Vu, Klievtsova, Leopold, Rinderle-Ma & Kampik 2025 — Practitioner Perspectives on Agent Governance in Business Processes

Empirical paper published at BPM 2025 (23rd International Conference on Business Process Management, pp. 29–43; arXiv preprint 2504.03693v2, 3 July 2025). Authors span SAP Berlin (Vu, Kampik), TU Munich (Klievtsova, Rinderle-Ma), Kühne Logistics University Hamburg (Leopold), and Umeå University (Kampik). Rare in the agentic-BPM corpus for being a practitioner-facing qualitative empirical study rather than a theoretical / architectural paper.

Summary

The paper addresses a gap noted implicitly by the APM Manifesto: the theoretical agentic-BPM literature (APM, ABPMS, A-BPMS keynote) is largely detached from current industry practice. The authors conduct 22 semi-structured interviews with BPM practitioners to investigate how organisations can effectively govern AI agents in business processes. The research question is phrased as: "How can BPM support the responsible integration of AI agents into organizations, ensuring both effective business outcomes and social accountability?"

§2 traces a historical timeline of agent concepts in BPM, organised along two streams:

Agents for BPM — agents as means to support BPM tasks (Agent-Oriented Programming '90s, RPA '10s, ABPMS '20s, GenAI agents mid-'20s).
BPM for Agents — BPM as a discipline that governs agent-based systems (Normative MAS, Process Choreographies, Agent System Mining, and finally Agentic BPM mid-'20s).

On this basis the authors introduce Definition 1 (Agentic Business Process Management): "ABPM describes (i) the deployment and execution of autonomous software agents to achieve business process goals and (ii) the application of agent-based abstractions for the process-oriented design and analysis of autonomous software agents." The definition is deliberately technology-agnostic and spans the BPM lifecycle's deployment/execution and design/analysis phases.

§3 describes the qualitative methodology: semi-structured interviews of 60–90 minutes, transcription, familiarisation, then deductive–inductive coding per Mayring (2019). 22 participants across Professional Services (10), Manufacturing (5), Telecommunications (3), and smaller sectors; Consultant (9), Application Owner (5), Center of Excellence Expert (3), and other roles. Notably none reported prior hands-on experience with agentic AI — answers reflect expectations informed by existing RPA/automation experience.

§4 reports findings under six themes:

Understanding. 10 of 22 were familiar with "agentic AI"; views ranged from "an evolution of RPA" to "digital assistant" to "orchestration layer".
Benefits. Enhanced efficiency (6), reduced manual work (6), improved data quality (5), error reduction (4), cost savings (4), compliance improvement (3), scalability (3), risk reduction (2).
Risks. Bias in AI models (5), over-reliance on automation (4), job displacement (4), lack of clarity (3), cybersecurity threats (3), governance issues (3), unauthorized decisions (3), data breaches (3).
Use cases. Task Automation (7), Process Monitoring (5), Customer Service Automation (4), Predictive Analytics (4), Quality Control (3), Data Structuring (3), Financial Risk (3), Supply-Chain Optimization (3), Master Data Maintenance (2).
Requirements. Governance Structures (8), Process Integration (7), Compliance with Guidelines (6), Employee Training (5), Cost Management (4), Preconfigured Use Cases (3), Early Testing (3), Access Control (3).
Autonomy and adaptability. Configurable autonomy — flexibility in low-risk areas, restriction in high-risk scenarios (system changes, external communications); start-simple-then-grow adoption.
Human involvement and responsibility. Agents as decision-support, not decision-makers, in complex scenarios. Responsibility most often attributed to organisations (10) and as shared responsibility (10), with AI developers (7), users (7), application owners (4), process managers (3), dedicated AI department (2) trailing.

§5 distils six adoption recommendations: (1) Business context — clear goals, expected benefits, cost assessment; (2) Guardrails — legal/ethical/organisational boundaries; (3) Human-agent collaboration — defined roles, intervention points; (4) Customization — tailor autonomy level, oversight, documentation, reasoning transparency; (5) Risk management — monitor performance, prevent undesired adaptation/evolution; (6) Adoption — seamless integration, fallback mechanisms.

Four BPM-specific actions are proposed: (1) Human-agent balance in workflow design; (2) Human involvement — peers vs supervisors; (3) Process structures evolving beyond static models toward real-time-adjustable structures; (4) Performance metrics — agent-impact metrics on efficiency, decision quality, organisational outcomes.

