Sample organization
Agent waste audit29 Sep – 28 Oct 2026 UTC, 30 daysReport readySynthetic data
Synthetic sample, not a customer result · No real organization, spend or savings are shown.Generated by our demo source and analysed by the same detectors that run on real audits.
Executive summary
$24,313.82 observed over 30 days. One anomaly worth reviewing and eight potential savings worth about $8,404 combined (overlap removed, an estimate) to validate; no savings confirmed.
Observed spendBilled cost reported by the provider.
$24,313.82
30 of 30 days with data
Potential savings, overlap removed (estimate)Levers can act on the same spend, so adding their amounts would count some dollars twice. Instead each one is applied to the spend the previous ones left: combined = T × (1 − Π(1 − aᵢ/T)), with T the observed spend and aᵢ each potential amount. An estimate under each lever’s stated assumptions, not a saving: only an approved, measured experiment verifies one.
$8,404
≈ 34.5% of $24,313 observed · 8 potential savings listed one by one below
Verified savingsMeasured in a recorded, approved experiment. A sample has no experiments, so nothing is confirmed.
None yet
Only an approved, measured experiment counts
Anomalous spendUnusual spend. Not a saving — may be legitimate. Anomalies can overlap, so they’re never added up.
1
Largest $1,406.35
Potential savingsModelled under stated assumptions. Not yet validated. Each is listed on its own; the combined figure above removes the overlap between them.
8
Largest $4,255.93
ValidatedFeasible and evidenced, confirmed by your team. Not yet measured. Validated amounts are totalled within their class only.
0
None yet
Key findings
- Anomalous spend: Spend spike on project proj_support_agentAnomalous spend$1,406.35
- Potential saving: Route simpler requests from frontier-large to efficient-smallPotential saving$4,255.93
- Potential saving: Shorter outputs on frontier-largePotential saving$2,158.92
- Observed spend: 95.4% of attributed spend is on model frontier-largeObserved spend$23,214.22
- Observed spend: 72.9% of attributed spend is on project proj_support_agentObserved spend$17,747.60
Potential savings are never simply added up: the combined figure applies each lever to what the others left. Potential and validated amounts aren’t savings until an approved experiment verifies them. Only verified savings are shown in green.
Coverage and data quality
30 of 30 days with data- Days with data
- 30 of 30
- Providers
- Synthetic demo source
- Currencies
- USD
- Attributed records
- 100%
- Outcome telemetry
- NoneWithout outcome telemetry, cost per successful task is unavailable.
- Provisional records
- Not applicableSample records aren’t imported or re-synced.
- Reconciliation conflicts
- Not applicableSample records aren’t imported or re-synced.
- Connections
- 1 connected, 0 failed
Spend profile
Charted from the spend ledger for the audit’s UTC days.Daily observed spend by project
$24,313.82 total
| Date | proj_support_agent | proj_sales_research | proj_doc_ingest | proj_internal_eval |
|---|---|---|---|---|
| 2026-09-29 | $582.21 | $183.02 | $40.58 | $12.09 |
| 2026-09-30 | $499.72 | $177.20 | $38.20 | $12.68 |
| 2026-10-01 | $588.53 | $189.95 | $40.97 | $11.29 |
| 2026-10-02 | $509.20 | $198.05 | $44.17 | $11.32 |
| 2026-10-03 | $333.92 | $108.40 | $24.92 | $7.32 |
| 2026-10-04 | $369.69 | $116.89 | $23.42 | $8.47 |
| 2026-10-05 | $589.38 | $180.34 | $44.92 | $11.50 |
| 2026-10-06 | $493.37 | $206.70 | $37.51 | $12.42 |
| 2026-10-07 | $590.80 | $192.78 | $41.57 | $13.50 |
| 2026-10-08 | $492.35 | $210.05 | $39.81 | $13.72 |
| 2026-10-09 | $542.86 | $182.35 | $37.73 | $12.59 |
| 2026-10-10 | $352.54 | $110.87 | $23.07 | $7.44 |
