Agentic Process Automation in 2026: Why Most AI Agent Projects Do Not Reach Production
Agentic Process Automation in 2026: Why Most AI Agent Projects Do Not Reach Production
Camunda’s 2026 State of Agentic Orchestration and Automation report surveyed 1,150 IT leaders. 71% said their organisation is using AI agents. 11% of agentic use cases reached production over the preceding year.
Gartner predicted in June 2025 that more than 40% of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.
Those two findings describe the same problem from different angles, and neither of them is a statement about model quality. Agentic process automation stalls between pilot and production for reasons that belong to process architecture: undefined escalation paths, absent human checkpoints, missing audit trails, and no deterministic boundary around a non-deterministic step.
The agentic process automation gap, in numbers
The Camunda survey puts additional detail around the gap. 73% of decision makers report a difference between their vision for agentic AI and current reality. 85% say they lack the process maturity to close it. 79% nonetheless plan to increase automation spend, by an average of 20% over two years.
Respondents also reported managing an average of 50 endpoints across their business processes, a figure growing at 14% year over year.
That last number is the one worth sitting with. An agent introduced into a process with fifty integration points is not a self-contained capability. It is a participant in a distributed system, and the failure modes it encounters are distributed-system failure modes: partial completion, timeout, duplicate execution, stale state, and a downstream service that returns success while doing nothing.
Pilots do not surface those conditions. Pilots run on curated inputs, in a single environment, with the person who built the agent watching. Production does not.
Four process failure modes
The reasons agent projects fail at the production boundary repeat across implementations.
No defined escalation path. The agent handles the cases it can and produces something for the cases it cannot. Without a modelled path for the second category, ambiguous outputs enter the process as though they were confident ones. The failure is silent, which makes it the expensive kind.
No human checkpoint with authority. Many designs include a review step. Fewer specify what the reviewer may do, what information they see, how long the case may wait, and what happens when the review does not occur. A checkpoint without an override path and a timeout is a queue, not a control.
No audit trail at decision granularity. The engine logs that a task completed. It does not necessarily log which inputs the agent saw, what it produced, which tool it called, or why the branch was taken. When a case is challenged six months later, the reconstruction is not possible, and in regulated sectors that alone blocks deployment.
No deterministic boundary. This is the structural one. A process step whose behaviour varies between runs is acceptable when it sits inside a frame that does not: defined inputs, defined output contract, defined timeout, defined compensation when it fails. Where the agent is the process rather than a step within one, there is nothing to test against and nothing to roll back to.
What BPMN still does that an agent does not
The value of a process model in an agentic architecture is not documentation. It is constraint.
BPMN gives you an explicit sequence with defined gateways, so the set of reachable states is knowable. It gives you boundary events, so a step that hangs or errors has a modelled consequence rather than an operational incident. It gives you compensation, so partially completed work can be unwound. It gives you message correlation, so a long-running case can wait for an external event without holding resources. And it gives you a token that either exists in a state or does not, which is what makes an instance inspectable at all.
None of that is displaced by an agent being good at its task. It is the frame the agent runs inside.
DMN carries the parallel argument for decisions. Where an agent scores, classifies, or recommends, the consequential rule applied to that output belongs in a decision table that a business analyst can read and an engine can execute. That separation is what preserves an explainable path from input to outcome when part of the pipeline is not explainable.
Where the agent belongs in the process model
The pattern that reaches production consistently places the agent as a task inside a modelled process rather than as the orchestrator of it.
Concretely: the process is modelled in BPMN. The agent occupies one or more service tasks, invoked with a defined input contract and returning a defined output contract. A gateway immediately downstream evaluates confidence, completeness, or a business rule, and routes low-confidence outputs to a human task. A boundary timer on the agent task handles the case where the model provider is slow or unavailable. An error boundary event handles malformed output. The decision rules that act on the agent’s output sit in DMN, versioned separately.
The inverse pattern, where an agent plans the sequence and calls tools as it sees fit, is legitimate for genuinely open-ended knowledge work and is where adaptive case management thinking applies. It is not appropriate for a claims process, a loan origination flow, or a regulated onboarding sequence, and the cost of discovering that in production is the reason a meaningful share of these projects are cancelled.
