LLM-avusteinen ostoreskontran maksuaikataulutus Odoossa

This thesis examines how large language models (LLMs) can support accounts payable (AP) scheduling within the Odoo 19 Enterprise ERP system. Following a design science approach, it develops PayOpt LLM, a prototype that integrates synthetic AP data, deterministic scheduling logic, and an LLM reasoning layer under strict schema validation and post-processing controls. The system generates draft-only payment plans within Odoo, enforcing liquidity constraints and due-date penalties through deterministic code. Discount-window logic is included at the data and interface level but is disabled in the final evaluation, and discount outcomes are therefore not assessed. The prototype is evaluated using 100 deterministic synthetic scenarios and compared against a FIFO-by-due-date baseline and a mixed-integer linear programming (MILP) benchmark. Results show that the LLM-assisted planner consistently preserves higher cash balances and reduces peak daily outflows, but does so primarily by deferring payments beyond the planning horizon. As a consequence, lateness outcomes under the LLM planner are substantially worse than under FIFO. The MILP benchmark achieves modest reductions in lateness relative to FIFO, providing a reference point for what is achievable under the same constraints. The findings demonstrate that a schema-constrained, tool-augmented LLM can reliably translate high-level payment policies into valid, auditable scheduling inputs, but that improved scheduling performance is not automatic. In its current form, the LLM planner functions as a conservative, liquidity-first decision policy rather than a due-date-optimising alternative. The thesis contributes an ERP-embedded artifact, a reproducible evaluation framework, and empirical evidence on the strengths and limitations of LLM-assisted decision support in accounts payable management.