| 2026-10-11 | $320.29 | $125.83 | $24.48 | $7.66 |
| 2026-10-12 | $580.35 | $187.18 | $42.72 | $12.97 |
| 2026-10-13 | $571.64 | $208.72 | $37.38 | $13.46 |
| 2026-10-14 | $564.19 | $181.62 | $38.14 | $12.69 |
| 2026-10-15 | $573.67 | $187.60 | $42.32 | $12.36 |
| 2026-10-16 | $553.39 | $191.26 | $44.00 | $13.41 |
| 2026-10-17 | $1,120.81 | $110.43 | $23.66 | $8.35 |
| 2026-10-18 | $1,060.72 | $119.89 | $24.99 | $8.49 |
| 2026-10-19 | $1,974.26 | $177.42 | $43.12 | $11.40 |
| 2026-10-20 | $500.78 | $187.06 | $43.34 | $11.88 |
| 2026-10-21 | $551.99 | $205.15 | $37.97 | $12.32 |
| 2026-10-22 | $543.10 | $210.41 | $43.63 | $12.73 |
| 2026-10-23 | $597.02 | $193.03 | $43.12 | $12.05 |
| 2026-10-24 | $351.30 | $126.60 | $24.42 | $7.29 |
| 2026-10-25 | $331.78 | $124.68 | $24.98 | $7.99 |
| 2026-10-26 | $600.58 | $176.54 | $40.66 | $11.42 |
| 2026-10-27 | $515.45 | $173.43 | $42.64 | $11.94 |
| 2026-10-28 | $491.73 | $188.45 | $41.17 | $12.00 |
Shaded: spend spike on project proj_support_agent · anomalous spend, not a saving
By provider
- Synthetic demo source$24,313.82100%
By project
- proj_support_agent$17,747.6072.9%
- proj_sales_research$5,131.8921.1%
- proj_doc_ingest$1,099.594.5%
- proj_internal_eval$334.731.3%
By model
- frontier-large$23,214.2295.4%
- efficient-small$1,099.594.5%
Findings
Grouped by evidence class. Each amount keeps its class. Only an approved, measured experiment turns money into a verified saving.12 findings · 11 openObserved spend3
Not summedBilled cost reported by the provider. Observations can describe the same spend from different angles (a project and its model), so they aren’t added up.- Info95.4% of attributed spend is on model frontier-large$23,214.22
- Info72.9% of attributed spend is on project proj_support_agent$17,747.60
- InfoNo outcome data: cost per successful task is unknownNot actionableUnavailable
Savings levers
Each amount is a potential saving under a stated scenario. Levers can act on the same spend, so they’re listed one by one and the combined line removes the overlap instead of adding them up.$8,404 combined, overlap removed (estimate)| Lever | Scenario | Amount |
|---|---|---|
| Prompt caching | 25% of uncached input served from cache | $750.84 |
| Prompt caching | 25% of uncached input served from cache | $35.56 |
| Batch processing | 20% of this workload can wait up to 24 hours | $1,774.76 |
| Batch processing | 20% of this workload can wait up to 24 hours | $513.19 |
| Batch processing | 20% of this workload can wait up to 24 hours | $109.96 |
| Model routing | 20% of premium-model tokens handled by the cheaper model | $4,255.93 |
| Output length | 15% fewer output tokens | $2,158.92 |
| Output length | 15% fewer output tokens | $102.26 |
Recommended next steps
Collected from every finding, duplicates merged, ordered by severity and evidence class.- 1Check deploy and incident logs for the affected dates1 finding
- 2Send run/outcome telemetry for this workload to see whether the extra spend produced successful tasks1 finding
- 3Create a daily spend alert for this dimension to catch the next occurrence early1 finding
- 4List the jobs in this project that nobody waits on (evals, enrichment, nightly summaries)3 findings
- 5Submit those jobs through the provider’s batch API (synthetic source: the discount is illustrated)3 findings
Show 3 moreShow less
- 6Move one job, compare cost per successful task for two weeks3 findings
- 7Set max output tokens per task from the 95th percentile of successful answers2 findings
- 8Ask for structured output (JSON) where a person doesn’t read prose2 findings
Methodology
11 deterministic detectors over provider-billed cost from Synthetic demo source. Each amount keeps one evidence class; classes are never added together.
