All casesNo. 058
KB Kookmin Bank Cuts Complaint Handling to 8.4 Days
An agent drafts complaint replies; then staff, not the AI center, start building agents
- Case no.
- 058
- Primary source
- Byline Network report of a Microsoft Industry Summit talk + KB Financial Group press releases
- Source published
- 2026-10-01

1. Why this case
KB Kookmin Bank is the banking arm of KB Financial Group in South Korea. At Microsoft Industry Summit 2026 in Seoul on 2026-09-30, Kim Ji-young, a general manager at the bank’s Financial AI Center 2 and head of its AI Dev Center, described how the bank uses AI agents. Byline Network reported the talk the next day. The bank already runs more than 60 AI agents in internal work, in private banking, corporate banking, financial consultation, loan review, and customer complaints.
The case has two parts. The first is one agent and one job. Complaint handling used to run in five steps, all done by people. The bank handed the middle of it, from searching for material to drafting the reply, to an agent. Staff kept the two ends: confirming the facts of the complaint and signing off the reply. As presented by the bank, average handling time fell from 19.7 days to 8.4 days, and open complaints fell from 1,204 to 278.
The second part is about scale. The bank’s Financial AI Center built agents on request from departments, and as demand grew, the requests piled up at the center. Kim’s own words: the center “became the bottleneck.” The bank’s answer was to change who builds. Business staff now build their own agents in an AI Lab and a group-wide agent challenge, and the AI center’s role became supporting technology and the platform. The platform, in turn, gathers security and compliance into one shared layer, which in a regulated bank is what makes many builders possible.
All figures here are the bank’s own, presented at a vendor event and reported by one outlet. The quotes are translated from Korean.
2. Problem
2.1 Complaint replies took too long, and the backlog grew
Kim described complaint work as a job where speed alone is not enough: “Complaint work means understanding the content precisely, checking the relevant information, and then preparing an answer, so to reply within 14 days, not only processing speed but careful review of the final answer becomes all the more important.” On what that meant in practice, Kim said: “We had to spend a long time on this process, and there were many unresolved cases.”
Before the agent, the process ran in five steps, in this order: intake, searching for material, organizing the content, writing the reply letter, and answering the customer. All five were done by staff.
2.2 The central AI center could not keep up with demand
Under the old model, a department asked for its internal work to be automated and the Financial AI Center built the agent. Kim said this model “produced many agents, but its spread was limited,” and explained why: “Because requests were concentrated in one place, the AI center, whose response slowed as demand grew, became the bottleneck to spreading agents.”
Kim named three more limits: development focused on flagship projects, so the variety of use cases was limited; ideas from business staff rarely turned into real experiments; and business staff had few chances to build and improve agents themselves. The article puts the underlying gap plainly: an agent built by the development side cannot be exactly the agent a business unit most needs, because each department works differently and wants to automate different parts.
2.3 Many builders in a regulated environment
Letting many people build agents raised four issues that the bank listed as core tasks for producing agents at scale: a platform anyone can build on; regulatory compliance and security for a safe development and operating environment; cost control while people experiment freely; and quality assurance for performance and trust. Any agent also has to fit Korea’s network-separation rules for financial systems and the conditions attached to the bank’s innovative financial services approval (Korea’s financial regulatory sandbox).
Confirmed facts: more than 60 AI agents in internal use (PB, RM, financial consultation, loan review, complaints); old complaint process = intake → material search → content organization → reply letter → customer answer; Kim: reply “within 14 days,” careful final review, long process, many open cases; central build-on-request model produced many agents but limited spread; AI center became “the bottleneck”; limits on use-case variety, experiments, business staff’s chances to build; four core tasks = platform, compliance and security, cost control, QA.
3. Solution
3.1 The complaint agent takes the middle of the process
The bank handed the steps between complaint intake and the final answer, from material search through reply drafting, to an agent. The agent reads the complaint and summarizes it, classifies its type, searches for relevant material, and drafts the reply letter. An employee confirms the facts of the complaint, does the final review of the agent’s draft, and answers the customer.
Kim summed it up: “With the introduction of the AI agent, we cut the complaint handling process from the previous five steps to three.” Going by the article’s description, the three are intake, the agent’s drafting stretch, and the employee’s fact check, review, and reply.
Other internal agents cover the remaining four areas. The PB agent supports market analysis, portfolio rebalancing, and report writing. The RM agent supports industry and financial analysis and writing credit opinions. The financial consultation agent supports consultations on products, rules, and procedures. The loan review agent is used for review support, corporate monitoring, and early warning.
