Privacy & guardrails
Protect sensitive customer data
You can now equip your agents with privacy and safety guardrails. Everything is off by default — enable what you need per agent. Find the settings under Agent → Advanced → Privacy & guardrails.
PII redaction
Sensitive caller data — BSN (citizen service number), IBAN, credit card, email address, phone number and date of birth — is automatically masked in the stored transcript, call logs, conversation analysis and outgoing webhooks. Instead of the actual value you'll see [BSN] or [DATE_OF_BIRTH]. The agent continues to work normally during the call; only what is stored and shared is protected.
Detection is precise: a BSN is validated with the eleven-test, an IBAN with the mod-97 check and a credit card number with the Luhn algorithm, so random digit strings are not unnecessarily masked. You choose per category what gets masked, and you can add your own patterns via a regular expression — useful for file numbers or internal customer IDs.
Also keep data away from the language model
Go one step further: you can also mask selected categories before they reach the language model. This setting is category-selective — for example, you can redact a date of birth from storage while keeping it visible to the agent so identity verification continues to work. Off by default.
Identity verification without the AI seeing the value
The new tool "Verify identity" lets the agent check a value — such as a date of birth — without the language model ever seeing it. The caller speaks the value, the platform compares it in code to the expected value (for example from your CRM), and the agent only hears whether it matches or not. Comparison is flexible: "March 12, 1985", "12-03-1985" and "1985-03-12" are treated as equal. This combines thorough verification with maximum privacy.
Find the tool under Agent → Tools → Verify identity.
Manipulation protection
Enable protection so your agent cannot be talked out of its role. The agent ignores attempts to change its instructions, reveal its internal workings, or impersonate another system.
Abuse filter
Limit abusive or threatening behaviour. Choose whether the agent calmly sets limits and continues, or politely ends the call after a warning.
Important notes
- Redaction applies to text (transcript, logs, analysis, webhooks). The audio recording still contains the spoken value.
- Phone numbers of the caller and called party are not redacted, as they are operationally required for callbacks or CRM lookups.
- Settings apply to new conversations from the moment you enable them.