🤖 AI

On the Hybrid Nature of ABPMS Process Frames and its Implications on Automated Process Discovery

This paper proposes conceptualizing the "process frame" of an AI-Augmented Business Process Management System (ABPMS) as a hybrid representation of semi-concurrently executed procedural and declarative models, arguing that adopting an open-world assumption for procedural models enables a new approach to automated process discovery.

Original authors: Anti Alman, Izack Cohen, Avigdor Gal, Fabrizio Maria Maggi, Marco Montali

Published 2026-04-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Anti Alman, Izack Cohen, Avigdor Gal, Fabrizio Maria Maggi, Marco Montali

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a highly advanced robot how to run a busy restaurant.

If you give the robot a strict, step-by-step manual (like a recipe), it might be great at making a specific pasta dish, but it will freeze up the moment a customer asks for a substitution or a waiter trips and drops a tray. If you only give it a set of vague rules (like "always keep the kitchen clean"), it might be too chaotic to actually serve food.

This paper explores a middle ground called an ABPMS Process Frame. Here is the breakdown of how it works using a "Restaurant Management" analogy.

1. The Concept: The "Guardrails" vs. The "Recipe"

In traditional systems, you give a computer a Recipe (a procedural model): "First boil water, then add salt, then add pasta." It follows the steps exactly.

In an AI-augmented system, we don't want to give it a recipe for every single tiny movement. Instead, we want to give it Guardrails (a process frame).

Think of the Process Frame as a combination of two things:

The Rigid Rules (Procedural): These are the "non-negotiables." For example: "To make a steak, you MUST sear it before you grill it." This is a specific sequence that shouldn't be messed with.
The Flexible Guidelines (Declarative): These are the "vibes" or general constraints. For example: "Every customer must eventually receive a drink," or "Don't let the kitchen get too crowded." It doesn't say when or how exactly, just that the outcome must happen.

The paper calls this a "Hybrid" approach. It allows the AI to have "Framed Autonomy." This means the AI is free to make decisions and optimize the restaurant (like moving a chef to a different station), but it stays within the "frame" of what is legal, safe, and professional.

2. The Problem: The "Messy Reality" of Data

The researchers wanted to know: Can we use AI to automatically "discover" these rules just by watching how a restaurant has been running in the past?

Imagine watching a video of a thousand shifts at a restaurant. You see people moving, cooking, and cleaning. The problem is that the video is "noisy." People skip steps, they do things out of order, or they repeat tasks.

If you try to write a strict "Recipe" based on that messy video, you'll end up with a manual that is impossible to follow because it's too specific to the mistakes people made.

3. The Solution: "Pockets of Rigidity"

The authors propose a clever trick. Instead of trying to turn the whole messy video into one giant, perfect recipe, they suggest looking for "Pockets of Rigidity."

Imagine looking at the restaurant footage and noticing:

Most things are flexible (people move around freely).
BUT, whenever someone handles a knife, there is a very strict, repeatable pattern of "Pick up knife $\rightarrow$ Cut vegetable $\rightarrow$ Put knife down."

The researchers developed a way to take "vague" observations (the declarative rules) and turn them into "solid" mini-recipes (the procedural rules). They call this mapping. By doing this, they can take a giant, confusing cloud of rules and organize them into a clean, manageable "Frame" that an AI can actually understand and use to make decisions.

Summary: The Big Picture

Instead of building a robot that is either a mindless follower (too rigid) or a clueless dreamer (too vague), this paper provides a mathematical way to build a smart manager.

This manager knows exactly which parts of the job are "set in stone" (the rigid pockets) and which parts allow for "creative freedom" (the flexible guidelines), allowing the AI to run the business efficiently without breaking the fundamental rules.

Technical Summary: On the Hybrid Nature of ABPMS Process Frames and its Implications on Automated Process Discovery

1. Problem Statement

The paper addresses a fundamental gap in the vision for AI-Augmented Business Process Management Systems (ABPMS). While traditional Business Process Management Systems (BPMS) focus on executing predefined models, an ABPMS is intended to exhibit autonomy, reasoning, and optimization capabilities. To achieve this, the system requires a "process frame"—a set of behavioral boundaries that define what the AI can and cannot do.

