Skip to main content

What is PII Redaction?

PII redaction automatically detects and removes sensitive information from both transcripts and audio recordings. This ensures compliance with privacy and security requirements while maintaining the overall context of conversations. This page covers server-side redaction: you send the call to Cekura with redact_fields, and Cekura detects and removes the entities from the stored transcript and recording. It is the option to use for audio, and for open-vocabulary entities such as names and locations that need model-based detection.
If raw PII must never leave your network, or your organization requires that its own model providers do the detection, see Client-Side Redaction — a script that detects and removes PII from both the transcript and the recording, using your own LLM and speech-to-text providers, before anything is sent. It also covers metadata, dynamic_variables and customer_number, which redact_fields does not reach. The two approaches can be combined.

Example

Original:
Redacted:
Audio Example:

How It Works

We identify the entities to be redacted from the call using state of the art models. We then replace the text with relevant placeholders.Example tokens:
  • <PERSON> - Person names
  • <LOCATION> - Geographic locations
  • <EMAIL> - Email addresses
  • <PHONE> - Phone numbers
  • <SSN> - Social Security Numbers
  • <CREDIT_CARD> - Credit card numbers

Using Redaction

Redaction can be configured at two levels:
Configure redact_fields on your agent via the agent API. Once set, redaction applies automatically to every call ingested for that agent — including auto-imported and webhook calls — without any per-call parameters.Set redact_fields inside the provider block when creating or updating an agent:
To disable agent-level redaction, set the field to an empty array:
When redact_fields is omitted from a per-call request, the agent’s configured redaction fields apply. If the agent has no redaction fields configured either, no redaction occurs.
redact_fields applies to transcript_json and the audio recording. Values in metadata, dynamic_variables and customer_number are stored exactly as you send them — scrub those before sending.

Available Fields

  • person_name - Person names
  • age - Age information
  • phone_number - Phone numbers
  • email_address - Email addresses
  • date - Specific dates
  • time - Time information
  • location - Physical locations and addresses
  • origin - Origin information
  • gender_sexuality - Gender and sexuality information
  • physical_attribute - Physical descriptions
  • occupation - Occupation details
  • username - Usernames
  • password - Passwords
  • ip_address - IP addresses
  • url - URLs
  • filename - File names
  • event - Event information
  • vehicle_id - Vehicle identification numbers
  • dob - Date of birth
  • healthcare_number - Healthcare identification numbers
  • medical_professional - Medical professional information
  • credit_card - Credit card numbers
  • account_number - General account numbers
  • bank_account - Bank account numbers
  • money - Financial amounts
  • ssn - Social Security Numbers
  • passport_number - Passport numbers
  • driver_license - Driver’s license numbers
  • numerical_pii - Other numerical PII

Pricing

Redacting any fields costs 0.1 credits per minute by default, regardless of how many fields are selected.

Implementation Examples