==================================================
 STORYGEN AI — STORY EXPORT
==================================================
Story ID:           20
Title:              Test_AF
Owner:              Ang Chwee Geok Karen - Digital Innovation Consultant <karen.ang@educlaas.com>
App Type:           Agentic AI App
Input Type:         file
Status:             features_generated
Source File:        PIL Capstone Project (DRAFT).pdf
Created:            2026-04-29 05:45:56 UTC
Updated:            2026-04-29 05:47:41 UTC
Features Generated: 2026-04-29 05:47:41 UTC
Total Clusters:     7
Total Features:     18

--- ORIGINAL INPUT ---
Friday, 27 March 2026
AI Citizens Projects
Group 
Financial
Reporting and Accounting
Group Tax
Finance Transformation

2
Definitions
CAC –Commercial & Agency Controlling (Agency Finance)
GFS –Oracle’s Global Financial System (Accounting System)
SOA –Statement of Accounts

3
AI Citizens Projects Overview
Power 
automate 
workflow for 
submission of 
Audited 
Financial 
Statements to 
HQ/CAC
Automated 
prepayments 
computations
Automated 
GFS AP 
payments 
review
AP payments 
dashboard
Automated 
lease schedule 
computations
Capital 
Accounting 
dashboard
Automated 
monthly SOAs 
to customers
Automated 
bank receipt 
classification 
(Trade vs 
non-trade)
Automated 
GFS data 
reconciliation 
for tax reporting

4
Problem Statement
-The current AFS submission 
process lacks a standardized 
workflow and centralized 
tracking, resulting in limited 
visibility, unclear accountability, 
and increased risk of late or 
missed submissions. 
-Manual coordination and 
follow-ups also weaken 
auditability and compliance 
oversight.
Financials
FTE savings: ~2.0 FTE 
annually
Benefits
-Improved visibility and control
-Reduced manual effort
-Stronger audit and 
compliance
Timeline
Target Go-Live: Q4 2026
Automated Workflow for 
Submission of AFS to GFRA/CAC
Problem Statement
-Prepayments calculations are 
currently performed manually
-High risk of formula errors, 
incorrect amortisationperiods, 
and inconsistent assumptions
-Manual process creates 
key-person dependency and 
weak audit trail
Financials
FTE savings: ~0.2 FTE 
annually
Benefits
-Automates monthly 
prepayment amortisation
based on defined rules and 
schedules
-Ensures timely, complete, 
and consistent month-end 
expense recognition
-Reduces risk of over-or 
under-recognition of expenses
-Improves auditability with 
standardisedlogic and clear 
computation trail
Timeline
Target Go-Live: Q4 2026
Automated Prepayments 
Computations

5
Problem Statement
-The AP payments review 
process is heavily manual and 
sample-based, requiring the 
team to manually open each 
invoice to verify bank details. 
-This approach is 
time-consuming, limits 
coverage to sampled 
transactions only, and 
increases the risk of errors or 
missed issues in payment 
details prior to execution.
Financials
FTE savings: ~2.0 FTE 
annually
Benefits
-Reduced manual effort in 
payment review
-Broader coverage of 
payment checks
-Improved accuracy over 
vendor bank details
-Stronger audit trail and 
payment assurance
Timeline
Target Go-Live: Q4 2026
Automated GFS AP Payments Review
Problem Statement
-AP payment information is 
currently spread across multiple 
systems and reports, with 
limited consolidated visibility 
across entities.
-Finance relies on manual 
extraction and tracking to 
determine payments made to 
date and outstanding payables, 
increasing the risk of oversight, 
delays, and inconsistent 
reporting across entities.
Financials
FTE savings: ~0.5 FTE 
annually
Benefits
-Improved visibility of 
payments made and 
outstanding across entities
-Reduced manual tracking 
and follow-ups
-Faster identification of 
overdue and upcoming 
payments
-Better payment planning 
and control
Timeline
Target Go-Live: Q4 2026
AP Payments Dashboard

