Earlier this week we saw Celonis announce the latest development in its process mining capabilities, its Process Intelligence Graph platform. Celonis’ shift to object centric process mining in recent years has allowed the company to develop a knowledge graph, which essentially maps an organization’s ‘process knowledge’, combined with the vendor’s knowledge of organizational operations, and creates a ‘digital twin’ of how a business is running.
Speaking with diginomica ahead of the announcement, Divya Krishnan, VP of Product Marketing at Celonis, said:
What’s really needed is this holistic end-to-end understanding of: how did the processes actually run? How did they get executed? These underlying systems also have a different means of engagement. They don’t play well together. They have their own vocabulary for how they record information about activities and events that are taking place. And they weren’t built to facilitate easy information sharing, from one tool or one system to another.
But for organizations to really be able to improve process performance, there is this need for a shared language to be able to go in and achieve the enterprise goal.
But what’s potentially more compelling about the announcement is how Celonis sees this playing into their customers’ generative AI plans. Speaking with Cong Yu, Celonis’ VP of Engineering (AI/ML), he explained that the Process Intelligence Graph could enable enterprises to have better control, and understanding, of what data they use for their AI models, providing greater governance. The argument being that this could reduce fears of ‘black box’ AI development and provide a level of traceability and governance, which will become more critical as AI regulations develop.
Because of Celonis’ object centric approach to process intelligence, the theory is that organizations will have a better approach to ‘master data management’, in that no matter what underlying systems are being used, there will be clearly defined ‘dynamic objects’ flowing through the enterprise, using a shared language within the Process Intelligence Graph.
I think generative AI has a lot of promise. You can improve workflows, get faster insights, and modernize everything that you do. Generative AI uses Large Language Models, which understand everything on the web, but they’re not built for customers specific data and knowledge, right? In order to ground these models, to prevent hallucinations and staleness, you really need to understand the process data and process knowledge…for these models to really work.
The one solution, probably the only solution, to help these models ground themselves in the right way, customized to the customer, as they choose the customer’s data and knowledge, is through this Process Intelligence Graph.
Yu pointed to OpenAI’s recent comments around ‘expanded knowledge’. OpenAI is responsible for kick starting the generative AI hysteria amongst buyers, following the launch of ChatGPT late last year. When OpenAI is talking about expanded knowledge, it’s focused on going deeper into industry specific data that isn’t readily available on the internet. Yu said:
Expanded knowledge is an acknowledgment that the world’s knowledge is not enough for your enterprise, right? You need to have expanded knowledge, which is exactly what Celonis provides through this Process Intelligence Graph.
Yu explained that the Process Intelligence Graph harmonizes and organization’s data into one place, so that the data can talk to each other, understand each other and are linked through this object centric model. Then on top of htat, you have the knowledge layer, which tells organizations about the important KPIs and metrics, where the opportunities are, where bottlenecks lie, and where there are efficiencies to be gained.
And as we know, enterprises are primarily going to be concerned with creating their own customized AI models, rather than just pulling information in from a model like ChatGPT. Yu explained why this knowledge layer and the desire to customize generative AI models is useful, centered around the idea of retrieval augmented generation. He said:
You really want to focus on the return on investment of your AI efforts. What you want to do is, you want to tailor the model to the use cases you care about. To the knowledge you care about.
And you want to do that in the most efficient way. Whether that’s cost wise or whether that’s speed wise.
Retrieval augmented generation is a lot more lightweight. You don’t need a lot of training data. You can basically extract the model to call some external capabilities.
Using the example of LLMs often failing when it comes to arithmetic. Yu added:
Instead of trying to force an LLM to be good at math, you just tell the LLM that whenever it sees a math formula, it should call a calculator and let the calculator do it for you.
The Process Intelligence Graph is essentially a sophisticated calculator that operates on your knowledge and assists the LLM to do the right thing for you. It’s the fastest time to value and the lowest cost from an enterprise perspective, and it’s probably going to generate better results as well. Because you have more control.
Essentially, what Celonis is pitching is providing enterprises with a high degree of control over their data. Master data management isn’t anything new, but in the context of AI it could become critically important. And Celonis sees its Process Intelligence Graph as a ‘next generation’ approach to doing this. For instance, if a company has five different KPIs all called ‘on time payment’, with the same description, this would be problematic in the context of AI models. Yu said:
Humans are going to get confused by this. Humans are not going to know how to pick the right one, so models are going to get confused too.
AI models can do things faster than humans, but they’re not smarter than a human. The data you have, the knowledge you have, is to your advantage, but you also need to transform that to understand how AI can really take advantage of this.
The master data you have is actually the core foundation, the quality of the master is actually the core foundation for you to build all these AI applications on top.
We have a unique advantage in that we’ve always been looking at this Process Intelligence Graph layer, even though we didn’t talk about it. This is the first time we are formalizing it, but we have all these connectors and extractors, taking all the data, and we have all these formulas and models we built on top. We’re really building the future for the next decades of power. We are also pretty open, we’ll have API’s to open this out for the business themselves to build applications, or their partners.
As the US and the EU, amongst other nations and regions, slowly put together their regulatory thinking, Celonis sees this approach as becoming even more critical – for compliance, governance and traceability. Yu said:
My view is that this is not a nice to have, this is a must have. Because, as the regulations come along, and all these safety issues become more important, you have to figure out a way of doing it. Whether that’s with the Process Intelligence Graph or fine tuning your own model, you have to do it. And the question is: what is the most efficient and fastest way for you to do it? And I think working with Celonis is the fastest way.
What I don’t know is what these regulations are going to look like, and therefore it’s going to impact how the AI models are being built and being applied and being generated.
The Process Intelligence Graph is definitely designed to help ease the regulation risk, because now it’s not like the knowledge is just inside a model or a black box, where you have no idea how it’s applied. The knowledge is outside. You can control the quality of the knowledge, you can audit the knowledge that you have.
It’s a compelling pitch from Celonis and one that will play well to enterprises that have anxiety around controlling their data when using large language models. However, this is still early days and the proof will be in the use cases. And Celonis is not the only vendor out there arguing their platform is the best approach for safely adopting generative AI models. But I do think Celonis customers, ones already getting control of their processes with the platform, will look at the Process Intelligence Graph with interest. If we see use cases of this emerging at Celonis’ user event this time next year, that will be telling.