Blog

Celonis, the German process mining startup with a $13 billion valuation, is taking a varied approach to AI. Like many enterprise software companies, it has been using machine learning models for a number of years, but with the rise of generative AI, it’s adding a copilot feature, announced today at Celosphere, the company’s customer conference taking place this week.

The company helps customers understand how work flows through different processes in a company, using software to find inefficiencies in the flow, something traditionally done by high-priced consultants. Last year it introduced a new feature that lets customers look at multiple processes by displaying them on a subway-style map.

This year, with the rise of generative AI, they are adding Celonis Copilot, a feature that sits beside the subway map and lets users ask questions about what they see.

Company CEO and co-founder Alexander Rinke says the company built Copilot on top of the OpenAI API. It had been something they had considered for some time, dating back to GPT-2, but when it really took off this year, they decided to build it into the product.

“It required quite a lot of orchestration and the right input and prompting and vector database work. So it’s not like it was super easy, but we definitely accelerated our investment in that area because we just really saw the potential,” Rinke told TechCrunch.

Beyond Copilot, the company is trying to help customers make the data in Celonis accessible to a large language model inside their companies, while also making it easier for other third party partners or customers to build applications on top of the process data stored on the Celonis platform. Rather than try to provide an LLM, they are focusing on how to process the variety of data types being tracked inside Celonis, which can be challenging for a large language model, especially when each customer might have different ways to describe elements of the same type of process.

It’s a complicated problem, so the company decided to take a multi-pronged approach to solve it. For starters, they are giving their customers a standard and structured way to process the data inside Celonis, what they call a process data model.

“We’re launching this process data model, so that customers can combine all of their processes and scenarios in one view, so it’s just naturally connected,” he said. That should make it easier for large language models to understand this data because it’s combined into a single entity.

The second entity involves defining the different elements of the process such as what you do mean by on time or late for an invoice, for example. “We’re taking all of the knowledge that we’ve gathered over the many, many years that we’ve been doing this to define these business definitions,” he said.

Then there is a whole API layer on that to expose it to the Celonis ecosystem to build applications on top of that, or expose the data to LLMs.

The power for the ecosystem, however, comes when you combine the process data model and the dictionary of process definitions to build what they are calling a process intelligence graph, which exposes connections between the different types of data.

“We’re doing this to structure process intelligence in a business across systems and departments in one connected product,” he said.

“And basically it gives you a common language to describe processes across the company, and is completely independent of the systems underneath. And this really enables a lot more value for customers. faster time to value, and also sets us up to build a network and a platform,” he said.

It also makes it easier for customers or third party partners to use the data in their own large language model implementations.

These products are mostly in private release for now, as they build them out and test them with customers, but should be released some time next year

Celonis has raised $2.4 billion, per Crunchbase, and was valued at $13 billion when it raised $1 billion in October 2022.