Spend trend & forecast — cloud vs. tokens
Actuals through today (Jul 16); dotted lines are P50 predictions to month end
Top recommendations
Ranked by dollar impact, one click to approve
Attribution snapshot
Top spenders this month
| ML Platform | $312k | |
| Payments Eng | $188k | |
| Risk & Fraud AI | $143k | |
| Data Platform | $76k | |
| Digital Channels | $54k |
Anomalies today
Flagged by the forecasting agent
Breach forecast
Probability of exceeding budget ceiling
Projected monthly spend — P10–P90 band vs. budget ceiling
Solid line is realized spend to date; dotted lines are the P10–P90 forecast from today forward
Per-service forecast — this month, P50 vs. ceiling share
Click a row to see which signals the model weighted most
| Service | P10 – P90 band | Breach risk |
|---|---|---|
| Risk & Fraud AI (GPU) | 73% | |
| ML Platform (tokens) | 41% | |
| Payments Engineering | 19% | |
| Data Platform | 8% |
Agent reasoning log
Live trace of the forecasting agent's decisions this week
Spend by team — this month
Drawn from the cost attribution graph, no manual tagging required — $773k total, split by team and by source
| Team | Cloud | Tokens | Total |
|---|---|---|---|
| ML Platform | $312k | ||
| Payments Engineering | $188k | ||
| Risk & Fraud AI | $143k | ||
| Data Platform | $76k | ||
| Digital Channels | $54k |
How ambiguous spend gets resolved without a tag
ml-shared-01.Optimize — cut what's already being spent
Four levers, ranked by confidence and dollar impact — no chart needed to see which ones matter
Idle GPU cleanup — $41k/mo
Clusters idle 6+ days with zero scheduled jobs. 97% confidence — safe to approve in one click.
Model routing & right-sizing — $23k/mo
Route ~60% of calls to a smaller model where quality holds up in testing — ~80% lower cost-per-answer, 88% confidence.
Semantic + context caching — $18k/mo
Serve repeated prompts from cache instead of re-computing them. Needs a small integration lift — 71% confidence.
Orphaned volume cleanup — $9k/mo
1.2k volumes with zero attachment events for 30+ days. 99% confidence — flagged auto-safe.
Monthly token cost — lever by lever
$63.9k → $22.4k, ≈65% lower, applied in sequence — the one chart worth keeping here
Pre-buy — commit ahead of demand
Once the steady-state floor is known, lock in the cheapest instrument for it
Commitment ladder
Procurement agent recommendation for the steady baseline, human-gated
| Instrument | Coverage | Discount vs. on-demand |
|---|---|---|
| Reserved instances (1-yr) | −42% | |
| Savings plan (compute) | −55% | |
| Provisioned throughput (LLM) | −72% |
Approval gate
No material commitment is placed automatically — every buy routes to a named approver before it executes.