Large Process Models — Kampik et al. 2024

Vision paper from SAP Signavio (Kampik, Warmuth, Rebmann, Agam, Egger, Gerber, Hoffart, Kolk, Herzig, Decker) with academic co-authors from Vienna (Han van der Aa), Melbourne (Artem Polyvyanyy), TUM (Stefanie Rinderle-Ma, Ingo Weber), and Humboldt-Universität zu Berlin (Matthias Weidlich). Published in Künstliche Intelligenz vol. 39 (2024). The paper articulates a conceptual framework — Large Process Models (LPMs) — for BPM software in the age of generative AI.

Summary

The paper's central claim is that pure LLMs are structurally unfit for business process contexts where decisions have critical operational implications: their statistics-based output is unpredictable, at times not desirable, and frequently illogical. Yet their correlation power over large, heterogeneous corpora is genuinely useful for BPM tasks such as drafting BPMN / DMN / CMMN models, generating analytical deep-dives, and recommending improvements. To reconcile these, Kampik et al. propose a neuro-symbolic architecture that fuses LLMs with symbolic data management (knowledge graphs), automated reasoning, and classical process analytics.

An LPM is explicitly not an LLM trained from scratch on process data. Instead it is a composite system with four layers:

Process data and knowledge sources — structured knowledge in knowledge graphs, language embeddings in vector stores, unstructured documents, tabular relational data (event logs included), and KPI / experience data.
Process atom layer — a layer of process atoms: atomic facts about a process (or relations between facts) that cannot be split without losing business meaning. Atoms are the bridge between natural-language LLM output and executable symbolic artefacts. They are conceptually equivalent to declarative-model constraints (DECLARE) and can be sourced by LLMs for the purpose of detecting semantic anomalies — behaviour unusual given the available knowledge.
Run-time contextualization — process query engines (over knowledge graphs and event data) and process prediction engines (classical ML + foundation models) provide the live grounding layer. On top sits a process-fine-tuned LLM contextualized per process vertical, region/regulation, and organisation.
BPM tools and integration — process modelling, analysis/mining, and execution tooling that consumes and emits artefacts the LPM can reason about. The human / machine process manager specifies BPM goals, consumes recommendations, and provides feedback.

The paper motivates the LPM from two failed extremes in BPM's history: purely imperative model-driven execution (BPMN with DMN rulebases) that rarely survives contact with real organisational change, and purely statistical ML approaches that struggle with BPM's knowledge intensity, cold-start training cost, and the reinforcement-learning credit-assignment delay characteristic of long-running processes. The LPM is positioned as the pars pro toto bridge between the two.

The authors explicitly align the LPM vision with the AI-Augmented BPM manifesto — the goal of making processes "adaptable, proactive, explainable, and context-sensitive" — and argue that achieving these properties requires the symbolic/subsymbolic fusion the LPM enacts. The paper closes by highlighting risks (data silos, knowledge representation, trust, maintenance) and concrete research challenges for each architectural layer.

Key claims

LLMs alone are insufficient for BPM. Unpredictability, undesirable output, and illogical behaviour are disqualifying in decision-critical business contexts.
Neuro-symbolic fusion is the path forward. Symbolic data management (knowledge graphs, automated reasoning) must complement LLM correlation power.
Process atoms are a new conceptual primitive. They generalise declarative constraints and provide the executable substrate the LLM layer writes to.
Fine-tune, don't train from scratch. Foundation-model fine-tuning + prompt contextualization is preferred for BPM verticals; scratch training is cost-prohibitive and brittle under concepts/concept-drift.
Classical BPM tooling stays. Conformance and performance assessments require hard guarantees; LLMs augment but do not replace process mining and conformance machinery.
The LPM is a framework, not a model. Depending on the task at hand it is instantiated from the general architecture — different queries, predictions, and tool chains compose different LPMs.
Agent-based and reasoning loops are first-class. The authors note that reasoning loops "have been at the center of AI research for decades" and are re-surging in the LLM context — but warn that nascent agent systems do not yet make full use of planning, reasoning, or RL capabilities.

