Artificial intelligence (AI) is becoming increasingly common in commercial lending, but banks adopting these tools risk overlooking a fundamental weakness: models can produce convincing answers without having access to the complete information needed to make reliable assessments.
Business lending requires financial institutions to understand borrowers across multiple entities, ownership structures and lending products. A single customer relationship may involve commercial and industrial (C&I) lending, commercial real estate (CRE), construction finance and treasury services. Platforms such as nCino, which provides cloud-based banking technology, support financial institutions in managing these relationships, where assessing borrowers requires visibility into financial performance, outstanding exposure, loan covenants, collateral and previous credit decisions.
AI is increasingly being deployed across these processes. Applications include intelligent document processing during loan intake, automated financial spreading, natural-language responses to credit policy queries and covenant monitoring. By automating administrative work, these applications can reduce the time lending teams spend entering information, requesting documentation and responding to application status enquiries.
However, the effectiveness of these tools depends on the information available to them. Models working with isolated applications, documents or systems may produce assessments without accounting for important details elsewhere in the borrower’s relationship with the bank. Unlike experienced lenders, who can recognise missing information and investigate further, AI systems may present incomplete findings with a level of confidence that makes their limitations difficult to identify.
This creates a challenge for financial institutions evaluating commercial lending technology. Individual AI applications may demonstrate strong capabilities, but their usefulness is restricted when they cannot access information across the wider credit relationship.
Source link