6
Problem Statement
-Lease schedule computations 
are highly manual and 
spreadsheet-driven, requiring 
significant effort to extract 
contract details, compute lease 
liabilities, right-of-use assets, 
and maintain schedules over 
time.
-This is time-consuming, 
error-prone, impacting the 
accuracy and completeness of 
lease accounting.
Financials
FTE savings: ~0.2 FTE 
annually
Benefits
-Reduced manual effort
-Improved accuracy and 
completeness
-Faster month-end close 
and reporting for lease 
accounting
-Stronger audit trail and 
compliance assurance
Timeline
Target Go-Live: Q4 2026
Automated Lease Schedule 
Computations
Problem Statement
-Capital accounting data is 
highly fragmented across 
sources.
-Significant manual effort is 
required to extract, validate, 
and consolidate data to obtain 
a complete view of capital 
assets. 
-This results in time-consuming 
reporting cycles, higher risk of 
inconsistencies, and limited 
timely visibility for review and 
close activities.
Financials
FTE savings: ~0.3 FTE 
annually
Benefits
-Improved visibility of 
capital assets, CIP, and 
related balances
-Reduced manual effort
-Faster month-end close 
and capital reporting
-Improved data 
consistency and reliability
-Stronger governance and 
audit readiness
Timeline
Target Go-Live: Q4 2026
Capital Accounting 
Dashboard

7
Problem Statement
-Monthly SOAs are manually 
downloaded for each customer 
and emailed individually 
-High manual effort during 
month-end closing 
-Risk of incorrect customer 
selection and missed / delayed 
SOA delivery
Financials
-Saves ~3 man-hours per 
month(2 staffs ×1.5 hours)
-Reduces manual processing 
risk and rework
Benefits
-Eliminates manual SOA 
downloading and emailing
-Ensures consistent and timely 
month-end SOA delivery
-Reduces risk of customer 
selection and emailing errors
Timeline
Target Go-Live: Q4 2026
Automated Monthly SOAs 
to Customers
Problem Statement
-Bank statements are 
downloaded regularly and 
reviewed manually 
-Credit transactions require 
manual filtering and AR vs 
Non-AR determination
-High effort and risk of 
misclassification or delays
Financials
-Current manual effort: 1.5 
hours per day
-Equivalent to ~33 hours per 
month
-Automation removes most 
recurring manual effort
Benefits
-Reduces manual filtering and 
rework 
--Improves timeliness and 
consistency of cash application
Timeline
To be implemented in 
collaboration with GL team
Target Go-Live: Q4 2026
Automated bank receipt 
classification (Trade vs 
non-trade)

8
Problem Statement
-Manual adjustments and 
classifications of GFS output for 
tax reporting 
-High risk of delayed 
identification of errors, 
misclassifications, and 
operational inefficiencies
-Increased scrutiny on data 
reconciliation in the coming 
yearsby IRAS
Financials
-93% of the Group’s total 
profits are reported in 
Singapore, which must be 
accurate to avail of the MPA tax 
exemption 
-Man-hours savings: up to 
1,000 man-hours annually 
(~90% of current time spent on 
annual Singapore tax reporting 
/ compliance) 
Benefits
-Lower risk of 
misclassifications and wrong 
adjustments
-Improved standardization for 
tax reporting for all Singapore 
entities
-Stronger audit trail and 
control
-Reducedoperational and 
compliance risk
Timeline
Target Go-Live: Q4 2026
Automated GFS data reconciliation for tax reporting

--- USER STORY ---
# Automated GFS Data Reconciliation for Tax Reporting

## Overview
This is an **Agentic AI App** that autonomously reconciles output from Oracle's Global Financial System (GFS) for Singapore tax reporting purposes. The solution replaces manual adjustments, classifications, and reconciliations performed by the Group Tax team with intelligent agents that extract GFS data, apply tax classification rules, identify discrepancies, and prepare audit-ready reconciliation packs — escalating only ambiguous cases to human reviewers. Given that 93% of the Group's profits are reported in Singapore (subject to MPA tax exemption requirements) and IRAS scrutiny is increasing, accuracy and traceability are mission-critical.