Framing distinctions introduced

LPM ≠ local process model. A footnote explicitly distinguishes the proposed LPM from the pre-existing "Local Process Model" notation (small frequent-behaviour patterns in event logs).
LPM ≠ BPM-LLM. An LPM is not an LLM trained on domain-specific data; it is a system in which a fine-tuned LLM is one component among several.
Imperative vs declarative BPM maturity. BPMN/DMN achieved mainstream maturity for imperative models; CMMN and declarative approaches did not (Camunda's "CMMN never lived up to its potential"). LPM sidesteps this by pushing declarative logic into the process atom layer rather than into a user-facing modelling language.
Object-level vs meta-level correctness. A symbolic rule can be logically correct (object-level) yet dated, inconsistent, or wrongly modelled from a domain perspective (meta-level). LPMs aim to continuously reconcile both levels using the data/knowledge layer.

§5 — How LPMs facilitate BPM (three benefits)

The paper articulates three concrete benefits an LPM-augmented BPM stack would deliver:

Reduction of effort & expertise required for knowledge-based BPM tasks. LPMs lower the entry bar for running BPM initiatives by (i) finding and contextualising information automatically, (ii) turning unstructured/semi-structured information into models and queries, and (iii) enriching context with logically inferred or statistically plausible facts. Specific capabilities envisioned: turning natural-language text into process models and queries; enhancing models/queries from natural-language feedback; recommending changes to process models; scaling generic insights via auto-generated query templates instantiated per organisation. The paper anticipates conversational process modelling and conversational process mining will reach production in the coming years.
Improvement of process observability. Process observability (analogous to data observability in distributed systems) is the extent to which a process is correctly observed and understood given the business objective. It tends to remain low when relying on a single method (modelling alone, or event-log mining alone) because event logs cover only a subset of process activity. By fusing knowledge and data from multiple sources, an LPM can: (i) turn unstructured informal process knowledge into actionable models and queries by traversing organisational knowledge silos; (ii) discover data sources in large information system landscapes and recommend ETL scripts; (iii) enable foundation-model-based forecasting that avoids per-organisation model training.
Convergence of process design, execution, and analysis. LPMs may eventually deliver continuous automated process improvement where the three lifecycle phases converge. This generalises the prescriptive-monitoring agenda (matching analysis insights to actions, fusing data to assess implications, continuously fine-tuning deployed changes). The vision aligns with Chapela-Campa & Dumas 2023's augmented-process-execution pyramid but pushes further into autonomous closed-loop adaptation.

§6 — Three-step feasibility roadmap

The paper offers a "feasibility-oriented" roadmap (not a research agenda):

Augmenting modelling and analysis with contextualised knowledge. Considered generally feasible given current state of the art. Key challenges are engineering: (i) rigid evaluation metrics for generative AI-supported model/query generation; (ii) LLM-friendly data exchange formats as middle-layer representations between symbolic and sub-symbolic systems; (iii) assessing fine-tuning vs. retrieval-augmented generation. → sources/2026-licardo-bpmn-assistant is a concrete instantiation of this step.
Fusing unstructured and tabular data for actionable insights. Considered substantial-challenge. Calls for: (i) conversational process-mining foundations; (ii) hypothesis-elicitation methods over tabular and unstructured data; (iii) scalable algorithms for evaluating those hypotheses.
Automating continuous improvement with the human in control. Considered blue-sky. Open questions: enterprise-software design for maximal modular flexibility on the knowledge level; how analysis-oriented BPM software changes when it becomes mission-critical to execution; how to define guardrails for autonomous process-execution systems so that standardisation versus tailored optimisation trade-offs can be shifted toward the latter.

The orthogonal feasibility challenges across all three steps are: data ingestion and curation under questionable LLM-generated quality; reliability and ethics oversight for stochastic-parrot / bullshit-generator failure modes; verifiability of inferences vs human cognitive load; choice-architecture effects on humans deferring to LPM recommendations; the cost frontier of foundation-model training, operation, and maintenance.

