The Automation You Built Will Hit a Wall. Most Already Have.
Your enterprise spent the last decade automating tasks. RPA bots scraping screens. Workflow engines routing approvals. Rule-based systems processing invoices. Billions invested to make existing processes move faster.
Here is the problem nobody in the vendor ecosystem will tell you directly: you automated the execution of work, not the intelligence behind it. And that distinction is about to become expensive.
When something unexpected happens, your current automation stops and waits for a human. Every time. Because it was never designed to reason. It was designed to follow a script. And scripts break the moment reality deviates from what whoever wrote them anticipated.
Researchers at the University of Tartu, presenting at the 2025 Workshop on AI for Business Process Management, published a blueprint for what replaces this. Their framework, Agentic Business Process Management Systems (A-BPMS), describes a fundamentally new class of enterprise platform. One that does not just execute processes but continuously senses, reasons about, and adapts them without waiting for a human to notice something has gone wrong. This is not an incremental upgrade to your workflow engine. It is a different theory of how operational intelligence should work.
Why Every Wave of Automation Has Left the Same Problem Unsolved
The paper opens with a historical observation that is more uncomfortable than it first appears. Every generation of business process technology has moved work further from human hands without moving intelligence closer to the work.
Paper gave way to spreadsheets. Spreadsheets gave way to workflow engines. Workflow engines gave way to RPA. Each wave made execution faster. None of them gave the system the ability to notice when the execution was wrong, or when the process itself needed to change.
The result is a specific kind of brittleness that every operations leader recognises. Your automation works until it does not. An invoice arrives in an unexpected format. A customer submits a request that falls between two process categories. A regulation changes and nobody has updated the script yet. At every one of these moments, the automation stops and a human steps in. The human fixes it. The automation resumes. Nothing was learned. The same exception will happen again next week.
Agentic BPM breaks this pattern by replacing pre-scripted rules with reasoning agents that evaluate context dynamically. When a situation arises that no existing rule covers, the system reasons through it rather than halting. It makes a decision, executes it, logs what it did and why, and continues. The script is no longer the ceiling of what the system can handle.
The Three Walls Enterprises Hit Before They Can Get There
The paper is direct about why organisations have not already made this transition. Three compounding constraints have blocked it, and none of them are purely technical.
The scripting problem is a maintenance problem. A single accounts payable process in a mid-size enterprise can have hundreds of exception paths. Scripting all of them is expensive, slow, and perpetually out of date the moment anything changes. Teams spend more time maintaining the automation than they would have spent doing the work manually. Agentic systems eliminate this by making the system responsible for handling novel situations, not the team responsible for anticipating them in advance.
The governance problem is a trust problem. CFOs and compliance officers have blocked autonomous AI from touching core processes because there was no accountability structure. When the system makes a bad decision, who is responsible? The paper's answer is the concept of a constitutional frame: a defined set of boundaries within which agents are authorised to act autonomously. The system can reason and decide freely within those boundaries. Outside them, it escalates. This is not a technical footnote. It is the governance architecture that makes board-level approval of autonomous operations possible.
The data fragmentation problem is a perception problem. Agents are only as capable as their situational awareness. Most enterprises have process data scattered across ERPs, CRMs, email threads, and legacy systems with no unified view of what is actually happening at any given moment. An agent operating without this unified view is not reasoning about your business. It is reasoning about a partial, stale approximation of it. The paper's proposed data layer aggregates structured and unstructured operational data, including historical event logs and live process states, into a single grounding layer. Without this foundation, every agentic capability built on top of it will underperform.
The Five-Layer Stack, Explained Through What Each Layer Makes Possible
The paper proposes five interconnected subsystems. The right way to understand them is not as technical components but as a progression of capability, each one unlocking what the next can do.
The data layer is the memory of the system. It stores historical event logs, current process states, past decisions and their outcomes, and all relevant documentation. Without it, every other layer is operating blind. With it, the system knows not just what is happening now but what has happened before in similar situations and what the consequences were.
The process intelligence layer is the analytical brain. It runs descriptive analytics telling you what happened, predictive analytics running simulations of proposed changes before they are made, and prescriptive analytics recommending specific interventions at the individual case level. This is the layer that transforms raw operational data into the kind of insight that currently requires a business analyst and a two-week reporting cycle to produce.
The action layer is where insight becomes change. It executes decisions by interacting with workflow engines, triggering downstream systems, updating records, and sending notifications. Critically, every action it takes generates new process data that flows back into the data layer. The system learns from what it does.
