AI Document Intelligence, KRA eTIMS Receipt OCR & Automated Three-Way AP Matching: Eliminating Manual Supplier Invoice Data Entry and Overbillings

In mid-to-large retail chains, hospitality groups, hospitals, construction firms, and manufacturing enterprises processing 1,500 to 30,000+ supplier invoices every month, Accounts Payable (AP) clerks spend 70% of their week typing PDF and paper supplier invoices line-by-line into the ERP. Worse still, manual AP processing fails to catch three expensive leaks: suppliers billing KES 1,480 per unit when the approved Purchase Order (LPO) agreed KES 1,420; suppliers invoicing for 200 cases when the warehouse Goods Received Note (GRN) only signed for 185 cases; and non-compliant tax invoices that lacks a valid KRA eTIMS Control Unit Invoice Number (CUIN)—exposing the company to a 30% Corporate Income Tax disallowance on non-eTIMS expenses.

We engineer custom AI Accounts Payable (AP) Automation, Tax Compliance & Deterministic 3-Way Matching Engines in Python, Django, Celery, and PostgreSQL that ingest supplier invoices via email, WhatsApp, or vendor portals, extract every header and line item with 99.4%+ field accuracy, verify the eTIMS QR/Control Unit signature live before booking the liability, and reconcile Purchase Order (LPO) ↔ Warehouse GRN ↔ Supplier Invoice down to the individual SKU and cent.

1. Why Legacy Template OCR & Manual ERP Checks Fail Accounts Payable

Traditional coordinate-based OCR tools (which break whenever a supplier shifts a column by 5 millimeters) and basic ERP invoice entry screens fail modern finance teams in four ways:

• Supplier SKU Naming Mismatches: Your internal ERP calls an item SKU-40912: Cooking Oil 20L Jerrycan, while Supplier A's invoice prints 20LTR GOLDEN FRY VEG OIL and Supplier B prints G/FRY 20L J-CAN. Without semantic embedding + historical vendor-SKU mapping, automation stalls on every line item.
• Multi-GRN & Partial Delivery Complexity: A single Blanket LPO for 1,000 bags of cement is delivered across 4 separate truckloads (4 separate GRNs over 2 weeks) and billed across 3 partial supplier invoices. Matching M invoices against N GRNs and 1 LPO requires a cumulative line-balance state machine, not a simple 1-to-1 document check.
• Invalid or Cancelled KRA eTIMS Tax Invoices: Under strict tax authority electronic invoicing mandates, if a supplier issues a commercial invoice without a valid eTIMS Control Unit serial and QR payload—or issues an eTIMS invoice that fails to declare your company's Buyer PIN—your finance team cannot deduct that expense for corporate tax purposes.
• Duplicate Invoice Submissions (Email + Physical Copy): A supplier emails a PDF invoice on Tuesday (INV-2026-884) and their driver hands a stamped physical copy of the exact same invoice to the receiving clerk on Friday (INV-2026/884). Without fuzzy invoice-number normalization and SHA-256 fingerprint checks, AP teams accidentally pay the same invoice twice.

2. End-to-End AI AP Ingestion, eTIMS Verification & 3-Way Match Architecture

Our pipeline combines multimodal Vision-LLM structured extraction with strict, deterministic PostgreSQL accounting tolerance gates—ensuring the AI never invents numbers and never posts an unbalanced journal entry:

A. Deterministic Three-Way Matching Engine (PO vs. Cumulative GRNs vs. Invoice)

While Vision LLMs excel at reading messy multi-page scans and mapping supplier item descriptions to your internal SKU catalog, all financial arithmetic and tolerance checks run in deterministic Python Decimal code. For every extracted invoice line, our engine verifies:

1. Cross-Foot Arithmetic Integrity: Does Qty × Unit_Price - Discount + VAT == Line_Total, and does the sum of all lines equal the invoice header total?
2. Buyer Tax PIN & eTIMS Control Number Check: Does the decoded eTIMS QR payload match the printed invoice total, Supplier PIN, and your corporate Buyer PIN?
3. Unit Price Tolerance Gate: Is Invoice_Unit_Price <= PO_Agreed_Unit_Price (within a configurable penny-rounding tolerance, e.g., ±0.5% or KES 2.00)?
4. Cumulative GRN Quantity Ceiling: Is Previously_Invoiced_Qty + Current_Invoice_Qty <= Total_Accepted_GRN_Qty?

