VISTA — Virtual Iskan Screening, Triage & Approval

Six-Step Application Workflow

Current state vs. AI-enhanced state at each stage. Click any step to expand the full specification.

Step 01

Application Intake

Manual data entry, missing fields, no pre-screening→AI Pre-Classification Agent scores risk and pre-identifies discrepancy patterns in real time

🤖 Pre-Classification Agent

Step 1 — Application IntakeAgent 1: Pre-Classification

Current State (Manual)

Manual data entry with no pre-validation
Missing fields cause downstream delays
No real-time eligibility checking
No risk scoring at entry point

AI Enhancement

Pre-Classification Agent runs at point of intake
Assigns risk score (0.0–1.0) to every application
Predicts likely discrepancy patterns
Triggers early data enrichment actions

Output

Risk score per application
Predicted discrepancy type flags
Recommended pre-screening actions
High-risk routing directive

Design Principle: Intercepting risk at the earliest possible stage significantly reduces the cost of remediation downstream. Every prediction generated at Step 1 gives the system a head start on evidence preparation before the rules engine runs at Step 5.

▼

Step 02

Data Enrichment & Validation

Manual checks against siloed government records→Automated multi-source API reconciliation with conflict map generation

🤖 Truth Verification Agent

Step 2 — Data Enrichment & ValidationAgent 4: Truth Verification

Current State (Manual)

Manual checks across siloed government databases
Reconciliation inconsistency between reviewers
No conflict map — contradictions go undetected
~30% of Step 6 exceptions could have been resolved here

AI Enhancement

Automated API calls: TAMM, WPS, DLD, MoF, Pension Fund
Structured conflict map generated for every application
Data inconsistencies surfaced before declaration stage
Reduces Step 6 caseload by ~30%

APIs Integrated

TAMM (Abu Dhabi identity & eligibility)
WPS (wage protection system)
DLD (Dubai Land Department property records)
MoF (Ministry of Finance pension)
ADHA internal case history

Key Impact: This step has the highest downstream leverage. Structured conflict detection at Step 2 eliminates approximately 30% of Step 6 manual review cases before they are created.

▼

Step 03

Document Upload & Review

Manual document inspection, Arabic OCR failure, unstructured data→Computer Vision OCR + bilingual NLP reads all documents into structured JSON

🤖 Evidence Extraction Agent

Step 3 — Document Upload & ReviewAgent 3: Evidence ExtractionAgent 7: Anomaly Detection

Current State (Manual)

Manual document review with no standardised extraction
Arabic OCR failure rates causing delays
Unstructured data not usable by downstream systems
Forgery detection entirely dependent on reviewer

AI Enhancement

Computer Vision OCR on all Arabic + English documents
NLP extracts structured fields into JSON
Anomaly Detection Agent runs simultaneously (forgery flags)
Document authenticity scoring via metadata analysis

Document Types

Emirates ID & passport
Salary certificates (Arabic/English)
Property ownership deeds
Marriage/divorce certificates
Retirement & pension letters
Bank statements

Bilingual Capability: The Evidence Extraction Agent handles Arabic-script documents natively, including right-to-left handwritten text. This is a core differentiator — most commercial platforms fail on Arabic OCR quality at government-document fidelity levels.

▼

Step 04

Applicant Declaration

Self-reported data unvalidated in real time→Real-time validation + citizen-facing bilingual clarification engine

🤖 Clarification Engine Agent

Step 4 — Applicant DeclarationAgent 5: Clarification EngineAgent 4: Truth Verification

Current State (Manual)

Self-reported data accepted without real-time validation
Conflicts between declaration and government data undetected
Applicants not prompted to resolve conflicts before submission
Errors discovered only at Step 6, causing costly delays

AI Enhancement

Real-time validation against government APIs at point of declaration
Citizen-facing clarification prompts resolve conflicts proactively
Bilingual (Arabic + English) clarification messages
Multi-round clarification state machine (max 3 rounds)

Clarification Flow

Round 1: Auto-generated clarification request
Round 2: Follow-up with specific document request
Round 3: Final escalation to human reviewer
Document receipt tracking per round
State machine manages all transitions

Strategic Insight: Resolving conflicts at the declaration stage (Step 4) is the most cost-effective intervention point. Every conflict resolved here eliminates a potential Step 6 manual review case.

