South Korean enterprise AI company Korea Deep Learning (KDL) is expanding its financial services ambitions after a proof of concept with Industrial Bank of Korea (IBK) showed an 88% reduction in financial document processing time. The project used KDL’s vision-language model (VLM)-based DEEP Agent for Finance to automate document handling for voice phishing relief cases, combining text recognition, information extraction and crime-type inference. The system was deployed within IBK’s closed network, with employees retaining review responsibilities for cases requiring additional verification.

VLM Agent Automates Voice Phishing Relief Paperwork

The IBK project addresses a time-sensitive process in which administrative delays can affect how quickly a bank handles support for voice phishing victims. Traditionally, employees had to review submitted documents, identify relevant information and enter details into internal systems. KDL’s system is designed to automate several of those steps, including document recognition and classification, information extraction and inference about the type of crime involved.

The technology uses a vision-language model (VLM), which processes visual and textual information together. This is relevant to financial documents because forms, identity records and supporting evidence can contain printed text, handwriting, tables and varied layouts. Conventional optical character recognition (OCR) primarily converts visible characters into machine-readable text; a VLM-based workflow can also use document context to support classification and information interpretation.

KDL reported 100% accuracy for printed-text recognition and document classification during the IBK evaluation, with handwriting recognition exceeding 97%. The company said all key performance indicators exceeded their targets using actual business data.

Those figures are promising, but they should be understood as results from a specific proof of concept rather than evidence of equivalent performance across every document type, institution or operating environment. The announcement does not disclose the size and composition of the evaluation dataset, the detailed test methodology or independent validation of the reported accuracy.

Closed-Network Deployment Addresses Financial Data Concerns

The system was deployed within IBK’s closed network, an important aspect for institutions handling identity documents and sensitive financial consumer information. Keeping processing within a bank-controlled environment can help address concerns around external data transfers, although security also depends on access management, system monitoring, model governance and data-retention controls.

KDL says employee review remains part of the process when additional verification is required. This human-review provision is significant in a consumer protection workflow: an AI system can help organize documents and identify likely information, but decisions affecting a customer’s claim or access to relief require appropriate accountability.

The distinction between automation and decision-making is especially important in fraud-related operations. Incorrect document interpretation or crime classification could delay assistance or trigger unnecessary investigation. Financial institutions evaluating such systems will need to examine confidence thresholds, escalation procedures, audit trails and how errors are corrected.

A separate report by South Korea’s EDaily said the IBK project ran from April to August 2026 and that the bank was planning to apply the technology in live operations in the first half of 2027. That points to a transition from technical validation toward operational deployment, although the live rollout remains a future step rather than a completed implementation.

KDL Builds an Asia-Pacific Financial Services Pipeline

The IBK result is part of KDL’s wider effort to expand beyond its domestic market. The company has also completed a separate proof of concept with Shinhan Vietnam Finance Company, a subsidiary of Shinhan Card, through a Shinhan Financial Group open innovation program.

The Vietnam project provides another indication of KDL’s attempt to adapt document automation to different financial operating environments. Seoul Economic Daily reported that the work tested the extraction and organization of financial document data and its connection to downstream task automation. Seoul Economic Daily

In Japan, KDL says it is discussing potential PoCs with nine financial institutions. It is also engaging prospective customers in banking, insurance and consumer finance across the region. Some institutions have signed non-disclosure agreements and are evaluating the technology using actual business data.

These discussions are not yet equivalent to customer contracts or confirmed deployments. KDL’s stated commercial plan is to convert technical evaluations into paid PoCs and supply agreements, while establishing local partnerships and sales channels from 2027. The company also plans to extend its document intelligence capabilities into insurance and securities, where institutions process large volumes of forms, applications and supporting records.

Document Intelligence Becomes an Enterprise AI Use Case

KDL’s approach reflects a wider enterprise AI trend: applying multimodal models to structured business processes rather than limiting generative AI to chat interfaces and content creation. Financial institutions have substantial volumes of documents that must be interpreted, checked and transferred into operational systems. Automating parts of that work could reduce processing time and free employees to handle exceptions and customer-facing tasks.

Research presented at the 2026 EMNLP Industry Track also highlights the importance of evaluating vision-language systems beyond headline accuracy. Its study of document extraction examines trade-offs involving robustness, cost and governance, underscoring why production suitability depends on more than a single performance score.

For banks, insurers and securities firms, document automation must work with varied formats, languages and legacy systems while preserving privacy and traceability. A successful pilot in one workflow does not automatically establish readiness for broader deployment. Institutions will need to assess model performance against their own documents, operational rules and regulatory obligations.

KDL’s IBK project gives the company a concrete financial-sector reference as it pursues regional opportunities. The next test will be whether its reported PoC results translate into sustained production use, measurable operating savings and repeatable deployments across different markets.

Market Landscape

Financial institutions are increasingly applying AI to document-heavy processes where manual review can slow customer service, compliance and operational decisions.

• Vision-language models in finance: VLMs can combine visual document understanding with text extraction, supporting workflows involving scanned forms, handwritten information and complex layouts.
• Fraud response and consumer protection: Automating document intake and classification may help banks process fraud-related claims more quickly, provided human review and escalation mechanisms remain in place.
• Private AI infrastructure: Closed-network and on-premises deployment options are relevant to institutions with strict requirements for data privacy, security and operational control.
• Regional financial technology expansion: KDL’s projects in South Korea and Vietnam, alongside discussions with Japanese institutions, illustrate its strategy to commercialize document intelligence across different financial markets.

The commercial challenge is moving from successful PoCs to repeatable deployments. Accuracy, integration costs, local-language support, auditability and measurable productivity gains will influence whether financial institutions adopt these systems at scale.

Top Insights

• KDL’s IBK proof of concept reported an 88% reduction in document processing time for voice phishing relief procedures.
• DEEP Agent for Finance combines document recognition, information extraction and crime-type inference using a vision-language model.
• The system achieved 100% printed-text recognition and document classification accuracy, with handwriting recognition exceeding 97% under project evaluation criteria.
• KDL has completed another financial services PoC with Shinhan Vietnam Finance Company and is discussing potential projects with nine Japanese institutions.
• The company’s next commercial objective is converting regional technical evaluations into paid PoCs, supply contracts and local partnerships from 2027.

Get in touch with our fintech expert