Accounts payable teams should ask whether sensitive invoice data needs to pass through a third-party generative large language model at all. That is the argument of a new position paper from an intelligent document processing (IDP) vendor.
The Position Paper on the Use of AI Techniques in Intelligent Document Processing was released on by San Diego-based ancora Software. It argues that purpose-built, continuously learning AI can match or exceed LLM-based extraction without routing confidential financial data through an external model.
"LLMs are impressive technology, and our position is not that LLMs are inherently insecure. The question is much simpler: Why does my sensitive invoice data need to go through a third-party LLM in the first place?" said Noel Flynn, CEO of ancora Software.
The paper points to the information invoices carry. It lists negotiated pricing, discounts, supplier relationships, banking information, payment terms and contractual terms.
It accepts that leading enterprise LLM providers have implemented security, privacy and tenant-isolation controls. When configured appropriately, these can help stop one customer's information being exposed to another, and may prevent its use in training shared models.
But the paper argues that these controls do not answer a more basic question. Once invoice data enters a third-party layer, where does it go, who can see it, and is that layer necessary?
Probabilistic versus exact
The paper notes that LLMs are probabilistic by nature. That makes them powerful in conversational, generative and reasoning applications, but financial document processing is a different challenge.
"Supplier banking information cannot be approximately correct," the paper states. Accounts payable automation requires information that is accurate, validated, consistent, traceable and auditable.
"Our objective isn't simply to produce an answer," Flynn said. "It is to produce accurate financial data that can be validated, traced, and confidently passed into downstream AP and ERP processes."
The paper also argues that a demonstration is not a platform. Generative AI has made it possible to show document extraction working in minutes or hours, rather than the weeks or months a prototype once took.
That has helped crowd the market. The paper cites industry analyst coverage tracking more than 300 providers across the global IDP landscape.
Production systems must handle multiple formats, from TIFF and PDF to spreadsheets and documents embedded in email bodies. They must also classify and separate documents, deal with attachments and integrate with ERP systems for validation, matching and field population.
Questions for every vendor
The paper sets out questions organisations should ask any prospective IDP or accounts payable vendor, including ancora itself.
Does the platform send invoices or extracted information to a third-party generative LLM? What does the model do, and can core invoice extraction operate without it?
Where is invoice information processed, stored and retained? Is it ever used to train, fine-tune, optimise or improve models or services?
When users correct an extraction, does the system remember that correction for future invoices from the same vendor? How is AI-generated output validated before it reaches the ERP or payment process?
What independent security and compliance credentials support the platform? What happens to processing if a third-party AI service becomes unavailable?
Finally, what execution time should be expected for each image, and what does LLM processing cost per image?
"The issue isn't whether a company can put an LLM into invoice processing. The question is whether doing so provides enough additional business value to justify the additional dependency, processing layer, and cost," Flynn said.
Learning from corrections
ancora's own approach uses patented unassisted and assisted machine learning built specifically for document processing. When a user corrects an extracted value, the system can apply that correction to the next invoice from the same vendor.
Its ancoraFusion architecture layers universal, customer-specific and vendor-specific intelligence with a small language model designed for document understanding.
The company says more than 2,000 customers use its technology. Based on its internal transaction data, it processes more than US$50 billion in annualised invoice transactions.
The paper frames the debate this way: "AI versus no AI is the wrong question. The right question is which AI is appropriate for the job."
https://ancorasoftware.com/