Skip to main content
Guardrails are automatic privacy controls that scan AI responses for sensitive personal information before they reach your users. When Prophecy Gov detects a PII type that is enabled for your city, it replaces the sensitive value with a redaction placeholder such as [SSN REDACTED]. This happens silently in the background — users see a clean response with the sensitive content removed rather than an error or a blocked response. Configuring guardrails is an admin-only action and the settings apply to all users in your city workspace.

​
Why guardrails matter for government use

Municipal documents often include residents’ personal information — names alongside dates, identification numbers in permit applications, contact details in public comments, and financial data in contracts. Even when the AI is summarizing a document rather than quoting it directly, there is a risk that sensitive details could surface in responses. Guardrails provide a city-wide safety net that runs automatically on every response, so staff do not need to manually review AI output for accidental PII disclosure.

​
PII types available

Guardrails are organized into four categories. Each individual type can be enabled or disabled independently.
LabelWhat is detected
Social Security Number (SSN)Nine-digit SSNs in XXX-XX-XXXX format
Driver’s license numberState-issued driver’s license numbers
Passport numberPassport document numbers
Taxpayer ID number (ITIN)Individual Taxpayer Identification Numbers (9xx-XX-XXXX)
Employer Identification Number (EIN)Federal EINs in XX-XXXXXXX format
LabelWhat is detected
Credit/debit card numbers16-digit card numbers validated with the Luhn algorithm
CVV/CVC codesCard security codes in an explicit card context
Bank account numbersAccount numbers explicitly labeled as a bank or deposit account
Routing numbersNine-digit ABA routing numbers validated by checksum
IBAN numbersInternational Bank Account Numbers
SWIFT/BIC codesBank identifier codes in a financial context
LabelWhat is detected
PasswordsValues following a password: or passwd= pattern
Authentication codes (OTP/MFA)One-time codes explicitly labeled as verification or login codes
LabelWhat is detected
Date of birthDates explicitly marked with a birth-date signal (DOB, “born on”, etc.)
Email addressesStandard email address format
Phone numbersPersonal phone numbers (not office extensions or reference codes)
Street addressesPrivate residential addresses
GPS coordinatesCoordinates tied to a specific individual’s location
License plate numbersVehicle license plates
The Personal Information category is disabled by default for new city workspaces. These types — especially email addresses, phone numbers, and street addresses — appear frequently in public municipal documents and can produce false positives that interfere with legitimate responses. Enable them only if your city’s documents contain private residential data that should not surface in AI responses.

​
Always-on labels

Two labels are permanently enabled and cannot be disabled:
  • Social Security Number (SSN)
  • Credit/debit card numbers
These are displayed with an Always on badge in the guardrails settings page. This ensures that the most sensitive identifiers are always protected, regardless of other configuration changes.

​
Configuring guardrails

1

Open Guardrails settings

Navigate to Settings → Guardrails. You will see all PII types organized by category, each with a toggle.
2

Enable or disable labels

Toggle individual labels on or off. Always-on labels (SSN and credit card) have a fixed toggle that cannot be changed.
3

Save your changes

When you have made your adjustments, click Save Changes. Your settings take effect immediately for all users in your city.
To discard unsaved changes, click Cancel before saving. Members can view the guardrails configuration but cannot modify it. The settings page displays a read-only view for non-admin users.

​
How guardrails affect responses

When an AI response contains content that matches an enabled PII label, Prophecy Gov replaces the matched text with a clearly labeled placeholder before the response is displayed. Examples:
  • 123-45-6789 becomes [SSN REDACTED]
  • A credit card number becomes [CARD REDACTED]
  • A personal phone number (when enabled) becomes [PHONE REDACTED]
The rest of the response is delivered normally. Users see a coherent answer with sensitive values replaced inline. There is no separate warning or blocked response — the redaction is applied silently.
If users report that legitimate content is being redacted — for example, a reference number that looks like a phone number — consider disabling the relevant label or narrowing which PII types are active. The Personal Information category is most prone to this in municipal contexts.