Automatically redact personally identifiable information from transcripts.

Overview

The PII redaction feature must be subscribed to by contacting our sales team.

Transcript with PII redaction and evaluation results
Review transcripts with sensitive customer information redacted.

PII redaction automatically detects and removes personally identifiable information from your production transcripts. When enabled, data like names, email addresses, phone numbers, and financial details are replaced with generic markers (eg: [NAME], [EMAIL_ADDRESS]) before transcripts are stored. Sensitive user information never persists in your conversation logs.

How it works

PII is detected using a machine learning model hosted within Voiceflow's infrastructure.

PII is detected using a machine learning model hosted within Voiceflow’s infrastructure. Your data is never sent to any third-party service during the redaction process.

Key behaviours:

  • Production only. Redaction applies only to production conversations. Development and test conversations are not redacted, so you can debug freely.
  • Not retroactive. Enabling redaction does not affect existing transcripts. Disabling it does not unredact previously redacted transcripts. Only future conversations are affected.
  • Structure preserved. The overall structure of your traces and requests remains the same. Only the content of specific text properties is modified.

Enabling PII redaction

You can unlock the PII redaction feature by subscribing the applicable add-on.

1

Subscribe to the PII redaction add-on

You can unlock the PII redaction feature by subscribing the applicable add-on. To do so, contact our sales team. If you’re on an enterprise plan, reach out to your account manager to have the feature enabled for your organization.

2

Enable PII redaction on your project

Open the project that you’d like to redact PII from. Then, head to Settings → Security and enable the PII redaction feature.

PII redaction toggle in security settings
Enable PII redaction to hide sensitive information in transcripts. Photo:
3

Repeat for other projects

PII redaction is a project-level setting. If you’d like to redact PII from multiple projects, enable it in Settings → Security on each individual project.

Viewing unredacted transcripts

After redaction is enabled, you can temporarily view the original unredacted version of a transcript using the toggle in the transcript detail view.

After redaction is enabled, you can temporarily view the original unredacted version of a transcript using the toggle in the transcript detail view. This is intended for debugging purposes.

Unredacted transcripts are only available for 48 hours after the conversation takes place. After that, the original data is permanently deleted and cannot be recovered.

What gets redacted

PII redaction scans both your agent’s responses (traces) and user inputs (requests), covering a wide range of PII types across different data structures.

Detected PII types

The redaction system detects over 35 types of personally identifiable information:

Protection methods

The system uses four approaches depending on the type of content:

PII redaction examples

Compare names, contact details and addresses before and after redaction.

These quick examples show how common types of PII appear in your transcript data before and after redaction is applied:

Text content:

  • Before: "Hello, I'm Sarah Johnson from Acme Corp"
  • After: "Hello, I'm [NAME] from [ORGANIZATION]"

Contact information:

  • Before: "Call me at 555-123-4567 or email sarah@acme.com"
  • After: "Call me at [PHONE_NUMBER] or email [EMAIL_ADDRESS]"

Address information:

  • Before: "I live at 123 Main St, Boston, MA 02101"
  • After: "I live at [LOCATION_ADDRESS], [LOCATION_CITY], [LOCATION_STATE] [LOCATION_ZIP]"

Redacting agent responses

Redact text and media in the traces sent by your agent.

Traces represent the messages and content your agent sends during a conversation, including text responses, cards, carousels, and media. PII redaction processes each trace type differently depending on the properties it contains.

Text trace

Redact message text and rich text, and erase the audio URL.

Modified properties:

  • payload.message: Censor redaction
  • payload.slate.content: Censor rich text redaction
  • payload.audio.src: Erase redaction
Before redaction
After redaction

Speak trace

Redact the spoken message and erase its audio URL.

Modified properties:

  • payload.message: Censor redaction
  • payload.src: Erase redaction
Before redaction
After redaction

CardV2 trace

Redact card text and button labels, and replace the image.

Modified properties:

  • payload.title: Censor redaction
  • payload.description.text: Censor redaction
  • payload.description.slate: Censor rich text redaction
  • payload.imageUrl: Override redaction
  • payload.buttons[].name: Censor redaction
  • payload.buttons[].request.payload.label: Censor redaction
Before redaction
After redaction
Apply card redaction to every card in the carousel.