Explore Concordia BPM, DNA’s open, AI-powered BPM platform built on open BPMN and DMN standards, with a transparent stack of Java, Angular, and PostgreSQL, and published pricing at €2,000 per month.
Human-in-the-loop as a design pattern
Human-in-the-loop is frequently written into agent designs as an assurance rather than a specification. As a design pattern it has parts.
The checkpoint sits at a step where intervention still changes the outcome, which usually means before the irreversible action rather than after it. The reviewer receives the agent’s output, the inputs it worked from, and the confidence or provenance signal that triggered the review. The reviewer’s authority is explicit: approve, reject, modify, or escalate, with each producing a different modelled path. A timer defines what happens if nobody acts. And the decision is recorded against the case, not just against the task.
Designed this way, the checkpoint also produces the evidence that regulated deployments need. Designed as a review queue with no override semantics, it produces delay and no evidence at all.
The platform landscape for agentic orchestration in 2026
The category has consolidated quickly, and the honest position is that several platforms handle this competently in different scenarios.
Camunda has moved decisively into this space, positioning 8.9 around agentic orchestration with agents modelled through BPMN ad hoc sub-processes, and shipping 8.10 alpha builds that add a processes MCP server and a provided LLM option. For enterprises already committed to Zeebe and cloud-native orchestration, this is a coherent path.
ServiceNow is strong where the data and the workflow both already live in ServiceNow. Its AI agents inherit that boundary. Outside the ServiceNow estate, orchestration across systems of record where ServiceNow is not the master becomes the constraint.
Microsoft Power Automate is the fastest route to a first working agent and remains an accessible entry point for teams inside the Microsoft estate. The governance ceiling arrives at scale, when flows proliferate across business units without a process owner or a shared model.
ARIS remains the modelling system of record in a large number of European enterprises, and its architecture repositories are where the authoritative process landscape lives. AI-assisted modelling extends that role rather than replacing it. The gap for most ARIS users is execution and end-user experience on top of the models they already maintain, which is complementary work rather than a migration argument.
Concordia BPM is DNA’s own platform in this category, built on open BPMN and DMN standards with AI capability in the platform rather than bolted alongside it.
For enterprises still running Camunda 7, the relevant point is that Community Edition reached end of life in October 2025, Enterprise support runs to April 2030, and paid extended support to April 2032. Adding agentic capability to a v7 estate is a different decision from adding it to a greenfield deployment, and DNA maintains a proprietary Custom BPA Automation Taskboard and Cockpit built on the Camunda v7 orchestration engine for clients staying on that path.
A governance checklist before the first agent goes live
Seven questions, each of which should have an answer that points at a model or a log rather than a policy document.
Which process does this agent participate in, and where is that process modelled?
What is the agent’s input contract and output contract?
What happens when the output is low-confidence, malformed, or absent?
Who reviews, with what authority, at which step, and within what time limit?
What is retained per case, at what granularity, and for how long?
What is the rollback or compensation path when the agent acts wrongly?
Who owns the process, as distinct from who owns the agent?
Question seven is the one that predicts outcomes. Agents built by a team that does not own the process almost always stall at the production boundary, because the people with authority to accept the operational risk were not part of the design.
The conclusion that the data supports
The gap between 71% adoption and 11% production is not going to be closed by better models. It closes when the process around the agent is defined well enough that the agent’s failure modes have somewhere to go.
That is orchestration work, and it is the same discipline enterprises have been doing with BPMN for twenty years, applied to a step that behaves differently. The organisations moving fastest are not the ones with the most sophisticated agents. They are the ones that had process models before they had agents.
DNA works vendor-neutrally across ARIS, Camunda, ServiceNow, SAP Signavio, and Power Automate, and recommends what is best for the client rather than what is licensed. See how discovery, design, and automation delivery fit together.
Book a meeting to review where agents fit in your process estate.
About DNA Automation Consulting
DNA Automation Consulting is an EU-based BPM consultancy specialising in process discovery, BPMN design, and automation delivery for enterprise clients in banking, insurance, energy, telecom, and manufacturing. 75% of client engagements last five years or more.
