AI Compliance Agents

The compliance AI system includes two agents that work together to reduce manual review burden: the Case Investigation Agent and the Match Verification Agent. Together, they can auto-resolve 50–70% of compliance cases that would otherwise require manual analyst review.

Agents Overview

Agent	Purpose	Auto-Resolution Target
Case Investigation	Investigates ReviewRequired cases by analyzing transaction patterns, customer history, and contextual signals	50–70% of review cases
Match Verification	Evaluates screening matches using Bayesian confidence scoring to distinguish true positives from false positives	60–80% of screening matches

Case Investigation Agent

When a transaction or customer screening produces a ReviewRequired status, the Case Investigation Agent evaluates the case automatically.

How It Works

Case Status: ReviewRequired
  │
  ▼
Case Investigation Agent
  ├── Analyze transaction patterns
  ├── Check customer risk history
  ├── Evaluate geographic risk signals
  ├── Cross-reference related cases
  │
  ▼
Decision
  ├── AUTO_CLEARED (low risk confirmed)
  ├── NEEDS_REVIEW (insufficient confidence)
  └── AUTO_ESCALATED (high risk confirmed)

Decision Factors

Factor	Weight	Description
transaction_pattern	0.25	Deviation from customer's normal behavior
customer_risk_profile	0.20	Historical risk level and previous alerts
geographic_risk	0.20	Source/destination country risk rating
counterparty_risk	0.15	Known counterparty risk signals
amount_anomaly	0.10	Statistical deviation from typical amounts
temporal_pattern	0.10	Unusual timing or frequency patterns

Example case decision:

{
  "case_id": "case-uuid",
  "agent": "case_investigation",
  "decision": "AUTO_CLEARED",
  "confidence": 0.93,
  "factors": {
    "transaction_pattern": { "score": 0.95, "note": "Consistent with regular supplier payment pattern" },
    "customer_risk_profile": { "score": 0.90, "note": "Low-risk customer, no previous alerts" },
    "geographic_risk": { "score": 0.88, "note": "Domestic transfer, low-risk jurisdiction" },
    "counterparty_risk": { "score": 0.97, "note": "Known counterparty, established relationship" },
    "amount_anomaly": { "score": 0.92, "note": "Within 1 std deviation of monthly average" },
    "temporal_pattern": { "score": 0.96, "note": "Regular business hours, consistent frequency" }
  },
  "reasoning": "Low-risk domestic transfer to established counterparty. Transaction pattern consistent with regular supplier payments over 12-month history.",
  "processed_at": "2026-04-07T14:30:00Z"
}

Match Verification Agent

The Match Verification Agent uses Bayesian confidence scoring to evaluate screening matches against sanctions lists, PEP databases, and adverse media sources.

Bayesian Confidence Scoring

Instead of relying solely on string similarity, the agent considers prior probability and multiple evidence signals:

P(true_match | evidence) = P(evidence | true_match) * P(true_match)
                           ────────────────────────────────────────
                                        P(evidence)

Evidence signals evaluated:

Signal	Description
name_similarity	Fuzzy string match score (Jaro-Winkler, Levenshtein)
date_of_birth	DOB match or near-match
nationality	Country of nationality match
id_numbers	Government ID number match
known_aliases	Match against known aliases
associated_entities	Connections to other matched entities

Example match evaluation:

{
  "match_id": "match-uuid",
  "agent": "match_verification",
  "decision": "FALSE_POSITIVE",
  "bayesian_confidence": 0.03,
  "prior_probability": 0.001,
  "evidence": {
    "name_similarity": { "score": 0.82, "note": "Common name, high base rate of coincidental matches" },
    "date_of_birth": { "match": false, "note": "DOB differs by 15 years" },
    "nationality": { "match": false, "note": "Different nationality" },
    "id_numbers": { "match": false, "note": "No matching ID numbers" }
  },
  "reasoning": "Name similarity is moderate but all secondary identifiers (DOB, nationality, ID) do not match. Bayesian posterior probability of true match is 3%."
}

note

The Bayesian approach significantly reduces false positives compared to name-only matching. A match with 85% string similarity but no corroborating evidence (DOB, nationality) will correctly score as a likely false positive.