Each card in the carousel is processed the same way as a CardV2 trace.

Modified properties (per card):

  • payload.cards[].title: Censor redaction
  • payload.cards[].description.text: Censor redaction
  • payload.cards[].description.slate: Censor rich text redaction
  • payload.cards[].imageUrl: Override redaction
  • payload.cards[].buttons[].name: Censor redaction
  • payload.cards[].buttons[].request.payload.label: Censor redaction
Before redaction
After redaction

Choice trace

Redact button names and their request labels.

Modified properties:

  • payload.buttons[].name: Censor redaction
  • payload.buttons[].request.payload.label: Censor redaction
Before redaction
After redaction

Visual trace

Replace the image URL with the redaction placeholder.

Modified properties:

  • payload.image: Override redaction
Before redaction
After redaction

Redacting user inputs

Redact messages, button labels and intent values sent to your agent.

Requests represent user inputs sent to your agent, such as typed messages, button presses, and intent matches. These are redacted in the same way as traces to ensure PII is removed from both sides of the conversation.

Text request

Redact personal information in the text payload.

Modified properties:

  • payload: Censor redaction
Before redaction
After redaction

Intent request

Redact the query and entity values while preserving intent names and entity types.

Modified properties:

  • payload.query: Censor redaction
  • payload.entities[].value: Censor redaction

Intent names and entity types are preserved.

Before redaction
After redaction

Path request

Redact personal information in the path label.

Modified properties:

  • payload.label: Censor redaction
Before redaction
After redaction

Message request

Redact personal information in the message text.

Modified properties:

  • payload.message: Censor redaction
Before redaction
After redaction

Action request

Redact the action label when it is present.

Modified properties:

  • payload.label: Censor redaction (if present)
Before redaction
After redaction

Redacting debug metadata

Keep allowlisted operational metadata and redact other debug properties.

Debug traces contain technical information about your agent’s operation, such as model usage, latency metrics, and navigation data. These are handled differently from regular traces using an allowlist system. All properties in payload.metadata are redacted by default unless explicitly allowlisted.

All other properties are replaced with PII detection markers or the message [Restricted due to potential PII]. This ensures new debug metadata properties are automatically protected until explicitly reviewed.

Full list of allowlisted properties

Look up debug metadata properties that remain unchanged.

The following properties are preserved as-is and never redacted:

  • userID - User identifier
  • versionID - Version identifier
  • tools - Tool registration and agent step data
  • model - AI model identifiers
  • provider - Service provider information
  • voice - Voice configuration
  • postMultiplierTokensConsumption - Token consumption metrics
  • variant - Message variant information
  • code - Execution code
  • latency - Performance timing metrics
  • maxTokens - Token limits
  • queryTokens - Query token counts
  • answerTokens - Answer token counts
  • temperature - AI temperature settings
  • tokens - Token counts
  • multiplier - Token multipliers
  • nextID - Navigation identifiers
  • diagramID - Diagram identifiers
  • intent - Intent information
  • confidence - Confidence scores
  • tokensMultiplier - Token multiplier values
  • tokensConsumption - Token consumption data
  • nodeID - Node identifiers
  • result - Result data
  • status - Status information
  • tokenConsumption - Token consumption data

Limitations

There are a few things to keep in mind when using PII redaction.

There are a few things to keep in mind when using PII redaction. These limitations are by design to balance privacy protection with usability.

  • Development conversations are not redacted. Only production conversations are processed.
  • Redaction is not retroactive. Enabling or disabling PII redaction only affects future conversations.
  • Audio is erased entirely. Audio URLs are set to empty strings because audio content cannot be selectively edited to remove only PII.
  • Images are replaced with a placeholder. All image URLs are overridden with a standard placeholder to prevent potential PII exposure through images.
  • Detection is English only. The PII detection model analyzes text in English.

Use the up and down arrow keys to select a result, Enter to open it, and Escape to close the search.