Dispute Refund Agent
Autonomous refunds, gated by policy. Human decides when it matters.
Rule-based processing. Same input produces the same outcome every time.
How it works
- Seed Disputes marks completed transactions as buyer disputes with realistic claim notes
- Run Agent sends each dispute to ArmorIQ's policy proxy which evaluates risk and decides to allow, hold, or block
- Check Dispute runs fraud detection (buyer history, amount thresholds, suspicious keywords) and blocks suspicious claims
- Cleared disputes proceed to resolve_dispute_refund which is held for human approval on the ArmorIQ dashboard
- The activity log polls every 5 seconds. Once approved, the agent retries and the Razorpay refund executes
Same inputs always produce the same outputs. Best for cases where policy thresholds are clear.
How it works
- Seed Disputes marks completed transactions as buyer disputes with realistic claim notes
- Run LLM Agent sends all disputes to gpt-4o-mini with a decision framework and three tools
- The LLM calls
list_open_disputes to read the database, then reasons about each claim
- For each dispute it calls
check_dispute to evaluate fraud risk before processing
- If cleared, it decides full refund, partial refund, or deny based on context
- Watch the AI thinking in the LLM Reasoning panel, then decisions execute through ArmorIQ
Smarter but less predictable. Handles nuanced context like buyer history and claim severity.
MCP: dispute-mcp
Policy: dispute-refund-guard
MCP: dispute-mcp
LLM: gpt-4o-mini via OpenRouter
Agent: dispute-llm-agent