Cost Attribution Errors

Issue: Costs Cannot Be Accurately Attributed to Users, Agents, or Use Cases

Frequency: Common

Symptoms

• Cannot determine which user/agent caused costs
• Chargeback calculations inaccurate
• Cost centers receive wrong allocations
• Cannot identify expensive operations
• Billing disputes unresolvable

Root Cause Multi-tenant agent systems often lack proper cost tagging. When multiple users, agents, or use cases share infrastructure, costs get aggregated without attribution. Without request-level tagging that flows through to billing, organizations cannot implement accurate chargebacks or identify optimization targets.

Example

Shared agent infrastructure:
  
Month-end bill: $15,000

Attribution attempt:
  User A: ??? (no tracking)
  User B: ???
  Agent Type 1: ???
  Agent Type 2: ???
  
Actual breakdown (if tracked):
  User A: $12,000 (power user, expensive queries)
  User B: $800
  Agent Type 1: $2,500
  Agent Type 2: $11,700
  
Problem: User A's department wasn't charged
         Wrong team optimized the wrong agent

Contributing Factors

• No request-level cost tagging
• Shared API keys across users/agents
• Async billing without request correlation
• Missing metadata in LLM calls
• Aggregated invoices only
• No tenant isolation in billing

Eval Recipes

Test Cases

Metrics

Mitigation Strategies

Prevention

1. Request tagging: Add user/agent/tenant ID to every LLM call
2. Correlation IDs: Flow IDs from request to billing
3. Per-tenant API keys: Separate keys for billing isolation
4. Real-time attribution: Calculate costs per-request

Architecture Pattern

Request → [Tag: user_id, agent_id, tenant_id]
              ↓
         LLM Call (tags in metadata)
              ↓
         [Cost calculated with tags]
              ↓
         Attribution DB

Production Signals

Key Metrics

Alerts

References

• OpenAI: Usage Tracking
• Anthropic: Message Metadata