„… we can extract 100% of the data correctly from all document types!”, up to this point the webinar presentation on the topic of “Digitalisation in Invoice Reception” was unspectacular, but this statement made me sit up and take notice. I eagerly awaited the challenging questions from the attendees at the concluding Q&A session. But nothing! Nothing at all! This 100% statement on the degree of automation in receipt capturing was apparently accepted. I hereby take this as an opportunity to face the facts and explain the BLU DELTA practice.
But first: What does the “market-standard” reality look like?
When it comes to automated receipt and document capture, automation rates—which refer to the percentage of a company’s receipts that are automatically returned as correct and subsequently processed without verification—cannot be generalized. Reconciliation with master data plays a crucial role here. However, reconciliation is only effective if the upstream data extraction has been successfully completed. Besides the variety of layouts (which usually corresponds to the number of suppliers), recognition rates depend primarily on the quality of the invoices (scanning, resolution, stamps, handwritten notes, etc., all influence quality). With standard systems, individual layouts are “trained.” This means that the fields of unknown layouts must be “taught” three to seven times. Therefore, 100% recognition is wishful thinking, and the actual recognition rate varies considerably depending on the variety of invoices and the amount of “training” effort one is willing to invest. If the layout variety is low, the risk and effort remain manageable. However, with high variety and large volume, close scrutiny is required!
Security through evaluation
Of course, it is better if the recognition rate can be determined in advance. With general AI models, the customer can evaluate the quality in advance with their own receipts. Individual invoices can be tested directly via our BLU DELTA system on our website. For a test with several invoices, online or API access can be requested free of charge, or a potential analysis can also be commissioned for sufficiently large quantities of receipts.
The BLU DELTA AI Potential Analysis – a statistically backed prediction
With the potential analysis, the customer receives a detailed report on the expected recognition rate per data field as well as on the overall recognition rate for the receipts circulating in the company within a statistically backed confidence interval.
Together with the customer, we determine a representative sample. Based on this invoice or document sample, the image data is analysed in a basic analysis and the target data is annotated and compared with the output of the BLU DELTA API (benchmark). Based on this, a detailed report on the quality of the documents and a prediction of the recognition rate of our AI model is created. In addition, the report also provides information about the optimisation potential. Thus, the customer has a secure basis to determine the degree of automation in advance.
Custom AI model as a by-product
And it gets even better. The output of the potential analysis – if helpful – is also an AI custom model optimised for the company. This model can be immediately tested live via our API and/or workflow components or directly integrated. Thus, the customer receives what BLU DELTA promises and there are no unpleasant surprises. But this much can be said in advance – it’s not 100%!
