Automating complex knowledge work through an agentic platform: our investment in DeepOpinion

At Red River West, our mission is to support groundbreaking companies that have the potential to become global leaders in their sectors. In…

At Red River West, our mission is to support groundbreaking companies that have the potential to become global leaders in their sectors. In line with this vision, we are thrilled to announce our investment in DeepOpinion: a generative automation platform that automates complex knowledge processes.

A problem that nobody has solved yet

Today, hundreds of millions of people work in what we call “work farms” doing a high volume of very repetitive tasks like processing insurance claims or gathering specific data from documents. These jobs are often the cause of burnout, are unfulfilling, etc. and they represent hundreds of billions of spending per year for insurers, banks, governments, etc.

That’s why for the past 20 years a lot of actors have tried to build software platforms to automate at least part of those jobs. Robotic Process Automation (RPA) in particular, stood at the forefront of this movement, promising to automate rule-based tasks with minimal human intervention.

However, while RPA has made strides in automating routine, structured processes, it never really delivered when applied to more complex tasks, especially those involving unstructured data. Processing these types of data remains one of the most significant challenges for automation.

The problem is that unstructured data (free-form text, emails, images, documents, etc.) constitute the bulk of information that is available in the business world. Without a solution to treat unstructured data, “work farm jobs” can’t be automated.

Traditional Intelligent Document Processing (IDP) platforms, which should have addressed this gap, have failed to deliver on their promise of accurately extracting and processing information from complex documents. Legacy players of the IDP market like ABBYY or Kofax were founded in the 80s/90s, and even if they evolved, they still fall short in terms of performance.

The RPA giants like UiPath and Automation Anywhere also developed their own IDP capabilities, but they face the same limitations. These platforms were built to automate repetitive, rules-based tasks such as processing structured spreadsheets or transferring files between systems. They are based either on predefined rules and templates or on what we call supervised learning.

Supervised learning is a type of machine learning where a model is trained on labeled data to make predictions or classifications (e.g. you train it with millions of examples of invoices, and then it learns how to extract information from specific invoices). Even if that works well on documents like IDs or invoices that are pretty standard, if you want to apply these algorithms to more complex documents or e-mail, you’ll need to train them for a long time and each time there is something out of the ordinary, it breaks. It’s why, for many tasks, it can’t have a good enough precision.

In recent years, a new generation of IDP companies like Rossom or Instabase have also emerged, challenging the dominance of these legacy platforms. But even if they often perform better than their predecessors, most of their core intellectual property was built on supervised learning too. They were all built in the pre-LLM (Large Language Model) era.

Enter LLMs: A paradigm shift

Large Language Models (LLMs) represent a significant leap forward in processing and understanding unstructured data. We’ve seen their power in action through tools like ChatGPT, where they excel in generating and interpreting text across various contexts. Unlike traditional systems that rely on rigid rules and templates, LLMs can comprehend complex information and make inferences that go beyond basic pattern recognition.

LLMs are changing everything by enabling zero-shot learning, where models can tackle new tasks without needing extensive, labeled datasets. This means businesses can quickly create proof-of-concept solutions for extracting data from diverse document types with minimal instruction. Coupled with specific techniques, LLMs can achieve performance levels far exceeding what was possible with traditional methods.

However, deploying LLMs into production workflows is more complicated than it initially seems. Despite their promise, LLMs face several challenges when used in real-world environments making it hard for legacy players to use them:

• Hallucinations: LLMs can generate incorrect or misleading information, which is problematic in industries like legal, finance, or healthcare, where accuracy is critical.

• Lack of stability and controllability: LLMs can be unpredictable, especially in complex workflows that require the coordination of multiple tasks. Ensuring that these models operate reliably without crashing or conflicting with other systems is a significant engineering challenge.

• Scalability issues: LLMs are resource-intensive and can be difficult to scale efficiently across various tasks, especially when those tasks require different levels of specialization or processing power.

• Confidence metrics: LLMs often lack the ability to self-assess the accuracy of their outputs. Without robust confidence metrics, it’s challenging to determine when the model’s predictions should be trusted or when human oversight is necessary.

• Human-in-the-loop needs: Even the most advanced models may struggle with complex or ambiguous tasks that require human judgment, especially when the consequences of errors are significant.

• Lack of specialization & task-specific actions: LLMs, being general-purpose models, may not handle highly specialized tasks effectively or interact directly with external systems. That’s why some companies like DeepOpinion have developed an approach based on agent-based systems. These systems can call on multiple specialized models or tools, ensuring that each task is handled by the appropriate expert system and checked multiple times allowing for more precise and efficient data processing.

