PDF invoice mismatches to AI capture + BirchStreet/NetSuite three?way match, fewer holds

Overview

A multi-brand hospitality chain was bogged down by invoice mismatches because supplier invoices arrived as PDFs with inconsistent line descriptions, pack sizes, and units of measure. Intelligex integrated an AI-assisted capture and three-way match layer into the client’s existing procure-to-pay stack, tuning optical character recognition (OCR) for hospitality categories and normalizing data against the item master. With category-specific tolerance rules and a clear exception queue, Accounts Payable (AP) saw fewer holds, buyers stopped chasing vendors for resubmissions, and approval cycles became predictable without changing core systems.

Client Profile

• Industry: Hospitality (multi-brand hotels and resorts)
• Company size: Multi-region, property-based operations
• Stage: Mature operations with centralized shared services for AP
• Department owner: Procurement, Supply Chain & Logistics (working closely with Finance/AP)
• Other stakeholders: Property-level receiving teams, Category managers, IT/Enterprise Applications, Internal Audit

The Challenge

The AP team was holding a significant portion of invoices due to mismatches between invoice PDFs, purchase orders (POs), and receiving records. Suppliers described the same items in different ways: “Russet 50 lb,” “Potatoes 50#,” and “Russet potatoes sack 50lb” all represented the same line, but didn’t match the item master. Units of measure varied by vendor and region. Freight and fees appeared on some invoices and not on others. The existing system required exact matches, so even routine variances led to holds and rework.

Most properties received goods and posted receipts promptly, but the matching depended on manual rekeying from PDF attachments or basic OCR that wasn’t tuned for hospitality. Month-end spikes and vendor-specific formats overwhelmed the AP queue. Buyers and property managers were pulled into invoice dispute resolution, and suppliers were asked to resubmit or reformat invoices—straining relationships and delaying payment runs.

Budget and time constraints precluded a rip-and-replace of the existing procure-to-pay stack. The company needed to work within its current tools—BirchStreet Systems for procurement and receiving, and Oracle NetSuite for AP and general ledger—while adding intelligence that could interpret invoices reliably and enforce practical tolerance rules.

Why It Was Happening

The root issue was fragmentation across document formats and item definitions. Distributors used their own SKUs, pack sizes, and text descriptions, while the hospitality chain’s item master reflected standardized categories and internal item codes. There was no robust crosswalk connecting external descriptions to internal items, and basic OCR struggled with hospitality-specific terms, weights, and abbreviations. Even when the right PO existed and the property had posted a receipt, free-text line descriptions prevented an automatic match.

Governance around tolerances had also become rigid over time. A blanket policy treated linens, produce, and beverages the same, even though weight-based categories naturally vary and should be matched by extended cost or by weight tolerance, not by exact unit count. Without category-aware rules and a structured exception process, teams spent more time adjudicating edge cases than processing clean invoices, creating unnecessary delays and inconsistent outcomes.

The Solution

Intelligex built an invoice capture and matching service that sits between the existing procurement and AP systems. We tuned OCR using category-specific dictionaries and mappings for hospitality items, normalized units of measure, and matched invoices to POs and receipts with category-based tolerance rules. The service integrated directly with BirchStreet for POs and receipts and with NetSuite for AP voucher creation, with a human-in-the-loop review queue for exceptions. Nothing in the core systems was replaced; we added an intelligent layer that standardizes inputs, applies policy consistently, and provides an auditable trail.

• Connectors to pull POs and receipts from BirchStreet Systems and to create bills in Oracle NetSuite.
• OCR pipeline using Azure AI Document Intelligence with hospitality-specific custom dictionaries (e.g., common abbreviations, pack sizes, varietals) and vendor layout profiles.
• Category classifier and item crosswalk that map vendor SKUs and descriptions to the internal item master.
• Unit-of-measure normalization aligned to GS1 units of measurement to reconcile pounds, cases, and eaches.
• Three-way match engine aligned to PO lines and receipts, with adjustable tolerance rules by category and vendor; freight/tax allocation logic; duplicate invoice detection.
• Human-in-the-loop exception queue with reason codes, side-by-side document view, and one-click resolution paths.
• Dashboards for AP and Procurement showing match rates, exception aging, top offending vendors, and rule performance.
• Granular permissions tied to roles (AP, Buyer, Property Manager) and an immutable audit log of changes and approvals.

