PII

Personally Identifiable Information (PII) refers to any data that can be used alone or in combination with other information to identify a specific individual, including direct identifiers and indirect identifiers such as IP addresses and device fingerprints.

Removing PII such as phone numbers from audit logs has a side effect: when a specific conversation needs to be traced back, the log itself cannot be directly linked to the customer's identity, and the message ID must be used for cross-database mapping in other systems.

PII minimization is the concrete implementation of the "data minimization" principle in Article 5 of the GDPR: collect and process only the minimum data necessary to achieve a specific purpose.

Test logs, DOM snapshots, and network traces often contain sensitive payloads, OAuth bearer tokens, PII, or internal infrastructure URLs, and passing raw context to an external LLM provider means significant compliance risk, requiring a strict sanitization layer before model ingestion.

If a platform stores content containing PII as-is and cannot configure filtering rules, then in production it will become a continuously expanding data breach surface.

AI Agent trace data often directly contains user prompts, model outputs, and tool return values, and this content may contain personally identifiable information (PII), credentials, or trade secrets.

The tools that a customer service Agent can invoke (such as order lookup, email sending, refunds) make the PII it holds (name, address, email, phone, order history) leakable through the tool invocations themselves.

Raw prompts often carry PII, proprietary source code, or credential keys, and full collection may run afoul of privacy regulations such as GDPR and HIPAA.

Data sensitivity is divided into several tiers: public data can be sent freely; the compliance of internal business data depends on industry regulation (such as finance, healthcare and government affairs, which often have hard rules); personally identifiable information (PII) and trade secrets generally cannot be sent out. Even if the provider promises not to use it for training, the data still leaves the local environment.

When enterprises call large models, prompts may carry sensitive information such as user names, phone numbers, ID card numbers and bank card numbers, and once sent to a third-party model provider it may constitute a data breach or violate regulations such as GDPR and the Personal Information Protection Law.

PII（个人身份信息）指可用于识别特定自然人的数据，如姓名、邮箱、电话、身份证号、IP地址等；在LLM场景中，prompt天然可能携带用户上传的文档、对话历史、账号凭据，因此span中记录原文的风险远高于传统API追踪。

除非有严格的PII脱敏管线，否则不应在span中记录原始prompt/completion内容，因为span虽然采样率通常低于指标，但仍然会被记录。

RAGGuardrails类实现输入和输出护栏：输入检查拦截Prompt注入（检测'ignore previous instructions'等触发词）和PII泄露（正则匹配SSN、邮箱、信用卡号、电话号码）；输出检查包含幻觉检测（提取答案中的数字，检查是否在上下文中，长度超过2位的数字若不在上下文则标记为未验证声明）和PII检测，并在PII出现时用[REDACTED]替换。

The three-step protection for incoming data flows includes tenant isolation, a desensitization middleware that scrubs PII, and an asset-level credential scanner that blocks tokens or keys from entering the output stream.

Common compliance technical measures include automatically detecting and masking PII before requests reach the gateway, recording tamper-proof audit logs, and role-based access control (RBAC).

Inqueria claims to complete PII (Personally Identifiable Information) desensitization within its own system before any model accesses the transcripts.

SCIM payloads typically contain employees' personal information such as name, email, department, job level, and phone number, which constitutes PII under regulations such as GDPR and CCPA.

Inputting data containing personally identifiable information (PII) into a model that will be used for training may touch the boundaries of GDPR data processing legality.

Customer conversations involve a large amount of sensitive information (PII), and the risks related to data storage, auditing, and compliance multiply after full deployment.