Practical interpretation of Woori Bank AI agent banking: Operational standards that must be designed first when putting 175 agents into the financial field

Publication date: 2026-04-07 | Category: Development information

1) Problem definition

The most difficult point when introducing generative AI in the financial sector is not choosing a model, but deciding how far to open business responsibility, internal control, and system connection. Chatbots only need to answer well, but agents are directly involved in customer service and internal work flow, so the cost of failure is much higher.

Woori Bank’s ‘AI Agent Banking’ promotion, which was unveiled on April 7, 2026, directly demonstrates this point. Samsung SDS was selected as the preferred bidder for a project that connects Woori Bank's various business systems and applies more than 175 AI agents to 29 core tasks in 5 major areas. It plans to first introduce about 90 by the end of the year and expand by August 2027.

This article is not a simple news summary, but rather summarizes implementation standardsnecessary for banking, insurance, securities, and fintech practitioners to make actual adoption decisions by referring to this case. In particular, we focus on the order in which tasks should be designed when handling tasks with different sensitivities such as customer service, corporate credit, asset management, and internal control at the same time. Conversely, this frame may be overkill for a personal finance chatbot experiment or single FAQ automation level.

2) Evidence and comparison

Compiling reports from AI Times, Money Today, and Maeil Business Newspaper, the key figures for this project are relatively clear. The target is 29 core tasks in 5 major areas, the scale is more than 175 AI agents, the expected effect is work processing speed of about 30% Improvement, the schedule is start in May 2026, first release of 90 items in December 2026, full expansion in August 2027. In addition, Samsung SDS announced that it will connect multiple language models and existing banking systems based on its own platform FabriX and build an operational data management system together.

The important point in practice is that this is not a ‘chatbot upgrade’. The bank's intention to add more than 175 agents is closer to subdividing role separation by work unit, approval boundary, log tracking, and exception handling rather than a single conversation UI.

• Cost basis: Costs for system connection, log storage, and quality monitoring may be greater than the model usage fee.
• Time basis: FAQ Chatbot takes weeks, RPA takes several months, and agent system takes longer when including authority/audit design.
• Accuracy criteria: In customer-facing work, hallucination prevention and approval flow are more important than answer quality.
• Difficulty level: AI agents are more difficult to connect to backend systems and control operations than models.

3) Step-by-step execution method

1. Separate work into three layers.
   Must be divided into customer contact work, judgment support work, and internal execution work. For example, customer consultation should be divided into explanatory type, corporate credit into analysis assistance type, and business automation into execution type to avoid forcing the same agent policy.
2. Document the scope of permissions per agent.
   It is safe to truncate permissions to four or more levels, such as View Only, Draft Only, Approval Request Only, and Limited Execute. In the financial sector, ‘automation of steps before human approval’ must come first rather than ‘automatic execution’.
3. Organize existing system connection methods with API priority.
   When an agent is attached to a credit system, CRM, consultation system, or internal control system, API linkage should be prioritized over screen automation to facilitate failure tracking and auditing. It is possible.
4. Fix the task success criteria in numbers.
   Example: Consultation summary accuracy 92% or higher, document classification error rate 3% or lower, approval conversion time reduced by 20%, tool call failure rate 2% or lower, missing sensitive information masking 0 cases.
5. Create logs and reproducibility first.
   You must leave all the prompt, input document hash, model version used, tool called, final output, and approver information. Without this item, expansion of the internal control area is virtually blocked.
6. The scope of the first release starts with low-risk tasks.
   The schedule to first release about 90 items by the end of the year is actually a strategy of putting low-risk counseling assistance, summary, and classification tasks first and putting high-risk tasks later. It’s reasonable.
7. Predefine outage conditions.
   Rollback conditions that automatically place the agent in a read-only or stopped state when an elevated hallucination rate, specific system response delay, policy violation output, or missing audit log occurs. Required.

#Financial sector AI agent PoC check example
- Layer A: Customer consultation assistance -> Answer draft/consultation summary/complaint classification
- Layer B: Judgment assistance -> Corporate loan document summary/risk point extraction
- Layer C: Internal execution -> Create approval request/issue ticket/routing follow-up work
- Common measurements: accuracy, acceptance rate, hallucination rate, call failure rate, audit log completeness

4) Mistakes/Pitfalls

1. Pitfall: Expanding on chatbot success experience
   Prevention: Answer-type AI and action-based agent must be treated as separate programs
   Recovery: If something goes wrong, first remove execution permission and revert to draft creation mode.
2. Pitfall: Mistaking the number of agents for performance
   Prevention: Make sure that the number 175 is actually a sustainable operating unit.
   Recovery: Consolidate or retire agents based on frequency of use and error rate.
3. Pitfall: Incorporating log-less automation into internal control tasks
   Prevention: Structure in which model input, output, and approval history can be monitored. Design:
   Recovery: Immediately revert missing log sections to a manual process and perform a re-audit.
4. Pitfall: Abandoning RPA completely
   Prevention: Rule-based procedures may actually make RPA more reliable
   Recovery: AI reseparates roles into interpretation and classification, while RPA reseparates roles into execution and formal repetition.

5) Execution Checklist

• Target tasks were divided into customer contact, decision assistance, and internal execution
• Permissions for each agent were divided into inquiry/draft/approval request/restricted execution
• Existing system linkage was designed with an API-first structure
• Defined KPIs for accuracy, hallucination rate, call failure rate, and approval turnaround time
• Designed to log prompts, model version, tool call, and approver
• The scope of the first release was limited to low-risk tasks
• Created a rollback condition that stops immediately in case of policy violation or log omission
• Documented standards for role division between RPA and AI agents

Definition of Done: If the scope of authority, approval flow, KPI, audit log, and rollback conditions for each primary target task are verified with documents and tests, the minimum completion criteria for the AI agent pilot in the financial sector have been met.

6) Reference

• AI Times - Samsung SDS, Woori Bank to build ‘AI Agent Banking’ (Publication date: 2026-04-07, Confirmation date: 2026-04-07)
• Money Today - Samsung SDS, Woori Bank to build ‘AI Agent Banking’ (Publication date: 2026-04-07, Confirmation date: 2026-04-07)
• Maeil Business Newspaper - Samsung SDS to plant ‘working AI agents’ at Woori Bank (Publication date: 2026-04-07, Confirmation date: 2026-04-07)

7) Author Viewpoint

My judgment is clear. The key point of this news is not that ‘banks also use AI’, but that the financial sector has now begun to verify the agent operation model for each business unit beyond answer-type AI. The number 175 is a promotional point, but it is also a warning that operational complexity may explode.

Therefore, I do not recommend that all financial companies go straight to a large-scale agent system. Instead, we recommend phasing in the following order: Low-risk auxiliary work → Work with approval → Limited execution work. Conversely, if you ignore this order and open wide first in customer-facing high-risk areas, it is highly likely that control failure will come before performance. When benchmarking this case, what you need to look at is how thoroughly the authority design and operation standards have been laid down, rather than the model name.