Document Anomaly Detection and Fraud Control
Data Alchemy runs anomaly detection on every document entering your business: it compares each invoice, delivery note, order and price list against your ERP master data and flags duplicates, off-contract prices, suspicious IBANs, inconsistent quantities and potentially fraudulent invoices before they are posted. Validation against up-to-date master data, 99.8% accuracy with Claude, ~12 s for a one-page document (Claude Sonnet), with no manual data entry.
What document anomaly detection is
Document anomaly detection is the process that checks every incoming document — invoices, delivery notes, orders, price lists — against the correct data already held in your ERP, to catch errors and fraud before posting. Done by hand it means spot-checking a few invoices while everything else goes through unchecked. Data Alchemy extracts the data with an LLM dedicated to each document model — Claude by Anthropic (the engine we recommend by default), GPT by OpenAI, Data Alchemy AI (Data Alchemy's proprietary model, running on its own servers and GPUs in Italy) or a compatible LLM through the European inference providers Scaleway and Seeweb Regolo.ai, on the customer's own subscription — and validates it in real time against supplier, item, price-list and contract master data. Every deviation becomes a flagged exception: a price that differs from the agreed one, an IBAN never used before by that supplier, an invoice already posted, a quantity that does not match the order or delivery note. Finance and controlling stop sampling and only handle the genuine anomalies.
Which document anomalies and frauds it detects
On every document Data Alchemy runs dozens of cross-checks against ERP master data and the posted history. Here are the typical anomalies caught automatically.
Price and list anomalies
- Unit price different from the agreed price list
- Discounts not foreseen by the contract
- Increases outside the agreed terms
- Line amounts inconsistent with quantity × price
Duplicates and double submissions
- Invoice already posted in history
- Same document number from the same supplier
- Double submission of the same invoice
- Credit notes with no reference invoice
Fraud signals
- IBAN never used before by that supplier
- VAT or tax code not in the system
- Supplier missing from the approved master data
- Payment details changed versus history
Inconsistencies and reconciliation
- Quantities that do not match order and delivery note
- Totals, VAT and taxable amounts not reconciled
- Inconsistent VAT rates or nature
- Items not present in the master data
Document anomaly detection in 5 steps
From document receipt to exception flagging, Data Alchemy automates every check by comparing the data against ERP master data, with no manual intervention.
1. Acquisition
The document arrives via email (from business mailboxes, through the Mail service installed at the customer's premises), scanner or upload. Only emails matching a model reach the platform.
2. Data extraction
A dedicated LLM reads the header and line items and extracts supplier, amounts, VAT, IBAN, items, quantities and prices, with no per-supplier template.
3. Master-data comparison
The extracted data is compared in real time with ERP master data: approved suppliers, items, price lists, contracts and payment terms.
4. Anomaly detection
Every deviation is classified and flagged: an off-list price, a duplicate, a suspicious IBAN, an inconsistent quantity or a failed reconciliation becomes a tracked exception.
5. Validation or block
Compliant documents are posted into SAP, Zucchetti or TeamSystem; anomalous ones are blocked and queued for review, together with the reason for the failed check.
Everything you need for document anomaly detection
Validation against ERP master data
Every document is compared in real time with SAP, Zucchetti or TeamSystem: suppliers, items, price lists and contracts become the reference for recognising what is anomalous.
Fraud and suspicious-IBAN detection
An IBAN never used by that supplier, a VAT number not in the system or changed payment details are flagged before the invoice is paid.
Duplicate check against history
Every invoice is compared with the posted history: duplicates, double submissions and identical document numbers from the same supplier are blocked before posting.
Price and price-list control
Unit prices, discounts and terms are checked against the agreed price list: deviations from the contract become an exception to approve, not a hidden cost.
Reconciliation and three-way matching
Totals, VAT and quantities are reconciled and cross-checked with order and delivery note, so lines that do not match surface before posting.
Tracked exceptions
Every anomaly is logged with its reason — the check it failed and the reference data — for audit and controlling.
