AI · Governance
AI policies
Thresholds, routing, escalation, model selection · last updated by Sam Hagen on 2026-05-02
Active policies
24
across thresholds, routing, escalation, models
Confidence threshold
Medium · 0.70
global default · per-policy overrides allowed
Auto-escalations
8
per week · trailing avg
Models in use
4
Opus 4.7 · Sonnet 4.6 · Sonnet 4.6 (vision) · Haiku 4.5
Recommended adjustments
- Risk flag confidence at 0.70 is producing 23% false positives this month — consider raising to 0.75.
- Brief auto-categorization confidence threshold of 0.65 means 1 in 8 intakes still goes to human review — within target.
AI grounded in last 30 days of policy outcomes
Model assignment per task
Each AI workflow has a primary and (optional) fallback model
| Task | Primary model | Fallback | Notes |
|---|---|---|---|
| Brief intake structuring | Claude Opus 4.7 | Claude Sonnet 4.6 | Highest accuracy on long ambiguous text |
| Vendor relationship intel | Claude Sonnet 4.6 | Claude Haiku 4.5 | Fast continuous refresh |
| Risk synthesis | Claude Sonnet 4.6 | Claude Haiku 4.5 | Multi-source aggregation |
| Anomaly detection | Claude Haiku 4.5 | none | Lightweight pattern matching |
| Contract sanity check | Claude Opus 4.7 | Claude Sonnet 4.6 | Legal-grade review |
| Visual mood-board match | Claude Sonnet 4.6 (vision) | none | Image embedding + similarity |
| Recap deck composition | Claude Sonnet 4.6 | none | Long-form generation |
| Burn forecast | Claude Haiku 4.5 | none | Numeric extrapolation |
| Pattern detection (cross-event) | Claude Sonnet 4.6 | none | Soft clustering |
| Standup brief composition | Claude Sonnet 4.6 | none | Daily generation |
This week's model spend
Open AI cost dashboardTotal tokens used
12.4M
Total cost
$487
Avg latency
2.1s
P95 latency
4.8s
AI flags, doesn’t fix · every threshold change is logged and reversible · 24 active policies govern the system