Dumas, Milani & Chapela-Campa 2026 — Agentic Business Process Management Systems

Position paper by Marlon Dumas, Fredrik Milani, and David Chapela-Campa (all entities/university-of-tartu), based on a keynote delivered at the 2025 Workshop on AI for BPM. Posted to arXiv on 25 January 2026 as arXiv:2601.18833v1. Distinct from the 18-author APM Manifesto (same year, same core of authors but different scope): this keynote paper proposes an architectural vision for A-BPMS platforms rather than a full research agenda.

Summary

The paper frames the rise of Generative and Agentic AI as the next wave in a five-decade succession of process-automation technologies (workflow management, BPMS, BRMS, RPA, ML-enhanced execution). What differentiates this wave is a shift from automation to autonomy and from design-driven to data-driven process management. Process mining provides the sensory substrate on top of which agents can sense process states, reason about improvement opportunities, and act.

Building on the ABPMS definition (Dumas et al. 2023), the authors define an Agentic BPMS (A-BPMS) as a process-aware information system leveraging agentic AI so that (i) execution flows are not fully pre-determined, (ii) adaptations do not require explicit software changes, and (iii) improvement opportunities may be autonomously discovered, validated, and applied.

The central conceptual contribution is the Agentic BPM Pyramid, a four-layer classification of data-driven BPM capabilities (adapted from Chapela-Campa & Dumas 2023):

Descriptive Process Analytics — automated process discovery, conformance checking, performance mining, variant analysis.
Predictive Process Analytics — what-if digital process twins and predictive process monitoring (case-level and process-level).
Prescriptive Process Optimization — automated process optimization and prescriptive process monitoring.
Agentic BPM (apex) — Automated Systems (a-priori predetermined) and Autonomous Systems (AI-agent orchestrated).

Axes cross the pyramid: Operational (day-to-day case-level recommendations) vs Tactical (weeks-to-months process change decisions).

The proposed A-BPMS architecture has five layers: data (event logs, model repositories, decision logs), process intelligence (the pyramid techniques), action (execution engines, RPA bots, planners, enterprise systems, IoT actuators, collaboration tools), orchestration (agentic and/or rule-based orchestrators), and conversational (generative-AI conversational agents + Model Context Protocol (MCP) tools for external-agent integration). The orchestration layer "reasons" on top of the process-intelligence layer; the conversational layer channels interaction with users and external agents.

§4 extends the traditional manual↔automated spectrum into a triangular autonomy spectrum with three vertices (human, rule-based, agentic) and seven agentic orchestration patterns: Sequential, Parallel, Routing, Managerial, Adaptive, Mesh, and Self-orchestration (increasing order of complexity / decentralisation). Four execution patterns describe how humans, rules, and agents interact at the sub-process level: Triage, Human-Assisted Agent, Agent-Assisted Human, Verification.

§5 (research implications) argues that activity-centric notations like BPMN are inadequate once agents become first-class performers: future formalisms need objective-based blocks, guard-rail annotations, and verification patterns. Verification-centric design is a concrete research direction: multi-layered validation across human, rule-based, and agentic components. The authors close by analogising to client-server and web-era automation — substitution alone yields modest gains; value emerges only once processes are redesigned around the new capability, suggesting the BPR heuristic catalogue needs an agentic extension.

Key claims

A-BPMS is formally defined by three properties (non-predetermined flow, no software-change adaptation, autonomous improvement discovery-validation-application).
The four-layer Agentic BPM Pyramid positions agentic BPM as resting on top of (and dependent on) descriptive/predictive/prescriptive process intelligence.
The five-layer A-BPMS architecture (data / process intelligence / action / orchestration / conversational) institutionalises Sense-Decide-Act at the platform level.
Agentic AI is distinct from generative AI: generative AI reactively produces content on prompt; agentic AI proactively pursues goals and takes actions.
The process-execution spectrum is triangular (human / rule-based / agentic) rather than linear (manual / automated).
Seven orchestration patterns and four execution patterns provide a design vocabulary for mixing actor types.
Activity-centric notations lack constructs for objectives, frames, and guard-rails; agentic process models will require new formalisms.
Verification patterns are structurally necessary for autonomous process design; redesign heuristics need an agentic extension.