Anomalies caught by querying your ERP directly
An anomaly can only be recognised if there is a correct reference to compare the document with. With the Business and Enterprise licences that reference can be read straight from your management system, via API, while the document is being processed.
Verifying AI-read data against external services
When reading a delivery note that references a purchase order, Data Alchemy queries the customer's ERP via API, retrieves the order contents and checks item code, quantity and the other fields line by line: every line that does not match becomes an exception, before posting.
Validation against a third-party source such as your ERP
Remote validation connects a third-party source such as your ERP to check item codes and retrieve the customer code from the company name or VAT number: a non-existent code or an unrecognised customer surfaces immediately as an anomaly.
External fields and remote validation are included in the Business and Enterprise licences. Compare the licences on the pricing and licences page.
The ROI of automated anomaly detection
Checking every document — not just a sample — protects margins and stops wrong or fraudulent payments. Here are Data Alchemy's reference figures, the same as on the home page.
of documents checked, no more sampling-only reviews
accuracy with Claude in extraction and comparison against master data
Per 1-page document (Claude Sonnet)
operational time (Data Alchemy average)
Solutions related to anomaly detection
Invoice reconciliation & three-way matching
Invoice-order-delivery note matching line by line with discrepancy detection before posting.
Learn more →AdministrationSupplier invoice automation
The solution built for the finance team: post invoices and entries with no re-keying, validated against the ERP.
Learn more →Data extractionInvoice data extraction
AI reading of header and line items from PDF, XML and FatturaPA, with duplicate checks and master-data validation.
Learn more →PlatformAll IDP solutions
The IDP platform that brings every module together: invoices, delivery notes, orders, contracts and price lists in one flow into your ERP.
Learn more →Frequently asked questions about document anomaly detection
What is document anomaly detection?
It is the process that checks every incoming document — invoices, delivery notes, orders, price lists — against the correct data already held in your ERP, to catch errors and fraud before posting. Data Alchemy extracts the data with a dedicated LLM and validates it in real time against supplier, item, price-list and contract master data, flagging every deviation as an exception to handle.
Which anomalies and frauds can it detect?
Data Alchemy catches prices different from the agreed price list, unforeseen discounts, duplicate or re-sent invoices, IBANs never used before by that supplier, VAT numbers not in the system, non-approved suppliers, quantities that do not match order and delivery note, and totals or VAT rates that do not reconcile. Every anomaly is classified and flagged with the reason for the failed check.
How does it detect invoice fraud?
Fraud detection is based on comparison with master data and history: the platform recognises whether the stated IBAN differs from the one usually used by that supplier, whether the VAT number or supplier are not approved in the system, whether payment details have been changed, or whether the same invoice has already been posted. These signals block the document before payment.
Does it integrate with the ERP (SAP, Zucchetti, TeamSystem)?
Yes. Master data (suppliers, items, price lists, contracts, payment terms) is read in real time from SAP, Zucchetti, TeamSystem and other ERPs through the AI agent installed at your premises (a database, CSV/JSON/TXT files or APIs). This master data is exactly what provides the correct reference Data Alchemy uses to recognise anomalies, and compliant documents are then posted straight into the ERP.
Does it check all documents or just a sample?
All of them. Unlike manual control, which for lack of time only spot-checks a sample of invoices, Data Alchemy runs the checks on every processed document; with Claude Sonnet or higher, reading takes on average around 12 seconds for a one-page document and around 18 seconds for two pages. Finance no longer samples but only handles the anomalies actually flagged by the platform.
How much does automated anomaly detection cost?
Data Alchemy works on a licence model — Basic, Business and Enterprise — plus pay-as-you-go credits, with a quote on request; to try the platform you start from the Demo (a trial Basic licence). External fields and remote validation, which check documents directly against your ERP, are included in the Business and Enterprise licences. The licence comparison is on the pricing page, and the ROI calculator lets you estimate the savings on your real volumes.
See anomaly detection on your documents live
Activate the free Demo with 10 documents and check the AI reading on your real documents, with 1 model chosen from the predefined ones in a dedicated environment. To see anomaly detection and validation against your ERP and calculate the ROI for your case, contact us to plan a call with a demo on your own documents.