{"id":"6f1197b3-5ae0-48ed-83d4-003f8aac33db","arxiv_id":"2507.23269","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper defines a taxonomy of explainability for autonomous business processes and lists challenges that must be solved before such systems can be trusted and audited.","lead":"This paper proposes eXplainable Autonomous Business Processes (XABPs), a framework that categorizes which parts of an AI-driven business process should be explained and to whom. It outlines 14 research challenges for making self-running workflows transparent, giving BPM researchers and practitioners a shared agenda for agentic systems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benefit claims rely on an unexamined causal link between explanations and trust/compliance; Section 4.5 Challenge 14 even concedes explanations can leak sensitive data or ease attacks, so the paper's core assertion is not established.","rationale":"The reader's conditional verdict correctly identifies the weakest assumption: the paper assumes opacity is the root cause of trust, bias, and compliance concerns, and that articulating rationale alleviates them. My stress-test concurs. The central claim is not the conceptual definition of XABPs, which is internally coherent and well structured, but the benefit claim that XABPs 'will help address' a list of serious concerns. That benefit claim is load-bearing because it supplies the paper's motivation and would justify future investment in the research agenda. It is unsupported by any empirical evidence, formal argument, or mechanism analysis. Moreover, the paper itself contains internal tension: Section 1 criticizes state-of-the-art XAI for producing nonsensical or uninterpretable explanations, Section 4.5 Challenge 14 acknowledges that explanations can harm privacy, IPR, and security, and no section explains why these failure modes are absent in XABPs. The illustrative example in Section 3 shows a JSON explanation but no evaluation of its fidelity or effect on users. The absence of support is not a fatal flaw for a research agenda, but it does mean the paper overclaims relative to what it shows. A conditional verdict, requiring the authors to soften the benefit statements or provide pilot evidence, is appropriate. My proposed test would settle whether the concern actually lands: if a feasible explanation improves bias detection and compliance without degrading trust, the causal link is at least plausible; if not, the central claim fails in its current form. Since the reader already recommended CONDITIONAL and this review does not point to a different outcome, the verdict remains UNCHANGED. I also agree with the reader's emphasis on compliance and user behavior as the most concrete aspect of the assumption; my check operationalizes both.","tokens_in":10288,"tokens_out":1918,"duration_ms":23951,"concrete_test":"Run a randomized, between-subjects study using a realistic ABP decision task based on Section 3's vendor-onboarding scenario. Prepare three conditions: (A) no explanation, (B) the XABP-style structured explanation, and (C) the structured explanation plus an explicit confidence and fidelity warning. Measure four outcomes: self-reported trust, willingness to override the system's recommendation, detection of an injected systematic bias (e.g., a subtle rule penalizing vendors from one region), and an independent compliance reviewer's judgment of whether the explanation satisfies GDPR Article 22 and AI Act transparency requirements. If condition B does not significantly outperform condition A on bias detection, trust, or compliance adequacy, the paper's central benefit claim is not supported. This directly tests the causal link the paper asserts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim, stated in the Abstract and Section 1, is that XABPs 'will help address' decreased trust, debugging difficulties, accountability gaps, hidden bias, and compliance issues by enabling systems to articulate their rationale. For this claim to hold, at least three conditions must be true: (1) the generated explanation faithfully reflects the actual reasoning of the ABP, (2) the intended explainee receives, understands, and acts on that explanation, and (3) that understanding translates into the promised outcomes (trust, bias mitigation, compliance). None of these conditions is demonstrated. Section 3's illustrative example shows a JSON explanation produced by an LLM-as-a-judge pattern, but it does not validate that the explanation corresponds to the true decision process, nor does it test whether recipients' behavior changes. Section 1's own list of XAI limitations, including explanations that are 'nonsensical or not interpretable for human users,' undercuts condition (1). Section 4.5, Challenge 14, explicitly concedes that explanations can reveal privacy-sensitive information, expose business-critical IPR, or make it easier to undermine system security, meaning that a well-formed explanation can itself create or worsen compliance failures. The paper neither engages with GDPR Article 22's right to meaningful information nor with AI Act transparency obligations, beyond asserting that compliance 'requires transparency.' This is structurally distinct from the paper's conceptual contribution: XABPs may be a coherent, well-specified framework while failing to deliver the benefit that is the paper's stated justification. The concern is not that the claim contradicts consensus; it is that the causal benefit is asserted rather than supported, and the paper's own challenges acknowledge mechanisms by which it could fail.