For legacy RPA and IDP players, adopting these technologies would require a fundamental shift from their traditional approaches and business models. It also requires tech talents that, most of them, don’t have and a high velocity to adapt to all the daily innovations around LLMs. That’s why, we believe that this paradigm shift will benefit a new kind of actors.

DeepOpinion: the complete platform for automating knowledge work

The deepOpinion team has built a generative automation platform that automates complex knowledge work processes through a multi-agent engine based on LLMs. The platform brings together all the necessary components for automating complex workflows and creating real ROI for its clients.

Several clients of DeepOpinion have reported that for certain use cases involving unstructured data, they previously engaged with multiple established providers such as UIPath. However, they found that even the best of these providers could only automate 10 to 15% of the processes after extensive months of training. In contrast, when these clients switched to DeepOpinion, they experienced an astounding 80 to 90% automation rate within just a few days, a rate that further improved over time. This has resulted in a greater than tenfold ROI across various applications, including insurance claim processing, customer support, underwriting, and loan processing.

Key technological differentiation

DeepOpinion uses the latest AI technologies and a cutting-edge approach to process unstructured data. What makes their differentiation is not just one particularly good algorithm or model but a combination of features, research work, and proprietary engines that allow them to tackle the limitations of LLMs and to automate knowledge work with high precision. Amongst those, here are some important aspects of the platform:

• Document understanding with LLMs: DeepOpinion uses LLMs to process unstructured text, extract key information, and classify data, outperforming traditional IDP solutions in accuracy and deployment time.

• A proprietary multi-agent engine: DeepOpinion built an engine that manages and coordinates multiple agents seamlessly allowing them to communicate with each other in a better way, thus ensuring the stability of their engine and avoiding infinite loops (one of the biggest issues of most agent systems).

• Proprietary confidence metrics: DeepOpinion has created a proprietary set of confidence metrics to benchmark LLMs. They use it to automatically guide their agents and allow companies to set confidence thresholds for processes to be fully automated.

• Human-in-the-loop capabilities: Businesses can easily label data with human language and provide feedback, allowing continuous model improvement and adaptation to new tasks. They can also set a minimum confidence level thanks to DeepOpinion’s proprietary confidence metrics.

• A proprietary fine-tuning approach: they built an automatic fine-tuning approach optimized for speed, costs, and minimal training data size.

• A proprietary OCR engine: they developed an OCR orchestration layer that uses multiple top-notch OCR systems to get the best accuracy out of them.

• No-code automation: A no-code interface enables non-technical users to design and deploy automation workflows, making them accessible to a wider audience.

A great go-to-market strategy

To build a world-leading AI company, great tech is not enough. Today, no AI company has a tech moat that is completely overwhelming. Open AI is only 6 months ahead of the competition (if it’s even still the case). So we believe that on top of a leading tech approach, what will make the difference between AI companies and allow a world leader to emerge will be their go-to-market strategy.

DeepOpinion chose a verticalised play by focusing first on Insurance and Financial Services which are two of the industries where unstructured data is the most present. By being vertically focused, DeepOpinion can provide greater ROI to its clients, seamless integration, and proven use cases.

On top of that, they developed a very organized strategy to land and expand great clients all across the world from Europe to the UAE. This strategy allowed them to grow significantly in the past year and land tier-one enterprise clients like Uelzner, Siemens, and many others.

To execute this strategy, they recruited top-tier talents from the previous generation of automation companies with former top sales & first employees at UI Path, Kofax, or BluePrism.

A top-tier team

The founders of DeepOpinion Stefan & Stefan (yep two Stefans 😅) are serial entrepreneurs who have already built an eight-figure business and have been working together for more than 15 years.

They successfully attracted top-tier talents including the former “Head of Speech” of Synthesia, the 2nd employee of BluePrism who grew it from 4 to 100M$ ARR in the US, etc.

We believe that they are building the best team to tackle this challenge.

The round

We are happy to co-lead this 11M€ Series A with Alpha Intelligence Capital which specialises in investing in AI companies across the world. Together, we will try to support DeepOpinion in its very ambitious vision of becoming the world-leading actor of intelligent automation, a $13.4 Billion market that should reach $50 Billion by 20230.

We will particularly support the DeepOpinion team towards their US expansion where we already see a lot of opportunities and a bright future.

We see a future where knowledge work automation is powered by LLMs, multi-agent systems, and a platform that brings all of the necessary components together in a cohesive, user-friendly package. With its forward-thinking approach, DeepOpinion is poised to lead the next generation of automation technology, and we’re excited to be part of that journey.

Abel, Olivier & Chloé