The central issue of this study arises from the deep gap between the realities of professional practice and the existing scientific literature; currently, we witness that in a significant number of organizations, artificial intelligence (AI) is used as a process executor at the core of accounting operational cycles (especially in the areas of accounts payable, accounts receivable, and bank reconciliation), while the lack of a coherent framework for designing these systems has seriously challenged their compatibility with internal control logic and auditability requirements. The aim of this study is to develop, present, and evaluate a control-compliant AI architecture; an architecture designed in a multi-layered manner and covering all aspects from data acceptance and preparation, the core of the machine learning (ML) model, and the mechanism of confidence thresholds, to the auditable logging layer and model lifecycle governance. Methodologically, the present study is based on a design science approach; That is, first, by purposefully combining the literature on accounting information systems (AIS), internal control, and explainable artificial intelligence (XAI), a conceptual model and design principles have been derived. Then, this architecture has been implemented in the form of an application case focused on automated invoice coding and bank reconciliation and evaluated on historical data in a quasi-experimental manner. The results clearly show that achieving high operational efficiency and a desirable level of control simultaneously is only possible if confidence thresholds are consciously defined, human-in-the-loop flow design is designed, and transaction granularity is recorded; on the contrary, the unconditional expansion of automation, although it may slightly improve speed, will significantly weaken traceability, segregation of duties, and audit reliability.