## Actors
- **Group Tax Analyst (Human)**: Reviews agent-prepared reconciliations, resolves flagged exceptions, and approves final tax reporting packs.
- **Group Tax Manager (Human)**: Provides final sign-off on reconciled tax data prior to submission and approves rule changes.
- **Finance Transformation Lead (Human)**: Maintains classification rule library and monitors agent performance.
- **Data Extraction Agent (AI)**: Autonomously pulls GFS trial balances, journal entries, and supporting transactional data on a defined schedule.
- **Classification & Adjustment Agent (AI)**: Applies tax classification logic (e.g., taxable vs. non-taxable, deductible vs. non-deductible, MPA-qualifying vs. non-qualifying) to GFS line items.
- **Reconciliation Agent (AI)**: Compares classified GFS output against expected tax-basis figures, identifies variances, and produces reconciliation summaries.
- **Exception & Audit Agent (AI)**: Flags anomalies, prepares audit trail documentation, and routes ambiguous items to human reviewers.

## Goals
- Reduce annual tax reporting effort by up to **1,000 man-hours (~90%)** for Singapore tax compliance.
- Ensure accuracy of Singapore tax reporting to safeguard **MPA tax exemption eligibility** (covering 93% of Group profits).
- Standardize tax reconciliation across all Singapore entities.
- Strengthen audit trail and reduce compliance risk in anticipation of heightened **IRAS scrutiny**.
- Detect classification errors and adjustments earlier in the reporting cycle.

## User Story
As a **Group Tax Analyst**, I want **AI agents to autonomously extract, classify, and reconcile GFS data for tax reporting**, so that I can **focus on reviewing exceptions and high-judgment items, ensure the Group meets IRAS requirements accurately and on time, and preserve our MPA tax exemption with a robust audit trail**.

## Detailed Workflow
1. **Scheduled trigger**: The Data Extraction Agent runs on a defined cadence (monthly close + annual tax reporting cycles), pulling GFS trial balances, journals, and supporting ledgers for all in-scope Singapore entities.
2. **Data validation**: The agent verifies completeness (all entities, all expected accounts, period boundaries) and checks for known data-quality issues. If critical data is missing or stale, it halts and notifies the Finance Transformation Lead.
3. **Classification & adjustment**: The Classification & Adjustment Agent applies the tax rule library to each GL line item, determining tax treatment (taxable income, non-taxable, deductible expense, disallowable expense, MPA-qualifying, capital vs. revenue, etc.). It uses confidence scoring on each classification.
4. **Decision logic**:
   - If classification confidence is **high (≥ defined threshold)** and matches historical patterns → auto-classify and log.
   - If confidence is **low** or the item is **novel/unmatched** → mark as exception and route to the Exception & Audit Agent.
   - If the value exceeds a **materiality threshold** → always require human confirmation regardless of confidence.
5. **Reconciliation**: The Reconciliation Agent compares classified GFS output to expected tax-basis figures, prior-period trends, and inter-entity consistency checks. Variances exceeding tolerance are flagged.
6. **Exception handling & hand-off**: The Exception & Audit Agent compiles flagged items into a reviewer dashboard with supporting evidence (source transaction, applied rule, rationale, confidence score, suggested classification) and notifies the Group Tax Analyst.
7. **Human review**: The Group Tax Analyst reviews exceptions, accepts/overrides agent recommendations, and provides feedback that feeds back into the rule library.
8. **Audit pack generation**: Once all exceptions are cleared, the Exception & Audit Agent produces a finalized reconciliation pack including rule version used, agent decisions, human overrides, timestamps, and full traceability from GFS source to tax-basis output.
9. **Approval & submission**: The Group Tax Manager reviews and approves the final pack. Approved data is locked and archived for IRAS audit readiness.
10. **Continuous learning**: Agent decisions, human overrides, and outcomes are logged; the Finance Transformation Lead reviews periodic performance reports to refine rules and thresholds.