Founded by Ante Gudelj and Nikola Dlaka. Clients include Proximus, Equinor, Merck, and NEXI.
Frequently asked questions
Camunda’s 2026 State of Agentic Orchestration and Automation report surveyed 1,150 IT leaders. 71% said their organisation is using AI agents. 11% of agentic use cases reached production over the preceding year.
Gartner predicted in June 2025 that more than 40% of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.
Those two findings describe the same problem from different angles, and neither of them is a statement about model quality. Agentic process automation stalls between pilot and production for reasons that belong to process architecture: undefined escalation paths, absent human checkpoints, missing audit trails, and no deterministic boundary around a non-deterministic step.
The agentic process automation gap, in numbers
The Camunda survey puts additional detail around the gap. 73% of decision makers report a difference between their vision for agentic AI and current reality. 85% say they lack the process maturity to close it. 79% nonetheless plan to increase automation spend, by an average of 20% over two years.
Respondents also reported managing an average of 50 endpoints across their business processes, a figure growing at 14% year over year.
That last number is the one worth sitting with. An agent introduced into a process with fifty integration points is not a self-contained capability. It is a participant in a distributed system, and the failure modes it encounters are distributed-system failure modes: partial completion, timeout, duplicate execution, stale state, and a downstream service that returns success while doing nothing.
Pilots do not surface those conditions. Pilots run on curated inputs, in a single environment, with the person who built the agent watching. Production does not.
Four process failure modes
The reasons agent projects fail at the production boundary repeat across implementations.
No defined escalation path. The agent handles the cases it can and produces something for the cases it cannot. Without a modelled path for the second category, ambiguous outputs enter the process as though they were confident ones. The failure is silent, which makes it the expensive kind.
No human checkpoint with authority. Many designs include a review step. Fewer specify what the reviewer may do, what information they see, how long the case may wait, and what happens when the review does not occur. A checkpoint without an override path and a timeout is a queue, not a control.
No audit trail at decision granularity. The engine logs that a task completed. It does not necessarily log which inputs the agent saw, what it produced, which tool it called, or why the branch was taken. When a case is challenged six months later, the reconstruction is not possible, and in regulated sectors that alone blocks deployment.
No deterministic boundary. This is the structural one. A process step whose behaviour varies between runs is acceptable when it sits inside a frame that does not: defined inputs, defined output contract, defined timeout, defined compensation when it fails. Where the agent is the process rather than a step within one, there is nothing to test against and nothing to roll back to.
What BPMN still does that an agent does not
The value of a process model in an agentic architecture is not documentation. It is constraint.
BPMN gives you an explicit sequence with defined gateways, so the set of reachable states is knowable. It gives you boundary events, so a step that hangs or errors has a modelled consequence rather than an operational incident. It gives you compensation, so partially completed work can be unwound. It gives you message correlation, so a long-running case can wait for an external event without holding resources. And it gives you a token that either exists in a state or does not, which is what makes an instance inspectable at all.
None of that is displaced by an agent being good at its task. It is the frame the agent runs inside.
DMN carries the parallel argument for decisions. Where an agent scores, classifies, or recommends, the consequential rule applied to that output belongs in a decision table that a business analyst can read and an engine can execute. That separation is what preserves an explainable path from input to outcome when part of the pipeline is not explainable.
Where the agent belongs in the process model
The pattern that reaches production consistently places the agent as a task inside a modelled process rather than as the orchestrator of it.
Concretely: the process is modelled in BPMN. The agent occupies one or more service tasks, invoked with a defined input contract and returning a defined output contract. A gateway immediately downstream evaluates confidence, completeness, or a business rule, and routes low-confidence outputs to a human task. A boundary timer on the agent task handles the case where the model provider is slow or unavailable. An error boundary event handles malformed output. The decision rules that act on the agent’s output sit in DMN, versioned separately.
The inverse pattern, where an agent plans the sequence and calls tools as it sees fit, is legitimate for genuinely open-ended knowledge work and is where adaptive case management thinking applies. It is not appropriate for a claims process, a loan origination flow, or a regulated onboarding sequence, and the cost of discovering that in production is the reason a meaningful share of these projects are cancelled.
