Banks Hand Fraud Victims' Paperwork to an AI Agent

Korea Deep Learning, an enterprise AI company, says its software cut financial document processing time by 88 percent at South Korea's Industrial Bank of Korea. The documents in question? Voice phishing relief cases, the paperwork of people who've just been robbed by scammers. The figure comes from a company press release dated Seoul, Nov. 10, 2026, distributed by PR Newswire. The release was issued by the company, and the performance claims are the company's own. No outside auditor signed off on any of it.

Who Has the Power

The system is called DEEP Agent for Finance, a VLM-based AI agent built to automate what the release calls "document-intensive procedures" for voice phishing relief. It handles document recognition, information extraction and crime-type inference. These are tasks that, per the release, "traditionally require substantial manual review by financial institution employees."

Read that again. The work being automated was done by people. Bank workers who sat with the files of scam victims and went through them. The release doesn't say what becomes of that labor. It says the solution was deployed inside IBK's closed network, with employee review kept for cases "requiring additional verification." Humans stay on for the exceptions. The routine goes to the machine.

The company reports 100 percent accuracy in printed-text recognition and document classification, and handwriting recognition above 97 percent, "based on the project's evaluation criteria." Whose criteria? The company's partner, a state-linked bank, and the company itself, grading its own exam with actual business data. All key performance indicators exceeded their targets, the release says.

What They're Calling "Efficiency"

CEO Ji-hyun Kim framed it this way: "Financial institutions handling voice phishing relief cases often spend considerable time processing documents before they can focus on supporting customers." The pitch is that the saved time flows back to victims. "We aim to help financial institutions automate repetitive tasks, improve operational efficiency and allocate more resources to consumer protection," Kim said.

Nothing in the release commits any bank to doing that. "Allocate more resources" is an aim, not a contract term. And the company's own roadmap shows where the energy is going: converting evaluations into paid proofs of concept and supply contracts, building local partnerships and sales channels from 2027, and pushing into insurance and securities.

Where This Is Headed

KDL has already completed a separate proof of concept with Shinhan Vietnam Finance Company, a local subsidiary of Shinhan Card, through the Shinhan Financial Group's open innovation program. It has supplied AI solutions in Vietnam and Taiwan region. In Japan, it's in discussions about potential proofs of concept with nine financial institutions. Some prospective partners have signed non-disclosure agreements and are evaluating the technology with actual business data.

Non-disclosure agreements. Actual business data. Behind closed doors, institutions are feeding their records to a vendor that serves more than 80 companies and public institutions in South Korea. Founded in 2019, the company is now seven years old and already eyeing the whole Asia-Pacific financial sector.

What's Missing

The release has plenty of numbers about speed and accuracy. It has nothing from the bank workers whose review the system replaces. Nothing from the scam victims whose files it processes. Nothing on how a victim's sensitive documents are protected, or who answers when the model misclassifies a case. The only voice quoted is the CEO's.

The people at the bottom of this chain are the ones who got scammed and the ones who processed their paperwork. Their experience didn't make the press release. The sales pitch did.