## Acceptance Criteria
- Data Extraction Agent successfully retrieves complete GFS data for all in-scope Singapore entities within the scheduled window, with zero missing periods.
- Classification & Adjustment Agent achieves ≥ 95% accuracy on auto-classified items (validated against human review sample) before go-live.
- All classifications include a **confidence score**, **rule reference**, and **rationale** in the audit log.
- **Guardrail – materiality**: Any item above the agreed materiality threshold is mandatorily routed for human approval, regardless of agent confidence.
- **Guardrail – novelty**: Any GL account, vendor pattern, or transaction type not previously seen is automatically flagged as an exception and never auto-classified.
- **Stop condition**: If data extraction fails, source data appears corrupted, or exception volume exceeds a defined threshold (indicating systemic issue), the agent halts the workflow and alerts the Finance Transformation Lead.
- **Escalation rule**: Any reconciliation variance exceeding tolerance, or any item potentially affecting **MPA exemption eligibility**, is escalated immediately to the Group Tax Manager.
- Final reconciliation pack is fully traceable from tax-basis output back to original GFS transaction (end-to-end audit trail).
- Group Tax Analyst can override any agent decision; overrides are captured and feed back into rule learning.
- System reduces annual man-hours on Singapore tax reporting by at least **80%** (target 90% / ~1,000 hours).
- Solution is operational and live by **Q4 2026**.
- All agent actions, model versions, and rule library versions are logged for IRAS audit purposes for at least the statutory retention period.

## Assumptions & Constraints
- GFS provides reliable API or extract access for the agents to retrieve data.
- A documented and maintained **tax classification rule library** for Singapore entities exists or will be developed prior to go-live.
- Materiality and confidence thresholds will be defined and approved by Group Tax leadership.
- Solution scope is initially limited to **Singapore entities** for tax reporting; expansion to other jurisdictions is out of scope.
- Final tax submissions to IRAS remain a human-approved activity; agents do not file returns autonomously.
- IRAS regulatory requirements and MPA exemption rules in effect through 2026 form the basis of classification logic; the rule library must be updated when regulations change.
- Solution must comply with internal data governance, security, and audit standards.
- Coordination required with the Finance Transformation team and GFS system owners for integration.

--- FEATURE LIST SUMMARY ---
This solution enables autonomous reconciliation of Oracle GFS output for Singapore tax reporting, replacing manual Group Tax effort with coordinated AI agents. Primary actors are the Group Tax Analyst, Group Tax Manager, and Finance Transformation Lead, supported by Data Extraction, Classification & Adjustment, Reconciliation, and Exception & Audit agents. The end-to-end flow moves from scheduled GFS extraction through tax classification, variance reconciliation, exception review, audit pack generation, and managerial approval, with continuous learning feeding the rule library. Master Data Configuration centralizes entities, users, tax rules, thresholds, and lookup catalogs. Across 18 rows the platform delivers automated tax classification, traceable audit packs, and IRAS-ready compliance safeguards.