Explore Concordia BPM, DNA’s open, AI-powered BPM platform built on open BPMN and DMN standards, with a transparent stack of Java, Angular, and PostgreSQL, and published pricing at €2,000 per month.
Human-in-the-loop as a design pattern
Human-in-the-loop is frequently written into agent designs as an assurance rather than a specification. As a design pattern it has parts.
The checkpoint sits at a step where intervention still changes the outcome, which usually means before the irreversible action rather than after it. The reviewer receives the agent’s output, the inputs it worked from, and the confidence or provenance signal that triggered the review. The reviewer’s authority is explicit: approve, reject, modify, or escalate, with each producing a different modelled path. A timer defines what happens if nobody acts. And the decision is recorded against the case, not just against the task.
Designed this way, the checkpoint also produces the evidence that regulated deployments need. Designed as a review queue with no override semantics, it produces delay and no evidence at all.
The platform landscape for agentic orchestration in 2026
The category has consolidated quickly, and the honest position is that several platforms handle this competently in different scenarios.
Camunda has moved decisively into this space, positioning 8.9 around agentic orchestration with agents modelled through BPMN ad hoc sub-processes, and shipping 8.10 alpha builds that add a processes MCP server and a provided LLM option. For enterprises already committed to Zeebe and cloud-native orchestration, this is a coherent path.
ServiceNow is strong where the data and the workflow both already live in ServiceNow. Its AI agents inherit that boundary. Outside the ServiceNow estate, orchestration across systems of record where ServiceNow is not the master becomes the constraint.
Microsoft Power Automate is the fastest route to a first working agent and remains an accessible entry point for teams inside the Microsoft estate. The governance ceiling arrives at scale, when flows proliferate across business units without a process owner or a shared model.
ARIS remains the modelling system of record in a large number of European enterprises, and its architecture repositories are where the authoritative process landscape lives. AI-assisted modelling extends that role rather than replacing it. The gap for most ARIS users is execution and end-user experience on top of the models they already maintain, which is complementary work rather than a migration argument.
Concordia BPM is DNA’s own platform in this category, built on open BPMN and DMN standards with AI capability in the platform rather than bolted alongside it.
For enterprises still running Camunda 7, the relevant point is that Community Edition reached end of life in October 2025, Enterprise support runs to April 2030, and paid extended support to April 2032. Adding agentic capability to a v7 estate is a different decision from adding it to a greenfield deployment, and DNA maintains a proprietary Custom BPA Automation Taskboard and Cockpit built on the Camunda v7 orchestration engine for clients staying on that path.
A governance checklist before the first agent goes live
Seven questions, each of which should have an answer that points at a model or a log rather than a policy document.
Which process does this agent participate in, and where is that process modelled?
What is the agent’s input contract and output contract?
What happens when the output is low-confidence, malformed, or absent?
Who reviews, with what authority, at which step, and within what time limit?
What is retained per case, at what granularity, and for how long?
What is the rollback or compensation path when the agent acts wrongly?
Who owns the process, as distinct from who owns the agent?
Question seven is the one that predicts outcomes. Agents built by a team that does not own the process almost always stall at the production boundary, because the people with authority to accept the operational risk were not part of the design.
The conclusion that the data supports
The gap between 71% adoption and 11% production is not going to be closed by better models. It closes when the process around the agent is defined well enough that the agent’s failure modes have somewhere to go.
That is orchestration work, and it is the same discipline enterprises have been doing with BPMN for twenty years, applied to a step that behaves differently. The organisations moving fastest are not the ones with the most sophisticated agents. They are the ones that had process models before they had agents.
DNA works vendor-neutrally across ARIS, Camunda, ServiceNow, SAP Signavio, and Power Automate, and recommends what is best for the client rather than what is licensed. See how discovery, design, and automation delivery fit together.
Book a meeting to review where agents fit in your process estate.
About DNA Automation Consulting
DNA Automation Consulting is an EU-based BPM consultancy specialising in process discovery, BPMN design, and automation delivery for enterprise clients in banking, insurance, energy, telecom, and manufacturing. 75% of client engagements last five years or more.
Founded by Ante Gudelj and Nikola Dlaka. Clients include Proximus, Equinor, Merck, and NEXI.
