Imagine a situation where a hospital operates without coordination or context. A patient arrives in the emergency room, explains their situation, gets assessed, and is sent to orthopedics.
The orthopedic team has no idea what the emergency room has already documented. The patient explains everything again. Next, they’re sent to a radiologist who has no context about what they’re looking for.
Eventually, the patient becomes the source of truth between disconnected departments, carrying their own medical information from room to room because no one else has the full picture. While the hospital’s care itself may be excellent, the system is fundamentally broken.
Even though this story may sound unrelated to automation, this is precisely what most organizations are doing when they deploy AI agents today.
For example, an organization may have an AI-based intake agent, a scheduling AI agent, a documentation AI agent, and an assessment AI agent, each operating on a single, isolated task. Meanwhile, none are sharing context or coordinating decisions, even related decisions. Without end-to-end management, deploying more agents only worsens fragmentation, leading to an automation ceiling.
It’s no surprise that recent research shows 71% of organizations have deployed AI agents. Yet last year, only 11% of those agentic AI initiatives actually made it to production. That gap between experimenting with AI and scaling it to deliver business value represents billions of dollars in unrealized potential (often sitting in proof-of-concept purgatory). Most often, agentic AI initiatives fail due to a lack of coordination as described above.
The automation ceiling explained
The automation ceiling occurs when individual AI agents stop delivering compounding value, and fragmentation generates chaos. This typically occurs when teams across marketing, sales, support, and finance each build their own agents using their preferred tools and frameworks.
At first, this strategy might demonstrate efficiency gains within each departmental silo, or at least generate considerable activity. Then the questions start: Which agents have access to customer data? How much are we spending on LLM use? Can we prove this decision followed proper approval workflows?
But nobody has the answers.
Agents are probabilistic by design, meaning they offer flexibility and adaptive reasoning, but that flexibility creates uncertainty in production. Organizations in regulated industries such as banking, insurance, and healthcare can’t operate within this unpredictable or uncontrolled system.
When an agent makes a decision that affects someone’s financial status or health coverage, leadership needs to demonstrate that the appropriate oversight was in place. However, when disconnected agents operate in silos without shared oversight or visibility, that demonstration is nearly impossible.
For many organizations, there is also a trust problem with AI agents. For example, an expense processing agent might run well for weeks, then approve a $12,000 team-building expense for a three-person department without flagging it, while rejecting a legitimate $200 software license as suspicious.
The worst part is that nobody can fully explain why. The reasoning happened inside the model. When operating otherwise capable agents, organizations need bounded risk, clear escalation paths, audit trails that explain decisions in business language, and the ability to correct mistakes quickly.
Why agents fail at scale
Most agent frameworks were created for short-lived reasoning loops, from the initial prompt to the final result, all measured in seconds or minutes. Yet, enterprise business processes look nothing like this.
A loan application waits weeks for document uploads, credit bureau responses, underwriter reviews, and committee approvals. An insurance claim pauses for repair estimates, medical assessments, and fraud investigation. A new-hire onboarding process coordinates across HR, IT, compliance, and facilities on entirely different timelines.
This becomes more complex when the process involves multiple agents operating without a shared understanding of what the others have done. Agents could duplicate work, conflict with one another, and make inconsistent decisions about the same customer or data. More autonomy only worsens this type of problem.
Break through the automation ceiling
Breaking through the automation ceiling requires coordinating AI agents, humans, and systems—also known as agentic orchestration. The orchestration layer determines how agents, existing systems, and employees work together across a complete business process, including the order of operations and when a human needs to step in. Agents retain their reasoning capabilities, while the orchestration layer enforces boundaries, ensures the process can run as long as necessary, handles error recovery, manages human escalation, and produces the audit trail for compliance.
Orchestrating agents effectively lets teams incrementally expand AI’s autonomy as they gain confidence in AI agents’ decisions. That means agents can start by analyzing and recommending, while humans review and decide.
Take an insurance claims process. An AI agent reviews a new claim, pulls the supporting documents, and proposes a settlement. Initially, every recommendation goes to a human adjuster with the orchestration layer enforcing a policy check and mandatory human approval above set payout thresholds.
Over time, as the agent’s recommendations prove consistent on straightforward cases, teams can expand AI’s autonomy, while continuing to route complex, high-value, or suspicious claims to a human.
3 real questions leaders must ask
AI agents at scale require organizations to ask a key question: How do we use them without creating coordination problems? To answer it, leaders should demand three things from the orchestration layer.
1. Do we have enforceable controls between an agent’s reasoning and its action? If an agent decides to issue a refund or change a customer record, the process should require a policy check or human approval before the action executes.
2. Can we capture a complete audit trail? This must be automatic so that when a regulator asks why something happened, the answer should be one query away.
3. Have we established agent-agnostic coordination? The model or framework a team chooses today might not be the one they’re using in three years. The orchestration layer should maintain process logic and governance while the underlying AI evolves. For leaders ready to move beyond the automation ceiling, getting these three things right is the difference between AI that transforms operations and AI that simply looks impressive on paper.
Jakob Freund is the CEO of Camunda.