---
title: "XABPs: Towards eXplainable Autonomous Business Processes"
type: source
tags: [xabp, xai, explainability, apm, autonomous-business-processes, dagstuhl-autobiz]
authors: [Fettke Peter; Fournier Fabiana; Limonad Lior; Metzger Andreas; Rinderle-Ma Stefanie; Weber Barbara]
year: 2025
venue: "PMAI@IJCAI'25 (co-located with IJCAI 2025); outcome of the 2025 AutoBiz Dagstuhl Seminar #25192"
kind: paper
raw_path: "raw/ABPS/Towards eXplainable Autonomous Business Processes.pdf"
key_claims:
  - XABPs (eXplainable Autonomous Business Processes) address trust, debugging, accountability, bias, and regulatory-compliance concerns that ABPs raise relative to traditional BPMS.
  - A generic explainability conceptualisation is adapted to BPM via four constructs — explainer, explainee, explanandum, explanans — each with BPM-specific actor types and aspects.
  - Four explanandum types are distinguished — Process Instance, Process Model, AI Component, Framed Autonomy Constraints — the last being ABP-specific.
  - State-of-the-art XAI techniques have four BPM-specific limitations: inability to express process-model constraints, failure to capture contextual richness, inability to reflect causal execution dependencies, and output that is non-interpretable for process users.
  - Fourteen research challenges are enumerated (C1–C14) grouped by explainability construct plus overarching concerns (quality, benchmarks, privacy).
  - An illustrative CrewAI-based vendor-evaluation agentic BPM system demonstrates how explainability is embedded by design in procurement.
created: 2025-09-01
updated: 2026-04-21
---

# 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 [[sources/2026-calvanese-agentic-bpm-manifesto|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

- **Generic XAI vs XABP.** Generic XAI explains AI components; XABPs additionally explain process instances, process models, and framed-autonomy constraints — three categories outside classical XAI's remit.
- **Technical vs user-centric explanation quality.** Fidelity/stability (technical) are orthogonal to usefulness/meaningfulness (user-centric); both matter.
- **One-shot vs conversational explanation.** Query-based and multi-round conversational explanations are distinct modes — the latter links to [[concepts/conversational-actionability]].
- **System-explainee vs human-explainee.** Systems (including other agents) can consume explanations for self-monitoring, reconfiguration, and inter-agent coordination — a use-case omitted from most XAI literature.
- **Explaining occurrence vs explaining non-occurrence.** Why did something happen (classical) vs why did something *not* happen (harder, requires counterfactual/causal machinery).

## Positioning vs related work in this wiki

- **Breakout report of the Dagstuhl seminar that produced [[sources/2026-calvanese-agentic-bpm-manifesto|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)?