Key claims

Agentic BPM (ABPM) is the governed deployment of autonomous software agents to achieve business-process goals, applying agent-based abstractions to process design/analysis.
The "Agents for BPM" vs "BPM for Agents" distinction frames the novelty of ABPM: BPM as a discipline that governs agent systems rather than just uses them.
Practitioners anticipate agentic AI benefits and risks largely through the lens of RPA, with no participant reporting hands-on agentic-AI experience.
Configurable autonomy with risk-proportional restrictions, not all-or-nothing autonomy, is the practitioner consensus.
Responsibility for agent failures is perceived as shared across organisations, developers, users, and domain owners — not locatable in a single role.
Static BPMN-style models are inadequate; process structures must evolve to support real-time agent-driven adjustments.
Agent-specific performance metrics are needed alongside traditional BPM KPIs.
Cultural acceptance of autonomy-to-agent transfer is a first-order adoption barrier — not just a technical issue.

Framing distinctions

Agents for BPM vs BPM for Agents. The historical "timeline" figure is the key conceptual device: the paper positions ABPM squarely in the "BPM for Agents" stream.
Configurable autonomy ≠ full autonomy. Practitioners uniformly reject across-the-board autonomy; they envisage a gradient scaled to task risk.
Agent as peer vs supervisor. Two distinct modes of human-agent collaboration surface in the data — a question the paper flags for future BPM research.
Benefits vs requirements are not symmetric. The most-cited benefits (efficiency, data quality) differ structurally from the most-cited requirements (governance, integration, compliance) — adoption is gated on governance, not on capability belief.
Decision-support vs decision-maker. Practitioners pull agents decisively toward decision-support in complex scenarios, preserving human decision authority.

Companion to APM Manifesto and A-BPMS keynote (Dumas et al. 2026). Where those papers are theoretical/architectural, this is the corresponding empirical practitioner study. The Dumas et al. 2026 keynote cites this paper as ref [12].
Contrasts with ABPMS Manifesto (2023): same agentic territory, but ABPMS is designer-oriented while this paper is practitioner-oriented. Cites ABPMS as ref [6] positioning ABPM as "BPM for Agents" continuation.
Connects to concepts/ai-adoption — adds a BPM-specific facet to the broader AI-adoption evidence base (sources/2025-korst-wharton-gen-ai-enterprise-adoption, sources/2025-handa-which-economic-tasks-ai). Unlike those, this paper captures BPM-specialist perceptions rather than cross-sector knowledge-worker usage.
Complements sources/2024-kampik-large-process-models (Kampik's other APM/ABPM contribution — the "Large Process Models" vision) and sources/2025-calvanese-autonomy-business-process-execution (framed-autonomy paper).
Explicitly references "Process Autonomization" (Janiesch, Kowalkiewicz & Rosemann 2025) and subject-oriented BPM (concepts/declarative-process-modelling lineage).
Sibling to sources/2025-fettke-explainable-autonomous-business-processes (the XABPs paper) — both 2025 BPM-community outputs shaping the empirical/theoretical scaffolding for the APM Manifesto.

Connections

Concepts (existing): concepts/agentic-bpm · concepts/framed-autonomy · concepts/self-modification · concepts/explainability-apm · concepts/conversational-actionability · concepts/perceive-reason-act · concepts/ai-adoption · concepts/abps-autonomy-levels · concepts/behavioral-variability · concepts/process-architecture · concepts/bpm-lifecycle

Entities: entities/timotheus-kampik · entities/stefanie-rinderle-ma · entities/henrik-leopold · entities/hoang-vu · entities/nataliia-klievtsova

Related sources: sources/2026-calvanese-agentic-bpm-manifesto · sources/2026-dumas-agentic-bpms-pyramid (cites this paper, ref [12]) · sources/2023-dumas-ai-augmented-bpms · sources/2025-calvanese-autonomy-business-process-execution · sources/2024-kampik-large-process-models · sources/2025-fettke-explainable-autonomous-business-processes (sibling workshop output) · sources/2025-korst-wharton-gen-ai-enterprise-adoption (cross-sector adoption comparator) · sources/2024-xu-the-agent-company-benchmark (agent-capability comparator)

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

Open questions raised by the source

How to quantitatively validate the qualitative findings across industries (sample was 5 countries, Germany-heavy)?
How do practitioner expectations evolve once hands-on agentic-AI experience accumulates?
How should BPM governance frameworks differ structurally from RPA governance (a specific BPM-research ask)?
What agent-specific performance metrics best complement traditional BPM KPIs?
How to operationalise "configurable autonomy" in BPMN-style notations lacking frame/guard-rail constructs?