Implementation

• Discovery: Mapped the current invoice flow end-to-end; sampled a diverse set of supplier PDFs (broadline distributors, linens, beverages) to catalog formatting patterns and common mismatch causes; reviewed tolerance policies and approval thresholds with Procurement and Finance.
• Design: Defined the data model for item crosswalks and units of measure; laid out category-specific tolerance rules; designed the exception queue and review gates; planned API integrations with BirchStreet and NetSuite without altering downstream posting logic.
• Build: Implemented the OCR pipeline and custom dictionaries; built mapping services for vendor SKUs to internal items; created a three-way match engine; developed connectors and webhooks to synchronize POs, receipts, and invoice status.
• Testing/QA: Ran the service in shadow mode, capturing invoices and simulating matches while AP continued its existing process; compared outcomes, tuned tolerance rules, and refined OCR layouts; added human-in-the-loop review steps for ambiguous lines.
• Rollout: Phased deployment by region and supplier group; started with high-volume, stable vendors; expanded to long tail after stabilizing classifications; kept a rollback path to the original process during each phase to prevent disruption.
• Training/hand-off: Conducted short sessions for AP, buyers, and property managers focused on exception handling and reason codes; provided playbooks for vendor-friendly invoice formatting; set up governance cadence with Procurement and Internal Audit.

Results

The AP team saw a clear drop in invoice holds tied to description mismatches and unit-of-measure errors. Clean invoices flowed straight through to vouchers in NetSuite, while exceptions were concentrated in a single queue with clear reasons and suggested resolutions. Buyers were no longer asked to decode vague line descriptions or ask suppliers to reissue invoices for minor variances; time previously spent on back-and-forth was redirected to sourcing and contract negotiations.

Finance gained a more consistent close process with fewer last-minute surprises. Audit trails improved through standardized reason codes and documented approvals. Category-aware tolerances reduced unnecessary disputes, especially for weight-based goods, and duplicate detection cut down on inadvertent double processing. Overall, invoice cycle time shortened, rework decreased, and visibility into bottlenecks improved without changing the client’s core systems.

What Changed for the Team

• Before: AP rekeyed line items from PDFs and paused invoices for unclear descriptions. After: OCR automatically captured lines and matched to the item master; AP focused on targeted exceptions.
• Before: Buyers fielded frequent vendor questions and resubmission requests. After: Exceptions were handled within the review queue with standardized reason codes; buyers were only pulled in for true discrepancies.
• Before: One-size-fits-all tolerances triggered holds for routine variances. After: Category and vendor-specific rules reduced false mismatches and made approvals more predictable.
• Before: Multiple tools and email threads to reconcile issues. After: A single exception workspace with side-by-side invoice, PO, and receipt context.
• Before: Limited insight into where and why invoices stalled. After: Dashboards showed top patterns by vendor, category, and property, enabling proactive fixes.

Key Takeaways

• Integrating an intelligent capture and matching layer can transform invoice processing without replacing your ERP or procure-to-pay systems.
• Category-aware tolerance rules reflect operational reality and prevent unnecessary exceptions, especially in weight- or pack-based categories.
• A maintained crosswalk between vendor SKUs and your item master is essential for reliable automation with diverse suppliers.
• Human-in-the-loop review maintains control and trust while allowing automation to handle routine cases.
• Start with high-volume vendors and run in shadow mode to tune models and rules before full rollout.

FAQ

What tools did this integrate with?
We integrated with BirchStreet Systems for POs and property-level receiving, and Oracle NetSuite for AP voucher creation and posting to the general ledger. OCR and document parsing were powered by Azure AI Document Intelligence. The approach is adaptable to other ERPs or procure-to-pay platforms with similar APIs.

How did you handle quality control and governance?
We implemented a human-in-the-loop exception queue with mandatory reason codes for overrides, periodic sampling of auto-matched invoices, and an immutable audit log. Tolerance rules were owned by Procurement and reviewed with Finance on a set cadence. Internal Audit had read-only access to logs and exception histories.

How did you roll this out without disruption?
We ran the system in shadow mode first, validating matches while the existing process continued unchanged. Deployment was phased by region and supplier group, with a defined rollback plan for each cutover. Training focused on exception handling, and vendor communications were coordinated to reduce format variability without demanding full reissuance of invoices.

How do tolerance rules affect supplier relationships?
Category-specific tolerances reduced unnecessary disputes and resubmission requests. Suppliers appreciated clearer, consistent feedback when exceptions did occur, and the client shared simple invoice formatting guidelines to improve straight-through processing.

What happens when menus change or new items are added?
The item crosswalk service listens for changes to the item master and allows rapid mapping of new vendor SKUs. When the system encounters an unmapped item, it routes that line to the exception queue for review, then persists the mapping so future invoices for that item match automatically.