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This position paper introduces eXplainable Autonomous Business Processes (XABPs), a conceptual framework for making the behavior of autonomous business processes (ABPs) explainable. It proposes a generic explainability model with explainer, explainee, explanandum, and explanans, and identifies four explanation subjects: process instance, process model, AI component, and framed autonomy constraints. The paper also presents a small illustrative example of an LLM-based vendor evaluation agent, lists fourteen research challenges organized by these concepts, and positions the work relative to autonomous BPM, XAI, and fairness/accountability/transparency research. The central claim, stated in the Abstract and Section 1, is that XABPs will help address concerns about trust, debugging, accountability, bias, and compliance by enabling systems to articulate their rationale.","tokens_in":10653,"tokens_out":3719,"duration_ms":43621,"significance":"If the framework is adopted, it could provide a useful shared vocabulary for explainability research in agentic BPM and help structure future work on autonomous processes. The paper's main strengths are its clear conceptual decomposition, the novel identification of 'framed autonomy constraints' as an explanation subject, the integration of prior BPM and XAI work into a single framework, and a concrete, if small, illustration of explainability-by-design in an agentic system. The list of challenges is broad and actionable. However, the paper is a position paper: the benefit claims are hypotheses rather than established results, and the illustrative example is not evaluated. The paper does not provide machine-checked proofs, reproducible experiments, or parameter-free derivations of the kind that would independently verify the central claim; for this genre, that is not a fatal defect, but it does mean the claims must be framed as a research agenda rather than as demonstrated conclusions.","major_comments":[{"comment":"The central claim that XABPs 'will help address' decreased trust, debugging difficulties, accountability gaps, hidden bias, and compliance issues is stated without qualification or an evaluation methodology. Since the paper is conceptual, this is acceptable as a research agenda only if the claims are explicitly reframed as expectations or hypotheses. Please (a) replace 'will help address' with 'are expected to help address' or similar, and (b) add a subsection or paragraph identifying the empirical questions that would test each benefit, e.g., whether explanations actually increase stakeholder trust, whether explainees understand and act on them, and whether explanations satisfy legal obligations. As written, the claim is load-bearing but unsupported, and Section 4.5, Challenge 14, itself concedes that explanations can reveal privacy-sensitive information or ease attacks, so the unconditional wording is too strong.","section":"Abstract and Section 1"},{"comment":"The illustrative vendor-evaluation example is presented as the only concrete realization of the XABP framework, but the generated JSON explanation is not evaluated in any way. The paper does not verify that the explanation faithfully reflects the actual decision process of the Vendor Evaluator agent, nor does it test whether the intended explainee (the vendor) understands, trusts, or acts on the explanation. If this example is meant only to illustrate the conceptual vocabulary, please state that explicitly. If it is meant to support the feasibility of the approach, report at least a basic fidelity check and a user study, or indicate that such validation is future work.","section":"Section 3, Figures 7 and 8"},{"comment":"The compliance claim is asserted as a need for 'an increasing level of transparency' (Section 1), but the paper does not engage with the specifics of GDPR Article 22(3) (meaningful information about the logic involved) or the AI Act's transparency obligations, nor does it explain how an articulated rationale would satisfy those legal requirements. Transparency may be necessary for compliance, but the jump from 'transparency' to 'compliance' is not demonstrated. Please either narrow the claim to state that explanations are a likely necessary component of compliance, or add a brief legal-analysis