3.2 Business staff build their own agents
The Financial AI Center changed direction. Instead of building on request, it set up a structure in which business staff find their own pain points and repetitive work and experiment on them. The AI center’s role became supporting technology and the platform.
- KB AI Academy and AI Lab. Staff learn methods in the AI Academy and build and run agents in the AI Lab. KB Financial Group launched KB AI Lab on 2026-07-14 as a project-based program for staff who completed the top course (AI Expert) of the group’s AI Academy. It runs two weeks of advanced training followed by about ten weeks of a real project, following the actual service development process. Graduates return to their units and keep developing their results. Byline reports 51 participants (45 from the bank, 6 from the card company), and that 19 AI tasks at the bank and 2 at the card company came out of it and were applied to real work. The group plans to extend the Lab to all affiliates step by step.
- Group AI agent challenge. To let business staff across the group find ideas in their own work and build them as agents, the group held the “2026 Group Integrated AI Agent Competition” on 2026-09-03. Per KB Financial Group, 116 teams and 316 people from seven affiliates took part, and 12 teams reached the final after affiliate preliminaries. Finalists covered security log analysis and anomaly detection, internal document analysis, customer feedback analysis and early detection of consumer risk, data analysis, insurance product development, customer consultation, and used-car purchase support. Judging weighed real-work applicability, extensibility to other work, and business effect. The grand prize went to a KB Insurance team whose “customer care AI control center” uses 20 AI agents to check consultation quality, compliance, and adherence to claims-adjustment standards, with integrated scoring that flags items needing immediate action. Winning tasks are to be applied to real work, and verified agents shared across affiliates as common group assets.
Kim on the challenge: “Through the competition, business staff found ideas themselves and showed them actually working, and we created a flow of spreading these verified best cases to all affiliates and applying them to work.”
3.3 Platform 2.0: common parts and common controls
To support many builders, the bank is moving its agent platform from a monolithic structure to microservices. Orchestration (agent-to-agent, A2A), model serving, memory, guardrails, evaluation, and authorization and authentication are separate modules that can be connected or swapped as needed. The platform being built supports both no-code and pro-code, manages document collection, parsing, and chunking, and can use multiple models.
Kim described the goal: “The core goal of Platform 2.0 is to connect agents based on the customer journey and to respond quickly to the pace of change in AI technology,” with a leader agent coordinating the specialist agents it needs. Functions are loosely coupled so that new technology can be added and modules deployed and replaced independently “without modifying the entire platform.”
A reuse system is being put in place so new agents can be built faster from the common elements and development standards of agents already built: a common agent library on the microservice structure. Multi-model support, A2A, flexible connection of specialist agents and tools, context memory, finance-specific guardrails and evaluation, observability, and FinOps have been introduced. All services run on an agent runtime on Microsoft Azure Kubernetes Service (AKS), and Azure services connect the data layer to customer business data.
An agent management system is the place to find which agents exist in the organization for A2A use. Kim: “The AI agent management system runs from pre-review of an agent through planning, development, operation, and monitoring,” managing the full lifecycle so agents can be used and connected within the group.
3.4 Security moved out of each agent into one layer
The bank separated security and control from business logic and gathered them into the platform’s single security system. Through it, the bank manages each agent’s regulatory compliance. The article says this is also how coding agents got past the regulatory barrier in software development.
- Common gateway. Security follows the approval conditions of the innovative financial services program and the network-separation rules, with common control through Microsoft Azure API Management (APIM).
- Secure vibe-coding platform. Developers on the bank’s internal business network can use Claude Code and other AI coding tools. A developer’s request passes through the platform, the APIM common gateway, and Private Link to the external model. The platform applies the same rules to every AI call: user authentication, permission check, personal-information (PII) masking, multi-model guardrails, and real-time block logs.
- Finance-specific guardrail. A guardrail co-designed by KB Kookmin Bank and the startup T-naps is being built into Platform 2.0 as a common trust layer: detecting and blocking personal information at input, defending against prompt injection and jailbreak attacks, storing audit logs, and monitoring violations in real time.
- LLM-as-a-Judge. A second model re-evaluates AI-generated answers on four criteria: accuracy, groundedness, regulatory compliance, and hallucination detection. Depending on the result, an answer is automatically routed to Pass, Review, or Block.
- Cost. The bank focused on securing stable processing capacity, choosing the model that fits each task, and cutting resource use with prompt caching.
3.5 Toward customer-facing agents
The bank has launched KB AI in its KB Star Banking app as a customer-facing agent. Kim said the bank is moving toward “intent-driven banking that understands the customer’s intent instead of the customer navigating menus directly,” connecting the specialist agents and services needed to resolve the request. With Microsoft, the bank is running a three-month Azure Innovation Hub project: business envisioning, use-case discovery and prioritization, solution envisioning, architecture design, and a rapid prototype.
