Banks Bear Responsibility for AI Loan Decisions, Forum Panelists Say

Tashkent, Uzbekistan (UzDaily.uz) — Lending banks remain fully responsible for credit decisions generated by artificial intelligence, even when portions of the technological infrastructure are outsourced to external vendors, according to panelists speaking at the Silk Road Finance and Technology Forum in Tashkent on 25 August 2026.

The session, titled "Can We Trust AI Without Trusting What Sits Beneath It?", was moderated by James Boey, Head of Asia and Co-Lead of Forums at the Global Finance & Technology Network.

Discussion focused on a hypothetical scenario involving a Samarkand entrepreneur whose loan application was rejected by an AI-based system. In the illustrative example, the bank verifies applicant identity digitally and utilizes multiple data sources, including a third-party model and offshore cloud infrastructure, but cannot explain the exact cause of rejection to the client.

Panelists debated which component of the technology stack requires prioritized attention. Aleksandr Simonenko, Chief Technology Officer at National Intellectual Solutions (Gorgona), highlighted data as the single layer that cannot be imported from abroad, noting that data errors propagate across all subsequent system levels.

Arvind Sankaran, Senior Fintech Expert at the Asian Development Bank, pointed to AI models and explainability as primary concerns, assessing data availability in Uzbekistan as already acceptable.

Sheruan Bashar, Deputy Head of the Main Information Center under the Central Bank of Uzbekistan, identified the lack of control over decision-making processes as the central issue. Davit Melikidze, CEO of UzCard, viewed data and models as equally critical, stating that data must remain localized and models must undergo local validation regardless of where they were developed.

Yanan Wu, CEO and Founder of Surfin Meta Digital Technology, also highlighted the model layer, emphasizing that modern systems built on foundation models with trillions of parameters must maintain transparency and ideally operate on open-source code.

Melikidze outlined five criteria guiding UzCard's AI implementation: understanding how the model operates, auditability, adaptability, the ability to override decisions, and process recoverability in case of failure.

Simonenko added that autonomous AI agents require protective safeguards that block or predict potentially problematic actions before execution or route them for human approval. He identified three essential management elements: control over user identification and localized data, full decision logging, and the capability to switch between model providers or payment rails.

Wu noted that his firm provides AI-driven fintech and financial inclusion services to 100 million clients across 13 countries, including Uzbekistan. He stated that rejections often stem from an inability to evaluate risk for applicants lacking credit histories, prompting the use of alternative data. He advocated for creditworthiness assessment access as a fundamental right, noting his service calculates social credit ratings in 30 seconds and credit scores within one minute.

Melikidze differentiated regulatory responsibilities, stating that regulators must oversee data protection, whereas individual lending institutions hold sole responsibility for client relationships.

Sankaran noted that business loan issuance involves data, underwriting, real-time decisioning, and post-issuance monitoring layers encompassing around 40 distinct applications utilizing AI.

Bashar stated that opaque decision-making systems must be avoided, as banks act as unexplainable "black boxes" during application rejections. He stressed that regulators cannot adopt a hands-off approach, though acknowledging that modern probabilistic AI models remain complex. He clarified that credit decisions are expected to rely on explainable traditional machine learning models rather than large language models.

Sankaran added that market participants combine AI models with rule-based systems and mandatory human-in-the-loop validation when evaluating first-time borrowers lacking credit history. Wu concluded that consumer data ownership serves as the foundation of trust, allowing clients to choose what data to share in exchange for financial services.