Skip to main content

Agent Tooling & Prompts

Three-Tier Framework​

BayanCore AI agents follow a three-tier framework. Each tier increases in autonomy but requires stricter guardrails.

Tier 1: Ask​

  • Natural language queries over company data
  • Read-only access to documents, reports, dashboards
  • RAG-powered responses with source citations
  • No system changes

Tier 2: Act​

  • AI performs actions on behalf of the user
  • Drafts documents, creates records, suggests workflows
  • All changes require human review before commit
  • Tool registry with JSON schemas and RBAC

Tier 3: Automate​

  • Fully automated workflows for repetitive tasks
  • Requires explicit user opt-in
  • Strict validation rules and rollback capability
  • Audit trail for every automated action

Tool Registry​

  • JSON schema definitions for all available tools
  • RBAC-based tool access per user role
  • Versioned tool contracts
  • Tool execution logging and audit

Safety Guardrails​

Input Validation​

  • Prompt injection detection
  • PII redaction before LLM processing
  • Rate limiting and abuse prevention
  • Content filtering for inappropriate requests

Output Validation​

  • Financial data cross-referencing
  • Deterministic rule validation
  • Hallucination detection
  • Confidence scoring and uncertainty flags

Execution Guardrails​

  • RBAC enforcement on all AI actions
  • Transaction limits per user/role
  • Time-based restrictions (e.g., no overnight automated actions)
  • Approval workflows for high-risk operations

Safety Principles​

  1. Human-in-the-Loop: All financial actions require human approval
  2. Source Attribution: Every AI response cites its data sources
  3. Permission Boundaries: AI cannot access data the user cannot
  4. Audit Trail: All AI interactions are logged and reviewable
  5. Rollback Capability: Every AI action can be undone