Limitations of this report3
- Amounts are provider-reported billed cost for complete UTC days; the current partial day is excluded from spike and growth detection.
- Cost data only: no outcome telemetry, so no cost per successful task and no failed-run analysis.
- All data here is synthetic and generated for demonstration. It is not real customer spend.
How each detector works11
- batch_processing 1.0.0
- Large standard-tier workloads on providers with a batch discount. Sizes moving a stated share to asynchronous batch processing.
- budget_burn 1.0.0
- Month-to-date spend run-rate projects above the organization’s configured monthly budget.
- failed_run_spend 1.0.0
- Estimated spend on runs whose failure was verified (customer-reported, system-verified or human-reviewed).
- model_routing 1.0.0
- A premium model carrying most spend while a much cheaper model from the same provider is already in use. Rates are measured from your own cost and token data.
- outcome_data_gap 1.0.0
- Material spend with no outcome telemetry — an analysis gap, never presented as waste.
- output_length 1.0.0
- Models where output tokens make up most of the cost. Sizes a stated reduction in output length.
- prompt_caching 1.0.0
- Models with material uncached input spend and a low cached share. Sizes moving a stated share of input to the provider’s cached rate.
- saving_guard 1.0.0
- A verified saving whose billing signal (cache share, routing share, output ratio or batch share) has given back half or more of its gain in the last 7 days.
- spend_concentration 1.0.0
- A single project/workspace/model accounts for at least 60% of spend in the period.
- spend_spike 1.0.0
- Days where spend on a project/workspace/model exceeds 2.5× its trailing 14-day median (minimum 7 days of history).
- sustained_growth 1.0.0
- Week-over-week spend growth above 20% for three consecutive complete weeks on one dimension.
Evidence classes5
- Observed spend
- Billed cost reported by the provider.
- Anomalous spend
- Unusual spend. Not a saving — may be legitimate.
- Potential saving
- Modelled under stated assumptions. Not yet validated.
- Validated opportunity
- Feasible and evidenced, confirmed by your team. Not yet measured.
- Verified saving
- Measured in a recorded, approved experiment.
Detectors produce only the first three classes. A person promotes a potential saving to a validated opportunity; only a recorded, approved experiment creates a verified saving.
What billing data can’t tell you4
- Whether the spend produced successful outcomes. That needs outcome telemetry, which billing data doesn’t contain.
- Why a spike happened. Billing data shows when and where, not which release, prompt or customer caused it. We never read prompts or outputs.
- Whether extra spend was waste or planned demand, such as a launch, a backfill or a load test.
- The cash effect of a change. Under prepaid commitments, lower usage frees capacity rather than reducing near-term invoices.
Full method: how agent loss prevention works. Same data over the API and MCP server.
Appendix: data sources and versions
- Organization
- Sample organization (synthetic data)
- Report
- Sample report, computed on request and not stored
- Period
- 2026-09-29 to 2026-10-29 (end exclusive, UTC)
- Generated
- 28 Oct 2026, 13:20 UTC
- Data sources
- Synthetic demo source
- Snapshot hash
- f69e9b483175…ca807fComputed the same way as a real audit, over the generated records.
- Money
- Exact integer microsStored in the billed currency and rounded only for display. Currencies are never converted or summed.
Detector versions11
| Detector | Version | Findings |
|---|---|---|
| batch_processing | 1.0.0 | 3 |
| budget_burn | 1.0.0 | 0 |
| failed_run_spend | 1.0.0 | 0 |
| model_routing | 1.0.0 | 1 |
| outcome_data_gap | 1.0.0 | 1 |
| output_length | 1.0.0 | 2 |
| prompt_caching | 1.0.0 | 2 |
| saving_guard | 1.0.0 | 0 |
| spend_concentration | 1.0.0 | 2 |
| spend_spike | 1.0.0 | 1 |
| sustained_growth | 1.0.0 | 0 |