§7 — Discussion: positioning vs other LLM-BPM proposals

The paper compares LPM to four parallel proposals:

BloombergGPT — domain-specific LLM trained from scratch for finance. LPM differs by not training from scratch; instead it fine-tunes plus prompt-contextualises a general LLM, accepting lock-in but avoiding the cost-frontier issue.
Vidgof et al. — vision and research agenda for LLMs in BPM — primarily aligned with the BPM lifecycle from a management view. LPM differs by being feasibility-oriented and software-architecture-centric.
ProcessGPT (Beheshti et al.) — transformer-based approach for recommending next actions in knowledge-intensive processes during execution. LPM differs in scope (entire BPM lifecycle vs execution-recommendation only) and posture (no proposal to train a process-execution-trace foundation model from scratch).
Berti & Qafari — off-the-shelf LLMs for process mining — answers user queries and generates symbolic queries on event data. LPM positions this work as a subset of LPM capabilities, providing first partial evidence of LPM feasibility.
Klievtsova et al., Grohs et al. — conceptual/experimental starting points for LPMs in process modelling. LPM frames these as supportive evidence of Step-1 feasibility while noting the field still lacks experimental work demonstrating LLM effectiveness in process-execution contexts.

§7.2 also flags Generative AI for BPM beyond LLMs: process models often live as figures in slide decks, and image-to-formal-model generation is an emerging direction; audio-to-process-model from interviews is feasible but speech-to-text pre-processing is more practical than direct foundation-model audio handling.

Predecessor: AI-Augmented BPMS (Dumas et al. 2023) — LPM explicitly adopts the four ABPMS characteristics (adaptable, proactive, explainable, context-sensitive) as design targets and contributes an LLM-era architecture for realising them.
Successor-in-dialogue: APM Manifesto (2026) — cites LPM as the LLM-centric BPM vision APM positions itself against, arguing that "while LLM potential facilitates some tasks, achieving self-improving processes… requires more than current-state LLMs." The two papers agree on neuro-symbolic hybridisation but disagree on the centre of gravity: LPM centres on a fine-tuned LLM with symbolic scaffolding; APM centres on goal-directed agents with LLMs as one reasoning substrate among several.
Parallel: Chapela-Campa & Dumas 2023 — takes the analytics-side view of the same transition (descriptive → predictive → prescriptive → augmented). LPM and the augmented-execution pyramid are complementary framings: LPM focuses on the software stack, Chapela-Campa on the analytical capability progression.
Textbook baseline: Fundamentals of BPM (2018), Van der Aalst Process Mining textbook — LPM treats these as the classical BPM tooling the system must integrate with, not replace.

Connections

Concepts: concepts/agentic-bpm · concepts/process-aware-information-system · concepts/explainability-apm · concepts/context-aware-bpm · concepts/conformance-checking · concepts/declarative-process-modelling · concepts/concept-drift · concepts/business-process

Frameworks: frameworks/bpmn · frameworks/dmn · frameworks/cmmn · frameworks/declare · frameworks/bdi-agents

Methods: methods/process-mining-basics · methods/agent-oriented-process-mining

Authors (entities): entities/timotheus-kampik · entities/artem-polyvyanyy · entities/stefanie-rinderle-ma

Related sources: sources/2023-dumas-ai-augmented-bpms · sources/2026-calvanese-agentic-bpm-manifesto · sources/2023-chapela-campa-augmented-process-execution · sources/2012-vanderaalst-process-mining-manifesto · sources/2011-vanderaalst-process-mining-book · sources/2026-licardo-bpmn-assistant (instantiates §6 Step 1) · sources/2025-varsani-neuro-symbolic-ai-sap-erp (concrete neuro-symbolic LLM/ERP integration) · sources/2024-kampik-large-process-models-correction (Fig. 1 erratum)