Executive Intelligence podcast - Celonis President Carsten Thoma on why enterprise AI needs operational truth before it can deliver

There's an obvious, but growing, disconnect in the enterprise AI market right now. On one side, you've got vendors shipping agents and automation tooling at pace. On the other, you've got CIOs telling us - consistently - that the returns aren't materializing as expected. What’s missing, according to Celonis, is operational context - not more features, not more agents, but a grounded understanding of how your business actually runs before you try to automate it.

That's the thread that runs through this week's episode of the diginomica Executive Intelligence podcast, where I sat down with Carsten Thoma, President of Celonis. Thoma is someone worth paying attention to, given his experience and his success in navigating previous technology disruptions. He co-founded Hybris in 1997, built it into the leading enterprise e-commerce platform, and saw it through to acquisition by SAP in 2013, where he subsequently led the CX division as President. He was Celonis's first external investor back in 2016, joined the board as an adviser, and took on the President role in 2023, where he now leads corporate strategy, platform ecosystem and innovation alongside co-CEOs Alex Rinke and Bastian Nominacher. In other words, he's been through multiple enterprise technology cycles and has strong views on where this one is heading - and where it's going wrong.

Thoma's core argument is that Large language models are very good at language - understanding what a user wants, working with unstructured data, generating code. But when you get down to the system level, where business processes actually run across enterprise landscapes, things get complicated quickly. Processes are spread across heterogeneous systems, the data is cryptic, and the event logs don't mean anything without translation and context. He described how Celonis has built its own transformer models that translate system-level event codes into a process language that can be understood and acted on - what the company calls an operational digital twin. An unbiased, system-agnostic view of how processes actually flow across the enterprise, not how anyone assumes they flow.

Without that grounding, Thoma argues, organizations are flying blind when they deploy agents and automation. They don't know whether a process is suitable for an agent, whether it needs a human in the loop, or whether the underlying process should be recomposed from scratch before any automation is applied. 

‘It belongs to the customer’

This connects to what Celonis has been calling "free the process" - which is essentially a challenge to the vendor lock-in dynamics that are emerging around AI. If vendors build walls around data and funnel customers towards their own agents and automation, the result will always be suboptimal - because those agents lack the broader, cross-system context needed to find the most effective solution. He made the point that core processes in large enterprises run across twenty-five or thirty systems on average. If you're only seeing part of that picture because a vendor has walled off access, your AI is working with an incomplete operational reality. And the data itself, Thoma stressed, belongs to the customer - it reflects their business reality, generated through systems they've already paid for. He said: 

The data that you feed automations, agents and recomposability with is actually the data that the business of the customer generates. The customer already pays for those systems in order to generate and host the data. But the data itself reflects the business reality of the customer. So who does it belong to? For sure, not the vendor.

As we see more often than not, the early days of a technology shift produce a tussle over value capture before the market settles into something more customer-friendly. Cloud went through exactly this cycle, and AI appears to be following the same pattern.

We also got into the SaaS-pocalypse conversation, which - as much as I dislike the term - is resonating with the CIOs in our network. Our own research suggests that while most enterprises aren't about to rip out their SaaS platforms, they are recognizing the ability to develop software internally with AI at lower cost, and they're bringing that conversation to their suppliers. 

Thoma's take was that tools lacking depth, differentiated data or a critical mass of business context are at risk. But he also argued that the large-scale vendors with deep operational knowledge and history will evolve - potentially into what he described as a next generation of business process outsourcing, powered by AI. It's a provocative framing, but if AI agents can increasingly take over defined business functions, the vendor relationship may start to look less like a software licence and more like an operational service.

A new social contract

Where Thoma was at his most passionate during the conversation was on the question of organizational and societal impact. He made the point that while change management is well understood when it comes to redesigning jobs and structures, it doesn't address the people who are navigating changing job roles. His advice to organizations was simple: honesty, and as early as possible. If you don't pursue AI-driven efficiency, you won't be competitive, and then you put more people at risk anyway, he added. 

But he went further than most technology executives are willing to go. Thoma argued that AI has the potential to change what income looks like for some people - but that economic security alone doesn't solve the question of purpose and belonging. What's required, he suggested, is a renegotiation of the social contract. He said: 

General income doesn't solve the state of happiness and stability for individuals. So I think we need to think hard about not only what is the shape of an organisation, but what is the shape of our society that people can find purpose and belonging.

We also touched on Europe's position in the AI value chain. Thoma argued that while the US has the pole position on model development, Europe and Asia have deep industrial and enterprise know-how that creates real opportunity in adoption and applied AI - provided there's the ambition and regulatory alignment to seize it. 

It was a wide-ranging conversation with someone who has the enterprise experience to back up the big claims. You can listen to the full episode now.