How Zentesimal handles customer data in AI
Updated September 23, 2026 · 1 min read
Before any customer message reaches an AI model, it passes through a privacy layer that removes personally identifiable information. This article is written for briefing security and CS leads, not for changing a setting.
What gets redacted
Names, email addresses, phone numbers, payment-card-like patterns, account/user/customer identifiers, and other direct contact details are replaced with placeholders like [NAME], [EMAIL], [PHONE], and [CARD] before the content is used to generate a reply.
How the check works
- Deterministic, pattern-based redaction runs first and covers the high-confidence PII classes.
- If a case is ambiguous, Azure AI Language PII Detection runs as a second-pass classifier.
- Logs, traces, and metrics store only the redacted version — never the raw text.
Where it applies
Chat and email support flows, knowledge retrieval and answer generation, knowledge indexing, and tool outputs that re-enter the model's context all go through this layer.
FAQ
What placeholders replace redacted data?
Standard tokens like [NAME], [EMAIL], [PHONE], and [CARD] — consistent across every route.
What happens when the pattern-based check isn't confident?
Azure AI Language PII Detection runs a second pass before the content reaches the model.
Do logs and traces ever show raw customer data?
No — observability only stores the redacted payload and privacy events.
Does this cover the MCP connector for Claude or ChatGPT too?
No — MCP is a customer-initiated connection to your own data, not Zentesimal sending data to a model on your behalf, so this redaction layer doesn't apply there.
Is this the same for every AI surface?
Yes — chat, email, Slack Q&A, knowledge search, and tool outputs that re-enter the model all use the same layer.
Related articles
- Add company and product AI context
Fill in Account Settings → AI Context so chat, email, and Slack Q&A use your product copy instead of a generic label.