Fettke, Fournier, Limonad, Metzger, Rinderle-Ma & Weber 2025 — XABPs: Towards eXplainable Autonomous Business Processes

Workshop paper at PMAI'25 (Process Management in the AI era, IJCAI'25 co-located workshop), authored by six contributors to the 2025 AutoBiz Dagstuhl Seminar #25192 that also produced the APM Manifesto. The paper is the explainability breakout report from that seminar — it sharpens the notion of eXplainable Autonomous Business Processes (XABPs) and enumerates research challenges specific to explaining autonomous business process behaviour.

Summary

The paper begins from the observation that Autonomous Business Processes (ABPs) — self-executing workflows driven by AI/ML — introduce five concerns relative to traditional BPMS: erosion of stakeholder trust, difficulty of debugging, hindered accountability, risk of perpetuated bias, and challenge of demonstrating regulatory compliance (GDPR, EU AI Act, finance/healthcare/HR applications). XABPs address these concerns by enabling systems to articulate the rationale behind their behaviour. Explainability is framed as a first-class citizen in the realisation of Agentic BPM systems, supporting autonomy from two perspectives: (i) enabling agents to independently resolve misalignments in other agents' behaviour, and (ii) reducing human intervention by making agent behaviour understandable and transparent.

Section 2 introduces a generic explainability conceptualisation adapted to BPM. The core quartet is:

Explainer ("who explains?") — AI agents, monitoring components, connected systems; domain experts, supervisors, trainers/annotators.
Explainee ("explained to whom?") — end users, process participants, managers, business analysts, compliance officers; also system explainees (the system itself, connected systems, other agents).
Explanandum ("what is explained?") — four BPM-specific types:
- Process Instance explanation (flow, decision points, resource assignment, outcome justification).
- Process Model explanation (model structure, policy compliance).
- AI Component explanation (AI decision, AI model behaviour) — classical XAI territory.
- Framed Autonomy Constraints — design autonomy, delegation rules, AI authority, escalation thresholds, compliance limits. This fourth type is ABP-specific and directly tied to concepts/framed-autonomy.
Explanans ("how is the explanation provided?") — five dimensions: explanation mechanism (feature attribution, example-based, rule-based, model simplification, counterfactual, visual), time of generation (ex-ante / run-time / post-hoc), presentation format (visual / verbal), interaction mode (one-shot / query / multi-round conversational), explanation quality (technical: fidelity, stability; user-centric: usefulness, meaningfulness).

State-of-the-art XAI techniques are insufficient for ABPs in four specific ways: (1) inability to express business-process-model constraints, (2) failure to capture contextual richness, (3) inability to reflect causal execution dependencies among activities, and (4) output that is often non-interpretable for process users. These map directly to research challenges in §4.

Section 3 provides a worked example — a vendor-evaluation agent in a procurement ABP, realised in CrewAI with LIME / SHAP / causal-inference tools. The Vendor Evaluator agent receives applications, produces numeric scores, and (on escalation) produces structured JSON explanations combining feature contributions, causal counterfactuals, and policy-aligned recommendations.

Section 4 enumerates 14 research challenges (C1–C14) structured along the four explainability constructs plus overarching concerns:

Explainee (C1): specifying preferences for explanations.
Explanandum (C2): taxonomy of ABP-specific explanation subjects.
Explainer (C3): new XAI techniques beyond standard XAI — what-if, causal, process-aware.
Explanans (C4–C11): actionable explanations preserving autonomy; generation-time decisions; adapting form to explainee; explaining non-occurring behaviour; synthesising heterogeneous data; identifying causal vs correlational explanations; evolving explanations over time; interaction between frames and explanations.
Overarching (C12–C14): quality assessment, benchmark datasets, privacy / IPR / security preservation.

The conclusion reiterates that ABPs are not simply sequential decision processes — they are distributed, non-sequential, only partially ordered by causality — a property that generic XAI literature has historically ignored.

Key claims

ABPs introduce five trust/debug/accountability/bias/compliance risks beyond traditional BPMS that require dedicated explainability support.
Explainability is a first-class citizen in Agentic BPM, not a bolt-on; it enables inter-agent alignment and reduces human-in-the-loop load.
Four explanandum types cover BPM explainability; the fourth (framed-autonomy constraints) is unique to ABPs.
Generic XAI techniques fail on four BPM-specific dimensions (model constraints, context, causal dependencies, interpretability for process users).
The explanation quality dimension is bifurcated into technical (fidelity, stability) and user-centric (usefulness, meaningfulness) properties.
Framing the explanans along five dimensions (mechanism, time, format, interaction, quality) provides a design space for XABP implementations.
14 named research challenges operationalise the XABP agenda; they sharpen the APM Manifesto's X1–X5 challenges.
Causal analysis is a prerequisite for explaining why certain behaviours did not occur — a uniquely hard XABP sub-problem.

Framing distinctions

Breakout report of the Dagstuhl seminar that produced the APM Manifesto. Paper opens by citing the seminar and notes that four breakout topics (framed autonomy, self-modification, conversational actionability, explainability) were discussed; this paper is the "explainability" deliverable.
Sharpens concepts/explainability-apm. The manifesto's explainability capability is operationalised here through a concrete actor/explanandum/explanans architecture and 14 named challenges that sub-divide and extend the manifesto's X1–X5 list.
Co-author continuity: all six authors are part of the AutoBiz community. entities/peter-fettke, entities/fabiana-fournier, entities/lior-limonad, entities/andreas-metzger, entities/stefanie-rinderle-ma, entities/barbara-weber all recur in the manifesto author list.
Cites process-mining XAI lineage (Mehdiyev & Fettke 2023 SLR on interpretable PPM; Stevens & De Smedt 2024 guidelines for outcome-prediction explanations; Harl et al. 2020 gated-graph XPPM) — connects to concepts/predictive-process-monitoring and sources/2023-cao-gated-rnn-explainable (if ingested later).
Connects causal BPM stream (Fournier, Limonad & Skarbovsky 2025 "The Why in Business Processes"; Alaee et al. 2024 causal reasoning for BPM) — linking explainability to concepts/causation and concepts/causal-process-discovery.
References ACM FAccT tradition as parallel, non-BPM stream — this paper brings FAccT concerns into BPM but notes that FAccT work "does not focus on the process perspective".

Connections

Concepts (existing): concepts/explainability-apm · concepts/agentic-bpm · concepts/framed-autonomy · concepts/conversational-actionability · concepts/self-modification · concepts/perceive-reason-act · concepts/causation · concepts/causal-process-discovery · concepts/predictive-process-monitoring · concepts/process-model-quality

Entities: entities/peter-fettke · entities/fabiana-fournier · entities/lior-limonad · entities/andreas-metzger · entities/stefanie-rinderle-ma · entities/barbara-weber

Related sources: sources/2026-calvanese-agentic-bpm-manifesto (parent manifesto) · sources/2023-dumas-ai-augmented-bpms (explainability as one of four ABPMS characteristics) · sources/2026-dumas-agentic-bpms-pyramid (conversational/orchestration layer counterpart) · sources/2025-elyasaf-self-modifying-abps (sibling Dagstuhl outcome) · sources/2025-fournier-agentic-ai-process-observability (sibling Dagstuhl outcome)

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

Open questions raised by the source

Does "explanation mechanism" become a contract that agents negotiate (C4)?
How to discard / update explanations as contexts drift (C10)?
How to benchmark XABP systems when process-level ground truth for rationale is rarely available (C13)?
Can explanations be withheld when they leak business-critical IPR, without violating regulatory transparency (C14)?