---
title: "Agentic Business Process Management Systems"
type: source
tags: [apm, abpms, agentic-ai, architecture, analytics-pyramid, autonomy-spectrum, keynote]
authors: [Dumas Marlon; Milani Fredrik; Chapela-Campa David]
year: 2026
venue: "arXiv:2601.18833v1 [cs.AI] (keynote, 2025 Workshop on AI for BPM)"
kind: paper
raw_path: "raw/ABPS/Agentic Business Process Management Systems.pdf"
key_claims:
  - An Agentic BPMS (A-BPMS) is a process-aware information system leveraging agentic AI such that (1) execution flows are not fully pre-determined, (2) adaptations do not require explicit software changes, and (3) improvement opportunities may be autonomously discovered, validated, and applied.
  - The Agentic BPM Pyramid classifies data-driven BPM approaches into four layers — Descriptive Process Analytics, Predictive Process Analytics, Prescriptive Process Optimization, and Agentic BPM at the apex — crossing operational and tactical use cases.
  - An A-BPMS architecture comprises five layers (data, process intelligence, action, orchestration, conversational) mirroring Sense-Decide-Act over process data; Model Context Protocol (MCP) mediates external-agent integration.
  - The classical manual-to-automated spectrum is extended with a third distinct mode (autonomous), producing a triangular autonomy spectrum whose vertices are human, rule-based, and agentic orchestration.
  - Seven agentic orchestration patterns are distinguished — Sequential, Parallel, Routing, Managerial, Adaptive, Mesh, Self-orchestration — plus four execution patterns (Triage, Human-Assisted, Agent-Assisted, Verification).
  - Agent-as-first-class-performer invalidates activity-centric notations like BPMN; future formalisms should support objective-based blocks, guard-rail annotations, and verification patterns.
  - Verification-centric design becomes necessary because autonomous decision-making lacks the determinism that made rule-based orchestration auditable.
created: 2026-04-21
updated: 2026-04-21
---

# 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 [[sources/2026-calvanese-agentic-bpm-manifesto|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 [[sources/2023-dumas-ai-augmented-bpms|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 **[[concepts/agentic-bpm-pyramid|Agentic BPM Pyramid]]**, a four-layer classification of data-driven BPM capabilities (adapted from [[sources/2023-chapela-campa-augmented-process-execution|Chapela-Campa & Dumas 2023]]):

1. **Descriptive Process Analytics** — automated process discovery, conformance checking, performance mining, variant analysis.
2. **Predictive Process Analytics** — what-if digital process twins and predictive process monitoring (case-level and process-level).
3. **Prescriptive Process Optimization** — automated process optimization and prescriptive process monitoring.
4. **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 [[sources/2005-reijers-limanmansar-best-practices-bpr|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 [[concepts/agentic-bpm-pyramid|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.

## Positioning vs related work in this wiki

- **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 [[sources/2023-chapela-campa-augmented-process-execution|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 [[sources/2026-calvanese-agentic-bpm-manifesto|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 [[sources/2025-calvanese-autonomy-business-process-execution|Janiesch et al. 2025 "Process Autonomization"]] (ref [5])** for the autonomy-spectrum idea.
- **References [[sources/2025-vu-practitioner-perspectives-agent-governance|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?