Blog / Security

Redact PII in Call Recordings and AI Summaries

A practical workflow for preventing and redacting sensitive data across call audio, transcripts, AI summaries, and CRM records. Includes PCI DSS guidance and a test checklist.

To redact PII in call recordings, stop sensitive details from entering the recording when possible, redact both audio and transcripts when they are captured, and check every downstream copy—including AI summaries and CRM notes. A reliable process combines prevention, post-call review, access and retention controls, and tests that reflect how customers actually speak.

Key takeaways

  • PII can include names, SSNs, birth dates, addresses, emails, phone numbers, and account identifiers; PCI DSS separately governs payment card data and sensitive authentication data.
  • PCI DSS prohibits storing a CVV after authorization, even encrypted, including in digital audio recordings.
  • A transcript edit does not necessarily change the audio, and sanitizing either one does not guarantee that AI summaries or CRM records are clean.
  • Manual pause and resume controls can miss spoken card details; test controls with realistic calls and inspect each output.
  • Teams using call summaries and HubSpot sync should verify CRM fields and downstream workflows before relying on redaction.

Why do call recordings and AI summaries need a redaction plan?

Call recordings can contain sensitive information in spoken audio, transcripts, voicemail, AI summaries, and CRM records. A redaction plan follows the data through each copy and derived output, because removing a number from one artifact does not prove it has been removed everywhere.

What counts as PII in a business call?

Personally identifiable information (PII) is information that can distinguish or trace a person, either by itself or when linked with other identifying information. NIST’s definition of PII includes examples such as names, Social Security numbers, and birth details.

In sales and support calls, sensitive details may include:

  • Card numbers and account identifiers
  • CVVs, SSNs, and dates of birth
  • Home or billing addresses
  • Email addresses and phone numbers

Whether a detail needs to be redacted depends on your purpose, policies, and obligations. Protect the full set of information your business handles; do not assume “PII” has one universal legal definition or that every category follows the same rules.

How is PII different from PCI card data?

Cardholder data and sensitive authentication data are payment-related categories with PCI DSS requirements. A CVV is sensitive authentication data; the card number is a different category. A payment detail may also identify a person, but general PII and PCI data are not interchangeable labels.

For PCI DSS scope and controls, map where payment data is captured, transmitted, processed, and stored. Ask your acquirer or a qualified assessor to review your specific environment rather than treating any single recording setting as a complete compliance determination.

Where should you stop sensitive data from spreading?

Use three control points: prevent capture, redact audio and transcripts after capture, and control downstream uses such as summaries, CRM syncs, exports, analytics, and backups. Prevention reduces exposure at the source; post-call redaction and lifecycle controls help address details that still get captured.

Prevent capture with secure keypad entry or payment transfer

For card payments, consider directing customers to a secure keypad flow or transferring them to a payment process designed to keep spoken card details out of the recording path. Evaluate how the transfer handles recording, transcripts, and any agent or customer audio before adopting it.

DTMF suppression masks keypad tones in the recording. It can reduce exposure when a customer enters digits by phone keypad, but it does not protect numbers spoken aloud or ensure that other systems never capture related data.

Should you pause or stop recording during payment?

Pausing suspends capture for a period; stopping ends recording. A pause preserves the rest of the call while excluding the payment segment, but the agent must start and resume it at the right time. Stopping avoids recording later audio, but also removes useful parts of the interaction and may not undo capture that already happened.

PCI SSC warns that manual pause and resume depend on agents acting at precisely the right time and can miss the beginning of card data. Its guidance calls for ongoing monitoring and verification. If you use agent controls, train staff on a clear script, define when to act, and test for timing mistakes.

What does post-call redaction protect?

Post-call redaction edits stored audio, transcript text, or both after a call. It can help when prevention fails, but coverage depends on the system and configuration. Confirm whether redaction affects audio, transcripts, or both, and whether it also reaches derived artifacts.

The call recording workflow should be evaluated alongside transcription, summaries, transfers, and retention. A redacted transcript does not automatically make the audio safe, and redacted audio does not guarantee clean text.

How do you redact recordings and summaries step by step?

A useful redaction workflow maps every data copy, defines what to protect, configures prevention and redaction controls, and tests every output. Assign an owner for each stage so a missed value can be traced from the original call through summaries, integrations, exports, and retention.

1. Map recordings, transcripts, summaries, and integrations

Draw the path from call to deletion. Include the recording platform, transcription service, AI summary tool, CRM, analytics, exports, support tickets, and any available backups. Note which system creates each copy, who can access it, and how long it is retained.

