Scaling AI agents without making a mess: simple, with Agent Orchestration

AI without governance is a disaster in slow motion

The biggest problem with AI agents is that they do a fantastic job when everything goes according to plan, but grind to a halt the moment something out of the ordinary happens.

Bernd Ruecker, co-founder and Chief Technologist at Camunda and someone who has spent years building large-scale automation platforms for enterprise environments, has a great and relatable anecdote about this: "You order something online, the payment goes through, the confirmation lands in your inbox and the product gets delivered. Everything runs smoothly up to that point. But the moment you want to return the product, things go wrong. The system doesn't recognize you, so you call customer service. They transfer you to another department, which sends you back to square one. End result: you're stuck with a product you're not happy with."

Bernd uses exactly this example to expose a fundamental problem: the so-called happy flow works fine, but the moment something runs just slightly differently, the system comes to a standstill.

This is not just a customer service problem, it's an architecture problem. Companies build processes that work for the standard situation, but nobody is steering the complete customer journey. Every department optimizes its own piece, while the whole falls apart.

AI won't fix it, orchestration will

Many companies believe AI agents will solve this problem for them. But Bernd is clear on this: "AI agents locked inside siloed systems without an overview of the whole, stay superficial." The result of these (expensive) agents? Lots of experiments, little scaling and no real ROI. He calls it the automation ceiling himself: you hit a plateau because the underlying structure is missing.

The problem is not that AI agents aren't smart enough. The problem is that they operate in silos like your CRM, ERP and customer service, without anyone overseeing the whole. Because if your systems aren't integrated, you're working with legacy systems and there's no end-to-end overview, you can keep building AI agents all you want. But all you'll create is chaos and a collection of disconnected islands that know nothing about each other.

BPMN as the backbone of intelligent workflows

The solution lies in combining fixed process structure with the flexibility of AI. Bernd calls that the and-and approach: "It's not about either deterministic or dynamic. It's about both." BPMN (Business Process Model and Notation) is the backbone in this: it defines which steps a process goes through, who or what is responsible and when an agent needs to bring a human into the loop. BPMN isn't the complete solution, but it is the structural layer on which you build your orchestration.

That is exactly what AI agents need to function well: context. Without that structure, an agent improvises and makes decisions without seeing the bigger picture. This leads to wrong decisions, errors and actions that don't connect. With orchestration, every agent knows its role, which systems it can touch and where its boundaries lie.

Less magic, more control: and that's good news

Agentic orchestration sounds complex, but Bernd breaks it down clearly: "You start small: one automated process, one agent that becomes part of it. From that foundation you build further. Departmental agents. And eventually a network of collaborating agents that runs end-to-end through the organization."

That requires more than just the right technology. It requires a different way of working: new processes, clear governance and collaboration between business and IT. Camunda is built to be technology-neutral and auditable, precisely for organizations that value control over chaos.

Whoever organizes it well, wins

Bernd's core message is as simple as it is sharp: AI only becomes truly valuable when you orchestrate it. Success isn't determined by AI itself, but by how well you organize your AI. And that starts with clear workflows, reliable data and systems that talk to each other.

Curious about how your organization takes that step? We're happy to think it through with you.

-- This blog is based on the keynote by Bernd Ruecker, co-founder and Chief Technologist at Camunda, at IncentroCon Agentic '26.