FBM Finance Hyperautomation

01

The Problem

State before automation — baseline from which we measure impact

1

Shared inbox for all Malta entities

7

Distinct legal entities receiving invoices

0%

Automation prior to this project

~1,400

Invoices processed manually per year

All seven Malta entities receive invoices to one shared mailbox — accounts.payable@fbm.mt. Every invoice was manually read, classified, and filed — one person, every working day, year-round. A single point of failure with no audit trail, no SLA, no scalability, and full exposure to misrouting risk.

02

The Solution

AI-driven, end-to-end invoice ingestion and routing pipeline

📧 M365 Outlook Email trigger

→

📄 PDF Extraction Base64 + metadata

→

🤖 Claude API Entity classification

→

⚖ Confidence Gate ≥90% auto-route

→

🗄️ Dropbox (DRB) Archived by entity

→

📊 Business Central Purchase draft (DBC)

→

📋 Notion Audit Full audit trail

Claude API reads each PDF invoice, identifies the correct recipient Malta entity using bill-to fields, VAT numbers, and supplier mappings, then automatically archives it to the entity's Dropbox folder and creates a draft Purchase Invoice in Microsoft Dynamics Business Central. Every event — success and failure — is logged to the Notion Audit Trail.

03

Pipeline Architecture

WF_AP_001_InvoiceIngestion — complete node map and decision tree

OUTLOOK TRIGGER accounts.payable@fbm.mt — polls every 5 min
  └─► FILTER_AttachmentIsPDF Pass only emails with PDF attachment
          └─► TRANSFORM_ExtractBase64 PDF + email metadata
                  └─► VALIDATE_TextExtractable Is text extractable?
                          ├─► NO  → QUEUE_ManualReview (ERR_IMAGE_ONLY_PDF)
                          │         └─► LOG_AuditNotion
                          └─► YES → TRANSFORM_BuildPrompt Inject entity list + VAT flags
                                    └─► AI_ClaudeClassifyEntity POST → Claude API → JSON
                                              └─► TRANSFORM_ValidateAPIResponse
                                                        └─► FILTER_ConfidenceThreshold
                                                                  ├─► ≥ 90% → TRANSFORM_NormalizeFields
                                                                  │           ├─► DRB_UploadInvoice /AP/TEST/[Entity]/YYYY/MM/
                                                                  │           ├─► HTTP_DBC_CreatePurchaseDraft
                                                                  │           └─► NOTIFY_APExecutive_AutoRoute
                                                                  └─► < 90% → QUEUE_ManualReview
                                                                              └─► LOG_AuditNotion ←── ALL branches

NODE 01

TRIGGER_M365_APMailbox

Email Trigger · 5 min poll

NODE 02

FILTER_AttachmentIsPDF

IF node · PDF only

NODE 03

TRANSFORM_ExtractBase64

Code node · PDF + metadata

NODE 3.5

VALIDATE_TextExtractable

IF node · Blocks image-only PDFs

NODE 04

TRANSFORM_BuildPrompt

Code node · Entity list injection

NODE 05

AI_ClaudeClassifyEntity

HTTP Request · Claude API

NODE 06

TRANSFORM_ValidateAPIResponse

Code node · Hallucination guard

NODE 07

FILTER_ConfidenceThreshold

IF node · ≥0.90 auto / <0.90 manual

NODE 08a

TRANSFORM_NormalizeFields

Code node · Build DRB path

NODE 08b

DRB_UploadInvoice

HTTP Request · Dropbox Business

NODE 08c

HTTP_DBC_CreatePurchaseDraft

HTTP Request · BC OData v4

NODE 10

LOG_AuditNotion

HTTP Request · ALL branches

04

The 7 Malta Entities

All invoices classified and routed to the correct BUHAY Group legal entity

BH

BUHAY Holding Limited

BHL · Holding company

FB

FBM Limited

FBM · Core operations

NT

NTT Limited

NTT · Technology

D2

D2R Limited

D2R · Operations

FD

FBM Digital Systems

FDS · Digital

DK

DRAKO Limited

DRK · Gaming

LS

LuckySix Limited

LSX · iGaming

Each entity has its own Dropbox folder and Business Central company. The AI classifier must correctly identify the recipient entity before any routing can occur. Classification accuracy is the #1 KPI for this module.