Feature Clusters & Features:
• Master Data Configuration
  - 1. User and Role Management — Maintains tax analysts, managers, and transformation leads with role-based access to reconciliation workflows and approvals.
  - 2. Singapore Entity Registry — Catalog of in-scope Singapore legal entities, GFS identifiers, and MPA-eligibility flags used throughout extraction and reporting.
  - 3. Tax Classification Rule Library — Versioned repository of Singapore tax rules covering taxable, deductible, MPA-qualifying, and capital vs revenue treatments.
  - 4. Materiality and Confidence Thresholds — Configurable thresholds that govern auto-classification, human review triggers, and escalation behavior.
  - 5. Reference Lookups and Chart of Accounts — Shared catalogs of GL accounts, tax categories, currencies, and period calendars referenced by every agent.
• Scheduled Data Extraction
  - 6. GFS Extraction Scheduler — Defines monthly close and annual tax cycle cadences that trigger autonomous GFS data pulls.
  - 7. Trial Balance and Journal Ingestion — Pulls trial balances, journals, and supporting ledgers from GFS via API into the reconciliation workspace.
  - 8. Data Completeness Validation — Verifies all entities, accounts, and periods are present and halts with alerts if critical data is missing or stale.
• Tax Classification and Adjustment
  - 9. Automated Tax Classification — Applies the rule library to each GL line item to assign tax treatment with confidence scores and rationale.
  - 10. Novelty and Materiality Guardrails — Forces human review for unseen accounts/vendors and any item exceeding the materiality threshold.
• Reconciliation and Variance Analysis
  - 11. Tax-Basis Reconciliation — Compares classified GFS output to expected tax-basis figures and prior-period trends to surface variances.
  - 12. Inter-Entity Consistency Checks — Validates consistency of treatments and balances across Singapore entities and flags tolerance breaches.
• Exception Handling and Human Review
  - 13. Exception Reviewer Dashboard — Presents flagged items with evidence, applied rules, suggested classifications, and confidence to the Group Tax Analyst.
  - 14. Override and Feedback Capture — Allows analysts to accept or override agent decisions and feeds outcomes back into the rule learning loop.
  - 15. MPA Eligibility Escalation — Immediately escalates any variance or item potentially affecting MPA exemption eligibility to the Group Tax Manager.
• Audit Pack and Approval
  - 16. Reconciliation Audit Pack Generation — Produces finalized, traceable packs with rule versions, agent decisions, overrides, and end-to-end source links.
  - 17. Manager Approval and Lock — Captures Group Tax Manager sign-off, locks approved data, and archives for IRAS audit readiness.
• Agent Orchestration and Continuous Learning
  - 18. Agent Orchestration and Performance Monitoring — Coordinates agent runs, logs model and rule versions, and reports accuracy and effort-savings metrics for tuning.

