Why UiPath Built Its Own Foundational Models

A few months ago, at the UiPath DevCon 2024 held in Bengaluru, co-founder and CEO Daniel Dines announced that the company is developing its own foundational models.

A month later, the enterprise automation and software company announced the new LLMs – DocPATH and CommPATH– to give businesses AI models extensively trained for their specific document processing and communications tasks.

While large language models (LLM) are very good with unstructured data, they are not suited to handle structured data.

The DocPath model developed by UiPath specialises in handling semi-structured and structured data, such as tables with intricate layouts spanning multiple pages, commonly found in documents like invoices and purchase orders.

CommPath, on the other hand, specialises in communication tasks, specifically understanding emails, tickets, and other forms of communication streams.

“Large language foundation models are very broad and great at summarising articles, etc, but we are an automation company. We want to make sure we’re getting the most accurate information out of those semi-structured and unstructured documents,” Mark Geene, senior VP and general manager, product, UiPath, told AIM.

Building Foundational Models

The models’ architecture is based on Flan-T5, an open source model developed by Google. When one is dealing with automation, confidence levels are critical.

Geene disclosed that the error rates on UiPath’s invoice dataset using GPT were nearly 40%. In contrast, the error rate for UiPath’s model is less than 5%.

However, it’s not like UiPath has no use case for LLMs. “For instance, LLMs excel at labelling a data set. When a customer wants to fine-tune a model, we’ll use GPT to do the labelling and training of that data set.”

UiPath trained the models with proprietary datasets, which the company has accumulated over the years. Interestingly, hundreds of UiPath customers also opted to share their data to train the models.

“We don’t use any data without the customer’s permission. Our datasets are built on purchase orders, invoices, and things like that, but we anonymise the data and ensure they are not linked to any individual customer or company.

“So we’ve been able to build a dataset with a lot of variants so we could train for a wide range of use cases,” Geene said.

Preparing for an RPA+AI Agent Future

UiPath recently laid off around 10% of its workforce. This comes at a time when the company is restructuring. In May, CEO Robert Enslin’s sudden resignation caught everyone off guard.

Even though Enslin cited personal reasons, the resignation came amidst disappointing Q1 results and lowered full-year revenue guidance. However, the company is betting big on AI Agents-led automation.

UiPath, as a company, also believes AI agents will play an important role in business processes. In fact, co-founder Dines believes around 80% of what humans do at work will be done by AI in the future.

If the rumblings in the industry are true, AI agents-led automation is the next big thing. Geene believes AI agents present new business opportunities for RPA firms.

“AI agents introduce new ways of doing things. It doesn’t have to do the same things repeatedly 10,000 times. Instead they introduce automation to tasks previously deemed not automatable, expanding the possibilities beyond traditional patterns of operation,” he said.

According to Geene, robot-led automation primarily handles tasks from high repeatability to moderate complexity. More straightforward tasks are increasingly managed by autopilots or task-level coding.

However, as we approach the peak of complexity, tasks become less repeatable. They will require an AI agentic approach to handle more nuanced and variable situations.

For instance, processing payments into an accounting system is repeatable and highly predictable, whereas creating unique sales proposals for each customer falls into the realm of moderate complexity with lower repeatability. AI agents will be able to handle complex, creative or dynamic tasks which software robots can’t.

“By categorising tasks into these quadrants based on complexity and repeatability, we can better select the appropriate tools for each job. This approach allows us to leverage automation where feasible and deploy agents where human judgement and adaptability are essential,” Geene said.

UiPath’s $35 Mn Investment in Holistic

To enable AI agents, earlier this year, UiPath invested about $35 million in AI startup Holistic. The startup, headquartered in France, is developing a multi-agent AI system designed to collaborate towards achieving goals within specific environments.

Their objective is to craft AI agents that exhibit higher performance, flexibility, and the ability to tackle more intricate challenges compared to current models.

“We view this partnership as a strategic investment that positions us at the forefront of advancements in agent-based processing, enabling us to automate a broader range of use cases,” Geene said.

The post Why UiPath Built Its Own Foundational Models appeared first on Analytics India Magazine.

Follow us on Twitter, Facebook
0 0 votes
Article Rating
Notify of
Inline Feedbacks
View all comments

Latest stories

You might also like...