Everything you wanted to know about the Agentic Approach to eInvoicing

Q&A with Fotis Drakos

I sat down with Fotis Drakos, Senior Product Manager of Integration at Hypatos.ai following a recent webinar, "The Agentic AI Approach to eInvoicing", to dive deeper into this important topic.  

As we know from years of continuous technological evolution, certain tasks are particularly well-suited for automation. Among these are processes that are essential to the business yet are prone to human error due to their repetitive nature, as well as those with well-defined rules that ensure accuracy in critical operations. eInvoicing is a prime example.

While the fundamentals of eInvoicing may be familiar, new and interesting questions continue to emerge from stakeholders in large, multi-national companies, where automation isn’t just beneficial—it’s becoming essential. Automating eInvoicing streamlines compliance and boosts accuracy, meeting the unique demands of complex organizational structures while minimizing costly manual errors.

Ana: Most global organizations elect to use English but naturally want to support local languages. Could a company use AI, for example, to support translation from a document in Chinese without involving a bi-lingual staff member?

Fotis: Absolutely. This basic feature has been available in Hypatos from the beginning. Translation is a must-have for large multinationals. So, if you receive a document in Chinese, it is possible for someone with no Chinese skills to read it once processed by Hypatos.  We have support for over 200 languages including those using Cyrillic and east Asian writing systems.  

Ana: There is a lot of talk about processing e-Invoices at 100%.  Is this even possible, or is it that processing is 90% complete, requiring intervention from human staff?  

Fotis: Believe it or not, 100% processing is a reality. For 100% processing accuracy to be true, the eInvoice needs to contain all necessary details. When that is not the case, Hypatos.ai is able to enrich the document and perform checks to move from 90% to 100% accuracy. Typically, 90% accuracy means key information is either incomplete or incorrect. For example, a tax code might be missing, or a PO line or Supplier ID required by the ERP system, might be incomplete. Inaccuracy and missing information hinder full processing. ‍

Ana: Most middleware or integration tools can extract information from an invoice and map it to an ERP system. Is an Agentic AI solution necessary?

Fotis:  Yes. An Agentic AI solution offers so much more than basic data mapping between XML and ERP systems. Data integration or middleware solutions focus on extracting information from the invoice and mapping it to the ERP fields, and managing some rule-based enrichment, such as matching supplier names with IDs.    

Agentic AI goes far beyond this by also enriching the invoice with master data, predicting accounting codes, such as GL accounts, and ensuring tax compliance. It also stays up to date with regulatory changes, validates eInvoice formats, and manages complex tasks like PO line matching and tax code compliance.

Middleware tools often require constant maintenance and manual intervention for edge cases that they cannot handle. By contrast, Agentic AI automates processes more intelligently and is capable of learning from the actions of expert teams, which reduces the need for ongoing manual adjustments and maintenance, especially as new invoice formats and regulations emerge.

Ana: How does Agentic AI keep pace with regulatory changes and how do you hold it accountable?

Fotis: AI agents comply with changes instantly if they are instructed to do so. The last point is important. An agent does not ‘care’ whether changes are subject to company policy updates OR regulatory policy updates such as VAT changes and ESG regulations.  The agent’s behavior and document processing decision-making are a function of human experts’ prompts. In the case of Hypatos, each agent comes with an easily editable prompt that allows the process owner to instruct in natural language.

Ana: Is an AI agent more or less likely to produce errors when it comes to eInvoice processing? Based on your last response I would say less likely – if the agent has been instructed correctly?

Fotis: Yes, that’s right. It’s difficult to give exact figures especially for critical fields like amounts, recipient names and addresses.  What I can say with confidence, is that AI agent errors are highly unlikely for several reasons:    

- Validation Rules: AI agents use specific validation rules to ensure data accuracy. For example, if an amount or recipient name is not confidently captured, the system can flag it or stop processing until the issue is resolved. This reduces the risk of mistakes going unnoticed.

- Consistent Data from XML: In eInvoicing, information from XML files is typically clean and structured, minimizing errors compared to PDF-based OCR processing, where misreads can occur.

- Continuous Learning: AI agents benefit from retrospective analysis. They can learn from past mistakes by comparing processed documents with ground truth data in the ERP system. This helps to improve future accuracy by refining prompts and enrichment processes over time.

While errors can occur due to incorrect document data or AI misinterpretation, the system is designed to catch these errors through validation and improve accuracy as it processes more documents. The error rate remains low, and success rates are expected to be even higher for e-Invoices compared to PDF processing.

Ana: Talking of PDF processing, is Optical Character Recognition a thing of the past now?

Fotis: Not quite. For standard PDF documents (those without embedded XML), we do use OCR to extract raw text, which is then processed by large language models (LLMs) for tasks like document understanding, extraction, and enrichment. When LLMs eventually become fully capable of processing raw PDFs directly in a cost-effective manner, OCR might no longer be needed in the future.

Ana: The German mandate for B2B e-Invoice processing becomes effective from January 1^st, 2025. How do you see AI and AI Agents contributing to the processing of specific invoice formats such as xRechnung, and ZUGFeRD?

Fotis: Recognizing the different e-Invoice formats that German suppliers may use, and being able to process them is something that can be easily done through Agentic tools. The tools can be instructed, in natural language, to properly identify each German eInvoicing format, and whether/how the invoice should be processed. This also applies to ZUGFeRD PDF/A-3 files (with embedded XML), which AI can properly identify, and perform the extraction, validation, and processing of the embedded XML content.

Ana: Besides German mandates, I’ve noticed the European Union is pushing for the use of a common e-Invoicing standard – EN16193.  Do you see the same happening on a global scale i.e., one global e-Invoicing standard? What role does AI play in this scenario?

Fotis: European countries are slowly adopting eInvoicing (in general) and the EN16193 standard (in particular). Given the effort required for non-European countries to catch up, it is highly unlikely that a global eInvoicing standard will be defined and adopted soon. When the time comes though, AI for eInvoice processing will still be relevant and a game changer. Processing with AI agents is about extraction and matching. It is also about providing deep contextual understanding of the eInvoice and deriving additional accounting information that is critical for the final posting in the ERP system. A tedious and demanding process that would still be relying on humans if not for AI.

Ana: I’ve seen many situations where suppliers send an eInvoice along with additional PDF attachments such as contracts and purchase orders.  Can AI process all this information and make sense of it?

Fotis: AI agents can manage all types of documents sent by a supplier and properly identify their type and purpose. For instance, if a supplier sends a contract or purchase order, along with an eInvoice, AI can associate these documents and be instructed to select all the relevant information from them such as contract, purchase order. This is used to enrich the information that is included in the eInvoice. The enriched information can be sent downstream to the ERP system, ensuring data completeness and that all the necessary information for the final ERP posting is in place.

Ana: One final question. If a supplier sends me a PDF document and an XML eInvoice through two separate workstreams and they both refer to the same invoice, can AI perform duplicate detection and ensure only one is processed and sent to my ERP system?

Fotis: AI agents can orchestrate the processing of incoming eInvoices and use advanced AI technologies (such as Retrieval-Augmented Generation) and other Agentic tools, to identify duplicates across different input streams and points in time. Exact logic for identifying such documents such as invoice number and invoice date exact match can be easily defined by using natural language.

‍

In summary, the speed and accuracy with which large companies can process documents in the area of Finance is incredible. Agentic AI and eInvoicing are a match.

Stay tuned for more information from Hypatos.

‍