Sample no-prompt audit
AI Spend & Data Exposure Report
ExampleCo is synthetic, but the report structure is the artifact AIRoute would deliver after reviewing usage exports, invoices, token counts, model names, API-key labels, workflow labels, sensitivity tags, latency, errors, and retry metadata.
Monthly AI spend reviewed$42,610
Estimated monthly opportunity$11,900
Data exposure findings6
High-sensitivity requests234k
Executive Summary
ExampleCo's AI usage appears concentrated in support summarization, RAG preprocessing, embeddings, batch classification, legal review, and finance extraction. The first-pass audit found likely overuse of premium models, duplicate retries, and sensitive workflows that need route policy before optimization.
AIRoute would not recommend immediate production rerouting. The next step is observe-mode instrumentation with workflow and sensitivity tags, then shadow benchmarking for approved low-risk workloads.
Recommended Pilot
- Start with support-ticket tagging.High volume, structured outputs, medium sensitivity, clear quality checks.
- Benchmark embeddings separately.Compare current provider against lower-cost embedding routes using retrieval quality tests.
- Hold legal and finance routes.Do not optimize sensitive workflows until private-route and retention policy are explicit.
Data Exposure Policy Findings
| Workflow | Sensitivity | Current Pattern | Recommendation | Policy Action | Route Guidance |
|---|---|---|---|---|---|
| Legal review | High | 52k requests through hosted model API | Require private route policy before savings tests | Hold or redact/tokenize | Private VPC, approved no-retention provider, or local model |
| Finance extraction | High | 182k invoice-related requests | Add sensitivity-aware routing rules | Approved providers only | No decentralized or cheaper external route until classified |
| Support workflows | Medium | 2.8M customer-ticket requests | Observe with retention and redaction settings | No prompt storage by default | Shadow only redacted samples first |
Opportunity Table
| Workflow | Current Pattern | Opportunity | Est. Monthly Savings | Risk | Next Step |
|---|---|---|---|---|---|
| Support tagging | Premium model for short classifications | Test smaller model or hosted open model | $3,100-$4,400 | Low | Shadow benchmark 2k samples |
| RAG preprocessing | Realtime calls for nightly document jobs | Batch execution and lower-cost summaries | $2,700-$3,600 | Low | Batch route test |
| Embeddings | Single provider, no retrieval QA | Compare lower-cost embedding route | $1,900-$2,800 | Medium | Measure recall@k |
| Customer chat | High-value live production flow | Observe only for now | Not counted | Hold | Collect metadata first |
Data Used And Not Used
Reviewed
- Provider invoice totals
- Usage export by model and timestamp
- Input/output token counts
- API-key and app labels
- Latency, error, and retry metadata
Not reviewed
- Raw prompts
- Model outputs
- Customer records
- Secrets or provider API keys
- Production traffic payloads