subsection that connects the XABP explanation elements to concrete legal obligations.","section":"Section 1, compliance bullet"},{"comment":"Challenge 14 explicitly concedes that explanations can reveal privacy-sensitive data, expose business-critical IPR, or make it easier to undermine system security. This is a direct qualification of the paper's central benefit claim and should be incorporated into Section 1 rather than relegated to a later challenge. The paper should discuss how XABPs would handle these risks, for example through selective disclosure, redaction, access control, or differential privacy, and should acknowledge that in some settings explanations may do more harm than good. Without this, the paper's framing suggests that explainability is unconditionally beneficial, which its own content contradicts.","section":"Section 4.5, Challenge 14"}],"minor_comments":[{"comment":"The phrase 'after a series of stimulating taks by experts' contains a typo: 'taks' should be 'talks'.","section":"Section 1"},{"comment":"There are several typographical errors in the tables: 'formalazing' in Table 1 should be 'formalizing'; 'resoning' in Table 2 should be 'reasoning'; and Section 2.4 contains a doubled comma in 'the explainee, , elaborated below'.","section":"Tables 1 and 2"},{"comment":"The phrase 'explainable ABPMN' appears to be a typo; it should likely be 'explainable ABPM' or 'explainable ABPs'.","section":"Challenge 9"},{"comment":"The new explanation subject 'Framed Autonomy Constraints' is introduced in Figure 2 but never given a formal definition in the text. A short paragraph distinguishing it from the process model and AI component explananda would improve clarity, especially since it is one of the paper's claimed novel contributions.","section":"Section 2.2"},{"comment":"The terms 'explanans' and 'explanantia' are used inconsistently; for example, Section 2.1 says 'the explanantia produced by the explainer should provide information' where a singular 'explanans' or a rephrased plural construction would be clearer. Please standardize the terminology.","section":"Throughout"},{"comment":"The paper mentions the 2025 AutoBiz Dagstuhl seminar as the origin of the ABP notion but does not include a formal reference for the seminar output. Adding a citation or a more detailed footnote would help readers trace the provenance of the concept.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a well-structured position paper from a strong group of authors, and the framework is likely to be useful to the BPM community. My main concern is that the central benefit claims are stated too assertively for the evidence provided; the authors should reframe them as a research agenda and add an evaluation roadmap. The heavy reliance on the authors' own prior work in Section 1's list of XAI limitations ([4]-[7], [13], [26], [29], [33]) is not improper, but independent evidence would strengthen the premise. If the journal's scope prioritizes conceptual contributions, the paper may be acceptable after the major revisions described; otherwise, it may be better suited to a venue that explicitly welcomes position papers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This one is worth a careful read if you work in BPM or XAI. The genuinely new bit is the treatment of \"framed autonomy constraints\" as a distinct explanation subject — why the system is allowed to behave as it does — alongside process instance, process model, and AI component. That gives the field a vocabulary it didn't have. The fourteen challenges are a useful checklist, and several are specific to BPM in nontrivial ways: explaining non-occurring behavior, causal dependencies in partially ordered process execution, and explanation timing relative to process state. The conceptual mapping of explainer/explainee/explanandum/explanans to BPM actors is done carefully, and the vendor example shows how the pieces fit, even if it is only illustrative.\n\nThe soft spot is the benefit claim. The abstract and intro say XABPs \"will help address\" trust, debugging, accountability, bias, and compliance, but the link between explanations and those outcomes is asserted, not demonstrated. No empirical evidence shows explanations change user behavior, and the GDPR/AI Act discussion stops at \"compliance requires transparency\" without engaging Article 22 or specific obligations. The paper's own Challenge 14 concedes that explanations can leak privacy-sensitive information or ease attacks, which cuts against the compliance benefit. That is not a fatal flaw — this is a position paper, and the link is a research hypothesis — but the authors should soften \"will help\" to \"may help\" or make clear that the benefit is an open research question. The heavy reliance on their own prior work to establish the insufficiency of current XAI is a mild concern, not a violation; the cited limitations are plausible, but independent sources would strengthen the premise.