Theodorakopoulos & Theodoropoulou 2026 — Business Intelligence and Business Process Management in the Era of Generative AI

55-page narrative-conceptual review published in MDPI Applied Sciences (May 2026). Authors are based at the Department of Management Science and Technology, University of Patras (Greece). The review's headline contribution is a layered integrative framework for connecting BI, BPM, big data analytics, process mining, GenAI, and DSS — explicitly tempered against techno-optimism.

This is the wiki's first ingested source on integrative BI ↔ BPM ↔ GenAI ↔ DSS framing. Complementary to (but lighter than) the APM Manifesto's pyramid framing.

What the paper does

A conceptual (not bibliometric, not PRISMA) review that:

Searches Google Scholar, Scopus, Web of Science for peer-reviewed work from 2015 onwards on BI, BPM, big data analytics, process mining, GenAI, DSS, and their intersections.
Thematically clusters the literature into conceptual foundations, data & computational infrastructure, process intelligence, GenAI in business processing, integrative decision-support architectures, application domains, and implementation tensions.
Synthesises the clusters into a layered architecture rather than a bibliometric aggregation.

Three research questions: (RQ1) how do BI/BPM/BDA/PM/GenAI/DSS relate within contemporary business-processing environments; (RQ2) what conceptual roles do they play as parts of an integrated decision-support architecture; (RQ3) what implementation tensions limit practical integration.

The 5-layer integrative framework (Section 7, Figure 4)

Layer	Domain	Role
Data & Computation	Big Data Analytics	Captures, stores, integrates, processes heterogeneous data streams (event logs, transactions, customer interactions, IoT, multimodal sources). Enabling base, not interpretation.
Organisational Insight	BI / business analytics	Transforms dispersed operational/transactional/contextual inputs into managerial interpretation. Distributed interpretive capability, not just dashboards.
Process Intelligence	BPM / process mining	Reconstructs how work unfolds across cases, activities, handoffs, delays, deviations. Goes beyond control-flow into multi-perspective + predictive + explanatory.
Augmentation	Generative AI	Mediates between layers — translates data, models, and analytical outputs into forms humans can query, summarise, recombine. Explicitly NOT the new core.
Decision	Decision Support Systems	Converts insight, process signals, generated explanations into alternatives, priorities, interventions, accountable choices. Action-conversion layer.

Cross-layer feedback loops: decisions trigger interventions → interventions generate new traces → new traces reshape data foundation. Framework is recursive, not linear.

Position on GenAI

A central thesis: GenAI is positioned as complementarity, not substitution. Conventional ML remains central for tightly-specified predictive and optimisation tasks ("classification, prediction, anomaly detection, resource allocation, narrow-task decision support"); GenAI's distinctive role is "summarisation, process knowledge extraction, conversational querying, draft model generation, explanation" (Table 1). The clearest evidence so far is in process-model drafting (Nivon & Salaün, Hörner et al., Klievtsova et al.) and document-level synthesis — not autonomous process authorship.

Section 6.5 catalogues five limit categories for GenAI in process-critical environments (Table 3):

Limitation	Why it matters	Typical consequence
Hallucination & factual instability	Fluent but unfaithful generation	Fabricated justifications, false confidence
Semantic imprecision in domain tasks	Near-correct outputs may fail sequence/rule logic	Process errors, compliance failures
Weak real-world evaluation	Benchmark success ≠ workflow success	Overestimation of deployment readiness
Privacy/security exposure	Data leaked/inferred/mishandled in prompts	Disclosure risk, regulatory exposure
Limited accountability/auditability	Outputs hard to trace to verifiable sources	Reduced contestability, weaker oversight

Implementation tensions (Section 9)

Data quality and interoperability accumulate across layers — weak source-data alignment propagates upward and makes integration appear performative rather than real.
Explainability ≠ effective oversight — Cecil et al. and Buçinca et al. cited: explainable AI advice does not significantly reduce people's tendency to follow incorrect advice; cognitive-forcing mechanisms outperform simple explanations.
Trust must be calibrated, not maximised — Afroogh et al. distinguish trustworthiness (system property) from trust (user disposition); over-trust is as problematic as resistance.
Governance is a design constraint — Papagiannidis et al. argue responsible-AI governance is operationalisation through structural/relational/procedural practices, not abstract principles.
Organisational readiness ≠ willingness to experiment — includes governance literacy, data stewardship, training, process-redesign capacity.

