Op-ed
The success of financial agents depends not on 'intelligence' but on document interpretation and accountability control
Korea Deep LearningPublished

Kim Dong-hyun, head of the Future AI Development Team at Korea Deep Learning, contributed "The success of financial agents depends not on intelligence but on document interpretation and accountability control" to AI Times.
The paradox of intelligence and the wall of operations
The piece defines 2025 as the year the potential of generative AI was explored through proofs of concept (PoCs), and 2026 as the time to work out a survival strategy after running into the walls of regulation and operations. Financial companies adopting LLMs hoped for scenarios in which AI would go beyond automating customer consultations to take on asset management, loan underwriting and risk management, but cases that fully automated real business processes are rare. The author finds the cause not in a lack of AI intelligence, but in general-purpose AI failing to meet the "data integrity" and "clear accountability" that finance requires.
"The rate of actual deployment in real work is still stuck at a threshold of around 10 to 20%."
— Kim Dong-hyun, Head of the Future AI Development Team, Korea Deep Learning
The weakness of RAG and deterministic data
The piece names the limits of document parsing as the biggest technical obstacle. How accurately PDFs, complex terms and conditions, and underwriting documents with dozens of merged cells are turned into data has become the bottleneck that decides the performance of RAG as a whole. RAG is good at finding context across documents, but when a table structure falls apart or a layout is misrecognized, misreading a single figure can turn into a major risk. The diagnosis is that what financial automation needs is not plausible search results but error-free "deterministic data."
The responsibility map and the governance of "approval"
The author names the responsibility map as another condition. The responsibility map is a system that holds executives clearly accountable when a financial incident occurs, and "black-box intelligence" with an opaque decision process becomes a critical risk under this system. The Financial Services Commission, too, is stressing the principle that AI is an aid and that employees bear final responsibility.
The piece explains that Korea Deep Learning's DEEP Agent recognizes both the visual structure of documents and the intent of the work to produce its results, freeing people from repeated review so they can focus on final approval. Instead of removing the human role, it raises people to the role of approver, creating automation whose results executives can sign off on with confidence.
A controlled structure becomes infrastructure
The conclusion is that the competitiveness of future financial AI lies not in the ability to generate text but in the ability to bring incomplete records under control as a complete data structure. The author names "determinism," which controls the probabilistic variability of generative models to produce a single value a system can accept, as the core value, and recommends that the financial sector invest in connectivity between systems and in data interpretation rather than in model intelligence.
"It is time to stop asking 'which retrieval technology do you use' and start asking 'what deterministic structure do you have.'"
The full piece was published in AI Times.
About Korea Deep Learning
Korea Deep Learning (KDL), founded in 2019, is an enterprise AI agent company. Built on KDL Frontier, its own vision-language model (VLM) specialized for documents, KDL makes AI agents that take on and complete the document work of businesses and public institutions. Its main products are DEEP OCR, DEEP Parser, DEEP Agent and DEEP Ops.







