The enterprise landscape is shifting from task-level AI automation to full process re-engineering. To discuss that evolution, VMblog spoke with Colin Teubner, Field CTO at Camunda, about the limits of bolting AI onto processes that were designed for a world without it, and how the AI intelligence layer of the Camunda platform – ProcessOS – can help organizations re-engineer and continuously improve end-to-end operations.
VMblog: To start off, can you give our readers a brief overview of what Camunda does and the role process orchestration plays in the modern enterprise?
Colin Teubner: Camunda is the enterprise platform for agentic orchestration, which means coordinating AI agents, people, and systems across end-to-end business processes, blending deterministic automation with dynamic AI reasoning in one governed runtime.
Most enterprises have accumulated automation over two or three decades, with RPA, integration platforms, and the workflow built into each system of record all automating a slice of the work. Meanwhile none of them owns the outcome a customer actually experiences. Orchestration is what decides how work moves between them, when to escalate, when to reroute, and when a person needs to make a decision. The rise of AI has added to this complexity and is the reason for Camunda’s focus on agentic orchestration.
VMblog: What is the primary operational problem enterprise leaders face in the market today, and how does ProcessOS help solve it?
Teubner: There is a widening gap between what AI can do and what enterprises get from it. Our research found 71% of organizations using AI agents, yet only 11% of use cases reached production last year, and 85% lacked the process maturity to make that leap. Adding an agent to a fragmented process adds capability and more fragmentation at the same time, so results flatten while costs climb. That is the automation ceiling. Underneath it sits a pace problem that predates AI, because updating processes can take years, and no organization can redesign operations on that clock when capability shifts every few weeks.
ProcessOS exists to compress that cycle, accelerating discovery, design, build, and continuous optimization. We pair it with an engineering team that works alongside the customer’s own people to take a critical process from legacy to AI-native in under 12 weeks, production-ready rather than a proof of concept. Early access is open now.
VMblog: Many companies have already added AI tools to their operations. Why does simply bolting AI onto existing processes often lead to failure?
Teubner: Every process running businesses today was designed for a world without AI; this poses a structural problem for organizations. The handoffs and review points exist because humans coordinate the work. Compared to agents, humans are slow to synthesize large amounts of documents and unstructured data, so attaching an assistant to individual tasks inside that structure usually produces an early productivity spike. But that’s followed by coordination problems. The systems on either side of the automated step never shared context, and the costs of the resulting exceptions and redundant work outpace any gains that were made.
Another challenge that organizations face is seeing success and trying to apply it to everything. The instinct for many organizations, after initial agentic successes, is to make everything an agent, but this is inherently expensive and slow. For example, for a financial services organization processing a payment, the same output should be produced from the same input every time, in milliseconds or less. Adding an agent step means model inference and often several reasoning loops, plus a governance obligation on every run because the behavior is probabilistic. We have customers with process volumes up to 4,000 times a second – that’s hundreds of millions per day. Agents are simply too slow and expensive to participate in such a process. The scalable path is to implement as much determinism as possible and as little agent autonomy as necessary. This makes the agentic steps better too, because the agent has fewer things to reason about and burns fewer tokens per case.
VMblog: How does ProcessOS actually work under the hood to discover, re-engineer, and optimize workflows?
Teubner: ProcessOS starts by reconstructing reality rather than documentation. It reviews operational telemetry, system logs, and existing documentation to build a picture of how a process actually runs, which is often different from the existing documentation. It will also search the web and review government regulations, grounding the effort in best practices and external requirements. From there, you state the outcome you want in plain language and it generates a proposed re-engineered process model, integrations, business rule decision tables, and agent prompts. The outputs are grounded back to evidence from the discovery effort, and scored on measures like cycle time, cost per case, and compliance rate, so a proposed change is backtested against live performance rather than argued about in a workshop. Nothing reaches production without human review of that evidence.
Organizational memory multiplies the benefits. As more processes run on the platform, ProcessOS reuses approved patterns and policies, so the tenth process moves dramatically faster than the first.
VMblog: With AI handling more operational decisions, how does ProcessOS maintain safety, compliance, and governance?
Teubner: Control is another key structural element of Camunda’s agentic orchestration, built into the process model rather than sitting in a policy document or a wrapper around the agent. Two patterns do that work. Outer orchestration coordinates the end-to-end journey across every system and person involved, so an agent participates but cannot skip an approval or act outside the boundaries the process sets. Inner orchestration governs what happens between the agent’s reasoning and its action, so before it sends an email or moves money, a policy check or an approval runs first.
The mechanism matters, because it is what allows us to make certain guarantees about agentic behavior. The agent calls a deterministic process as a tool, so governance executes as part of the agent’s own runtime rather than as a check layered on top, and a misbehaving agent cannot reason its way around a boundary the process enforces. In loan origination for example, the journey is orchestrated from the outside while inside the underwriting agent a policy check and an approval gate run before any decision is acted on. The agent handles large unstructured document inputs and makes reasoned decisions, but the process holds the line for regulations and required checks.
VMblog: Looking ahead, what key question should C-level executives ask themselves to determine if their organization is truly ready for AI?
Teubner: The key question executives should ask is whether they would build their current operations the same way if they were starting today. The answer is almost always no, and the gap between that and the operation they actually run is the size of the opportunity. It also reframes the work, because readiness is not a procurement question about which models to buy. Organizations that overlay AI on legacy operations accumulate complexity faster than they remove it, which is how you end up with more agents but no more throughput. The harder question is who owns the end-to-end outcome, as distinct from the functions it passes through. Often, in large enterprises, nobody does, and that lack of ownership leads directly to stalled pilots.
For anyone evaluating platforms rather than diagnosing their own organization, two questions cut through the category claims. Can it govern agents from both the outside and the inside, since policing the surrounding process while leaving the agent’s own loop unsupervised is only half a guarantee. And, is the process logic legible to the business rather than buried in code, because if only engineers can read what runs the operation, nobody else can verify it.
##
Colin Teubner is Field CTO at Camunda, helping enterprises rethink how they design, run, and improve their most critical business processes. He brings more than 20 years of experience in process automation, spanning enterprise software, consulting, and industry analysis. Colin began his career as an analyst at Forrester Research, advising Fortune 500 companies on business process management and automation strategy, then moved into software, building and taking to market automation technologies that power some of the world’s largest enterprises. Today, he works with Camunda’s largest customers to apply that experience to process re-engineering, agentic orchestration, and AI adoption.
