Celonis

Context data. Not a new concept. But it's suddenly clear why it matters. Enterprises have been building AI strategies around better models, better algorithms, better infrastructure. But none of that works without one thing: understanding how their business actually runs. The processes. The exceptions. The rules that nobody wrote down but everyone knows. The knowledge that lives in people's heads. Marina F Bellini from Mars on our Leading Indicators series: "It's not about AI creating insights. It's AI acting on those insights. And for you to act, you really need to understand the context." We're just starting to operationalize this. But this is the difference between AI projects that deliver and AI projects that stall.

In the telecom boom, infrastructure companies built the pipes. Netflix and Amazon built on top of them. They took the prize. Same arc with AI: foundation models are becoming the pipes. Whoever owns the infrastructure rarely captures the most value. That goes to the companies that use it to change how people live and work. The next decade's winners will be the ones that use the AI models to rewire enterprise operations. That's the real value. That's what we're building at Celonis.

Want to make AI work for your company? See how the creator of Ozempic did it. https://lnkd.in/eQMecZA4

They say the “All-you-can-eat era of AI” is over at Accenture — and I have thoughts on it. You might have seen 404 Media’s article (then picked up by Melissa Rosenthal at State of AI) that said the firm that once made AI use a condition of promotion is now capping central funding of Claude token spent at $29 a month per employee. Everything above that needs to be charged to a specific client project or team budget, meaning usage needs to be justifiable (and authorized). Now, you could shrug and say: “Not surprised. Seen it before.” ➡️ Uber burned through its 2026 AI coding budget by April and and set strict monthly per-tool caps. ➡️ Microsoft has given every division a token budget and shifted simpler workflows to cheaper default models. And it is indeed a rather sensible strategy, especially when some workers were reportedly racking up to $1,000 in token costs per month. But to me, the really interesting part of the story is this: If we focus on limiting token consumption alone, we’re blaming teams for high bills when we should be looking at how we build AI systems. As long as AI needs to do ‘reasoning gymnastics’ (e.g. read SOP handbooks, evaluate policy edge-cases, and “figure out” operational rules from scratch) to answer even a simple question like “Can we expedite Order #402 for Customer X?”, no token budget in this world will suffice. CIOs might say now that having a data mesh is the answer. But that still means AI is inferring business logic from static data tables rather than real process execution. Worse, the agent might STILL recommend the wrong action if it doesn’t understand all interdependencies. (Don’t get me started about what that looks like if several agents run a process and cross-check each other’s work, the $$$!) So to me, there’s only ONE valid solution (and a certain #MIT professor named Devavrat Shah might agree with me here 😉). If you want to know what that is, check below. 👇 [Image credit: State of AI]

Europe starts more VC-backed tech companies than anywhere else. Only four of the world’s top 50 are European. Plenty of talent. Not enough scale. I wrote about four ways to fix that for World Economic Forum: One rulebook for scaling up, later-stage capital, sovereign infrastructure, and AI that makes it past the pilot. 🔗 https://okt.to/M8p1Kd

Boy, I wish this AI playbook existed back in my days at Merck I fell into every trap you can think of when my team and I ran the first AI pilots across GBS: -We worked for months on use cases that didn’t even need an agent. -We ignored our existing tech stack to build from scratch. -We confused "time saved" with actual money saved. -We completely underestimated the human side of adoption. Classic build the plane while flying it. If you’re in the middle of an AI rollout and want to measure P&L impact (not just vanity metrics), I highly recommend giving this book a read (link in comments 👇).