Governance · Policy configuration
Escalation Rules & Autonomy Settings
Define what the agent network may do on its own, what requires an associate's approval, and what must go straight to a human specialist. Changes are versioned and reviewed by the governance board.
Policy v4.2 · effective Jul 28, 2026
Autonomy posture
Calibrated
Balanced between throughput and control
Action authority split
4·4·3
Auto4(36%)
Approval4(36%)
Human3(27%)
Straight-through rate
61.4%
Trailing 30 days · target 55–70%
Override rate
4.1%
Human reversals of AI recommendations
Risk tiers
Every case is scored into a tier at triage. The tier sets the baseline autonomy before action-level rules apply.
| Tier | Entry criteria | Volume | Baseline autonomy |
|---|---|---|---|
R1Routine | Value < $100 · verified identity · no fraud flags Full autonomy across informational and servicing actions. | 62% of case volume | |
R2Standard | $100–$500 · single anomaly signal · verified identity AI drafts a complete plan; associate approves before execution. | 24% of case volume | |
R3Elevated | $500–$2,500 · conflicting identity or geo signals Approval gate plus mandatory policy citation and evidence pack. | 11% of case volume | |
R4Critical | > $2,500 · suspected account takeover · vulnerable customer No automated execution. Direct routing to a named specialist queue. | 3% of case volume |
Confidence thresholds
Composite confidence is the weighted agreement of all specialist agents on the proposed plan.
Auto-execute floor0.92
Composite confidence required for unattended execution.
Approval-gate floor0.75
Below this, AI may only recommend — never pre-select actions.
Mandatory human floor0.60
Below this, the orchestrator escalates without proposing a plan.
Agent disagreement delta0.15
Confidence spread across specialists that forces a review.
Case CS-48291 resolved at composite 0.90 — below the 0.92 auto floor, which is why the $420 provisional credit was routed to associate approval and then to L2 Fraud.
Action categories
Authority is set per action and always intersected with the case risk tier — the more restrictive rule wins.
| Category | Action | System | Reversibility | Authority |
|---|---|---|---|---|
| Informational | Read transaction & card history | Core Banking · CRM | N/A | |
| Informational | Cite policy or fee schedule to customer | Policy Engine | N/A | |
| Protective | Temporary card block / freeze | Card Services | Reversible | |
| Protective | Step-up identity verification | Socure · OTP | Reversible | |
| Servicing | Issue digital replacement card | Card Services | Partially | |
| Servicing | File Reg E dispute claim | Disputes Platform | Reversible | |
| Financial | Post provisional credit ≤ $500 | Ledger | Clawback req. | |
| Financial | Post provisional credit > $500 | Ledger | Clawback req. | |
| Financial | Fee reversal or goodwill credit | Ledger | Irreversible | |
| Account control | Close or restrict account | Core Banking | Irreversible | |
| Account control | Change contact details on file | CRM | Irreversible |
Policies by issue type
Disputes & chargebacks
Owner · Resolution & Escalation Agent
Auto ceiling
$500
Human SLA
3–5 min
Regulatory
Reg E · 10-day provisional credit clock
Auto
Single unrecognized charge under ceiling with clean history
Approval
Charge above ceiling or 2+ disputed items in 30 days
Human
Merchant on internal watchlist or prior dispute reversal
Account access
Owner · Triage Agent → Identity
Auto ceiling
n/a
Human SLA
2–4 min
Regulatory
KYC · CIP re-verification
Auto
Password / passkey reset after full-factor verification pass
Approval
Lockout release with one failed factor
Human
Device + geo mismatch, or suspected account takeover
Payments & transfers
Owner · Account & Transaction Agent
Auto ceiling
$1,000
Human SLA
5–8 min
Regulatory
NACHA · OFAC screening
Auto
Status lookup, trace request, or stop-payment on pending item
Approval
Recall of a settled ACH under ceiling
Human
Wire recall, new payee, or OFAC name hit
Fraud & account takeover
Owner · Resolution & Escalation Agent
Auto ceiling
$0
Human SLA
< 90 sec
Regulatory
SAR review · FinCEN referral
Auto
Protective card freeze and alert suppression
Auto
Evidence pack assembly and specialist briefing
Human
Any monetary movement, SAR-eligible pattern, or elder-abuse signal
Autonomy pyramid
Risk tiers stack by volume. Higher tiers are narrower and more restricted.
Hard escalation triggers
These override every tier and action rule. When one fires, the orchestrator stops and routes to a human.
Locked by compliance
Reg E provisional credit timing cannot be shortened
OFAC hits always route to sanctions review
SAR-eligible patterns cannot be auto-closed
Customer-visible AI disclosure always on
Change history
Jul 28 · 14:02
Raised auto-execute floor 0.90 → 0.92 after Q2 calibration review.
R. Osei · Ops Director
Jul 21 · 09:41
Provisional credit ceiling held at $500 pending Reg E audit.
Governance Board
Jul 09 · 16:20
Fraud issue type moved to R4 default; zero monetary autonomy.
D. Ruiz · L2 Fraud Lead