Guardrails & allowed models
Guardrails establish the operational boundaries within which AI workers execute. Allowed models determine which foundation language models are permitted across your workspace. Together they let administrators construct a safe, enforceable operating envelope that governs all enterprise automation.
Organization-wide guardrails
Section titled “Organization-wide guardrails”Individual agents and functions have their own task-specific parameters, but administrators can set organization-wide guardrails that apply on top of every workload. Global guardrails act as a security floor that no individual user or agent can bypass or override.
- Failsafe security net — if a builder creates a custom agent and omits a safety constraint, global guardrails step in to enforce compliance automatically.
- PII and privacy protection — intercepts, masks, or blocks sensitive customer data and personally identifiable information before it leaves company network boundaries. Credentials themselves are handled separately; see Secrets.
- Unauthorized system isolation — prevents agents from querying or transmitting payloads to unapproved third-party APIs or un-vetted external connectors. See Connector governance.
- Consistent platform safety — global rules are what make broad, self-serve building safe across non-technical departments.
Allowed models and model routing
Section titled “Allowed models and model routing”Rival Enterprise is model-agnostic: agents and workflows can run across a range of leading foundation models. Enterprise Admins curate the roster of allowed models to balance capability, speed, cost, and compliance.
- Capability allocation — approve high-reasoning, top-tier models for complex multi-step analysis, technical problem solving, and critical decisions.
- Cost optimization — route routine, high-volume tasks such as classification or data extraction to faster, low-cost models to minimize compute spend. See the Cost Optimization Center.
- Data policy and residency — restrict sensitive tasks to models that comply with regional data-hosting regulations or strict zero-data-retention agreements.
How enforcement works
Section titled “How enforcement works”Administrators configure global guardrails and model permissions centrally, in the AI Governance module of the Command Center. Because guardrails are enforced by the platform runtime rather than by prompt engineering or user diligence, they hold consistently on every run.
-
A user or agent initiates a task payload.
-
The system evaluates model eligibility — is the requested model permitted for this user or team?
-
The system checks global safety guardrails — is sensitive data or PII present? Is an unapproved connector requested?
-
A compliant payload executes on the approved foundation model.
-
A non-compliant payload is intercepted — the platform blocks the run, masks the sensitive fields, or routes execution to an approval queue.
Guardrails and model governance overview
Section titled “Guardrails and model governance overview”| Governance area | Administrative control | Enterprise outcome |
|---|---|---|
| Model whitelisting | Curate approved foundation models | Enforces data compliance and prevents vendor lock-in |
| Model routing | Steer tasks by model capability and cost | Prevents unnecessary spend on routine, high-volume tasks |
| Global guardrails | Enforce PII masking and connector restrictions | Establishes a security floor across all agents |
| Runtime enforcement | Apply policy at the platform execution level | Removes reliance on individual memory or prompt engineering |
A worked example: multi-model governance
Section titled “A worked example: multi-model governance”Consider an enterprise balancing high-volume support with advanced data research.
- Model whitelist setup — an IT administrator permits two foundation models: a highly capable reasoning model for complex analytics, and a light, cost-effective model for high-frequency data processing.
- Global policy deployment — the admin sets an org-wide guardrail blocking data transmission to unapproved external connectors, and mandating PII masking across all customer-facing workflows.
- Unrestricted departmental building — Marketing, Support, and Finance teams build agents freely with RivalBot. Every agent operates inside the allowed models and guardrails, with no per-builder setup.