▼

Step 05

Rules Engine Check

Raw flags only, no classification→AI Classifier enriches every flag with type, document checklist, resolution pathway & risk score

🤖 Classifier Agent

Step 5 — Rules Engine CheckAgent 2: ClassifierAgent 1: Pre-ClassificationAgent 7: Anomaly Detection

Current State (Manual)

Rules engine produces raw flags only — no intelligence
No classification of discrepancy type
No required document checklist generated
No resolution pathway suggested
No risk scoring on flags

AI Enhancement

Classifier Agent enriches every flag with:
→ Discrepancy type (20+ types / 6 categories)
→ Required document checklist
→ Resolution pathway (Exception / Stop Collection / Loan Deferral)
→ Risk score (0.0 – 1.0)
→ Confidence band (High / Medium / Low)

Classification Output

Discrepancy category (6 categories)
Specific type (20+ types)
Business resolution path
Evidence set required
Routing directive to correct agent

Model Note: The Classifier Agent uses a fine-tuned model trained on ADHA historical cases. Classification accuracy improves continuously as more cases are processed — a structural advantage that compounds over time.

▼

Step 06

Discrepancy Resolution

100% manual, 30–60 min per case→Full multi-agent resolution, sub-2-minute end-to-end. PoC validated March 2026.

✓ PoC Validated · <2 min · 92% confidence

Step 6 — Discrepancy Resolution✓ PoC ValidatedAll 8 Agents Active

Current State (Manual)

100% manual processing
30–60 minutes per case
No structured evidence extraction
No audit trail or XAI report
Reviewer judgment inconsistency
No fraud detection running in parallel

AI Enhancement

Full multi-agent resolution pipeline
Sub-2-minute end-to-end processing
Classification → Evidence → Verification → Decision → XAI
Anomaly Detection running in parallel
ADHA reviewer approves or overrides every final decision
Complete audit-ready XAI report generated

PoC Results (March 2026)

Case type: EXCEPTION_RETIREMENT
Resolution time: <2 minutes (vs 30–60 min manual)
Confidence score: 92%
Rounds to resolution: 2 clarification rounds
Outcome: Validated & approved for production

Governance Principle: Step 6 is the most complex and highest-value stage. ADHA reviewers retain full governance — the AI recommends, the reviewer decides.

Nine-Agent AI Architecture

Master Orchestrator coordinates 8 specialised agents. Click any agent to expand its full specification.

⬡ Agent 0 — Master Orchestrator

Coordinates all agents · Routes data flows · Maintains full decision audit log · Manages round-based clarification state machine

Azure AI FoundryLangChain OrchestrationAzure Service BusRedis StatePostgreSQL Audit Log

Agent 0 — Master Orchestrator

Core Responsibilities

Coordinates all 8 specialised agents
Routes data flows based on case type
Maintains full decision audit log for every case
Manages round-based clarification state machine
Handles parallel vs sequential agent execution
Consolidates outputs into unified case record

Technology Stack

Azure AI Foundry (orchestration runtime)
LangChain (agent chaining & memory)
Azure Service Bus (message routing)
Redis (state management)
PostgreSQL (audit log persistence)

Output

Workflow state object
Agent routing directives
Consolidated audit log
Case status updates
Escalation triggers

Design Principle: The Master Orchestrator is the single source of truth for every case. Individual agents can be upgraded or audited without touching the orchestrator logic.

Agent 01

Pre-Classification Agent

Analyzes Steps 1–4 data in real time to pre-identify likely discrepancy patterns before the rules engine runs

Agent 02

Classifier Agent

Interprets rules engine flags; maps to business resolution path; determines required evidence set

Agent 03

Evidence Extraction Agent

Computer Vision OCR + NLP reads all uploaded documents (Arabic/English) into structured JSON

Agent 04

Truth Verification Agent

Triangulates government API data, applicant declarations, and extracted document evidence; flags all conflicts

Agent 05

Clarification Engine Agent

Auto-generates bilingual clarification requests; manages multi-round clarification state (max 3 rounds)

Agent 06

Decision Engine Agent

Applies ADHA policy-as-code to map evidence to correct resolution: Approve, Defer, Stop Collection, Escalate, or Resubmit

Agent 07

Anomaly Detection Agent

Runs parallel to all agents; detects fraud patterns, document manipulation, identity spoofing, unusual patterns

Agent 08

XAI Report Generator

Generates complete audit-ready justification report for every recommendation with confidence scores and policy citations

Agent 1 — Pre-Classification Agent

Function

Analyzes Steps 1–4 data in real time to pre-identify likely discrepancy patterns before the rules engine runs. Assigns risk score and predicted discrepancy types to every application.

