More startups are using GenAI for IDP

last updated:

More startups are using GenAI for IDP

last updated:

Within 12 months of the ChatGPT earthquake, an astonishing 75% of IDP software companies have added generative AI functionality to their product line.

We surveyed the industry in December and asked vendors if they were using generative AI (GenAI) in their products. 75% said yes. (Full results of the research are available in the new IDP market report.) This is perhaps the best proof of our proclamation last year that GenAI will be the single most disruptive technology in the history of IDP.

GenAI innovation is happening all across the industry, spanning from the newest startups to the oldest incumbents like Tungsten Automation (formerly Kofax). Some of the most innovative ideas come from nimble startups who have both the room to experiment and a motivational sense of urgency. I talked recently to four such startups using GenAI to create new use cases, challenge the incumbents and win business from much larger competitors. Here are the briefs on Deep Cognition, Skwiz,, and Sensible. 

Trainingless IDP from Deep Cognition

Founded in 2017 as a deep learning platform for anyone who wanted to create and train their own AI models, this Texas-based software company pivoted a few years back to build an IDP platform for invoice processing. In a crowded market, Deep Cognition created differentiation by handling the most complex commercial invoices. The company has since gained traction with key accounts in the cross-border freight forwarding and customs brokerage industries. In April 2023, the company raised $1.2 million in a venture round that included renowned investor Mark Cuban. 

Deep Cognition recently announced its “Trainingless IDP” solution. More than just a clever slogan, the team demonstrated how that works in a customs brokerage use case. When a commercial invoice from a new supplier arrives, the software performs the mandatory checks and extracts the data without any training or annotating. In other words, training-less. The software comes with about 100 pre-configured AI models fine-tuned for logistics and other use cases. The company said it is using proprietary generative AI models to achieve high accuracy out of the box, because it is well-known that foundational LLMs such as OpenAI’s GPT suffer from hallucination and accuracy limitations.

The team cited one new customer who achieved 90% accuracy out of the box before any fine-tuning was done. That is very impressive, considering the bewildering complexity and variability of cross-border customs and freight forwarding documentation. For exception handling, there’s always a human in the loop to correct errors or low threshold results. The AI model learns from that.

Squeezing data out of documents with Skwiz

This is definitely a candidate for my “weirdest company name” award. Skwiz mean squeeze, as in squeeze the data out of the document like you squeeze the juice from an orange. It’s actually a fitting metaphor for IDP. In a familiar origin story, the Brussels-based founders of Skwiz met at university as computer science students working on machine learning projects. Then, just as many other IDP startups have, the team met a customer who asked them to solve an invoice processing problem that another well-known IDP company failed at. Next thing you know, they launched Skwiz as a cloud invoice data extraction API using machine learning models.

After ChatGPT came out, the team decided to rebuild Skwiz from the ground up around GenAI. The product was relaunched in 2023 with a new end user interface and the AzureOpenAI LLM under the hood to handle the generative functions. Before we could say it, Skwiz was quick to point out that LLMs are not replacing machine learning models. The team provided the following slide to show where each method is best deployed. LLMs can be used alone for some simple document use cases; but more complex and variable documents require a combination of LLM and ML models to achieve the rigorous accuracy requirements of a business process.

Document Processing Infrastructure by

This Amsterdam-based startup was founded in 2021 as AutoPilot. With several other companies and products using the same name, to avoid market confusion the company rebranded last month as Before starting the business, the founders worked together at a Dutch IT consultancy that specialized in RPA projects.’s first customer was a large international ship broker. The broker employed several people to read and process up to 400 emails a day from clients with urgent questions. This was not fast enough to meet the demand. developed software to read and parse the emails and extract the necessary data. Now client emails are processed much faster with less human involvement.

Fast forward to today, is a 4th Wave early-stage startup with a minimal viable product (MVP) currently in beta test with several customers. The team just raised $2.4M in a pre-seed funding round from Google Ventures and others based on a business plan to offer what they call a “Document Processing Infrastructure”. The team found that foundational LLMs such as OpenAI and Google are able to achieve about 70% data extraction accuracy out of the box, which is far from acceptable for their customers. They decided to use open-source LLMs and customize solutions around specific data sets. is focusing initially on the Netherlands and Belgium markets, where they say there is  much less competition from the big IDP vendors. We agree with that strategy; too many funded startups try to go big from the moment the seed money hits the bank account then are crushed by competitors in other regional markets.

DevOps for Document Automation from Sensible

Josh Lewis experienced firsthand the challenge of extracting structured data from PDFs at Newfront Insurance, where he built internal tools for managing quotes, applications, and certificates of insurance. He couldn’t find any developer tools that gave him sufficient control over the parsing process while still working at a layer of abstraction above the raw OCR or text dump from a PDF. Josh began to develop a cloud-based API to extract data from PDFs. At first, he thought, how hard can this be? Then he found out what all IDP insiders know: it’s devilishly hard to accurately extract data from all kinds of different documents with variable layouts.

So, Josh started Sensible in 2020 along with Ming Lu, formerly head of products at Lattice. Their mission is to provide document automation APIs for developers of SaaS products and make it easy for them to integrate and manage – and also provide friendly support. The main target audience are SaaS developers whose own target customers happen to have document-centric workflows; this turns out to be a much larger market than one would think. Sensible raised $6.5 million in 2022 and now has 20 employees. The company goes way beyond just docAI APIs, providing its customers with a management dashboard to manage, test and orchestrate their IDP projects. Their goal is to become the “DevOps for Document Automation” company. Makes sense: developers helping other developers to develop.

That’s the latest startup news. Stay tuned for much more to come in 2024. If you are a startup using generative AI for document processing and would like to brief us, please contact us at [email protected].

Leave a Comment

We use cookies to ensure that we give you the best experience on our website. If you continue to use this site we will assume that you are happy with it.

Work Intelligence Market Analysis 2024-2029