Artificial intelligence has been one of the most talked about topics recently. New AI tools are rapidly appearing on the market and are also starting to be integrated directly into corporate systems that process various administrative documents or other internal data – including the iNVOiCE FLOW solution, which now uses the improved Aurora AI module. However, there is much less talk about how these applications actually work internally and what all has to happen before the result reaches the user.
Andrej Gono , founder of the startup Citymind , an expert in artificial intelligence and a doctoral student at Mendel University in Brno, connects AI development, academic research and practical use of artificial intelligence in companies in his work. Thanks to this connection, he is able to describe very precisely how today's AI systems are designed, where they make sense and where, on the contrary, they hit their limits.
How does AI work in the background in apps like Aurora?
Traditional large language models have been trained primarily on continuous texts. This means they understand the meaning of sentences and paragraphs well, but administrative documents are not structured in this way. Invoices or orders have key information scattered throughout the page – in tables, headers, footers or notes.
Therefore, in the background of applications like Aurora, there is not a general LLM, but a specialized model called Transactional Large Language Model, or T-LLM for short. This is optimized for working with semi-structured documents. In this case, the AI not only reads individual data, but also understands its meaning in the context of the entire document - for example, it compares items with orders or checks whether totals make sense.
It is also essential that the document is not viewed as just text. It is perceived as an image. It recognizes the layout of the page, lines, tables, graphic elements or logos. It is this combination of visual and textual view that allows the document to be truly understood.
The key element is the so-called discriminative decoder. This means that the model cannot generate any new text, but only selects and interprets the data that is actually contained in the document. Thanks to this, the system cannot "invent" any data, which is absolutely essential for a business environment.
What needs to happen “behind the scenes” before AI reaches the user layer?
Even before the model itself has a say, a number of steps take place that are crucial to the quality of the result. Documents come from a wide variety of sources – from emails, via APIs, shared folders, EDI gateways to physical scanners. Each of these channels has different characteristics and different risks of errors.
The first task is therefore normalization. All inputs need to be converted into a uniform format that other system components can work with. This step includes content detection and extraction, format conversion, and input quality control – for example, whether the document is too dark or light, blurry, in insufficient resolution, or with missing pages.
In addition, for EDI messages, it is necessary to unify various structured formats, such as XML, JSON or EDIFACT, to create a consistent data base.
Then comes computer vision. Documents are not always in text form, so AI must first “read” the image content. Neural networks are used to convert scans or images into text, but at the same time analyze visual elements – the position of fields, table borders or relationships between individual parts of the document.
The recognized text is then converted into a form that the neural network understands. It undergoes tokenization , i.e. dividing the text into smaller units, and an inference process , during which T‑LLM not only extracts data, but also understands their mutual relationships. The model is also continuously adapted to specific company data, without the need for fixed templates.
Before being passed to the user layer, the data is validated. Each extracted field is given a confidence score that determines whether human review is needed. The information is then adapted to the formats of the target systems, typically ERP ( Enterprise Resource Planning ), and validated against orders, for example, via API ( Application Programming Interface ).
The backend also includes continuous improvement mechanisms . If the user makes a correction, the system learns from it and uses this knowledge for subsequent documents. In the context of iNVOiCE FLOW, this means that AI not only extracts data, but also prepares it for a smooth and reliable flow into accounting systems.
How does the process work that allows documents to be read, understood, classified, and passed on?
The process begins with document reading using OCR (Optical Character Recognition). The system extracts text from PDFs, scans or structured formats and identifies key fields such as supplier, amount or date.
This is where T-LLM and NLP tools come into play. The model understands the relationships between individual data – for example, whether the amount corresponds to the items, whether it is a duplicate invoice or a non-standard situation. Documents are classified by type and routed to the correct queues.
Based on rules, for example by supplier or amount, data is transformed into the required ERP format. After validation, it is passed on to the workflow – in the case of iNVOiCE FLOW, this means automatic approval, generation of notifications for discrepancies, or direct import into the accounting system.
How does this type of technology fit into corporate workflows and where do you see its main benefit?
The main benefit is the replacement of manual steps with automated and adaptive processes . The workflow in iNVOiCE FLOW starts with document receipt, continues with AI extraction and validation, follows the approval flow and ends with import into ERP and archiving.
This allows one person to oversee volumes that would previously have required an entire team. Not because AI makes decisions for them, but because it removes routine and repetitive work and leaves room for humans to handle exceptions and control.
As part of your collaboration with GRiT, you are also working on developing their AI assistant. How is this assistant designed and how will it work in practice?
Technically, the solution is built on an architecture called RAG – Retrieval-Augmented Generation . Citymind’s AI is not trained on data from the entire internet and is strictly limited to information from a specific organization, such as GRiT.
The system first goes through web pages, PDF documents or internal materials and divides them into smaller logical sections. Each of them is converted into a vector form that represents its meaning. When a query is made, the system first finds the most relevant parts of the data and only then sends them to the language model with instructions to answer from them.
Thanks to this approach, answers are always based on specific sources and traceable. In addition, a mechanism has been implemented for GRiT, where the assistant recognizes a specific company based on the entered IČO and places the answer in its context - for example, explaining how GRiT can help it with faster data flow within the supply chain.
At GRiT, we see artificial intelligence as a tool that should help companies manage data in their daily operations – from documents to orders to the traceability of entire processes. We are interested in how technologies accelerate data flow, where they bring greater accuracy, and where, on the contrary, there must be room for human control and decision-making. If the topic of automation and data flow has interested you from the perspective of your company, please contact us .
Are you dealing with invoice approval?
30 minutes, no obligation. We will show you how iNVOiCE FLOW fits into your ERP.