Technology Stack

LLM (GPT-4o)
Policy Rules Engine
Redis Cache (real-time scoring)

Output

Risk score (0.0–1.0)
Predicted discrepancy types
Recommended data enrichment actions

Note: This agent runs before the rules engine, giving the system a head start on classification. High-risk applications are pre-flagged so downstream agents can prioritise evidence extraction accordingly.

Agent 2 — Classifier Agent

Function

Interprets rules engine output flags and maps each flag to a structured business discrepancy type with confidence score, required document checklist, and resolution pathway.

Technology Stack

Fine-tuned LLM on ADHA historical cases
Policy Rules Engine
Discrepancy taxonomy (6 categories, 20+ types)

Output

Discrepancy category & type
Required document checklist
Resolution pathway directive
Risk score (0.0–1.0)
Confidence band (High / Medium / Low)

Note: The Classifier is the intelligence layer on top of the rules engine. Without it, flags are unactionable. With it, every flag becomes a structured work item with a clear resolution path.

Agent 3 — Evidence Extraction Agent

Function

Computer Vision OCR + NLP pipeline that reads all uploaded documents in Arabic and English and outputs structured JSON for downstream processing.

Technology Stack

Azure Document Intelligence (OCR)
GPT-4o Vision (bilingual extraction)
Apache Tika (multi-format parsing)

Output

Structured JSON per document
Field-level extraction (name, date, amount, etc.)
Document confidence score
Extraction completeness flag

Arabic-First: Native Arabic OCR at government-document fidelity is a core differentiator. Most commercial platforms degrade significantly on handwritten Arabic text.

Agent 4 — Truth Verification Agent

Function

Triangulates three data sources — government API records, applicant declarations, and extracted document evidence — and generates a structured conflict map.

Technology Stack

Multi-source API integration (TAMM, WPS, DLD, MoF)
LLM for conflict reasoning
Structured conflict schema

Output

Conflict map (field-level discrepancies)
Conflict severity scores
Resolution recommendations per conflict
Verified facts list

Note: Truth Verification is the most evidence-intensive agent. It must handle partial data, API timeouts, and conflicting ground truths gracefully.

Agent 5 — Clarification Engine Agent

Function

Generates bilingual (Arabic + English) clarification requests to applicants based on identified conflicts. Manages a state machine of up to 3 clarification rounds before escalating to a human reviewer.

Technology Stack

GPT-4o (bilingual message generation)
State machine (Redis)
Email/SMS notification gateway

Output

Bilingual clarification request (Round 1–3)
Document request checklist
Round completion flag
Escalation trigger (Round 3)

PoC Result: EXCEPTION_RETIREMENT case resolved in 2 clarification rounds (vs Round 3 maximum). This is the agent that communicates directly with the citizen.

Agent 6 — Decision Engine Agent

Function

Applies ADHA policy-as-code rules to the verified evidence set and maps the case to one of five resolution outcomes with a confidence score and XAI-ready justification chain.

Technology Stack

Policy-as-code engine (Drools / OPA)
LLM for edge-case reasoning
Confidence scoring model

Decision Outcomes

APPROVEDEFERSTOP COLLECTIONESCALATERESUBMIT

Governance: Agent 6 recommends — ADHA reviewer decides. Decision confidence <60% automatically triggers human escalation.

Agent 7 — Anomaly Detection Agent

Function

Runs in parallel to all pipeline stages. Continuously monitors for fraud patterns, document manipulation, identity spoofing, and statistical outliers.

Detection Types

Document forgery (metadata + visual analysis)
Identity spoofing (cross-reference checks)
Unusual income/property patterns
Repeat application patterns

Escalation Logic

Fraud risk score >0.75 → Immediate freeze + senior reviewer
Document forgery detected → Case hold + evidence preservation
All anomalies logged with full evidence chain

Architecture Note: Agent 7 runs parallel to every other agent simultaneously. It can freeze a case at any point if fraud risk exceeds threshold — architecturally enforced, not configurable.

Agent 8 — XAI Report Generator

Function

Generates a complete, audit-ready explainability report for every decision recommendation. The report provides full justification for ADHA reviewers and satisfies regulatory audit requirements.

