поток подтверждения, при котором модель предлагает вопрос, а пользователь подтверждает его отправку, обеспечивая контроль над авторизацией и ограничивая передачу данных.

Задача

Ensuring that a visitor’s decision, not the AI model’s suggestion, authorizes forwarding a question to the representative’s creator.

Результат

The implementation separates model suggestion from user-initiated write, enforces context validation, limits data exposure to the short question, and ensures idempotent confirmations.

Рабочий процесс

• Model proposes question
• Server creates signed preview
• Client displays question with confirmation control
• Visitor confirms via authenticated request
• Server validates preview, context, and eligibility
• Server stores the request

Доказательства

The model can prepare a question, but the visitor decides whether to send it.

Источник

The preview expires after one hour and is signed.

Источник

Only the short question shown in the preview is stored for the creator.

Источник