Deviation Intelligence

The core product innovation — classifying overrides in real time

The system classifies overrides in near real time and applies selective intervention — preserve agent autonomy on corrective deviations, add friction only for cost-risk and policy-risk subsets, and continuously improve from outcomes. This is the product.

Deviation Taxonomy

Five types of override, each with a distinct intervention strategy.

Corrective Override

Tier L

Agent corrects a weak copilot suggestion — wrong policy cited, missing context, or tone mismatch.

Allow + capture rationale + feed learning loop

Empathy-Within-Policy

Tier L

Agent uses judgment while staying inside policy and cost guardrails — softer tone, extra explanation.

Allow + light logging

Cost-Risk Override

Tier M–H

Compensation likely above tolerance or precedent band for this case type.

Justification required; escalate by threshold

Policy-Risk Override

Tier H–C

Possible breach of mandatory process, safety, or compliance boundary.

Supervisor approval or block per tier

Quality-Risk Closure

Post-audit

Fast closure with weak resolution quality — short interaction, no confirmation, unresolved signals.

Post-audit flag + coaching / escalation checks

Risk Scoring Model

Weighted multi-factor score, range 0–21. Computed in <500ms.

RiskScore = W₁·P + W₂·F + W₃·Fr + W₄·R + W₅·C
W₁ = 2.0Policy Criticality
0, 1, 3, 5

From policy tier: 0 = no match, 1 = advisory, 3 = mandatory, 5 = zero-tolerance

W₂ = 1.5Financial Exposure
0–5

Normalized: (proposed_compensation − market_median) / threshold_band

W₃ = 1.5Fraud Signal
0–5

From upstream fraud detection: 0 = no signal, 5 = confirmed indicators

W₄ = 1.0Recurrence Pattern
0–3

Agent's high-risk deviation count in trailing 30 days

W₅ = 1.0Confidence Gap
0–3

|copilot_confidence − agent_action_distance|

Intervention Tiers

Risk score maps to four intervention levels. Friction is concentrated on the ~10% that matters.

L
Low
Score 0–4
Allow + log
~70%
of overrides
M
Medium
Score 5–9
Allow + reason tag
~20%
of overrides
H
High
Score 10–15
Supervisor approval
~8%
of overrides
C
Critical
Score 16+
Block / safety escalation
~2%
of overrides
Low frictionHigh friction

Learning Loop

Five feedback signals that continuously improve the system.

Approved corrective overrides
Suggestion improvement backlog
Rejected risky overrides
Guardrail rule updates
Suggestion → close (no override)
Self-serve graduation candidates
QA audit labels
Risk threshold calibration + training data
Agent reason-tag text
Policy gap identification

Calibration Cadence

Weekly

High-risk deviation review; threshold adjustment proposals; FP/FN rates

Biweekly

Suggestion quality review; approved override batch promotion

Monthly

Full threshold and policy mapping update with compliance sign-off

Quarterly

Model retrain evaluation; self-serve graduation review; cross-market comparison

Classifier Error Handling

The classifier will make mistakes. Each error type has a UX recovery path and systemic correction.

Design principle: Err toward Tier M (reason tag) rather than Tier H (block) when uncertain. A reason tag costs seconds; a false block costs minutes and trust.

False PositiveCorrective flagged as risky
Impact

Agent frustrated; resolution delayed; supervisor queue noise

UX Recovery

Agent sees risk rationale inline; can submit structured appeal; appeal auto-logged for calibration

Systemic Correction

Weekly calibration: FP rate tracked per rule; rules with FP >20% flagged for threshold adjustment

False NegativeRisky passes as corrective
Impact

Risk leaks through; caught only at post-audit

UX Recovery

Post-resolution audit catches pattern; agent receives coaching feedback

Systemic Correction

Audit-flagged misses fed into calibration cycle; persistent misses trigger rule additions

AmbiguousMedium classifier confidence
Impact

Uncertain intervention — defaults conservatively

UX Recovery

Defaults to Tier M (reason tag required); does not block agent

Systemic Correction

Ambiguous cases reviewed weekly; labeled and added to training set