Report Contents

Decision recommendation & confidence score
Full evidence chain (document references)
Policy citations (specific rule IDs)
Conflict resolution logic
Anomaly flags (if any)

Output Formats

PDF (Arabic + English, reviewer-facing)
JSON (system-to-system, audit log)
Dashboard widget (reviewer UI)

Regulatory Principle: No AI recommendation is presented to a reviewer without a complete XAI report. Every recommendation is explainable, challengeable, and auditable.

Exception Type Taxonomy

20+ discrepancy types across 6 business categories. Click any category to view the full detail.

PoC Validated (Phase 1)

Phase 2 Expansion

Anomaly / Fraud Detection

🪪 Identity & Eligibility

EXCEPTION_CITIZENSHIP_STATUS

EXCEPTION_AGE_MISMATCH

EXCEPTION_MARITAL_STATUS

EXCEPTION_DEPENDENT_COUNT

💰 Income Verification

EXCEPTION_INCOME_MISMATCH

EXCEPTION_EMPLOYMENT_STATUS

EXCEPTION_RETIREMENT

EXCEPTION_SECONDARY_INCOME

🏠 Property Ownership

EXCEPTION_PROPERTY_OWNED

EXCEPTION_PROPERTY_UNDISCLOSED

EXCEPTION_PROPERTY_TRANSFER

🏦 Loan & Financial

EXCEPTION_EXISTING_LOAN

EXCEPTION_LOAN_DEFAULT

EXCEPTION_DEFERRED_INSTALMENT

EXCEPTION_STOP_COLLECTION

📄 Document Integrity

EXCEPTION_DOCUMENT_EXPIRED

EXCEPTION_DOCUMENT_MISSING

EXCEPTION_DOCUMENT_FORGERY

🚨 Fraud & Anomaly

EXCEPTION_IDENTITY_SPOOFING

EXCEPTION_PATTERN_ANOMALY

EXCEPTION_REPEAT_APPLICATION

EXCEPTION_COORDINATED_FRAUD

Identity & Eligibility — Exception Detail

Agents Involved

Agent 2: Classifier
Agent 4: Truth Verification
Agent 6: Decision Engine

Resolution Paths

Document resubmission requested
Government API cross-check (TAMM)
Senior reviewer escalation if unresolved

Phase Status

Scheduled for Phase 2 production deployment. Foundational infrastructure validated in PoC.

Income Verification — Exception Detail

PoC Validated Case

EXCEPTION_RETIREMENT — validated March 2026
Resolution time: <2 minutes
Confidence score: 92%
Clarification rounds: 2 (of max 3)

APIs Used

WPS (Wage Protection System)
Pension Fund API (MoF)
TAMM employment history

Resolution Paths

Salary certificate cross-check
Pension letter validation
Employer verification request
Loan Deferral pathway (if applicable)

PoC Significance: EXCEPTION_RETIREMENT was selected for the PoC because it is one of the most common complex exceptions at ADHA, requiring coordination between pension data, employment history, and declaration fields.

Property Ownership — Exception Detail

Agents Involved

Agent 3: Evidence Extraction (deed OCR)
Agent 4: Truth Verification (DLD API)
Agent 6: Decision Engine

APIs Used

Dubai Land Department (DLD)
Abu Dhabi Land Registry
TAMM property records

Resolution Paths

Property transfer documentation review
Eligibility re-assessment post-transfer
Stop Collection pathway if criteria unmet

Loan & Financial — Exception Detail

Agents Involved

Agent 2: Classifier (path routing)
Agent 6: Decision Engine (policy-as-code)
Agent 8: XAI Report (regulatory docs)

Resolution Paths

Loan Deferral (MoF API integration)
Stop Collection directive
Instalment restructuring recommendation
Senior reviewer escalation

Phase Status

Phase 2 target. Stop Collection and Loan Deferral pathways require MoF API integration — scoped and estimated for Phase 2.

Document Integrity — Exception Detail

Agents Involved

Agent 3: Evidence Extraction (OCR confidence)
Agent 7: Anomaly Detection (forgery scoring)
Agent 5: Clarification Engine (re-upload request)

Forgery Detection

Metadata inconsistency analysis
Visual pattern anomaly detection
Cross-document field consistency check
Issuer verification where API available

Resolution Paths

Expired: Re-upload requested via Clarification Engine
Missing: Document checklist sent to applicant
Forgery: Case freeze + senior reviewer + legal flag

Fraud & Anomaly — Exception Detail

Agent Responsible

Agent 7: Anomaly Detection (always parallel)
Runs simultaneously with every pipeline stage
Cannot be disabled or bypassed

Detection Triggers

Fraud risk score >0.75 → Immediate escalation
Document forgery detected → Case freeze
Identity spoofing flag → Legal referral workflow
Pattern anomaly → Portfolio-level review

Escalation Path

All escalations routed to senior ADHA reviewer
Full XAI context provided at escalation
Evidence chain preserved for legal use
Case cannot proceed without explicit senior approval

Architecture Note: Fraud detection is the one pathway that can stop any case at any moment. This is by design — architecturally enforced, not a configurable parameter.

