AI Decision Agents

The IDV Decision Intelligence Agent automates the manual review queue by applying confidence-weighted scoring to verification cases. Instead of simple pass/fail boolean flags, the agent evaluates multiple signals and produces a nuanced confidence score that accounts for OCR errors, image quality, and data consistency.

How It Works

Manual Review Queue
  │
  ▼
Decision Intelligence Agent
  ├── OCR confidence analysis
  ├── Face match scoring
  ├── Document authenticity checks
  ├── Data consistency validation
  │
  ▼
Confidence-Weighted Decision
  ├── HIGH confidence (>0.90) ──▶ Auto-approve
  ├── MEDIUM confidence (0.70-0.90) ──▶ Flag for human review
  └── LOW confidence (<0.70) ──▶ Auto-reject

Key Capabilities

Capability	Description
Confidence-weighted scoring	Multi-signal scoring (0.0–1.0) instead of binary pass/fail
OCR error tolerance	Handles common OCR misreads (O/0, I/1, diacritics) without false rejections
False positive detection	Identifies cases that would be incorrectly rejected by rule-based systems
Shadow mode	Agent runs alongside human reviewers without taking action — for validation
Audit trail	Every decision includes full reasoning and signal breakdown

Confidence Scoring

The agent produces a composite confidence score from multiple signals:

Signal	Weight	Description
face_match	0.30	Biometric face comparison score
ocr_data_match	0.25	Extracted text matches provided data
document_authenticity	0.20	Document tamper detection score
data_consistency	0.15	Cross-field consistency (DOB, name, ID number)
image_quality	0.10	Image clarity and completeness

Example scoring output:

{
  "verification_id": "ver-uuid",
  "composite_score": 0.92,
  "decision": "APPROVE",
  "signals": {
    "face_match": { "score": 0.95, "weighted": 0.285 },
    "ocr_data_match": { "score": 0.88, "weighted": 0.220 },
    "document_authenticity": { "score": 0.97, "weighted": 0.194 },
    "data_consistency": { "score": 0.90, "weighted": 0.135 },
    "image_quality": { "score": 0.86, "weighted": 0.086 }
  },
  "ocr_corrections": [
    { "field": "id_number", "extracted": "A12O45678", "corrected": "A1204567B", "confidence": 0.82 }
  ],
  "reasoning": "High confidence across all signals. OCR correction applied to ID number (O→0 substitution)."
}

Shadow Mode

Before enabling autonomous decisions, run the agent in shadow mode. In this mode, the agent processes every case in the manual review queue and records what it would decide, but does not take any action. Human reviewers continue to make all decisions.

This allows you to:

• Compare agent decisions against human reviewer outcomes
• Measure agreement rate and identify edge cases
• Build confidence before enabling autonomous mode

tip

We recommend running shadow mode for at least 2 weeks and targeting a >95% agreement rate with human reviewers before enabling autonomous decisions.

Enable the Agent

Configure the IDV Decision Intelligence Agent via the API.

PUT /api/v1/ai/agents/idv_decision_agent/config

cURL

curl -X PUT https://api.korastratum.com/api/v1/idv/ai/agents/idv_decision_agent/config \
  -H "Authorization: Bearer $TOKEN" \
  -H "X-Tenant-ID: demo_bank" \
  -H "Content-Type: application/json" \
  -d '{
    "enabled": true,
    "mode": "shadow",
    "auto_approve_threshold": 0.90,
    "auto_reject_threshold": 0.70,
    "ocr_error_tolerance": true,
    "max_daily_auto_decisions": 500,
    "excluded_document_types": ["PASSPORT"],
    "notify_on_override": true
  }'

Node.js

const config = await idvRequest("PUT", "/api/v1/ai/agents/idv_decision_agent/config", token, {
  enabled: true,
  mode: "shadow",
  auto_approve_threshold: 0.90,
  auto_reject_threshold: 0.70,
  ocr_error_tolerance: true,
  max_daily_auto_decisions: 500,
  excluded_document_types: ["PASSPORT"],
  notify_on_override: true,
});

Python

config = idv_request("PUT", "/api/v1/ai/agents/idv_decision_agent/config", token, {
    "enabled": True,
    "mode": "shadow",
    "auto_approve_threshold": 0.90,
    "auto_reject_threshold": 0.70,
    "ocr_error_tolerance": True,
    "max_daily_auto_decisions": 500,
    "excluded_document_types": ["PASSPORT"],
    "notify_on_override": True,
})

Go

config, err := idvRequest(ctx, "PUT", "/api/v1/ai/agents/idv_decision_agent/config", token, map[string]interface{}{
    "enabled":                  true,
    "mode":                     "shadow",
    "auto_approve_threshold":   0.90,
    "auto_reject_threshold":    0.70,
    "ocr_error_tolerance":      true,
    "max_daily_auto_decisions": 500,
    "excluded_document_types":  []string{"PASSPORT"},
    "notify_on_override":       true,
})

Configuration Options

Field	Type	Description
enabled	boolean	Enable or disable the agent
mode	string	shadow (observe only) or autonomous (take action)
auto_approve_threshold	number	Minimum composite score for auto-approval (0.0–1.0)
auto_reject_threshold	number	Maximum composite score for auto-rejection (0.0–1.0)
ocr_error_tolerance	boolean	Enable intelligent OCR error correction
max_daily_auto_decisions	integer	Safety cap on autonomous decisions per day
excluded_document_types	string[]	Document types that always require human review
notify_on_override	boolean	Send notification when agent decision differs from what a human would decide

warning

When switching from shadow to autonomous mode, the agent will begin auto-approving and auto-rejecting cases. Ensure your thresholds are validated against shadow mode results before making this change.

Get Agent Status

GET /api/v1/ai/agents/idv_decision_agent/status

Response:

{
  "agent_id": "idv_decision_agent",
  "enabled": true,
  "mode": "shadow",
  "today": {
    "cases_processed": 142,
    "auto_approved": 0,
    "auto_rejected": 0,
    "flagged_for_review": 142,
    "agreement_rate_with_humans": 0.96
  },
  "last_30_days": {
    "cases_processed": 4260,
    "agreement_rate_with_humans": 0.94,
    "false_positive_rate": 0.02,
    "false_negative_rate": 0.01
  }
}

Next Steps

• Verification Flow — End-to-end identity verification process.
• Scoring Reference — Verification scoring methodology.
• Webhooks — Subscribe to agent decision events.