Neuro-Symbolic AI in SAP ERP — Varsani 2025 (year unverified)

Industry-flavoured paper from a SAP Solution Architect (Sparsh Varsani, Cognex; varsanisparsh@gmail.com) presenting a neuro-symbolic middleware that contextualises LLMs with structured SAP ERP data. The paper has IEEE-style conference layout but no explicit venue or publication date on the PDF; references include 2025 IEEE conference papers, so the work is most likely a 2025 industry preprint or workshop paper. Treat citations as unverified until venue is confirmed.

Core idea

Pre-trained LLMs alone cannot operate over rule-driven, schema-strict ERP environments such as SAP S/4HANA — they hallucinate against structured master data, mis-map user intent to schema entities, and produce outputs that fail authorisation and rule-compliance checks. Symbolic-only systems are rigid and fail under naturalistic user input. The proposed solution is a hybrid three-layer architecture:

Neural layer — pre-trained LLM (GPT-4 + LLaMA 2 variants, BERT) that parses natural-language queries.
Middleware — schema alignment, context enrichment via RAG, embedding-based grounding against the SAP ERP schema graph.
Symbolic backend — SAP HANA + ABAP rule engine with forward-chaining rules over SAP Reference Models, authorisation roles, and process ontologies.

LLM outputs flow through the middleware to be grounded into schema-validated triples (table, attribute, constraint) via a grounding function G(q) = {(t_i, a_j, c_k) | φ(q, S)} where φ performs hybrid embedding-similarity + symbolic validation. Rules of the form r: ⋀ Pᵢ(x) → Q(x) are then forward-chained; only validated outputs reach SAP HANA execution. On rule violation, the system returns correction suggestions rather than executing.

Methodology

Schema linking and grounding

Contrastive learning between query embeddings and ERP graph embeddings establishes the (query → table-attribute pair) mapping.
RAG retrieves similar historical ERP queries from a vector store; relevant examples are appended to the prompt context.
The grounding equation φ(q, S) = arg max cos(q⃗, e⃗_{t,a}) · 𝟙_valid(t, a) selects only semantically similar AND domain-valid table-attribute pairs.

Reasoning layer

A symbolic rule base R = {r_1, ..., r_n} enforces business logic via forward chaining. SAP-specific symbolic knowledge sources are: SAP Reference Models, authorisation roles, process ontologies. Example: an order-amount > $1000 rule triggers an approval action only if all prerequisite conditions are satisfied.

Training and fine-tuning

The LLM is fine-tuned on a curated dataset D from SAP user logs, support tickets, and ERP documentation. Optimisation uses Reinforcement Learning from Human Feedback (RLHF) with a reward function R_user quantifying both correctness and user satisfaction.

Pipeline (Algorithm 1)

Generate prompt from query + schema context + historical examples.
Generate neural output q′ = π(p).
Ground q′ into ERP schema via the grounding function.
Validate via symbolic rules.
If valid → execute on SAP HANA, return result r.
Else → return error with correction suggestions.

Evaluation

Setup: SAP S/4HANA (12.1M records, 30 structured tables) across MM (Materials Management), SD (Sales & Distribution), FI (Financial Accounting). LLM inference: Azure OpenAI GPT-4 + Hugging Face LLaMA 2. Middleware: Python 3.10, FastAPI, SAP Cloud SDK.

Metrics: Query Translation Accuracy A_trans (%); Explainability Score E_score (1-5 expert-rated); Execution Latency T_exec (s); User Trust Rating U_trust (1-10).

Baselines vs Proposed (Table II):

System	Accuracy	Explainability	Latency	Trust
Rule-Based Bot	62.4%	3.9	1.8s	5.6
LLM-Only	74.8%	2.7	1.2s	6.1
Symbolic-Only	71.2%	4.3	2.3s	6.4
Neuro-Symbolic	89.6%	4.6	1.5s	8.7

The neuro-symbolic system trails only LLM-Only on latency (the symbolic validation step adds ~0.3s) but dominates every other dimension. Ablation confirms that removing RAG, symbolic validation, or schema grounding each significantly degrades accuracy — they are core architectural components.

Per-module breakdown: Highest accuracy is achieved on MM (~91%) due to dense schema structure and well-defined ontologies; FI lags slightly (~86%).

Case studies:

Procurement query resolution — "List pending purchase orders over $100,000" correctly translated to a compound SQL join between EKKO and EKPO with amount + approval-status filters.
Financial audit trail generation — produced full traceability chains for expense postings, rated highly by auditors on interpretability.
Sales forecast augmentation — RAG-enhanced retrieval significantly improved report contextuality vs symbolic-only baseline.

The paper explicitly cites Kampik et al. 2024 (Large Process Models) [refs 2 and 7] as conceptual precedent — Varsani's system can be read as a concrete enterprise-system instantiation of the LPM vision: a process-aware (here: ERP-aware) software stack fusing fine-tuned LLM, knowledge graph, RAG, and symbolic reasoning. The differences:

Scope: LPM addresses the entire BPM lifecycle; Varsani's system addresses ERP query/operation translation specifically.
Symbolic backend: LPM remains generic about the symbolic substrate; Varsani commits to SAP HANA + ABAP rule engines as the concrete reasoning layer.
Maturity: LPM is a vision paper; Varsani reports an empirical evaluation with quantitative baselines.

Other prior art positioned in §II:

NSEA (Lin & Lin 2025) — quantum-safe cryptography + neuro-symbolic for ERP; conceptual, no SAP-specific deployment model.
Bosch industrial neuro-symbolic AI (Zhou et al. 2022) — manufacturing context, static-document focus, no live transactional integration.
Miladinovic 2024 — LLMs over customer-specific SAP knowledge objects via static document embeddings; no real-time data sync or symbolic validation.

Limitations and open questions

Year and venue unverified — cannot confirm peer-review status.
Author works at "Cognex" but writes as "SAP Solution Architect"; potential vendor bias in evaluation framing.
Manually curated domain ontologies and schema mappings — future work proposes graph-based ontology induction to reduce setup overhead.
No multi-tenant / federated-learning evaluation despite future-work mention of multi-subsidiary deployments.
No comparison to Kampik et al.'s LPM as a baseline despite explicit conceptual citation.

Connections

Concepts: concepts/agentic-bpm · concepts/business-process · concepts/concept-drift · concepts/normative-frame

Frameworks: frameworks/bpmn (via process ontology grounding) · frameworks/dmn (via business-rule reasoning)

Related sources:

sources/2024-kampik-large-process-models — directly cited and extended; Varsani's architecture is a SAP-specific LPM instantiation.
sources/2026-licardo-bpmn-assistant — parallel: BPMN Assistant grounds LLMs in BPMN syntax; Varsani grounds them in ERP schema. Both rely on a deterministic intermediate validation layer.
sources/2026-calvanese-agentic-bpm-manifesto — APM Manifesto's conversational-actionability and explainability capabilities; Varsani operationalises both for ERP contexts.
sources/2023-dumas-ai-augmented-bpms — ABPMS adaptable/explainable/context-sensitive properties.