For the complete documentation index, see llms.txt. This page is also available as Markdown.

PII Redaction for Past Tickets

Turn on PII redaction to remove credit card numbers, SSNs, names, addresses, and more from past helpdesk tickets before they're persisted or indexed.

What it does

When PII redaction is enabled, every past ticket your agent syncs is run through an AI-based detector before anything is written to the database or sent to the search index. Personally identifiable information is replaced with a typed mask like [REDACTED:CC] or [REDACTED:EMAIL].

This means:

  • Your agent cannot see the original PII when retrieving from past tickets

  • Your stored ticket history does not contain the original PII

  • The search index does not contain the original PII

The masked content still has plenty of context for the agent to learn from — only the sensitive substrings are removed.

What gets redacted

The detector looks for fifteen categories of PII:

Category
Examples

Credit cards

4532-1234-5678-9010

Social Security Numbers

123-45-6789

Date of birth

"DOB: 1984-07-19"

Bank account / IBAN

GB82 WEST 1234 5698 7654 32

API keys & secrets

AWS keys, Stripe sk_live_*, OpenAI sk-*, GitHub ghp_*, etc.

JSON Web Tokens

eyJhbGciOi…

Email addresses

customer@example.com

Phone numbers

+1 (415) 555-0182, +44 20 7946 0958

Person names

"John Smith called…"

Physical addresses

"742 Evergreen Terrace, Springfield, IL 62701"

IP addresses

IPv4 and IPv6

Passport numbers

6–9 alphanumeric

Driver's licence numbers

National IDs

Aadhaar, NHS number, etc.

Tax IDs

EIN, VAT, etc.

The detector knows to ignore well-known non-PII lookalikes, including:

  • Product, brand, or company names (Salesforce, Stripe, Acme Co.)

  • Test card numbers (4242 4242 4242 4242, 5555 5555 5555 4444, etc.)

  • Order numbers, invoice IDs, ticket IDs

  • Dates that aren't introduced as birth dates

Where it applies

PII redaction runs on past helpdesk tickets that are synced into your agent's knowledge for training. It is opt-in per integration.

Surface
Redaction applies?

Past tickets synced from your helpdesk

Yes, when enabled

Help Center articles, macros, knowledge base pages

No

Files uploaded to your workspace

No

Conversations the agent has at runtime

No

Tickets that were indexed before you enabled the toggle

No (see below)

Turning redaction on or off

Today PII redaction is configurable on the Zendesk Tickets path-settings page. (Support for Intercom, Freshdesk, Gorgias, and HubSpot is coming.)

  1. In your agent's platform view, open the connected Zendesk integration

  2. Click Tickets in the integration's sub-menu, then open its settings

  3. Scroll to the Redact PII before indexing toggle

  4. Flip it on (or off) and click Save Settings

The default for new Zendesk integrations is on. The next time you trigger a past-tickets sync (manually, or automatically after a settings change), tickets that come in will go through the redactor.

What this doesn't do

  • Existing indexed tickets are not retroactively redacted. Flipping the toggle only affects tickets ingested from now on. To re-redact older tickets, trigger a fresh past-tickets sync.

  • Runtime tool calls aren't affected. When the agent fetches a Zendesk ticket live (e.g., to draft a reply), the raw ticket content reaches the model. Redaction only runs on the persistent training corpus.

  • Non-helpdesk knowledge sources aren't affected. Articles, files, and other sources are not run through the redactor.

How accurate is it?

Detection runs on a fast, modern LLM (Gemini 3 Flash) with a tuned prompt and a deny-list of common false positives. In our benchmarks against synthetic support tickets the redactor catches the vast majority of obvious PII without flagging things like product names or test card numbers.

That said, no automated detector is perfect:

  • A new or unusual PII format may slip through

  • A name that's also a common noun may be missed

  • Very long tickets may have lower recall than short ones

If you're indexing tickets that contain especially sensitive data (e.g., regulated health information), pair this feature with the usual access controls and consider clearing especially sensitive tickets from your helpdesk before sync.

FAQ

Does redaction slow down ticket sync?

There's an additional ~1–2 seconds per ticket for the LLM call. Past-ticket sync is a background job, so this doesn't affect interactive operations.

What happens if the redactor's AI service is unavailable?

The sync fails closed — individual tickets that can't be redacted are skipped rather than indexed in their raw form. They'll be retried on the next sync.

Can I see what was redacted from each ticket?

Not in the UI today. The number and types of redactions per ticket are logged on our side so support can investigate specific cases on request — contact us.

Do you store the original ticket content anywhere after redaction?

No. The redacted version replaces the original before anything is persisted in our database or sent to the search index.

Will redaction be available for Intercom, Freshdesk, Gorgias, and HubSpot?

Yes — the underlying detector already supports those integrations. We just need to expose the toggle in each one's past-tickets settings page. Coming soon. In the meantime, the past-tickets sync on those integrations does not redact.

Questions?

For security or privacy questions about PII redaction, or to request the feature on a specific integration ahead of general availability:

Last updated