The development of an in‐house accounting bot—an artificial intelligence (AI) assistant capable of generating internally structured bookkeeping double‐entry posting schemes—is explored in this paper. The processes of curating a suitable dataset, selecting, and fine‐tuning a seven‐billion‐parameter language model, categorized as a small language model (SLM) (SLMs typically refer to models with fewer than 10 billion parameters, whereas medium‐sized models often have 14B parameters, and large‐scale models exceed 70B), are described. A human‐evaluated benchmark is also presented to assess model performance. To achieve efficient supervised fine‐tuning (SFT), low‐rank adaptation (LoRA) was employed, significantly reducing memory requirements by using a small set of trainable parameters while maintaining model expressiveness. The process of backpropagation was further optimized using Unsloth, a high‐performance training framework designed for efficient video memory usage and flash attention mechanisms, which accelerates adaptation and reduces memory overhead. The model whose layers were updated is called QwenCoder2.5. It was selected with the presumption that it would be able to learn how to generate and examine bookkeeping patterns generated by accounting information system (AIS) over a 17‐year history. This proof of concept aims to support researchers and practitioners exploring the integration of generative AI in accounting by providing insights into both the benefits and challenges of AI‐driven automation in bookkeeping tasks. The study demonstrates how an SLM can be fine‐tuned on a proprietary dataset of journal posting schemes to assist accountants, auditors, and financial analysts while also facilitating synthetic data generation. Challenges related to AI, data preprocessing, fine‐tuning optimization, and evaluation methodology are introduced and examined.

This paper presents a systematic empirical evaluation of Retrieval-Augmented Generation (RAG) across 22 open-weight language models spanning five families—Qwen3.5, Llama 3.x, Gemma 3, Phi-4, and DeepSeek R1 distilled—on a domain-specific Croatian accounting benchmark derived from the publicly available mariozupan/rrif dataset. All models were deployed locally via Ollama on a single GPU server, reflecting privacy-preserving enterprise conditions. Two retrieval corpora were evaluated: a structured English Q&Aderived corpus (Corpus A, 522 accounts) and a Croatian corpus parsed from the RRiF chart of accounts (Corpus B, 2462 entries), the latter indexed under three embedding models, yielding five conditions and 116,600 queries in total. Four findings are reported. First, on Corpus A, RAG raises lookup Exact Match from near-zero to 94.9–96.2% with 18 of the 22 models converging into that band, with equivalence testing confirming practical equivalence for 87% of within-band pairs at a two-point margin; the four exceptions lie at 1B–3B parameters, but the boundary tracks instruction-following quality rather than scale. Because Corpus A shares its source with the evaluation questions, these results are an upper bound rather than an estimate of field performance. Second, embedding model choice dominates language model choice on the Croatian corpus, where a multilingual encoder raises average Exact Match by approximately 43 percentage points. Third, reasoningdistilled models show no consistent advantage on this task. Fourth, decomposing responses into mutually exclusive outcome categories shows that retrieval reduces the invention of account codes from 33.6% to below 0.3% under every condition tested, while residual error consists almost entirely of wrong-but-existing accounts whose source shifts from retrieval to selection as retrieval improves—a distinction that a binary hallucination measure conflates. Non-generative baselines outperform every model on Corpus A and ten of the 22 on Corpus B, indicating that a language model is warranted only conditionally for this task.

Supervised fine‐tuning (SFT) enables large language models (LLMs) to acquire domain‐specific knowledge, yet the resulting trade‐off between specialization and general capability preservation remains poorly understood. This paper investigates this trade‐off systematically by fine‐tuning three open‐source LLMs on double‐entry bookkeeping posting schemes and evaluating them against a comprehensive benchmark suite. The study contrasts full parameter fine‐tuning with parameter‐efficient QLoRA across dense and Mixture‐of‐Experts architectures, under real‐world GPU memory constraints. An industry‐standard evaluation framework ensures reproducible and transparent benchmarking alongside a custom domain‐specific task. Results suggest that the selected fine‐tuning configuration can have a more pronounced effect on outcomes than model size alone, with implications for practitioners selecting strategies for domain‐specific LLM deployment.