Include voicemail and call transfers. A transferred customer might repeat payment information to another agent; voicemail transcription can create a text copy even when nobody saves the audio. Teams using a HubSpot calling integration should check which call fields, notes, and activities are synced.

2. Define categories and rules for cards, IDs, and contact details

Write down which values should be prevented, masked, or removed. Define the expected handling for card numbers, CVVs, SSNs, birth dates, account IDs, addresses, phone numbers, and emails. Decide how much context to keep—for example, a payment outcome without the card digits.

Specify what happens when detection is uncertain. A human review queue can catch ambiguous cases; a documented fallback should explain when recording pauses, when a call transfers, and how staff continue the conversation.

3. Test realistic calls and inspect every output

Use scripted test calls with synthetic details rather than real customer data. Include ordinary speech and difficult cases:

Test scenarioWhat to check
Digits spoken one by oneAudio and transcript coverage
Number read in chunks, then correctedOriginal, correction, and final value
Two speakers or interruptionsWhether the system detects both voices
Poor audio or an unfamiliar accentMisses and false detections
Voicemail or an unanswered callAudio, transcript, and notifications
Transfer to payment or another agentRecording behavior before and after transfer
Keypad entryDTMF suppression and any transcript output

Inspect the recording and transcript separately, then review AI summaries, CRM fields, exports, and available copies. Record the test case, expected result, actual result, and reviewer so failed cases lead to a configuration or workflow change.

4. Apply access, retention, deletion, and audit controls

Limit access to people who need recordings for a defined job. Set retention periods for each artifact and confirm how deletion applies to recordings, transcripts, summaries, exports, and backups. Keep an audit trail of configuration changes, access, redaction actions, and review outcomes.

Retention should match your business need and applicable requirements. Avoid keeping sensitive material indefinitely because storage is cheap or because no one owns the deletion schedule.

How do you keep card numbers and IDs out of AI call summaries?

Generate summaries from sanitized transcripts when the system supports that flow, or scan and redact summaries independently before they reach a CRM or export. A summary can reproduce a card number or ID even when another artifact has been cleaned, so verify each derived output on its own.

Use sanitized source text where possible

A card number can be absent from the final audio but remain in an unredacted transcript that an AI model uses to write a summary. If redaction happens after summary generation, the summary may already contain the number.

Prefer a workflow that redacts the source transcript before summarization. If that is unavailable, apply a separate sensitive-data scan to the generated summary and any synced CRM note.

Preserve useful context with typed placeholders

Replace sensitive values with consistent placeholders, such as [CARD NUMBER REDACTED] or [ACCOUNT ID REDACTED]. This preserves the fact that a customer supplied a value without preserving the value itself.

For example, “Customer provided card number 4111…” can become “Customer provided [CARD NUMBER REDACTED]; payment completed.” The second note keeps the service outcome while removing the sensitive digits.

Teams reviewing summary quality can use a repeatable 20-call summary evaluation and the summary-versus-transcript guide to decide which artifact belongs in each workflow. Remove the space before the URL in the first link when publishing.

What should you ask a vendor about call redaction?

Ask vendors to describe supported data types, languages, processing timing, artifact coverage, limits, review options, deletion behavior, audit logs, and incident handling. Test those claims with your own scripted calls; product documentation describes a specific implementation, not a universal guarantee for every platform or configuration.

Use this checklist during evaluation:

  • Which categories and languages can the system detect?
  • Does processing happen live, after the call, or both?
  • Are audio and transcript redacted separately, and can both be covered?
  • Are AI summaries, CRM notes, exports, and integrations scanned?
  • What are the call-length, timing, and redaction-duration limits?
  • Can a reviewer inspect and correct uncertain results?
  • How are redaction events, access, and configuration changes logged?
  • What does deletion cover, including derived copies and backups?
  • What is the process for reporting and investigating a missed value?

As dated, product-specific examples, Dialpad documents separate transcript, audio, or combined redaction options, while Salesforce documents a 30-minute limit for its described mechanism and warns that automated redaction can miss sensitive information. Recheck current documentation and availability before comparing platforms. Do not assume those examples apply to other systems.

For a broader quality process, see the AI call summary review workflow. A summary can be useful for sales coaching or support follow-up only if the data handling matches your policies.

What do PCI DSS and U.S. recording rules require?

PCI DSS Requirement 3.3.1 prohibits storing sensitive authentication data, including card validation codes, after authorization—even encrypted—and PCI SSC says this applies to digital audio recordings. Recording a card number does not by itself establish that a system is noncompliant; scope and controls depend on the environment and should be reviewed with your acquirer or a qualified assessor.

