Camunda's CEO blames stacks built one function at a time for the agent production gap

The two headline figures rest on different denominators. The 71% counts organizations; the 11% counts use cases ^[1][2][17]. Multiply them and you get 7.8%, a meaningless number, because a firm with forty agent use cases and one in production sits inside the same 71% as a firm with two. On its own, the second figure supports a ratio: 89% of use cases did not reach production last year, roughly eight stalled for every one live ^[14][15]. The column does not state the survey's sample size or method ^[16].

The 11% is also a one-year measure ^[2]. A use case started in the fourth quarter falls into the same denominator as one started in the first, so part of the 89% is a project still inside its first year.

Freund's account of the cause is about wiring. Enterprises built the stack one function at a time: a claims system for the claims team, a procurement system for procurement, an ERP for finance ^[5]. A loan application crosses several of them, and something has to carry it across each handoff ^[6]. In the hospital case he describes, the intake agent captures information the scheduling agent never sees, and the imaging agent finishes its analysis without telling the assessment agent it is done ^[7]. "Adding intelligence to a broken process does not fix the process," he wrote ^[8].

The obvious objection is that Camunda sells end-to-end process orchestration with agentic AI, so the missing layer in the diagnosis is the layer on its price list ^[4]. The precedent is the part of the argument that holds regardless of the product line. Freund wrote that enterprises hit the same wall years ago with robotic process automation, and before that with point-to-point integrations, each wave layered onto processes nobody had redesigned ^[10]. He also wrote that "the uncomfortable truth is that the ceiling was never really about AI" ^[9], and that AI is revealing the structural gap faster and at greater scale than the earlier technologies did ^[18].

The trade-off in front of a buyer this quarter is between funding another agent and funding the redesign that would let the existing agents hand work to each other. The first is cheap and demonstrable, and on an unchanged legacy process Freund says it tends to deliver modest gains, far short of what leaders were promised when they approved the AI budget ^[13]. The second is slow. Its first step is understanding the process as it runs today, which Freund says usually differs from the documented version, and its second is designing the target process around the outcome, asking what it would look like built from scratch with agents available from the start ^[12].

Sequencing decides which number a team reports next year. An agent approved now onto the old handoffs is a candidate for the same 89%; a discovery exercise approved now skips the demo this quarter and yields a redesigned process to put agents into later ^[14].