05

Classification Logic & Confidence Routing

5 priority signals · 90% threshold · Auto-route or Human Review

Classification Signals — Priority Order

P1 Bill To field Exact entity name match

P2 VAT number Match against Entity Mapping DB

P3 Supplier name Known entity → supplier mapping

P4 Document header Letterhead / company text

P5 Email subject / sender domain Fallback only

≥ 90%

AUTO-ROUTE

Normalize fields & build DRB path

Upload to Dropbox /AP/TEST/[Entity]/

Create Purchase Invoice Draft in DBC

Notify AP Executive via Teams/email

Log to Notion Audit Trail

< 90%

HUMAN REVIEW QUEUE

Route to manual review queue

Notify AP Executive with context

Log to Notion Audit Trail (DLQ)

AP Executive classifies manually

Feedback loop for prompt tuning

06

Error Handling

Every failure mode defined — Saga/DLQ pattern, Notion as official audit log

ERR_NO_ATTACHMENT

Stage: Ingestion

Skip + log · No user notification

No notification required

ERR_EXTRACTION_FAILED

Stage: Pre-processing

Retry ×1, then manual queue

Notify: AP Executive

ERR_IMAGE_ONLY_PDF

Stage: Pre-processing

Block AI call → Manual Review + Audit Log

Notify: AP Executive

ERR_API_TIMEOUT

Stage: AI Call

Retry ×3 with exponential backoff (5s→15s→30s)

Notify: Data Engineer

ERR_LOW_CONFIDENCE

Stage: Post-AI Classification

Manual review queue

Notify: AP Executive

ERR_HALLUCINATED_ENTITY

Stage: Post-AI Validation

Manual review queue + critical alert

Notify: Data Engineer

ERR_DRB_UPLOAD_FAILED

Stage: Storage (Dropbox)

Retry ×1, then alert

Notify: Data Engineer

ERR_DBC_FAILED

Stage: ERP (Business Central)

Log + manual entry flag

Notify: AP Executive

⚠️ LOG_AuditNotion must be connected to every terminal branch — including all error paths. The Notion Audit Trail is the official Dead Letter Queue (DLQ) for this pipeline.

07

Storage Structure & Environments

Dropbox Business folder hierarchy · TEST vs PROD isolation

Dropbox Folder Structure (DRB)

/AP/
  /TEST/  ← Active · UAT
    /BUHAY-Holding-Limited/
    /FBM-Limited/
    /NTT-Limited/
    /D2R-Limited/
    /FBM-Digital-Systems/
    /DRAKO-Limited/
    /LuckySix-Limited/
  /PROD/  ← Post-UAT
    [same structure]

File naming convention:
INVREF – INVOICE – SUPPLIER.pdf
e.g: INV-2026-0042 – INVOICE –
     Supplier Name.pdf

Environment Strategy

TEST — ACTIVE

Pipeline validation + prompt tuning

Real anonymised invoices · DBC Test Company

PROD — POST-UAT

Live operations · Real invoices

⚠ Never run PROD until UAT complete

Go-live strategy: Shadow Mode first (Strangler Fig — Phase 1). Auto-route activation only after ≥90% accuracy confirmed over test corpus.

Finance
Hyperautomation

What we're building
and why it matters

The Problem

The Solution

The Scale

AP Process
Automation

The Problem

The Solution

Pipeline Architecture

The 7 Malta Entities

Classification Logic & Confidence Routing

Error Handling

Storage Structure & Environments

Automation Modules

Technology Stack

n8n

Claude API

Dynamics BC

Dropbox Business

Power BI

Microsoft 365

Cloudflare

Orion

ACE

Project Documentation

Architecture Design

Workflow Specs

UAT / Test Reports

Academic Deliverables

Notion Tracker

AP Prompt Library

FinanceHyper­automation

What we're buildingand why it matters

The Problem

The Solution

The Scale

AP ProcessAutomation

The Problem

The Solution

Pipeline Architecture

The 7 Malta Entities

Classification Logic & Confidence Routing

Error Handling

Storage Structure & Environments

Automation Modules

Technology Stack

n8n

Claude API

Dynamics BC

Dropbox Business

Power BI

Microsoft 365

Cloudflare

Orion

ACE

Project Documentation

Architecture Design

Workflow Specs

UAT / Test Reports

Academic Deliverables

Notion Tracker

AP Prompt Library

Finance
Hyperautomation

What we're building
and why it matters

AP Process
Automation