LLM-Assisted Process Modelling

The use of large language models to create, edit, refine, or validate business process models — typically BPMN, occasionally Petri nets, DECLARE constraints, or process-tree intermediate forms. Distinct from LLM-as-execution-engine work (where an LLM operates within a process) and from agentic process mining (where LLMs analyse logs).

Why this is hard

LLMs handle natural language fluently but produce verbose, error-prone XML when asked to emit BPMN 2.0 directly: invalid IDs, broken sequence flows, unmatched gateways, missing DI metadata, hallucinated elements. The verbosity also costs latency and tokens. Licardo et al. (2026) document a representative failure: DeepSeek V3 reached only 8% editing success when prompted to manipulate BPMN XML directly, vs 50% via a structured intermediate.

The complexity is compounded for editing: most LLM tools target one-shot generation from a textual description, but real BPM work is iterative — modellers refine, fix, restructure. Iterative XML manipulation amplifies syntax-error compounding.

Architectural patterns

Three architectural patterns are emerging:

1. Direct XML generation

Prompt the LLM with the BPMN 2.0 spec subset and ask for XML. Brittle for editing, viable for one-shot creation with frontier models (Claude 4.5 Sonnet, GPT-5.1 — see BPMN Assistant evaluation).

2. Structured intermediate representation

LLM emits a constrained format (JSON tree, Python POWL code, abstract syntax tree); a deterministic translator converts to BPMN XML. Reduces output complexity, enables schema validation. Examples in this wiki:

3. Atomic editing functions

Constrain the edit operation surface to a small fixed set (delete, add, move, redirect, update); the LLM uses function calling to compose atoms. Stronger guarantee of structural integrity. Currently only BPMN Assistant in the ingested corpus implements this.

Validation as a first-class layer

The architectural pattern that empirically dominates is structured IR + schema-validation retry loop: the validator's error message is fed back to the LLM as a new prompt, allowing iterative self-correction. This is the programmatic guardrail that distinguishes reliable LLM-modelling tools from brittle one-shot prompting.

Quality dimensions

Licardo et al.'s survey of competing tools partitions the evaluation problem along these axes:

Tool	Goal	Evaluation focus
ProMoAI (Kourani 2024)	NL → model + refinement	Conformance-based model quality
BPMN-Chatbot (Köpke & Safan 2024)	Efficient NL → BPMN	Correctness + token efficiency + acceptance
BPMNGen (Hörner 2026)	NL → BPMN + human-centred quality	Semantic alignment + cognitive load + acceptability + comprehension
BPMN Assistant (Licardo 2026)	Reliable NL → BPMN + edit robustness	GED/RGED structural fidelity + failure rate

No single tool yet covers all dimensions. Production deployment will likely combine structural-correctness validation (BPMN Assistant style) with human-centred usability evaluation (BPMNGen style).

Positioning vs Large Process Models vision

Kampik et al.'s LPM vision (§5, §6 Step 1) explicitly identifies LLM-augmented modelling as the most feasible near-term capability. LPM frames this work as one capability among many in a fused neuro-symbolic stack, not as a standalone product. Concrete tools like BPMN Assistant are evidence that LPM Step 1 is achievable now.

Positioning vs APM

APM Manifesto does not centre on LLM-modelling but acknowledges it as one of the LLM-augmented BPM capabilities. APM's contribution is the agent-execution paradigm; LLM-modelling sits in the parallel "modelling and analysis" lane.

Open challenges

Multi-pool / collaboration diagrams — current tooling (incl. BPMN Assistant) cannot lay out cross-organisational diagrams reliably.
Real-world enterprise narratives — public corpora (Friedrich 2011, PET) lack validated gold-standard models. Synthetic descriptions overstate model performance.
Soundness verification — token-replay conformance ≠ deadlock-freedom; soundness checks (concepts/soundness) are not yet integrated into LLM-modelling pipelines.
Usability validation — most evaluations are automated; no comprehensive HCI studies with non-technical analysts.
Multilingual / domain-specific terminology — performance degrades on non-English process descriptions or domain-heavy terminology.
Integration with process mining — closing the loop between LLM-generated models and discovered models from event logs is largely unexplored.

concepts/agentic-bpm · concepts/process-discovery · concepts/process-model-quality · concepts/soundness · concepts/declarative-process-modelling · concepts/conformance-checking

frameworks/bpmn · frameworks/dmn · frameworks/declare

sources/2026-licardo-bpmn-assistant — BPMN Assistant: JSON IR + atomic function-calling editing
sources/2024-kampik-large-process-models — LPM vision §5/§6 Step 1
sources/2025-varsani-neuro-symbolic-ai-sap-erp — parallel grounding pattern in ERP context
syntheses/llm-bpm-reading-list — reading list hub
syntheses/bpmn-modelling-practical-guide — constructive builder's guide; LLM tools should respect this rule set