Enable the Agents

Configure both compliance agents via the API.

Case Investigation Agent

PUT /api/v1/ai/agents/case_investigation/config

cURL

curl -X PUT https://api.korastratum.com/api/v1/compliance/ai/agents/case_investigation/config \
  -H "Authorization: Bearer $TOKEN" \
  -H "X-Tenant-ID: demo_bank" \
  -H "Content-Type: application/json" \
  -d '{
    "enabled": true,
    "mode": "autonomous",
    "auto_clear_threshold": 0.90,
    "auto_escalate_threshold": 0.85,
    "max_daily_auto_decisions": 200,
    "excluded_risk_levels": ["CRITICAL"],
    "require_dual_signal": true
  }'

Node.js

const config = await complianceRequest("PUT", "/api/v1/ai/agents/case_investigation/config", token, {
  enabled: true,
  mode: "autonomous",
  auto_clear_threshold: 0.90,
  auto_escalate_threshold: 0.85,
  max_daily_auto_decisions: 200,
  excluded_risk_levels: ["CRITICAL"],
  require_dual_signal: true,
});

Python

config = compliance_request("PUT", "/api/v1/ai/agents/case_investigation/config", token, {
    "enabled": True,
    "mode": "autonomous",
    "auto_clear_threshold": 0.90,
    "auto_escalate_threshold": 0.85,
    "max_daily_auto_decisions": 200,
    "excluded_risk_levels": ["CRITICAL"],
    "require_dual_signal": True,
})

Go

config, err := complianceRequest(ctx, "PUT", "/api/v1/ai/agents/case_investigation/config", token, map[string]interface{}{
    "enabled":                  true,
    "mode":                     "autonomous",
    "auto_clear_threshold":     0.90,
    "auto_escalate_threshold":  0.85,
    "max_daily_auto_decisions": 200,
    "excluded_risk_levels":     []string{"CRITICAL"},
    "require_dual_signal":      true,
})

Match Verification Agent

PUT /api/v1/ai/agents/match_verification/config

{
  "enabled": true,
  "mode": "autonomous",
  "false_positive_threshold": 0.10,
  "true_positive_threshold": 0.80,
  "require_secondary_identifier": true,
  "max_daily_auto_decisions": 500
}

Configuration Options

Field	Type	Description
enabled	boolean	Enable or disable the agent
mode	string	shadow (observe only) or autonomous (take action)
auto_clear_threshold	number	Minimum confidence to auto-clear a case
auto_escalate_threshold	number	Minimum risk confidence to auto-escalate
false_positive_threshold	number	Below this Bayesian score, mark as false positive
true_positive_threshold	number	Above this Bayesian score, mark as true positive
max_daily_auto_decisions	integer	Safety cap on autonomous decisions per day
excluded_risk_levels	string[]	Risk levels that always require human review
require_dual_signal	boolean	Require at least 2 corroborating signals for auto-clear
require_secondary_identifier	boolean	Require non-name identifier match for true positive

warning

CRITICAL risk level cases should always require human review. We recommend keeping this in excluded_risk_levels even in autonomous mode.

Agent Performance Metrics

GET /api/v1/ai/agents/performance

Response:

{
  "case_investigation": {
    "enabled": true,
    "mode": "autonomous",
    "last_30_days": {
      "cases_processed": 3420,
      "auto_cleared": 2052,
      "auto_escalated": 171,
      "sent_to_human": 1197,
      "auto_resolution_rate": 0.65,
      "accuracy_vs_human_baseline": 0.97
    }
  },
  "match_verification": {
    "enabled": true,
    "mode": "autonomous",
    "last_30_days": {
      "matches_processed": 8940,
      "false_positives_cleared": 6258,
      "true_positives_flagged": 894,
      "sent_to_human": 1788,
      "auto_resolution_rate": 0.80,
      "accuracy_vs_human_baseline": 0.96
    }
  }
}

Next Steps

• Case Management — Manual case review workflow.
• Batch Screening — Bulk customer and transaction screening.
• Transaction Monitoring — Real-time transaction surveillance.
• Matching Algorithms — Details on fuzzy matching and scoring.