Framing distinctions

Automated vs Autonomous. An automated system presupposes that the full decision space can be pre-specified; an autonomous system handles unforeseen situations through agentic AI. Automated systems may halt or escalate when outside their spec; autonomous systems decide within a frame.
Generative AI vs Agentic AI. Generative AI is reactive and content-producing; agentic AI is proactive, goal-pursuing, and action-taking. Agents are autonomous by definition.
Operational vs Tactical. Both axes cross the full pyramid; agentic BPM is not limited to operational use (case-level execution) but also addresses tactical process change.
Orchestrator vs Performer. Activities and processes can each be executed/orchestrated by a human, a rule-based system, or an agentic AI — yielding nine positions in the 3×3 actor-role matrix, collapsed visually to a triangle.

Direct extension of [[sources/2023-dumas-ai-augmented-bpms|ABPMS (Dumas et al. 2023, ref [6] in paper)]]. The three-property A-BPMS definition explicitly builds on the 2023 ABPMS definition — sharpening property (2) (adaptation without software changes) and adding property (3) (autonomous improvement).
Pyramid adapted from Chapela-Campa & Dumas 2023, ref [8]. The four-level analytics pyramid in that paper (descriptive → predictive → prescriptive → augmented) is restructured here with agentic BPM replacing augmented at the apex.
Complements the APM Manifesto (2026). Whereas the 18-author manifesto enumerates capabilities (framing, explainability, conversational actionability, self-modification) and 24 named challenges, this keynote provides an architectural and classificatory scaffold. The manifesto's "frame" (normative) is not foregrounded here; conversely, this paper's pyramid is absent from the manifesto.
References Janiesch et al. 2025 "Process Autonomization" (ref [5]) for the autonomy-spectrum idea.
References Vu et al. 2025 (ref [12]) as companion practitioner-side evidence.
Acknowledges BPR-heuristic lineage (sources/2005-reijers-limanmansar-best-practices-bpr) — proposes agentic extensions of the classic redesign heuristics.

Connections

Concepts (existing): concepts/agentic-bpm · concepts/framed-autonomy · concepts/perceive-reason-act · concepts/explainability-apm · concepts/conversational-actionability · concepts/self-modification · concepts/predictive-process-monitoring · concepts/prescriptive-process-monitoring · concepts/process-discovery · concepts/conformance-checking · concepts/process-aware-information-system · concepts/abps-autonomy-levels · concepts/bpr-heuristics

Concepts (new): concepts/agentic-bpm-pyramid

Entities: entities/marlon-dumas · entities/fredrik-milani · entities/david-chapela-campa · entities/university-of-tartu

Related sources: sources/2023-dumas-ai-augmented-bpms (ABPMS predecessor, ref [6]) · sources/2023-chapela-campa-augmented-process-execution (pyramid basis, ref [8]) · sources/2026-calvanese-agentic-bpm-manifesto (companion full manifesto) · sources/2025-calvanese-autonomy-business-process-execution (autonomy spectrum via Janiesch et al. ref [5]) · sources/2025-vu-practitioner-perspectives-agent-governance (practitioner-side companion, ref [12]) · sources/2005-reijers-limanmansar-best-practices-bpr (BPR redesign heuristics — proposed agentic extension)

Syntheses: syntheses/abpms-to-apm-evolution · syntheses/apm-manifesto-core-messages · syntheses/llm-bpm-reading-list

Open questions raised by the source

Which BPMN / DECLARE constructs survive agent-as-performer, and which must be replaced?
How are verification patterns best embedded across mixed human / rule-based / agentic sub-processes?
What is the "agentic extension" of the classical Reijers & Liman-Mansar redesign heuristic catalogue?
How do the seven orchestration patterns map to the APM Manifesto's capability stack?