Guardrails and Governance

This page explains what you can change safely, what requires care, and when you should contact Airgentic support. Understanding these boundaries helps you operate confidently without risking unintended consequences.

The safety model

Airgentic is designed to be customer-configurable, but some changes have more impact than others. We group changes into three categories:

Safe to change

These changes are low-risk and designed for routine use:

Curated answers

• Add, edit, or remove curated answers at any time
• Test in Admin Chat to confirm the override works as expected
• Changes take effect after saving (may take a short time to propagate)

Service messages

• Welcome message, no-answer message, follow-up message
• Chat window title, chat input placeholder, chat icon tooltip
• Voice welcome prompt and preamble
• Changes are cosmetic and don't affect AI behaviour

User interface labels

• Agent name displayed to users
• Contextual welcome messages toggle
• Collapsed citations toggle

Data sync scheduling

• Enable/disable automatic syncs
• Change sync times and days
• Run manual syncs

Privacy settings

• Retention period (days)
• Delete conversations only toggle

Uploading documents

• Add PDFs, Word documents, or HTML files
• Remember to index after uploading

User management

• Add or remove admin console users
• Change user roles
• Scope users to specific services

Change with care

These changes can significantly affect service behaviour. Always test thoroughly in Admin Chat before applying to live users.

Agent prompts

Prompts control how agents behave. Even small wording changes can affect responses.

Best practices:
- Make incremental changes rather than rewriting entire prompts
- Test extensively in Admin Chat with varied questions
- Use the AI Prompt Review feature to get feedback
- Remember that prompts are version-controlled — you can restore previous versions if needed

Related: Editing a Prompt

Agent roles and routing

The agent role description determines when the Frontline delegates to a specialist. Poorly written roles can cause routing errors.

Best practices:
- Be specific and clear about what the agent handles
- Avoid overlapping role descriptions between agents
- Test delegation behaviour with various question types

Related: Editing an Agent

Agent status (Testing → Live)

Moving an agent from Testing to Live makes it available to all users. Be sure it's ready.

Best practices:
- Test thoroughly in Admin Chat while status is Testing
- Verify routing, prompts, and function behaviour
- Consider timing — avoid going live during high-traffic periods without monitoring

Search settings

Category mappings, boosts, and synonyms affect what users find.

Best practices:
- Understand current search behaviour before changing
- Test search results in the widget after changes
- Changes to category mappings may require a reindex

Related: Search Settings

Crawler scope settings

These control what pages are indexed. Incorrect settings can exclude important content or include unwanted pages.

Best practices:
- Document your current settings before changing
- Test with a crawl run and check the logs
- Review indexed page counts to confirm expected coverage

Related: Crawler Settings

Higher risk

These changes can impact the entire service if applied incorrectly. Please proceed with caution.

Identity provider configuration

Changes to OIDC settings, client IDs, tenant IDs, or redirect URIs can break authentication for secure services.

Why:
- Incorrect configuration can lock users out
- Changes need to be coordinated with your identity provider settings
- Recovery may require Airgentic assistance

Service-wide prompt components

Global prompt components (Identity, Tone, Guardrails, etc.) affect all agents. Changes here have broad impact.

Why:
- Affects all agents simultaneously
- May override service-specific customisations
- Requires understanding of the modular prompt system

Advanced function code

Custom functions that call external APIs, send emails, or perform complex logic.

Why:
- Functions run code — bugs can cause failures or unintended behaviour
- API credentials are involved
- Testing functions requires understanding of AirScript

Note: Using global functions provided by Airgentic is safe. Writing or modifying custom functions requires care.

Account-level settings

Account deletion, billing changes, or partner-level access.

Why:
- Destructive actions cannot be undone
- May affect multiple services

Testing before changes go live

The Admin Chat is your primary testing tool:

1. Make your change with the agent status set to Testing (for agent changes) or apply the setting
2. Test in Admin Chat with varied questions
3. Use the Trace Log to verify routing, search results, and prompts are working correctly
4. Check edge cases — what happens with unusual or adversarial inputs?
5. Move to Live only when confident

Related: Admin Chat

Version control and rollback

Airgentic provides version control for prompts:

• Every prompt save creates a version with timestamp, author, and commit message
• View previous versions via the Version History modal
• Restore any previous version if needed

This means you can experiment knowing you can always roll back.

Related: Editing a Prompt

Governance considerations

Who can make changes

Use role-based access to control who can modify configuration:

Consider whether everyone needs full admin access or whether analyst roles are sufficient for some team members.

Related: User Management

Change communication

When multiple people administer a service:

• Communicate before making significant changes
• Document what you changed and why (commit messages help)
• Coordinate testing so changes don't conflict

Audit trail

Airgentic logs changes with timestamps and user information. If you need to investigate what changed, contact support for assistance.

Privacy and security

Data retention

You control how long conversation data is kept:

• Set the retention period in Privacy settings
• Choose whether to delete only content or all data
• Understand your organisation's data policies

Related: Privacy

Allowed origins

The allowed origins list controls which websites can embed your widget:

• Keep this list limited to legitimate sites
• Remove origins that are no longer in use
• Incorrect settings can allow unauthorised embedding or block legitimate use

Related: Security

API keys

The Insights API key grants access to raw analytics data:

• Regenerate if you suspect a key has been compromised
• Store keys securely
• Delete unused keys

When something goes wrong

If you make a change that causes problems:

For prompts

1. Go to Edit Prompts → select the prompt → View History
2. Find the previous working version
3. Restore it and save

For agent status

1. Set the agent back to Testing or Off
2. The Frontline will stop delegating to it immediately

For other settings

1. Revert to previous values if you remember them
2. Contact support if you need help identifying what changed

If you're locked out

Contact Airgentic support immediately — we can help restore access.

Summary table

Related pages

• Admin Chat — Testing changes
• Editing a Prompt — Prompt editing and version history
• User Management — Access control
• Privacy — Data retention
• Security — Origins and API keys
• Contact Support — Getting help

Next: What to Use When