Application domains (Section 8, Table 4)

Seven business domains illustrate the framework: operational process optimization · customer-facing/service processes · supply chain & logistics · finance, compliance & risk · knowledge-intensive admin workflows · procurement, sourcing & contracts · human–AI managerial decision-making. Each domain highlights which layers are most involved, what practical value the framework adds, and the main implementation concern.

Future-work agenda (Section 10.2)

Cross-layer evaluation frameworks — move beyond isolated component benchmarks toward integrated business-processing evaluation.
Human-centered & explainability-aware GenAI for BPM — focus on oversight conditions, cognitive burden, contestability.
Privacy-preserving and federated process analytics — for cross-organisational, regulated environments.
Process-aware multimodal LLM systems — reason across documents, event logs, interfaces, images while staying grounded.
Real-time adaptive DSS under governance constraints — adaptive recommendations remaining auditable and accountable.

Substantial overlap with the open gaps catalogued in syntheses/llm-bpm-reading-list §E.2.

Limitations

Acknowledged by the authors (Section 10.1):

Narrative review, not PRISMA — coverage is selective and conceptual rather than exhaustive/statistical.
Framework not empirically validated — should be read as conceptual architecture, not tested implementation model.
Evidence base is uneven — BI/BPM/PM literature is comparatively mature; GenAI-in-process-critical evidence is newer, less stable, often prototype-level.

Wiki-maintainer additional caveats:

MDPI Applied Sciences is a high-volume open-access journal; not a core BPM venue. The authors are from a management-science department, not the established BPM research community.
The "BI / BPM / PM / BDA / GenAI / DSS as layered architecture" framing is useful as vocabulary but overlaps substantially with prior layered framings (Process Mining 2.0 from Dumas 2021; the APM Manifesto's framed-autonomy → explainability → conversational-actionability → self-modification capability stack). Treat as integrative/teaching reference, not as a primary theoretical contribution.

Connections

Concepts:

concepts/agentic-bpm — 5-layer framing as a complementary lens on the same problem APM addresses; useful contrast.
concepts/explainability-apm — calibrated-trust + cognitive-forcing argument; relevant to APM's explainability capability.
concepts/predictive-process-monitoring — situates PPM within Process Intelligence layer.
concepts/prescriptive-process-monitoring — situates PrPM in the Decision layer.
concepts/llm-assisted-process-modelling — strongest documented GenAI-BPM evidence cluster.
concepts/llm-based-ppm — relevant to the Augmentation layer's process-aware future direction.

Sources cited that exist in the wiki:

[5] Weinzierl ML in BPM systematic review (not yet ingested).
[6] sources/2024-kampik-large-process-models — LPM vision.
[45] Ceravolo PPM Challenges (not yet ingested).
[47] Teinemaa outcome-PPM (not yet ingested — flagged as priority in syntheses/ppm-landscape §6).
[54] Carmona conformance checking textbook (not yet ingested).
[76] Bernardi RAG + process-aware DSS (not yet ingested).
[77] Rebmann LLM-on-semantics-aware-PM (not yet ingested — also priority in syntheses/llm-bpm-reading-list §B.L8).
[78] sources/2024-kampik-large-process-models — Kourani LLM in BPM (partial overlap with Licardo et al. 2026 already ingested).
[79] Grohs LLMs in BPM (not yet ingested — flagged in reading list §B.L6).
[81] Berti PM-LLM-Benchmark (not yet ingested — flagged in reading list §B.L8 vicinity).
[101] sources/2023-chapela-campa-augmented-process-execution — augmented process execution (currently a stub).
[122] van der Aalst beyond control-flow PM (not yet ingested).
[126] van der Aalst beyond workflow-like representations (not yet ingested).
[129] Rizzi explainable PPM user evaluation (not yet ingested).
[148] Graafmans DMAIC + PM (not yet ingested).
[154] Middelhuis learning-based resource allocation (not yet ingested).
[217] sources/2020-rama-maneiro-deep-learning-ppm-review — DL-PPM benchmark.

Syntheses:

syntheses/llm-bpm-reading-list — new entry in §A.7 (Integrative reviews); future-work directions feed §E.2 gaps.
syntheses/abpms-to-apm-evolution — alternative integrative framing alongside ABPMS → APM evolution.