Current research lacks a formal way to represent these boundaries. Traditional procedural models (like BPMN or Petri nets) are often too rigid and "closed-world" (requiring every possible behavior to be explicitly modeled), while declarative models (like Declare) are highly permissive but can become complex and difficult to interpret for specific, rigid sub-behaviors. The core challenge is: How can we represent a heterogeneous, multi-model process frame that allows for both rigid procedures and flexible, autonomous reasoning?

2. Methodology

The authors propose a hybrid business process representation that combines procedural and declarative modeling paradigms under a unified framework.

Conceptualization (The Process Frame): Instead of a single model, the process frame is defined as a set of overlapping process specifications. Each specification is "local" to a subset of activities.
Formalization (Open-World Assumption & Trace Projection): The authors adopt the open-world assumption for both paradigms. They use trace projection ( $\rho|\Sigma'$ ) to formalize how a global execution trace is viewed by individual local specifications. A global trace is accepted by the process frame if and only if every local specification accepts the projection of that trace onto its own set of activities. This allows specifications to execute semi-concurrently, ignoring activities that do not belong to them.
Automated Discovery Approach: The authors investigate whether common procedural behaviors (sequence, parallelism, XOR, OR) can be discovered from event logs using declarative constraints. They use Deterministic Finite Automata (DFA) to represent the state-space of Declare constraints and a systematic mapping process (creating Petri nets $\rightarrow$ generating logs $\rightarrow$ discovering Declare models $\rightarrow$ validating via DFA) to identify which Declare templates correspond to specific procedural constructs.

3. Key Contributions

Hybrid Process Frame Framework: A formal definition of a process frame as an amalgamation of multiple, potentially heterogeneous, process specifications that synchronize on shared activities.
The "Pockets of Rigidity" Concept: A novel approach to simplifying declarative models. Instead of using complex declarative constraints to describe rigid sequences, one can selectively "plug in" procedural fragments (like Petri nets) into the declarative model to represent those specific, rigid behaviors.
Expanded Declare Mapping: The paper extends previous work by accounting for activity repetitions (loops). It introduces/utilizes advanced Declare templates (e.g., Alternate Succession, Interposition, and a novel Balanced Enablement) to capture complex behaviors.
Discovery Feasibility Analysis: A rigorous demonstration of how declarative constraints can be used to reconstruct procedural control-flow structures (Sequence, AND, XOR, OR) even in non-block-structured and overlapping scenarios.

4. Results

Mapping Success: The authors successfully mapped 14 out of 16 common procedural variants (covering various cardinalities: mandatory, optional, repeatable) to Declare constraints.
Handling Complexity: The methodology proved capable of capturing:
- Sequence Flows: Including optional and repeatable segments.
- Parallelism (AND): Using the novel Balanced Enablement constraint to ensure all branches complete before a join.
- Exclusive Choice (XOR): Using Not Chain Succession to replace traditional Not Co-Existence constraints, which fail in the presence of loops.
- Inclusive Choice (OR): Using Interposition to manage branch completion.
Non-Block-Structured Discovery: The study demonstrated that the proposed method can discover behaviors that are not neatly contained in "blocks," such as overlapping parallel branches and complex exclusive choice paths.

5. Significance

This paper provides the theoretical and practical foundation for the "process awareness" required by AI-augmented systems. By moving away from a single, monolithic model toward a hybrid, multi-model frame, it allows:

Scalability: Complex systems can be managed by composing smaller, manageable specifications.
Autonomy: The open-world assumption provides the "permissiveness" necessary for an AI agent to find optimal paths within defined boundaries.
Practicality: It bridges the gap between the mathematical elegance of declarative modeling and the intuitive, structured nature of procedural modeling, offering a roadmap for the next generation of automated process discovery tools in AI-driven environments.