--- FEATURE LIST (18 features across 7 clusters) ---

#1 | Cluster: Master Data Configuration | Feature: User and Role Management
  Description: Central authority for user accounts, roles, and permissions across the reconciliation platform. Ensures segregation of duties between preparers, reviewers, and approvers.
  Workflow:
    1. Admin creates user accounts for tax analysts, managers, and transformation leads.
    2. Roles and permissions are assigned per function.
    3. Access scopes (entities, modules) are defined.
    4. Periodic access reviews are conducted.
    5. Audit logs capture all access changes.
  Table:       users_roles
  Columns:     id (bigint, pk), user_name (varchar 255), email (varchar 255), role_code (varchar 50), entity_scope (json), is_active (boolean), created_at (timestamp)
  Actor:       System Administrator
  AI Agent:    None
  ----
#2 | Cluster: Master Data Configuration | Feature: Singapore Entity Registry
  Description: Catalog of in-scope Singapore entities with GFS mappings and MPA-eligibility flags. Used by all agents to determine extraction and tax-treatment scope.
  Workflow:
    1. Finance Transformation Lead registers each Singapore legal entity.
    2. GFS entity identifiers and ledger codes are mapped.
    3. MPA-eligibility flags and exemption parameters are recorded.
    4. Effective dates and statuses are maintained.
    5. Updates are versioned for audit.
  Table:       sg_entities
  Columns:     id (bigint, pk), entity_code (varchar 50), entity_name (varchar 255), gfs_entity_id (varchar 50), mpa_eligible (boolean), effective_from (date), status (varchar 30)
  Actor:       Finance Transformation Lead
  AI Agent:    None
  ----
#3 | Cluster: Master Data Configuration | Feature: Tax Classification Rule Library
  Description: Versioned repository of tax classification rules covering taxable, deductible, MPA-qualifying, and capital vs revenue treatments. Drives the Classification & Adjustment Agent's decisions.
  Workflow:
    1. Lead authors classification rules with conditions and tax treatments.
    2. Rules are versioned and tied to effective periods.
    3. Group Tax Manager approves rule changes.
    4. Rules are published for agent consumption.
    5. Historical versions are retained for audit.
  Table:       tax_rules
  Columns:     id (bigint, pk), rule_code (varchar 50), description (text), conditions (json), tax_treatment (varchar 50), version (varchar 20), effective_from (date), approved_by (bigint, fk)
  Actor:       Finance Transformation Lead
  AI Agent:    None
  ----
#4 | Cluster: Master Data Configuration | Feature: Materiality and Confidence Thresholds
  Description: Configurable thresholds governing auto-classification, mandatory human review, and stop conditions. Centrally tuned to balance automation and risk.
  Workflow:
    1. Group Tax leadership defines materiality amounts per category.
    2. Confidence score thresholds are set for auto-classification.
    3. Exception-volume stop thresholds are configured.
    4. Thresholds are versioned with approval records.
    5. Agents read thresholds at runtime.
  Table:       thresholds_config
  Columns:     id (bigint, pk), threshold_type (varchar 50), category (varchar 50), value (decimal 18,2), version (varchar 20), approved_by (bigint, fk), effective_from (date)
  Actor:       Group Tax Manager
  AI Agent:    None
  ----
#5 | Cluster: Master Data Configuration | Feature: Reference Lookups and Chart of Accounts
  Description: Shared catalogs of GL accounts, tax categories, currencies, and period calendars used by every agent. Provides consistent reference data across the workflow.
  Workflow:
    1. GFS chart of accounts is imported and mapped.
    2. Tax categories, currencies, and period calendars are maintained.
    3. Vendor and counterparty reference data is synced.
    4. Mappings are validated for completeness.
    5. Lookups are exposed to agents via API.
  Table:       reference_lookups
  Columns:     id (bigint, pk), lookup_type (varchar 50), code (varchar 50), description (varchar 255), parent_code (varchar 50), is_active (boolean), updated_at (timestamp)
  Actor:       Finance Transformation Lead
  AI Agent:    None
  ----
#6 | Cluster: Scheduled Data Extraction | Feature: GFS Extraction Scheduler
  Description: Defines and triggers cadences that automate GFS data pulls aligned with close and tax cycles. Drives the autonomous start of each reconciliation run.
  Workflow:
    1. Lead defines extraction cadences (monthly close, annual tax cycle).
    2. Schedules are linked to entity scope and period definitions.
    3. Scheduler triggers Data Extraction Agent runs.
    4. Run status and next-execution times are tracked.
    5. Failures trigger retry and alerts.
  Table:       extraction_schedules
  Columns:     id (bigint, pk), schedule_name (varchar 100), cadence (varchar 50), entity_scope (json), next_run_at (timestamp), last_run_status (varchar 30), is_active (boolean)
  Actor:       Finance Transformation Lead
  AI Agent:    Data Extraction Agent
  ----
