Enterprise AI Agents Are Everywhere, but Few Reach Production

AI agents have spread across the enterprise faster than most organizations can move them into production. Camunda’s 2026 State of Agentic Orchestration & Automation report found that 71% of surveyed organizations already use AI agents, yet only 11% of agentic use cases reached production over the previous year.

Nearly three-quarters said their ambitions for agentic AI remain ahead of what they can operate today. Most agents still work as assistants, handling tasks such as summarization and information retrieval. The obstacle is often integrating agents into processes that require clear ownership, oversight, and accountability.

Practitioners from Audi, Danica, and Provinzial discussed the challenge at CamundaCon 2026, and Camunda is responding through its agentic orchestration platform, which coordinates agents, people, and systems across end-to-end processes.

The Production Gap Begins in the Process

The report from Camunda points to process readiness as the first barrier. Survey respondents said their business processes already span an average of 50 endpoints, from ERP and CRM applications to automation tools, devices, and AI systems. That footprint is expanding by 14% each year, giving every new agent more dependencies to navigate.

Camunda found that 85% of organizations have not reached the process maturity required for agentic orchestration. An agent may successfully classify a document or recommend an action, but its value depends on whether the surrounding process can use its output.

Half of respondents said uncontrolled agentic AI could make poorly implemented processes and automations worse. An agent can speed up one activity without removing delays elsewhere in the workflow. Its recommendation may still wait in an approval queue, or its classification may pass into a disconnected application that slows the next step.

Enterprises therefore cannot close the production gap by adding intelligence to fragmented work. They must first make the process stable enough to absorb it.

Agent Islands Break at Process Boundaries

At CamundaCon 2026, practitioners described the same obstacle from different angles: Agents can work on their own, but production requires connecting them to the full process.

Audi has agents developed by different business units, but they are not yet coordinated across the processes they support. Jonas Viehof, a consultant at Audi, said the process design team should define how those agents fit into the wider workflow. Audi currently uses an agent to read incoming customer messages and route them into the process. It does not send replies, and Viehof said a human would remain in the loop for that step initially.

Lee Kiew Seng, head of orchestration and integration at Danica, said the insurer had spent 12 months building the foundations for agentic work. Its next step is to realign stakeholders and develop a stronger implementation plan.

Dr. Simon Eisbach, head of AI and automation platforms at Provinzial, said employee-assistance agents moved faster because human oversight was built into the process from the start. That made them easier to introduce under regulatory constraints.

Together, the examples show that the production gap is partly an ownership problem. Individual teams can build agents, but the enterprise still needs one operating model for assigning roles, approvals, and accountability across the full process.

Camunda Orchestration Moves Agents Into Production

Camunda’s platform gives agents defined roles inside live workflows. The process determines when their output can move forward, when a person must review it, and how the organization records the resulting action. This allows an agent to contribute judgment without taking responsibility for the entire process.

ProcessOS builds on that platform by addressing how the process is designed. The early-access product uses four agents to discover how work currently happens, redesign the workflow, generate the assets needed for deployment, and improve performance after launch. Human reviewers must approve proposed changes before they enter production.

Camunda tested the approach on its own quote-to-cash process. The company reports that cycle time fell from 115 to 80 days, while manual touchpoints dropped from as many as 80 per contract to two or three.

The case shows what Camunda means by moving agents into production: redesigning and operating a complete process around controlled agent participation.

What This Means for SAPinsiders

• Human oversight becomes a path to scale. Keeping people at selected decision points lets enterprises deploy agents without granting full autonomy. Those review points also generate evidence for where controls can later be reduced safely.
• Process ownership becomes an AI investment decision. Organizations may gain more from funding process mapping and governance than launching additional pilots. The team controlling end-to-end workflow design will increasingly determine which agents scale.
• Process redesign becomes the real ROI test. Camunda’s internal results suggest value comes from changing the full workflow, not accelerating one isolated task. Enterprises should measure cycle time and handoffs alongside agent performance.