\n\nThe stress-test note treats the asserted benefit as load-bearing in a way that feels too harsh for a roadmap paper. A motivating claim backed by an argument and a challenge list is within normal bounds. The reader's CONDITIONAL verdict is about right.\n\nWho it's for: BPM researchers, XAI people wanting a process-level frame, and anyone building agentic workflows. It deserves a serious referee; with minor revisions — softening benefit claims, adding a brief legal note, and trimming self-citation reliance — it becomes a useful agenda-setting paper.","headline":"A well-structured agenda paper that gives BPM explainability a useful vocabulary, provided the authors soften the causal claims about trust and compliance.","tokens_in":11182,"tokens_out":1972,"would_cite":false,"duration_ms":23748,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Autonomous workflows should be built to state their own rationale, says this paper.","keywords":["Business Process Management","Autonomous Business Processes","Explainability","Agentic BPM","Explainable AI","Process Mining","Accountability","XABP"],"falsifier":"A concrete check: deploy an agentic vendor-onboarding system of the kind the paper illustrates, then have auditors and applicants work only from the system's generated explanations. If auditors cannot determine why a vendor was rejected, whether escalation thresholds were respected, or whether a GDPR-related exception was legal, and if applicants' trust or behavior does not improve relative to receiving no explanation, the paper's promise is undercut.","tokens_in":10049,"feed_emoji":"🤖","tokens_out":5140,"duration_ms":55036,"temperature":0.7,"pith_summary":"This paper proposes that autonomous business processes (ABPs), self-executing workflows that use AI and machine learning with minimal human oversight, should be designed so that they can articulate the rationale behind their own actions. It argues that this capability, called eXplainable ABPs (XABPs), would address five named concerns: eroding stakeholder trust, difficult debugging, hindered accountability, hidden bias, and regulatory compliance. The paper contributes a conceptual framework that fixes the central terms of an explanation—who explains, to whom, about what, and with what content—and it identifies four kinds of explanation subjects plus a roadmap of fourteen research challenges. If the framework is adopted, explainability becomes a first-class design property of agentic business process management rather than a post-hoc add-on.","feed_headline":"Make autonomous workflows explain themselves","feed_subtitle":"A new framework maps every 'why' question to four explanation subjects and a 14-point research agenda.","key_machinery":"The load-bearing object is the explanandum, organized into four understanding subjects: process instance, process model, AI component, and framed autonomy constraints (the rules that define what autonomy the system is allowed to exercise, such as delegation rules and escalation thresholds). Around this, the framework specifies the explainer (system or human that generates the explanation), the explainee (human or system that receives it), and the explanans (the explanation itself, characterized by mechanism, generation time, interaction mode, presentation format, and quality). Together these carve every 'why' question in an autonomous process into a defined slot, which is what turns explainability from an aspiration into a design problem with named research tasks.","core_discovery":"On the paper's own terms, the central claim is that ABP opacity is the shared root of the five concerns, and that building ABPs as XABPs—systems that can state why a process instance followed a path, why a process model is structured as it is, why an AI component recommended or decided something, and why the system is permitted to behave as it does—will alleviate them. The paper does not prove this empirically; it establishes the conceptual apparatus needed to ask and answer such questions systematically. It structures explainability along four axes—explanandum, explainer, explainee, and explanans—and derives fourteen challenges that would have to be solved for XABPs to be realized.","pith_inferences":["This inference goes beyond the paper: if the framework is right, the hardest test case is framed-autonomy explanation, because it requires the system to expose its own governance rules; a concrete experiment would check whether auditors can use such explanations to verify GDPR or AI Act compliance.","This inference goes beyond the paper: the taxonomy implies that process-instance explanations are causally loaded—they need counterfactual or causal methods, not just feature attribution—because users ask why something happened rather than which features were influential.","This inference goes beyond the paper: the authors' assumption that explanation restores trust is testable; a field study measuring adoption behavior after deployment of