#7 | Cluster: Scheduled Data Extraction | Feature: Trial Balance and Journal Ingestion
  Description: Pulls trial balances, journals, and supporting ledgers from Oracle GFS into the platform. Stores raw extracts with control totals for downstream agents.
  Workflow:
    1. Agent calls GFS APIs for in-scope entities.
    2. Trial balances, journals, and supporting ledgers are retrieved.
    3. Data is staged into the reconciliation workspace.
    4. Row counts and control totals are recorded.
    5. Run metadata is logged for traceability.
  Table:       gfs_extracts
  Columns:     id (bigint, pk), run_id (bigint, fk), entity_id (bigint, fk), period (varchar 10), data_type (varchar 30), record_count (int), control_total (decimal 18,2), extracted_at (timestamp)
  Actor:       Data Extraction Agent
  AI Agent:    Data Extraction Agent
  ----
#8 | Cluster: Scheduled Data Extraction | Feature: Data Completeness Validation
  Description: Verifies completeness, period alignment, and quality of GFS extracts before classification begins. Halts the workflow and notifies stakeholders if critical issues are found.
  Workflow:
    1. Agent compares extracted entities and accounts against expected scope.
    2. Period boundaries and balance integrity checks are run.
    3. Known data-quality rules are applied.
    4. If critical gaps exist, workflow halts.
    5. Alert is sent to Finance Transformation Lead.
  Table:       extract_validations
  Columns:     id (bigint, pk), run_id (bigint, fk), check_type (varchar 50), status (varchar 30), issue_details (text), halted (boolean), validated_at (timestamp)
  Actor:       Data Extraction Agent
  AI Agent:    Data Extraction Agent
  ----
#9 | Cluster: Tax Classification and Adjustment | Feature: Automated Tax Classification
  Description: Applies tax classification logic to every GFS line item with confidence scoring and rule rationale. Auto-classifies high-confidence items and logs full audit detail.
  Workflow:
    1. Agent loads active rule library version.
    2. Each GL line item is evaluated against rules.
    3. Tax treatment, confidence score, and rationale are assigned.
    4. High-confidence items are auto-classified.
    5. Decisions are logged with rule references.
  Table:       classifications
  Columns:     id (bigint, pk), run_id (bigint, fk), gl_line_id (bigint), tax_treatment (varchar 50), confidence_score (decimal 5,4), rule_id (bigint, fk), rationale (text), status (varchar 30)
  Actor:       Classification & Adjustment Agent
  AI Agent:    Classification & Adjustment Agent
  ----
#10 | Cluster: Tax Classification and Adjustment | Feature: Novelty and Materiality Guardrails
  Description: Enforces mandatory human review for novel patterns and material amounts to protect MPA eligibility. Prevents auto-classification on items requiring judgment.
  Workflow:
    1. Agent checks each item against historical patterns.
    2. Unseen accounts, vendors, or transaction types are flagged as novel.
    3. Items above materiality threshold are flagged regardless of confidence.
    4. Flagged items are routed to Exception & Audit Agent.
    5. Guardrail decisions are logged for audit.
  Table:       guardrail_flags
  Columns:     id (bigint, pk), classification_id (bigint, fk), flag_type (varchar 30), reason (text), threshold_value (decimal 18,2), routed_to (varchar 50), created_at (timestamp)
  Actor:       Classification & Adjustment Agent
  AI Agent:    Classification & Adjustment Agent
  ----
#11 | Cluster: Reconciliation and Variance Analysis | Feature: Tax-Basis Reconciliation
  Description: Compares classified GFS output to expected tax-basis figures and flags variances beyond tolerance. Generates summaries for analyst review.
  Workflow:
    1. Agent aggregates classified GFS output.
    2. Compares totals against expected tax-basis figures.
    3. Computes variances by category and entity.
    4. Highlights items beyond tolerance.
    5. Produces reconciliation summary.
  Table:       reconciliations
  Columns:     id (bigint, pk), run_id (bigint, fk), entity_id (bigint, fk), category (varchar 50), gfs_amount (decimal 18,2), tax_basis_amount (decimal 18,2), variance (decimal 18,2), within_tolerance (boolean)
  Actor:       Reconciliation Agent
  AI Agent:    Reconciliation Agent
  ----
#12 | Cluster: Reconciliation and Variance Analysis | Feature: Inter-Entity Consistency Checks
  Description: Validates classification and balance consistency across Singapore entities and prior periods. Surfaces anomalies that may indicate misclassification or data issues.
  Workflow:
    1. Agent compares classifications across Singapore entities.
    2. Identifies inconsistent treatments for similar items.
    3. Compares against prior-period trends.
    4. Flags anomalies for review.
    5. Logs findings to reconciliation pack.
  Table:       consistency_checks
  Columns:     id (bigint, pk), run_id (bigint, fk), check_type (varchar 50), entities_compared (json), finding (text), severity (varchar 20), created_at (timestamp)
  Actor:       Reconciliation Agent
  AI Agent:    Reconciliation Agent
  ----
