A multiagent framework for trustworthy AI financial advisory for displaced Ukrainians in Poland

Large language models (LLMs) offer opportunities for scalable financial guidance, but hallucinations, numerical errors, insufficient grounding, and unstable decisions limit their use in high-stakes financial settings. This study develops and evaluates a multiagent AI framework for retirement-oriented financial guidance using synthetic profiles modelled on displaced Ukrainians in Poland. The framework combines retrieval-augmented generation (RAG), structured profile generation, deterministic budget calculations, and an architecturally separate Evaluator Agent. The primary computational evaluation comprised 500 synthetic advisory cases. Additional experiments included a paired no-RAG ablation on 50 fixed profiles and a repeated-run stability analysis of 100 fixed profiles replayed five times. Under the predefined automated routing criteria, 327 cases were accepted, 170 required review, and 3 were flagged. After the deterministic budget overwrite, a separate deterministic arithmetic check found zero profile-to-answer mismatches. In the paired ablation, the mean final composite score was 8.40 with RAG and 7.08 without RAG. The difference was driven primarily by larger deterministic penalty deductions and grounding-gate failures without RAG, while the mean of the five LLM rubric dimensions did not improve with RAG. In repeated runs, only 14% of profiles retained identical routing across all five replays. The deterministic arithmetic pathway remained invariant, while generation, retrieval, model-based evaluation, and threshold-based routing varied across repeated execution. The framework is therefore interpreted as a controlled computational testbed and quality-assurance architecture rather than a validated autonomous financial advisor.