AI Large Models Redefine Non‑Retail Credit: A Deep Dive into Core Banking Processes

1. Building an Enterprise‑Grade AI Infrastructure

Faced with limited resources, the bank concentrated its technology budget on a unified AI foundation rather than scattered applications. It adopted a hybrid architecture, layered evolution, and platform empowerment to create a sustainable digital base. The infrastructure covers three pillars:

Compute: Private‑cloud deployment secures core data while exploring shared industry models for non‑sensitive scenarios, balancing compliance and cost.

Platform: An enterprise‑level intelligent agent development platform abstracts model differences and offers standardized APIs, reducing duplicate effort and lowering development barriers.

Data: Standardized collection and storage of multimodal data such as loan‑related images lay the groundwork for later asset conversion.

This staged, scenario‑validated approach moves AI from laboratory prototypes to scalable production.

2. Vertical Deep‑Diving into Non‑Retail Credit

Non‑retail credit, with large loan amounts, complex transaction structures, and heavy image material, lags behind retail credit in digitalization. The bank identified three pain points: manual due‑diligence report writing, labor‑intensive loan approval, and delayed post‑loan monitoring. To address these, AI was anchored to the entire credit lifecycle.

2.1 Pre‑Loan Due Diligence Reconstruction

Traditional due diligence required analysts to manually extract information from dozens of document types, taking over a week per case and risking omissions. The bank replaced end‑to‑end black‑box solutions with an engineered, modular pipeline: document preprocessing → retrieval → key‑value extraction → analysis generation . The framework includes:

Standardized template tooling: Zero‑code configuration splits reports into ~20 modules (company info, financial statements, top suppliers, etc.). A prompt engine translates Word templates into AI agents that automatically populate each module.

Quality‑control triangle: Each extracted field is tagged with confidence scores and color‑coded for quick review; one‑click traceability links results back to source files, enabling audit and model feedback loops.

Task‑scheduling engineering: Chained asynchronous calls, caching, and timeout retries ensure high performance and resilience, while human review validates AI‑extracted core data.

2.2 Loan Issuance Automation

Centralizing loan approval shifted workload from branches to the head office, increasing review volume and compliance demands. The bank built an automated chain: image classification → key‑element extraction → intelligent matching → penetrative analysis . Multimodal models classify and verify image completeness in real time; a KV extraction engine identifies invoices, contracts, and other critical elements, cross‑checking them against structured loan system data. Real‑time risk alerts flag duplicate invoices, abnormal contract terms, or mismatched amounts. Engineering measures such as asynchronous parallel processing, result caching, confidence grading, and one‑click traceability support high‑concurrency scenarios. The solution processes over 2,000 model calls daily, achieving 98% invoice extraction accuracy and 96% usage‑type classification accuracy, while feeding structured data into downstream risk monitoring.

2.3 Post‑Loan Risk Early‑Warning

Post‑loan management traditionally relied on periodic manual checks and rigid rule triggers, missing timely signals of operational anomalies. Leveraging the accumulated data assets and multimodal capabilities, the bank plans a proactive risk‑prediction system that combines large‑model pattern recognition with machine‑learning quantitative assessments. Core initiatives include:

Building a forward‑looking risk‑warning engine that penetrates corporate group transactions, supply‑chain cash flows, and industry‑level shocks.

Establishing a quantitative risk‑return evaluation loop that continuously back‑tests model accuracy, false‑positive rates, and interception effectiveness, driving iterative model and rule updates.

Integrating post‑loan warning data with earlier due‑diligence and issuance assets to create a full‑lifecycle customer risk view.

3. Joint Innovation with Technology Vendors

Recognizing that small‑to‑mid‑size banks cannot shoulder large‑model R&D costs alone, the bank co‑developed an explainable, auditable AI application framework with leading technology partners. The collaboration embeds multimodal models directly into paper‑less workflows, captures native digital documents (financial statements, contracts, invoices), and yields two benefits: vendors gain vertical scenario expertise, while the bank internalizes core multimodal and KV‑model capabilities, amassing high‑quality training data for future fine‑tuning.

4. Data Asset Construction

The competitive edge in the model era lies in high‑quality, business‑embedded data assets. The bank has curated tens of thousands of annotated financial statements, invoices, and contracts, forming a comprehensive multimodal training set and relational graph. These assets not only boost current AI precision but also enable rapid model fine‑tuning as base models evolve, compressing business response cycles to a weekly cadence and establishing a sustainable strategic advantage.

5. Roadmap to Full‑Bank AI Enablement

Following a single‑point breakthrough, the bank will extend AI vertically across the entire credit chain and then horizontally to other scenarios. By 2026, the “Xing Xiao Zhi” enterprise‑level AI agents (43 already deployed) will support front‑office, risk, R&D, and operations roles, driving labor‑hour substitution, coding efficiency, test coverage, and architecture decisions. An AI‑R&D management system will automate code assistance, test case generation, and end‑to‑end log analysis, ensuring stable, long‑term AI adoption.