XABP-style explanations would settle it, and the paper does not run such a study.","This inference goes beyond the paper: the illustrative vendor-evaluation example suggests the framework can double as a specification checklist for building agentic BPM systems today, even before the open challenges are solved."],"forward_implications":["If XABPs are realized, stakeholders can ask standard questions of a running process—why this path, why this decision, why this resource assignment, why this outcome—and receive answers tied to the specific process instance.","Explainability shifts from an optional XAI wrapper to a first-class design property of Agentic BPM systems, shaping how agents justify behavior to humans and to other agents.","The four explanation subjects give BPM researchers a target list: process-instance, process-model, AI-component, and framed-autonomy explanations each need dedicated techniques.","The fourteen challenges define a concrete research agenda, spanning preference specification, causal explanation, timing and preservation of explanations, adaptation to explainee, and benchmarking.","Quality assessment of explanations becomes a dual problem of technical fidelity and user-centric usefulness, requiring both objective and subjective measures."],"supporting_citations":[{"why":"Defines AI-augmented business process management systems, the foundation XABPs build on.","marker":"[1]"},{"why":"Systematic review of interpretable ML for predictive process monitoring that anchors the XAI-in-BPM context.","marker":"[2]"},{"why":"Survey of explainable AI techniques that the paper treats as the state of the art to be extended.","marker":"[3]"},{"why":"Shows a model-informed XAI extension and exemplifies the limitation of expressing process model constraints.","marker":"[4]"},{"why":"Provides causal execution dependency discovery, cited as missing from standard XAI.","marker":"[6]"},{"why":"User study showing LLM-based process explanations often fail users, motivating the XABP quality agenda.","marker":"[7]"},{"why":"Supplies the philosophical grounding that an explanation should give the causes of the explanandum.","marker":"[8]"},{"why":"Local explanation approach and overview of explainable AI for process mining that the paper extends toward ABPs.","marker":"[13]"}],"fun_headline_variants":["Explainable AI for autonomous workflows: a four-axis framework","Four explanation subjects to demystify autonomous business processes","A research agenda for explainable autonomous business processes","Making every 'why' answerable in autonomous workflows","From opaque to transparent: a systematic path to XABPs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that opacity is the root cause of the trust, debugging, accountability, bias, and compliance problems in autonomous business processes, and that making a system articulate its rationale will materially reduce those problems; if explanation does not change behavior or satisfy legal transparency requirements, the central claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Explainable AI for autonomous workflows: a four-axis framework","Four explanation subjects to demystify autonomous business processes","A research agenda for explainable autonomous business processes","Making every 'why' answerable in autonomous workflows","From opaque to transparent: a systematic path to XABPs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00056,"raw_usage":{"total_tokens":2579,"prompt_tokens":781,"completion_tokens":1798,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":397,"completion_tokens_details":{"reasoning_tokens":1720}},"tokens_in":397,"tokens_out":1798,"duration_ms":13553,"temperature":1.0,"reasoning_tokens":1720,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:52:21.953279+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: deploy an agentic vendor-onboarding system of the kind the paper illustrates, then have auditors and applicants work only from the system's generated explanations. If auditors cannot determine why a vendor was rejected, whether escalation thresholds were respected, or whether a GDPR-related exception was legal, and if applicants' trust or behavior does not improve relative to receiving no explanation, the paper's promise is undercut.","supporting_citations":[{"cited_title":"Interpretable and Explainable Machine Learning Methods for Predictive Process Monitoring: A Systematic Literature Review","cited_arxiv_id":"2312.17584","evidence_quote":"Systematic review of interpretable ML for predictive process monitoring that anchors the XAI-in-BPM context."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows a model-informed XAI extension and exemplifies the limitation of expressing process model constraints."},{"cited_title":"Mehdiyev, P","cited_arxiv_id":null,"evidence_quote":"Local explanation approach and overview of explainable AI for process mining that the paper extends toward ABPs."}],"review_version":1}