Intelligent Document Processing (IDP) Review - Planet Ai

I thoroughly enjoy meeting with an organization and finding out something that pleasantly surprises me. When I met with Dennis Seemann and Holger Prohl from PLANET AI to discuss their Intelligent Document Processing solution set, I learned that Planet AI is one of the oldest IDP manufacturers, just behind ABBYY (founded in 1989) and Kofax (founded in 1985). Planet AI was founded in 1992 as an intelligent document analysis research and development company, and was recently acquired by Bechtle in 2023. Given their 30+ years of market experience, they offer a unique perspective on IDP, and an existing pedigree as a market leader in OCR accuracy and advanced recognition modules.

Planet AI has been the first IDP solution that I have found that's offered strictly on Linux servers as a JAVA application, and while Planet AI will run on Windows, there are some limitations. Planet AI is strictly an on-premise solution and does not offer SaaS. This provides resellers with size and scope to white label the product and self host, allowing for extra profit margin and retain control of customers data and documents.

User Interface

The Planet AI capture solution comes with a cockpit User Interface (UI), which includes a testing web client, web control where you design the system elements from models, and recognition and extraction attributes. It also includes a dashboard for reporting, model, and neural net building. It is important to note that Planet AI does not have an end user data verification UI, which allows users to validate extraction data, manipulate documents, and configure inputs and outputs.?

Despite this, most of the admin and setup for the underlying models, modules, and training do have a UI. Planet AI has prioritized configuration of the models and training over the data verification experience.?

Their approach is to allow resellers and partners to create their own capture UI and access the API through GRPC (Google Remote Procedure Call). GRPC is a cross-platform high-performance remote procedure call framework. Initially created by Google, GRPC is open source and is used across many organizations. Custom development can be successful with the prerequisites of Java skills with some GRPC knowledge.?

Given the Planet AI tech stack likely predates RESTful API’s, the newcomer IDP manufacturers who created API-first platforms have a leg up on the underlying architecture. Integrators would need to dust off their #Java and #GRPC skills to make this platform work. That means other ERP and CRM applications that share a Java core, would be prime candidates for embedding Planet AI vs. newer Microsoft .Net architected platform products.

Web Client

The web client, essentially a harness to test document extraction, is very limited. But, for handwriting recognition, the UI shows the computer generated text above the handwriting to show the engine's interpretation. The web client allows you to test and consume previously created modules.?

There are 4 inputs for each test:

• Load Module - this establishes inputs, where the documents are coming into, and any pre-processing. Planet AI has a handful of UI modules for aspects of the product, but these were developed internally to support deployment.

• Recognition Module - recognition of computer generated text or handwriting
• Extractor Module - what extraction parameters are set for the document type. Multiple methodologies offered by the Planet AI IDA engine.?

• Write Module - what is the format of the data thats getting exported. Currently output of data is limited to JSON or hOCR files whereas document output is image and PDF.

Web Control

Web Control is where you define workflows, models, and notifications in a shared development environment, where the workflow container can use any one of the previously created modules. Here you can also stitch together these modules into an end to end process workflow.

This is where PLANET AI shines with so many cool features that capture practitioners need to create end to end IDP solutions. Clearly given they have been in business 30+ years, their roots still show through with more actionable models than most products I see. Here are a few features:

• Document Split & Classification - classification of documents or pages as well as document separation based on neural net or on rules.?
• Understanding - understanding of content from structured or unstructured documents
• Switch - document routing and workflow based on already generated results
• Image Normalization - image enhancement features to cleanup documents

Dashboard

The dashboard UI is for high-level, reporting statistics associated with batches of documents. It also includes data labeling to define data extraction fields within a document, training data and neural nets. While the product demo was high level we didn't get too much into the inner workings of data labeling and training. But from what I can glean from the features, the underlying model building for extraction and classification provides reseller and systems integrators with tools to build use case customer models.

Outputs and Inputs

Currently, The Planet AI solution only supports image file formats and does not support electronic files (e-files). The data output is strictly JSON. There is no automated service to poll for importing jobs from email or SMB folders. Their existing load module tool allows browsers to select files on an ad-hoc basis.

Data and document output is also not available in a batch export service. The test harness allows manual download of the extraction results in JSON, but does not show the data in a standard data verification UI that's typical of other IDP solutions.

The Engine

Accuracy

Having interviewed many IDP solutions, the number one question I ask is, “Why your product over another fairly capable IDP product?” and the answer is typically, accuracy. We are the most accurate which means less manual human labor. I haven't found any company that actually backs up this statement with data, until Planet AI. If you search online for IDP benchmarking you won't find too much empirical data with known commercial products.

Benchmark Testing - Planet AI vs. Commercial and Open Source OCR Engines

Planet AI published a white paper in 2023 comparing commercial, open source and their own internal IDA engine. The analysis used these OCR engines in comparison to the PlanetAI IDA solution.?

• Open-source Engines: easyOCR, MMOCR, PaddleOCR, Tesseract
• Commercial Engines: Amazon, Azure, Google, Planet AI
• Visual-based Engine: GPT4 Vision Preview (OpenAI)

The main metric comparison was on CER (character error rate), where a select number of handwriting and computer text were analyzed by character. Obviously the lower error rate the better the engine. Demonstrated by the chart below, there is a huge gap between cloud engines and Planet AI over visual gen ai tools like ChatGPT 4 and open source engines such as Tesseract.

Commercials

Since Planet AI is an on-premise based system, the commercials are different than most SaaS IDP solutions. While Holger explained they negotiate based on number of pages and use case, they primarily sell site licenses with unlimited capture. For more specific use cases and lower volumes, its consumption based on the number of pages but also the modules. Classification, recognition and extraction are the three modules. At lowest tier volume with all modules would be ~$0.40 per page.

Conclusion

Planet AI based on an hour and half interview style demo, I took away that this is a feature rich OCR and machine learning platform more than its a turnkey, commercially ready product customers can pick off the shelf. But unlike other API-first IDP solutions, that have no or limited user interfaces, this solution provides UI for the setup, admin and learning modules.?

So Planet AI has focused on what they deem important and have focused on the underlying technology “engines” to make it so much more robust and configurable than focusing on data verification UI’s. They are inventing tools like Entity Finder, which can identify the entity/company associated with unstructured documents by leveraging their advanced ai and machine learning algorithms. They focus on their partner network to build the pieces they are missing, while consuming their stable, advanced tech stack and have placed less focus on user interface, inputs and outputs like traditional batch capture tools.

As a consultant, where does this solution make sense? First, given the price point for the solution, you need to have very high volumes to reduce the per page price or enter into an unlimited site license. That commercial model just doesn't make sense for smaller volumes.?

Second, there are some challenges with consuming Java applications through GRPC, hosting on your own equipment and trying to make it cloud ready. So a customer that already has their own infrastructure, has a Linux expert on staff and supports Linux applications today are best primed to reap the benefits.?

Third, there is no user interface, so customers looking to embed this IDP technology within their ERP application would be best. Given the legacy nature of the Planet AI API the best target application is a Java core running on Linux. That Linux/Java application target market is very narrow, but I can see a great synergies between Planet Ai and open source ERP systems like Apache OFBiz or ERPNext.