The AI Invoice Analysis-as-a-service Landscape
The world of automatic invoice analysis has transformed in the last 3 years. All major cloud providers are now offering AI-as-a-service products, and business document analysis tools take a big part of their offering. Automatic document analysis can be transformative for many industries, and the VAT reclaim world is no different. Extracting information from invoices and receipts can drive decisions and tracking that are crucial for organizations.
We looked up the offering from 4 major cloud-AI providers and compared their document analysis products. Each solution has different parameters, such as cost, API, language bindings, metrics and integration with other data services. We hereby present our findings.
Microsoft Azure Cognitive Services
“Form Recognizer” identifies key-value data and tables from form-like documents, including invoices and receipts. The model can do so without labeled data as it leverages pre-trained models, but one can also provide manually labeled data for better performance metrics on harder document images.
The receipt identification API provides output fields such as: receipt type (e.g. “itemized”), merchant name, address, phone number, transaction date-time, line items and their prices, as well as totals and tax. The API supports HTTP REST call as well as a simple Python binding. This service is open and ready to be used by 3rd parties, and has a working online demo.
Google Document AI
Google’s “Document AI” is claimed extract and identify common fields e.g. addresses, phone numbers, merchants, amounts, and other key values such as tax in receipts and invoices. However, Google is not yet releasing the Document AI as a ready service API now, and request potential users to call the sales team.
The AWS service called “Textract” is able to perform OCR as well as extract key-value pairs and identify data types such as dates and amounts. While it doesn’t offer a direct all-in-one service to analyze an invoice or receipt fully, like Azure’s “Form Recognizer”, it does extract all the data to do so semi-manually. Online examples of using Textract for invoice analysis contain some regular expression or further steps to transform the output according to the use-case.
IBM Watson Invoice Understanding
IBM’s Watson AI services suite offers a direct API to analyze invoices. The service is able to extract the following fields: amounts, currencies, order and invoice IDs, dates and items. The service is accessible via a REST API with standard request tools such as curl. The service is offered in a beta state, and interested parties must email the IBM business partnership team to get access to the service API.
Analyzing invoices with AI is now offered as an automated service. This was not the case 3 years ago. This change is transforming the data pipelines for many organizations that rely on analyzing business documents. As far as we can ascertain, Microsoft’s Azure Cognitive Services invoice analysis service is the only readily available product that can be used immediately and fully, and this service is also in its 2nd version that speaks to its maturity. However, we are certain the other providers are hard at work to get their services up-and-running to join the race. IBM and Google are closer to releasing their API to the public than Amazon. With such services in place, invoice analysis can now be put into existing data processing pipelines to support use-cases such as VAT reclaim.