#13 | Cluster: Exception Handling and Human Review | Feature: Exception Reviewer Dashboard
  Description: Presents Group Tax Analysts with prioritized exceptions and full supporting evidence. Enables efficient human review focused on judgment items.
  Workflow:
    1. Agent compiles flagged items into a review queue.
    2. Each item displays source transaction, applied rule, rationale, and confidence.
    3. Suggested classification is presented.
    4. Analyst is notified of pending items.
    5. Analyst opens dashboard to begin review.
  Table:       exception_queue
  Columns:     id (bigint, pk), run_id (bigint, fk), classification_id (bigint, fk), reason (varchar 100), suggested_treatment (varchar 50), confidence_score (decimal 5,4), status (varchar 30), assigned_to (bigint, fk)
  Actor:       Group Tax Analyst
  AI Agent:    Exception & Audit Agent
  ----
#14 | Cluster: Exception Handling and Human Review | Feature: Override and Feedback Capture
  Description: Captures analyst overrides with rationale and timestamps for full traceability. Feeds learning loop to improve rules and agent accuracy.
  Workflow:
    1. Analyst reviews exception with evidence.
    2. Accepts or overrides agent recommendation.
    3. Provides reason and corrected treatment.
    4. Override is logged with timestamp and user.
    5. Feedback feeds rule learning queue.
  Table:       decision_overrides
  Columns:     id (bigint, pk), exception_id (bigint, fk), original_treatment (varchar 50), new_treatment (varchar 50), reason (text), overridden_by (bigint, fk), overridden_at (timestamp)
  Actor:       Group Tax Analyst
  AI Agent:    Exception & Audit Agent
  ----
#15 | Cluster: Exception Handling and Human Review | Feature: MPA Eligibility Escalation
  Description: Immediately escalates variances or items potentially impacting MPA exemption to the Group Tax Manager. Protects 93% of Group profits subject to MPA rules.
  Workflow:
    1. Agent identifies items affecting MPA eligibility or exceeding tolerance.
    2. Escalation ticket is created for Group Tax Manager.
    3. Manager is notified immediately.
    4. Manager reviews and decides resolution path.
    5. Decision is logged on the run record.
  Table:       escalations
  Columns:     id (bigint, pk), run_id (bigint, fk), trigger_type (varchar 50), item_reference (varchar 100), severity (varchar 20), escalated_to (bigint, fk), status (varchar 30), resolved_at (timestamp)
  Actor:       Group Tax Manager
  AI Agent:    Exception & Audit Agent
  ----
#16 | Cluster: Audit Pack and Approval | Feature: Reconciliation Audit Pack Generation
  Description: Produces a finalized, end-to-end traceable audit pack from tax-basis output to GFS source. Includes rule versions, agent decisions, and human overrides.
  Workflow:
    1. Agent verifies all exceptions are cleared.
    2. Compiles classifications, overrides, rule versions, and timestamps.
    3. Links each line back to GFS source transactions.
    4. Generates finalized audit pack document.
    5. Stores pack with version metadata.
  Table:       audit_packs
  Columns:     id (bigint, pk), run_id (bigint, fk), pack_reference (varchar 100), rule_version (varchar 20), generated_at (timestamp), document_uri (varchar 500), status (varchar 30)
  Actor:       Exception & Audit Agent
  AI Agent:    Exception & Audit Agent
  ----
#17 | Cluster: Audit Pack and Approval | Feature: Manager Approval and Lock
  Description: Captures Group Tax Manager sign-off and locks approved reconciliation data. Archives packs to meet IRAS audit retention requirements.
  Workflow:
    1. Manager reviews finalized audit pack.
    2. Approves or requests revisions.
    3. On approval, run data is locked.
    4. Pack is archived for IRAS retention.
    5. Approval event is logged immutably.
  Table:       approvals
  Columns:     id (bigint, pk), audit_pack_id (bigint, fk), approver_id (bigint, fk), decision (varchar 30), comments (text), approved_at (timestamp), locked (boolean)
  Actor:       Group Tax Manager
  AI Agent:    None
  ----
#18 | Cluster: Agent Orchestration and Continuous Learning | Feature: Agent Orchestration and Performance Monitoring
  Description: Coordinates multi-agent runs and captures full operational telemetry for audit and tuning. Provides KPI dashboards on accuracy and effort savings to refine the system.
  Workflow:
    1. Orchestrator coordinates agent run sequencing and dependencies.
    2. Logs model versions, rule versions, and run telemetry.
    3. Tracks accuracy, override rates, and time-savings KPIs.
    4. Lead reviews periodic performance dashboards.
    5. Insights drive rule and threshold tuning.
  Table:       agent_runs
  Columns:     id (bigint, pk), agent_name (varchar 100), model_version (varchar 50), rule_version (varchar 20), started_at (timestamp), completed_at (timestamp), status (varchar 30), metrics (json)
  Actor:       Finance Transformation Lead
  AI Agent:    Data Extraction Agent
  ----

==================================================
 END OF STORY EXPORT
==================================================
