Skip to main content

ToolboxAI Privacy Layer

Before any prompt reaches an external provider like OpenAI or Claude, ToolboxAI's Privacy Layer scans it for personally identifiable information and neutralizes it — without breaking the AI's ability to reason about the record. It's on by default, for every org, with no setup required — this page covers how it actually works.

Two different treatments, depending on what's found​

  • Emails and phone numbers are pseudonymized, not just hidden. Each real value is swapped for a consistent fake one before the prompt is sent, and swapped back to the real value in the AI's response before you ever see it. The AI still gets to reason about "is this the same person" or "same company" across a record's activities — it just never sees the real address or number to do it.
  • SSNs and credit card numbers are redacted outright. These are replaced with a generic label and never restored, since there's no legitimate reason an AI response needs to echo a real SSN or card number back to you.

Fake values are deliberately, unambiguously fake​

  • Fake emails always use example.com, example.net, or example.org — the domains IANA reserves under RFC 2606 specifically for documentation and placeholder use, guaranteed never to resolve to a real, registrable business. That's a stronger guarantee than an invented company-sounding domain, which could coincidentally match a real one.
  • Fake phone numbers always fall in the 555-0100 through 555-9999 range — the block the North American Numbering Plan reserves for fictional use in film and TV (the same convention behind every "555" movie phone number). A fake number generated this way can never be mistaken for a real, reachable one.
  • On phone numbers specifically, the area code is left untouched — a bare area code identifies a broad region, not a specific person, so hiding it has no privacy benefit and only costs the AI useful regional context. Only the exchange and subscriber digits (the part that could actually reach someone) are replaced.

The same real value always maps to the same fake value​

Within a single request, every occurrence of the same real email or phone number resolves to the same fake one — so if a record's activity history mentions the same person five times, the AI still recognizes it as one person, not five different placeholders. Identity is matched case-insensitively for emails and by normalized digits for phone numbers, so formatting differences (Sarah@Acme.com vs. sarah@acme.com, or (415) 555-1111 vs. 415-555-1111) don't create false duplicates.

Where this applies​

The Privacy Layer runs the same way across every feature that builds a prompt from your Salesforce data — ToolboxAI and Field Extractor both go through it, not just one or the other, and it's active from the moment you install the package. It's controlled by a single org-wide toggle, on by default, under Settings → Data Settings — see ToolboxAI Data Options if your organization has a reason to turn it off.