• Evgenii Zharnakov, at IFSG


• 09.09.2026 12:15 pm
#banking

Financial organisations see multiple complex digital systems running in parallel, with large volumes of operational data needing to be processed every day. In these operations, different systems may contain different versions of the same fact, leading to inconsistencies that require painstaking manual verification. AI agents can help with this operational burden by implementing clear rules for how the system evaluates conflicting evidence and what happens when ERP, CRM, and primary documents do not agree. 

    When an ERP contains one transition amount, date, or deal parameter while CRM contains another, an employee is typically needed to inspect records or source documents to understand the problem. Separate ERP and CRM systems may record account balances differently, or track different dates for the same transaction. If an AI agent can continuously monitor and compare discrepancies between systems, hours of time can be saved. The tricky part is understanding what evidence the agent should rely on when systems disagree. 

    In this environment, AI agents should use original documents as a source of verification. When information across internal systems is inconsistent or incomplete, primary documents such as contracts, invoices, and statements can help establish a more reliable version of the facts. Modern open-source multimodal models with vision capabilities make this approach increasingly practical because they can process visual document content directly, extract key information, and support further reasoning over it. These models can be combined with workflow logic so that AI agents not only read documents, but also compare the extracted information with internal records, identify discrepancies, and support data correction. 

    The agent should never assume that one system is correct by default. Its first step should be comparing systems to search for discrepancies. If one is located, the agent should check additional evidence and search primary documents where necessary. Once all relevant information has been collected, the agent can send this summary and recommendation to an employee for implementation. The output should explain what information conflicted, which sources the agents checked, what evidence informed the agent’s recommendation, and what the employee should review or correct. The final decision should, of course, remain human.  

    In order to work effectively, agents need access to source data, layered model architecture, logging and manual override mechanisms, and must be careful about data protection. Fast and lightweight models should be used for instant checks, and more robust models for comprehensive operations. Using open-source models that can also be fine-tuned on the bank's own datasets within the bank's closed perimeter reduces the risk of data leakage and simplifies access management. 

Banks need explicit rules for how an agent evaluates conflicting evidence inside operational systems before they can rely on its recommendations. The most substantial benefits will come when you deploy an AI agent into a specific, well-described process. High-risk processes, especially those involving confidential data, regulatory judgments, or material financial impact, are better served by a closed-loop architecture based on open-source models deployed inside the bank’s perimeter. Low-risk or non-sensitive processes, by contrast, may justify the use of the most advanced external models. In this way, the AI strategy is not uniform across the bank but calibrated to the risk profile of each process and the security architecture of the organization. In my experience, the implementation of these agents has helped reduce manual reconciliation effort by 27% and improved early discrepancy detection by 95%. With clearly regulated advisory-only AI agents and human final decisions, financial institutions can more easily identify discrepancies in their information and become overall more productive.