Toward Self-Modifying Autonomous Business Process Systems

Summary

Elyasaf et al. present the conceptual groundwork for self-modifying Autonomous Business Process Systems (ABPSs) — systems that change their own structure/behavior with minimal human intervention. The paper is explicitly positioned as output from the AutoBiz Dagstuhl Seminar 25192 working group on self-modification, complementary to the group outputs on framed autonomy, conversational actionability, and explainability.

The authors adopt the adaptation vs evolution distinction: adaptation is short-term and instance-specific (rerouting one order around a malfunctioning robot), evolution is long-term and model-level (introducing a preventive maintenance rule after patterns emerge over thousands of picks). They ground adaptation in the software-engineering MAPE-K self-adaptation loop (Monitor, Analyze, Plan, Execute, over a Knowledge base) and evolution in software-evolution literature.

They organize the problem space across five dimensions: (D1) adaptation vs evolution; (D2) task / control-flow / process granularity; (D3) reactive vs proactive (where proactive draws on predictive and prescriptive monitoring); (D4) human-driven vs autonomous; (D5) planned vs emergent (bottom-up learning).

A key contribution is the autonomy-level roadmap adapted from SAE J3016: Level 0 No Automation, Level 1 Process Assistance (recommendations, anomaly flags), Level 2 Partial Autonomy (bounded automation), Level 3 Contextual Autonomy (autonomous orchestration with human fallback only in edge cases). Level 3 aligns with agentic BPM. Table 1 cross-multiplies levels by modification object (task/flow/process) to sketch capability requirements per cell.

Section 5 elaborates three core research challenges: Governance, Oversight, and Human Interaction (learning-to-defer, XAI-based justification, multi-objective optimization over performance/cost/compliance/satisfaction, LLM-as-a-judge for quality assurance); Continual Learning and Adaptation Management (closed-loop recording of adaptation effectiveness, safe generalization, bounded knowledge representations for long-running processes); and Modeling and Measuring Uncertainty (separating aleatoric randomness from epistemic lack-of-knowledge, Bayesian/fuzzy representations, explainable uncertainty communication).

Connections

Deepens concepts/self-modification with the 5D classification and autonomy-level roadmap; the adaptation/evolution distinction matches the APM manifesto (sources/2026-calvanese-agentic-bpm-manifesto).
New concept: concepts/abps-autonomy-levels (SAE-inspired three-level roadmap).
New concept: concepts/mape-k-loop (self-adaptation canonical loop) — cross-refs software engineering.
New concept: concepts/aleatoric-vs-epistemic-uncertainty.
New entities: entities/achiya-elyasaf, entities/estefania-serral-asensio.
Existing entities updated: entities/andreas-metzger, entities/sebastian-sardina, entities/arik-senderovich, entities/niek-tax.
Relates to concepts/predictive-process-monitoring and prescriptive monitoring (Metzger et al.; Bozorgi/Dumas/La Rosa/Polyvyanyy 2023).
Complements concepts/framed-autonomy and concepts/explainability-apm as sibling AutoBiz outputs.
Complements sources/2025-fournier-agentic-ai-process-observability — observability is an input to the MAPE-K Monitor/Analyze phases.