End-to-End Resolution Flow

Step 6 multi-agent pipeline: from rules engine flag to ADHA reviewer decision. Anomaly Detection runs in parallel throughout.

Primary Resolution Pipeline

① Intake & Pre-Classification

Rules engine flag received by Master Orchestrator

Agent 0: OrchestratorAgent 1: Pre-Classification

▼

② Classification

Discrepancy type, confidence score, required evidence set

Agent 2: Classifier

▼

③ Evidence Extraction

OCR + NLP on all documents → structured JSON (Arabic + English)

Agent 3: Evidence Extraction

▼

④ Truth Verification

Triangulate: Gov API + Declaration + Documents → conflict map

Agent 4: Truth Verification

▼

⑤ Clarification (if needed)

Bilingual citizen request · Max 3 rounds · State machine managed

Agent 5: Clarification Engine

▼

⑥ Decision

Policy-as-code → Approve / Defer / Stop Collection / Escalate / Resubmit

Agent 6: Decision Engine

▼

⑦ XAI Report Generation

Full audit report (PDF + JSON) · Policy citations · Evidence chain

Agent 8: XAI Report Generator

▼

⑧ ADHA Reviewer Decision

Human reviewer: Approve or Override · Full governance retained

Human in the Loop

Parallel & Exception Paths

⚡ Agent 7 — Anomaly Detection (Always Parallel)

Runs simultaneously with every pipeline stage.

Monitors: Fraud patterns · Document manipulation · Identity spoofing

Fraud risk >0.75 → Immediate freeze + senior reviewer escalation

🔄 Clarification State Machine (Agent 5)

Round 1 → Clarification request sent (bilingual)
Round 2 → Specific document request
Round 3 → Final escalation to human reviewer

PoC: EXCEPTION_RETIREMENT resolved in 2 rounds

🚨 Escalation Triggers

→ Fraud risk score >0.75 (Agent 7)
→ Clarification Round 3 reached (Agent 5)
→ Decision confidence <60% (Agent 6)
→ Policy rule not found in rule engine
→ Document forgery detected (Agent 7)

All escalations routed to senior ADHA reviewer with full XAI context

✅ Decision Outcomes (Agent 6)

APPROVE DEFER STOP COLLECTION ESCALATE RESUBMIT

✓ PoC Result — March 2026

<2 min

Resolution Time

vs 30–60 min manual

92%

Confidence Score

EXCEPTION_RETIREMENT

PoC Results, Impact Metrics & Production Targets

Validated March 2026. All figures based on PoC data and ADHA workflow analysis.

<2 min

Resolution Time

vs 30–60 min manual

92%

PoC Confidence Score

EXCEPTION_RETIREMENT

~30%

Step 6 Caseload Reduction

Via Step 2 reconciliation

20+

Discrepancy Types

Across 6 categories

AI Agents

1 Orchestrator + 8 Specialists

Max Clarification Rounds

PoC resolved in 2 rounds

Languages

Arabic + English (bilingual)

100%

Human Governance Retained

Reviewer approves every decision

Phase 1 — PoC Validated (March 2026)

Step 6 full multi-agent resolution pipeline
EXCEPTION_RETIREMENT at 92% confidence
Sub-2-minute end-to-end resolution
XAI report generation
Bilingual clarification engine
Anomaly Detection (parallel)

Phase 2 — Production Expansion Targets

Steps 1–5 AI enhancement deployment
All 20+ discrepancy types activated
Stop Collection integration
Loan Deferral integration (MoF API)
Identity & Eligibility full coverage
Income Verification (WPS + pension APIs)

Governance Principle: The AI system recommends — ADHA reviewers decide. Every recommendation is accompanied by a full XAI report with confidence scores, conflict resolution logic, evidence chain, and policy citations. Human oversight is not optional; it is architecturally enforced at every decision point.