The orchestration layer is the executive function. It coordinates all other subsystems, manages the reasoning process for complex decisions, and determines when to act autonomously versus when to escalate. This is where the constitutional frame is enforced. It is also the layer that can be configured differently for different process types: rule-based for predictable workflows, agent-based for high-variability situations, hybrid where appropriate.
The conversational layer is the interface. It exposes the system to humans through natural language: a VP of Operations asking "why did cycle time increase 18% this quarter?" gets a reasoned, data-backed answer in seconds rather than commissioning a two-week analysis. It also exposes the system to external tools and platforms through standardised integration protocols.
The critical architectural point: you cannot skip layers. Agentic orchestration built on top of fragmented data and no intelligence layer will fail. The stack must be built from the bottom up.
What This Looks Like Inside a Commercial Loan Underwriting Operation
A mid-size bank runs commercial loan underwriting across a process that has not fundamentally changed in fifteen years. Application received. Documents routed to analysts. Credit checks triggered. Risk scored. Exceptions escalated to senior underwriters. Decision communicated. Total cycle time: seven to fourteen days. Human bottlenecks at every escalation point. Senior underwriter time consumed by cases that could have been decided algorithmically.
With A-BPMS deployed across this workflow, the operation changes at every layer.
The data layer aggregates every historical loan application, decision, outcome, and exception into a unified event log. The system learns that 73% of applications from a specific industry segment with a certain revenue profile resolve identically. It learns which application characteristics predict escalation before the analyst has opened the file.
The intelligence layer runs a continuous simulation of the current process. It identifies the three highest-cost bottlenecks, quantifies them in analyst hours and cycle time per quarter, and surfaces them to operations leadership without anyone asking.
The orchestration layer stops making recommendations and starts making decisions. Straightforward applications are processed and decided autonomously within the constitutional frame defined by risk, legal, and compliance leadership. Complex cases are routed to the right senior underwriter with a pre-assembled analysis dossier, not a raw document pile.
The conversational layer lets the head of underwriting ask the system directly: "Show me every application where we deviated from the recommended path in the last 90 days and what the outcome was." The answer returns in seconds.
The measurable outcome: cycle time drops from fourteen days to under 48 hours for 60% of applications. Analyst capacity shifts from data gathering to judgment. Routine decision error rates drop. The operation has not automated a process. It has redesigned it around intelligence.
How to Sequence This Without Betting Everything on a Single Deployment
This transition does not happen in a single programme. The paper's architecture implies a clear sequencing, and rushing any stage undermines what follows.
Build the foundation before anything else (Months 1 to 3). The goal in this phase is not to deploy agents. It is to establish the data and intelligence layers that make agents viable. Identify two or three core processes where execution data is already being captured digitally. Deploy process mining to run descriptive analytics on them. You will find waste you did not know existed. Quantify cycle times, exception rates, cost per case, and deviation from designed flow. This becomes both your baseline and your business case. A serious process mining engagement for one process should surface identified waste that exceeds its cost within the first discovery sprint.
Introduce autonomy within one department before scaling it (Months 4 to 9). Select a pilot process with high volume, clear performance metrics, and manageable compliance risk. Before any agent touches a live case, your legal, compliance, and operations teams must agree on the constitutional frame: what decisions agents can make autonomously, what requires human review, and what escalates immediately. Start with intelligent triage rather than autonomous decision-making. Have the system route cases to the right executor rather than make final decisions. Build organisational trust in the system's judgment before expanding its authority. Measure agent-handled cases against human-handled cases on accuracy, cycle time, and exception rate.
Move to enterprise-wide infrastructure only when the foundation is proven (Months 10 to 24). Unify the data layer across departments. This requires integration with your ERP, CRM, and systems of record, and it is where you will encounter the most organisational resistance because data ownership is political. Expand orchestration patterns based on process complexity: sequential for predictable workflows, adaptive for knowledge-intensive processes, mesh patterns for high-variability environments. Deploy the conversational layer to operational leaders. Establish a process governance council to own the constitutional frames, review agent performance quarterly, and approve expansions of autonomous authority.
The organisations that sequence this correctly will have a compounding advantage. Each process that moves onto the agentic layer generates data that improves every subsequent deployment. The system gets better at understanding your business the longer it operates within it.
The gap between rule-based automation and agentic process management is not a software upgrade. It is a different theory of what enterprise operations should be capable of. The component technologies exist today. The architectural blueprint now exists. The question is which organisations build on it deliberately and which ones discover it reactively when their process infrastructure becomes a competitive liability.