# Python / Django: Deterministic 3-Way AP Matching & Automatic Debit-Note Discrepancy Engine
from decimal import Decimal, ROUND_HALF_UP
from django.db import transaction

class ThreeWayAPMatchEngine:
    UNIT_PRICE_TOLERANCE_PCT = Decimal("0.002")  # 0.2% rounding tolerance

    @classmethod
    @transaction.atomic
    def evaluate_supplier_invoice_match(cls, extracted_invoice_id: int) -> dict:
        inv = ExtractedSupplierInvoice.objects.select_for_update().get(id=extracted_invoice_id)

        # 1. Hard Gate: Verify KRA eTIMS Control Unit Invoice Number (CUIN) & Buyer PIN
        if not inv.etims_verified or inv.extracted_buyer_pin != inv.company.kra_pin:
            inv.match_status = "BLOCKED_ETIMS_NON_COMPLIANT"
            inv.save(update_fields=["match_status"])
            return {"status": inv.match_status, "reason": "Missing/invalid eTIMS CUIN or Buyer PIN mismatch."}

        discrepancies = []
        total_overcharge_kes = Decimal("0.00")

        for line in inv.lines.select_related("po_line").all():
            po_line = line.po_line
            # 2. Check cumulative accepted GRN quantity vs. cumulative invoiced quantity
            unbilled_grn_qty = po_line.total_grn_accepted_qty - po_line.total_qty_already_invoiced
            if line.invoiced_qty > unbilled_grn_qty:
                qty_excess = line.invoiced_qty - unbilled_grn_qty
                overcharge = (qty_excess * po_line.agreed_unit_price).quantize(
                    Decimal("0.01"), rounding=ROUND_HALF_UP
                )
                total_overcharge_kes += overcharge
                discrepancies.append({
                    "sku": po_line.sku.code,
                    "type": "QUANTITY_EXCEEDS_GRN",
                    "invoiced_qty": str(line.invoiced_qty),
                    "available_grn_qty": str(unbilled_grn_qty),
                    "overcharge_kes": str(overcharge),
                })

            # 3. Check Unit Price against Approved PO Unit Price
            max_allowed_price = po_line.agreed_unit_price * (Decimal("1.0") + cls.UNIT_PRICE_TOLERANCE_PCT)
            if line.invoiced_unit_price > max_allowed_price:
                unit_diff = line.invoiced_unit_price - po_line.agreed_unit_price
                billable_qty = min(line.invoiced_qty, max(Decimal("0.0"), unbilled_grn_qty))
                price_overcharge = (unit_diff * billable_qty).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)
                total_overcharge_kes += price_overcharge
                discrepancies.append({
                    "sku": po_line.sku.code,
                    "type": "UNIT_PRICE_EXCEEDS_PO",
                    "invoiced_price": str(line.invoiced_unit_price),
                    "po_agreed_price": str(po_line.agreed_unit_price),
                    "overcharge_kes": str(price_overcharge),
                })

        if not discrepancies:
            inv.match_status = "THREE_WAY_MATCHED_READY_FOR_POSTING"
        else:
            inv.match_status = "EXCEPTION_DEBIT_NOTE_DRAFTED"
            inv.discrepancy_overcharge_kes = total_overcharge_kes
        inv.save()

        return {
            "invoice_number": inv.invoice_number,
            "match_status": inv.match_status,
            "total_overcharge_caught_kes": str(total_overcharge_kes),
            "discrepancies": discrepancies,
        }

B. Automated Supplier Discrepancy Debit Notes & Monthly Statement Reconciliation

When the 3-Way Match engine catches a price or quantity overcharge (e.g., 10 damaged units rejected at the receiving bay on the GRN that the supplier still billed on the invoice), your AP clerk does not need to spend 25 minutes writing an email. With one click, the system generates an itemized Supplier Discrepancy Claim / Request for Credit Note attaching the signed GRN shortage report and PO price excerpt, holds the disputed delta, and clears the undisputed portion for payment upon receipt of the supplier's eTIMS Credit Note.

3. Build vs. Manual AP Entry & Per-Page OCR SaaS Comparison

4. Development Timeline & Cost Guide

• Phase 1: Omnichannel Ingestion, Vision-LLM Line Extraction & eTIMS QR Validator (Weeks 1–4): Dedicated AP email/portal ingestion, structured line extraction, cross-foot math check, and KRA eTIMS compliance gate.
• Phase 2: Semantic Vendor-SKU Alias Resolver & 3-Way Match Workbench (Weeks 5–8): Self-learning vendor SKU catalog mapping, cumulative PO/GRN/Invoice 3-way matching, and side-by-side exception review UI.
• Phase 3: Supplier Statement Auto-Reconciliation & ERP/Bank Payment Sync (Weeks 9–11): Automated monthly supplier statement reconciliation, withholding VAT/WHT calculation, and bidirectional sync with SAP, Oracle, Dynamics Business Central, Odoo, or custom ERPs.

5. Frequently Asked Questions (FAQ)