PCI SSC also says suppression or redaction should be enabled where technology exists. For phone payments, review the full path: capture, authorization, recording, transcription, storage, and any summaries or synced records. A pause control alone does not settle compliance scope.

In the United States, recording notice and consent requirements depend on jurisdiction and situation. Give callers an appropriate notice for your use case and check the rules that apply to your business; the state-by-state recording law guide can help you identify questions for counsel.

Set a retention schedule and an incident process before rollout. The FTC reports that the Consumer Sentinel Network received more than 1.1 million identity theft reports in 2024; those reports do not show that call recordings caused those incidents. The FTC breach response guide notes that all states, the District of Columbia, Puerto Rico, and the Virgin Islands have breach-notification laws covering breaches involving personal information.

What should sales and support teams check before and after rollout?

Before rollout, sales and support teams should map data flows, configure prevention and redaction, and test audio, transcripts, summaries, and CRM syncs. After rollout, they should sample results, review missed detections, check access and retention, and retest after changes to prompts, integrations, or call workflows.

Before enabling recording and AI summaries:

  • Identify sensitive categories and where each artifact is stored.
  • Choose secure payment entry, transfer, pause, stop, or redaction controls.
  • Test spoken digits, corrections, interruptions, transfers, voicemail, and poor audio.
  • Check recordings, transcripts, summaries, CRM fields, exports, and deletion.
  • Assign owners for reviews, access, retention, and incident response.

After rollout:

  • Sample redacted calls and downstream records on a set schedule.
  • Track false negatives, false positives, and review completion.
  • Retest when integrations or call flows change.
  • Confirm that retention and deletion work across connected systems.

For example, a sales or support team using Voicalyst call summaries with a HubSpot sync should verify that redacted details do not reappear in synced notes, fields, or exports. Voicalyst’s AI call summaries and HubSpot integration can be evaluated as parts of that workflow; confirm the current behavior that matters to your controls before relying on it. Explore Voicalyst.

Frequently Asked Questions

How do you redact PII from a call recording?

Use layered controls: prevent capture with secure keypad entry or payment transfer, pause or stop recording when needed, and redact both audio and transcripts after capture. Then check summaries, CRM notes, exports, and integrations for copied values. Test with scripted calls that include corrections, interruptions, poor audio, voicemail, and transfers; review misses and keep an audit trail.

Is it PCI compliant to record a phone payment?

Recording a phone payment is not automatically noncompliant, but PCI DSS requirements apply to systems that store, process, or transmit account data. Sensitive authentication data such as a CVV must not be stored after authorization, even encrypted, including in digital audio. Ask your payment acquirer or a qualified assessor to evaluate your scope and controls.

Should I pause or stop recording when taking payment details?

A pause can protect the payment segment if it starts and ends at the right time, but manual controls can miss spoken digits or a correction. Stopping recording avoids capturing later call audio, while secure keypad entry or a payment transfer can prevent spoken details from reaching the recording path. Choose controls that fit your workflow and verify their behavior.

Does redacting a transcript also redact the audio recording?

Not necessarily. Audio and transcript redaction may be separate controls, and a transcript-only change can leave the spoken value audible in the recording. Confirm whether the system redacts audio, text, or both, then inspect each output and any derived summaries or CRM records.

How do I keep credit card numbers out of AI call summaries?

Where possible, generate summaries from a redacted transcript. Independently scan summaries and synced CRM fields as well, because they may have been generated from an unredacted source or may reproduce a number. Use a typed placeholder such as [CARD NUMBER REDACTED] while keeping the useful outcome, such as “payment completed,” in the note.

What information counts as PII in a customer service call?

PII includes information that can distinguish or trace a person alone or when linked with other identifying information. In calls, that can include names, SSNs, dates of birth, addresses, emails, phone numbers, account identifiers, and payment details. Cardholder data and sensitive authentication data have specific PCI DSS implications beyond general PII.

How can I verify that a call recording has been fully redacted?

Test calls using realistic speech patterns, then inspect the audio and transcript around each sensitive segment. Check AI summaries, CRM notes, exports, and available copies or backups too. Record the test result, reviewer, and any missed values; retest after configuration changes and monitor samples after rollout.

Put this into practice with Voicalyst

Voicalyst is the cloud phone system US sales and support teams set up in minutes: a business phone system with cloud telephony, AI call summaries and texting.

Start free trialBook a demo
VE

Voicalyst Editorial Team

Guides and research from the team building Voicalyst.