Frequently asked questions
What is agentic process automation?
The use of AI agents that can plan, call tools, and act with some autonomy inside business processes, as opposed to rule-based automation that executes a fixed sequence. In production architectures the agent normally occupies specific tasks inside an orchestrated process rather than replacing the process.
Why do most AI agent projects fail to reach production?
Camunda’s 2026 research reports 71% of organisations using AI agents but only 11% of use cases reaching production, with 85% citing insufficient process maturity. The recurring causes are process-level: undefined escalation paths, human checkpoints without override authority, audit trails that lack decision granularity, and no deterministic boundary around the agent’s step.
Do I need BPMN to run AI agents?
No, but the alternative is an execution model that is harder to test, inspect, and roll back. BPMN supplies defined gateways, boundary events, compensation, and message correlation, which are the mechanisms that make an agent’s failure modes recoverable rather than operational incidents.
What is the difference between agentic orchestration and process orchestration?
Process orchestration coordinates people, systems, and services across an end-to-end process. Agentic orchestration is the same discipline extended to include AI agents as participants, with the additional requirement that non-deterministic steps are contained by deterministic structure.
Where should human oversight sit in an agentic process?
Before the irreversible action, at a step where intervention still changes the outcome, with the reviewer seeing the agent’s inputs and output, holding explicit approve, reject, modify, and escalate authority, and with a timer defining what happens if nobody acts.
What is agentic process automation?
The use of AI agents that can plan, call tools, and act with some autonomy inside business processes, as opposed to rule-based automation that executes a fixed sequence. In production architectures the agent normally occupies specific tasks inside an orchestrated process rather than replacing the process.
Why do most AI agent projects fail to reach production?
Camunda’s 2026 research reports 71% of organisations using AI agents but only 11% of use cases reaching production, with 85% citing insufficient process maturity. The recurring causes are process-level: undefined escalation paths, human checkpoints without override authority, audit trails that lack decision granularity, and no deterministic boundary around the agent’s step.
Do I need BPMN to run AI agents?
No, but the alternative is an execution model that is harder to test, inspect, and roll back. BPMN supplies defined gateways, boundary events, compensation, and message correlation, which are the mechanisms that make an agent’s failure modes recoverable rather than operational incidents.
What is the difference between agentic orchestration and process orchestration?
Process orchestration coordinates people, systems, and services across an end-to-end process. Agentic orchestration is the same discipline extended to include AI agents as participants, with the additional requirement that non-deterministic steps are contained by deterministic structure.
Where should human oversight sit in an agentic process?
Before the irreversible action, at a step where intervention still changes the outcome, with the reviewer seeing the agent’s inputs and output, holding explicit approve, reject, modify, and escalate authority, and with a timer defining what happens if nobody acts.
What is agentic process automation?
The use of AI agents that can plan, call tools, and act with some autonomy inside business processes, as opposed to rule-based automation that executes a fixed sequence. In production architectures the agent normally occupies specific tasks inside an orchestrated process rather than replacing the process.
Why do most AI agent projects fail to reach production?
Camunda’s 2026 research reports 71% of organisations using AI agents but only 11% of use cases reaching production, with 85% citing insufficient process maturity. The recurring causes are process-level: undefined escalation paths, human checkpoints without override authority, audit trails that lack decision granularity, and no deterministic boundary around the agent’s step.
Do I need BPMN to run AI agents?
No, but the alternative is an execution model that is harder to test, inspect, and roll back. BPMN supplies defined gateways, boundary events, compensation, and message correlation, which are the mechanisms that make an agent’s failure modes recoverable rather than operational incidents.
What is the difference between agentic orchestration and process orchestration?
Process orchestration coordinates people, systems, and services across an end-to-end process. Agentic orchestration is the same discipline extended to include AI agents as participants, with the additional requirement that non-deterministic steps are contained by deterministic structure.
Where should human oversight sit in an agentic process?
Before the irreversible action, at a step where intervention still changes the outcome, with the reviewer seeing the agent’s inputs and output, holding explicit approve, reject, modify, and escalate authority, and with a timer defining what happens if nobody acts.
Written by:
Nikola Dlaka
Managing Partner
Share with friends: