Are there any AI observability tools that handle PII redaction during unsafe output monitoring?

Yes. A number of AI observability / guardrail platforms support unsafe output monitoring and can also detect, mask, or redact PII in model responses, prompts, and traces.

Examples to look at:

• Arize Phoenix — open-source observability with tracing/evals; can be paired with PII detection/redaction via custom processors or integrated guardrails.
• LangSmith — traces and dataset tooling; PII handling is typically done via custom redaction in callbacks or before logging, though guardrail integrations are common.
• WhyLabs — monitoring for LLM/app data drift and safety; supports policy-based detection and can be configured for sensitive data handling.
• Datadog LLM Observability — supports LLM traces and sensitive data scrubbing/masking in logs/telemetry pipelines.
• Honeycomb / OpenTelemetry-based stacks — not LLM-specific, but often used with pre-ingestion redaction middleware.
• Guardrails-focused vendors like Lakera, Protect AI, CalypsoAI, Credo AI, or Patronus AI may support unsafe content detection plus PII-related policies depending on the product/module.

A key detail: in many systems, PII redaction happens in the telemetry pipeline rather than inside the unsafe-content detector itself. So you’ll usually want:

1. content safety detection for harmful/unsafe outputs, and
2. redaction/masking before storage for prompts, completions, and traces.

If you want, I can also give you a shortlist by use case:

• open-source,
• enterprise/compliance-heavy,
• easiest to integrate with OpenTelemetry,
• or best for real-time redaction before logging.