AI-Driven Approvals¶

Use AI models to automatically evaluate tool call requests and make approval decisions based on configurable guidelines.

Availability

AI-driven approvals are available in both Open Source and Enterprise editions.

Overview¶

Instead of requiring a human to review every approval request, you can configure an AI model to evaluate requests automatically:

Low-risk requests → AI approves instantly
Uncertain requests → Escalate to human reviewers
Clearly risky requests → AI denies with reasoning

The AI evaluates each request against your guidelines and returns a confidence score. Requests above the threshold are auto-decided; requests below it follow the fallback behavior you configure.

Configuration¶

Via Policy YAML¶

version: "1.0"
metadata:
  name: ai-safety-review

approval_workflows:
  - name: ai-reviewer
    approval_type: ai_driven
    ai_model: claude-sonnet-4-20250514
    ai_guidelines: |
      Approve routine operations (file reads, searches, list operations).
      Deny any destructive operations (delete, drop, truncate).
      Escalate anything involving production infrastructure or billing.
    ai_confidence_threshold: 0.8
    ai_fallback_behavior: escalate
    escalation_workflow: human-review

  - name: human-review
    timeout_seconds: 600
    required_approvals: 1
    approvers:
      - team: sre-team

tools:
  - name: bash
    source: builtin
    approval_workflow: ai-reviewer

Via Web UI¶

Navigate to Approval Workflows → Create Policy
Select AI-Driven as the approval type
Configure the AI model, guidelines, and threshold
Set the fallback behavior
Click Save

Configuration Reference¶

Field	Type	Default	Description
`approval_type`	`"ai_driven"`	`"standard"`	Must be `"ai_driven"` to enable AI evaluation
`ai_model`	string	—	AI model to use (e.g., `claude-sonnet-4-20250514`, `gpt-4o`)
`ai_guidelines`	string	—	Instructions for the AI when making decisions
`ai_confidence_threshold`	float	`0.8`	Minimum confidence (0.0–1.0) for auto-decision
`ai_fallback_behavior`	string	`"escalate"`	Action when uncertain: `escalate`, `approve`, or `deny`
`escalation_workflow`	string	—	Name of the policy to escalate to (for `escalate` fallback)

Fallback Behaviors¶

Behavior	When AI confidence is below threshold
`escalate`	Forward to human reviewers via the `escalation_workflow`
`approve`	Auto-approve the request (use with caution)
`deny`	Auto-deny the request (conservative approach)

Writing Effective Guidelines¶

Your ai_guidelines string is the system prompt that tells the AI how to evaluate requests. Write them like instructions for a security reviewer:

✅ Good Guidelines¶

ai_guidelines: |
  You are reviewing tool call requests for a production SaaS application.

  APPROVE if:
  - The operation is read-only (file reads, searches, list/get operations)
  - The target is a development or staging environment
  - The operation is idempotent and low-risk

  DENY if:
  - The operation deletes data or drops tables
  - The operation modifies production infrastructure
  - The operation involves secrets or credentials

  ESCALATE if:
  - You are unsure about the impact
  - The operation involves billing or payments
  - The operation affects customer-facing systems

❌ Poor Guidelines¶

ai_guidelines: "Be careful with approvals"  # Too vague

Audit Logging¶

Every AI decision is logged with full context:

Decision: approve, deny, or escalate
Confidence score: 0.0–1.0
Reasoning: AI's explanation for the decision
Model used: Which AI model evaluated the request
Guidelines applied: The guidelines text at decision time

View AI decision history in the Audit Log section of the dashboard.

Examples¶

Example 1: Conservative Production Guard¶

approval_workflows:
  - name: prod-guard
    approval_type: ai_driven
    ai_model: claude-sonnet-4-20250514
    ai_guidelines: |
      Deny all write operations in production.
      Approve read-only operations.
      Escalate deployment requests.
    ai_confidence_threshold: 0.95
    ai_fallback_behavior: deny

Example 2: Fast Development Flow¶

approval_workflows:
  - name: dev-fast
    approval_type: ai_driven
    ai_model: gpt-4o
    ai_guidelines: |
      Approve most development operations.
      Only deny clearly destructive operations like rm -rf or DROP TABLE.
    ai_confidence_threshold: 0.6
    ai_fallback_behavior: approve

Example 3: Tiered with Human Escalation¶

approval_workflows:
  - name: human-sre
    timeout_seconds: 300
    required_approvals: 1
    approvers:
      - team: sre-team

  - name: ai-triage
    approval_type: ai_driven
    ai_model: claude-sonnet-4-20250514
    ai_guidelines: |
      Approve routine operations instantly.
      Escalate infrastructure changes to the SRE team.
    ai_confidence_threshold: 0.85
    ai_fallback_behavior: escalate
    escalation_workflow: human-sre