Autonomy in Business Process Execution: Why We Need First-Class Abstractions for Goals and Normative Frames

Summary

This position paper — derived from Dagstuhl Seminar 25192 AUTOBIZ — argues that framing the autonomy of AI agents is the central challenge for BPM in the LLM era, and that current BPM abstractions are structurally inadequate. Classical process specification languages (BPMN, DECLARE) are operational (they describe how activities are sequenced or constrained). Goals are implicit and normative constraints (obligations/permissions/prohibitions) are typically buried in informal policy documents. For autonomous, goal-oriented agents to operate safely at scale, the authors propose elevating goals and normative frames to first-class BPM abstractions alongside operational specifications.

A normative frame is deontic in nature: it states what agents must, may, or must not do without prescribing procedural execution. This leaves room for agents to synthesize their own operational strategies, maximizing autonomy within deontic bounds. The authors ground the proposal in deontic logic, temporal reasoning (LTLf/LDLf), automated planning, and normative multi-agent systems.

Section 2 develops three blueprint scenarios for framed autonomy: (1) a single decision-maker constrained by a process-level frame (centralized intelligence), (2) multiple decision-makers with frames over individual agents, and (3) multiple decision-makers with frames over process behavior or parts thereof (distributed intelligence). Distributed scenarios add complexity because resources may be mutually inconsistent and global frames must be projected to local agents.

Section 3 outlines three practical challenges: defining a pragmatic agent notion that BPM practitioners can use; establishing a frame meta-model and specification languages (potentially hybrid symbolic/subsymbolic, with LLM-assisted frame extraction from text and rule mining from well-behaved traces); and operationalizing frames on real-world data with the explainability required for large-scale deployment.

The call to action proposes (i) introducing first-class goal/frame abstractions, (ii) building synthesis algorithms guaranteeing frame-compliant and performant operational specifications, and (iii) demonstrating applicability in real BPIS.

Connections

Directly refines and extends concepts/framed-autonomy from the APM manifesto.
Cites and builds on sources/2023-dumas-ai-augmented-bpms (ABPMS research manifesto).
Reinforces first-class status of goals: relates to concepts/agentic-bpm capability stack.
New concept: concepts/normative-frame (distinct from operational frame).
New entities: entities/giuseppe-de-giacomo (already exists), entities/yves-lesperance, entities/andrea-matta.
Existing entities updated: entities/diego-calvanese, entities/timotheus-kampik, entities/andrea-marrella.
Referenced but not ingested: Acitelli, Alman, Maggi, Marrella 2025 on synthesizing framed autonomy via automated planning; Kampik & Okulmus 2024 SIGNAL; Chopra et al. 2018 Handbook of Normative MAS.
Related frameworks: frameworks/bdi-agents, DECLARE (unverified — no concept page yet).