Confirmed facts: complaint agent summarizes, classifies, searches, drafts; employee confirms facts, final-reviews, answers; five steps to three; AI center role changed to technology and platform support; KB AI Lab launched 2026-07-14 (2 weeks + ~10 weeks; AI Expert graduates); 51 participants, 19 bank + 2 card tasks (Byline); challenge 2026-09-03, 7 affiliates, 116 teams, 316 people, 12 finalists (KB release); Platform 2.0 microservices with A2A, models, memory, guardrails, evaluation, auth; AKS runtime; agent management system across lifecycle; APIM common control; secure vibe-coding platform with Claude Code; T-naps guardrail; LLM-as-a-Judge with Pass/Review/Block; KB AI in KB Star Banking.
4. Impact
All figures are as presented by the bank. No measurement period, baseline date, or complaint volume is given.
| Metric | Before | After | Basis |
|---|---|---|---|
| Average complaint handling time | 19.7 days | 8.4 days | Complaint agent, Byline 2026-10-01 |
| Open (unresolved) complaints | 1,204 | 278 | Complaint agent, Byline 2026-10-01 |
| Steps in the complaint process | 5 | 3 | Kim Ji-young, Byline 2026-10-01 |
Average handling time fell by more than half. The old average, 19.7 days, was longer than the 14 days Kim mentioned for replying; the new one, 8.4 days, is shorter. Open complaints fell to less than a quarter of their earlier count.
On spreading agent building:
| Program | Figure | Source |
|---|---|---|
| KB AI Lab | 51 participants (bank 45, card 6); 19 bank and 2 card AI tasks applied to work | Byline 2026-10-01 |
| Group AI agent competition | 116 teams, 316 people, 7 affiliates; 12 finalists | KB Financial Group, 2026-09-07 |
On answer quality controls, there are two separate measures, both from corporate-lending tests. Kim said that in an evaluation on 60 corporate-lending scenarios, the KB–T-naps finance-specific evaluation model scored 90 out of 100, 28.2 points above the average of general frontier models, and that its lightweight structure made real-time inference two to three times faster. KB Financial Group’s 2026-06-25 press release, announcing a Financial Services Commission chairman’s award for the same partnership, says a pilot in corporate-lending consultations blocked more than 90% of hallucinations in advance, and that the real-time answer evaluation and control system is to be extended to deposits, retirement pensions, and other consultation areas.
Kim’s summary of the overall approach: “We are applying AI quickly to the work with the biggest effect first, cutting processing time and raising the quality and consistency of AI agents, and verified cases are at the stage of expanding to all group affiliates.”
5. Insight
The work that moved was the middle of the complaint. Searching for material, organizing the content, and drafting the reply moved to the agent. People kept the two ends that carry accountability: confirming what actually happened to the customer, and signing off the final answer. Kim’s emphasis on careful review of the final answer explains why the bank cut steps from the middle rather than the end. The time saved came from not having to assemble each case from scratch.
A working agent creates a queue at whoever builds agents. The bank’s first model, departments ask and the AI center builds, did produce many agents. Its limit was the queue. The fix was not a bigger center but a different builder: business staff who know the work find the problem and build the agent, and the center supplies the platform. The AI Lab, with its two-plus-ten-week format and return of graduates to their units, and the group competition, with real-work applicability as a judging criterion, are both ways to turn staff ideas into running agents.
In a bank, many builders need one set of controls. The bank did not ask each new agent to solve security, privacy, and compliance on its own. It moved those into one platform layer: one gateway, one set of checks on every AI call, one guardrail, one evaluation step, one registry from pre-review to monitoring. That is what the article credits for getting coding agents past the regulatory barrier, and it is the precondition for letting business staff build.
Limits. All figures come from the bank, as presented at a Microsoft event and reported by one outlet. The complaint figures have no stated period, baseline date, or volume, and the article does not say whether other changes in the same period contributed. It also does not say who built the complaint agent, so its results should not be read as results of the business-led model, which the bank presents as its next stage. The AI Lab’s 21 tasks are reported as applied to work, but their effects are not given. Parts of Platform 2.0 are described as under construction. The two answer-quality figures (90 points on 60 corporate-lending scenarios; more than 90% of hallucinations blocked) come from the bank’s own tests on corporate-lending scenarios, not on complaints. No staffing or cost figures are published.
One-line takeaway: KB Kookmin Bank cut complaint handling time by giving an agent the drafting in the middle and keeping staff on fact checks and sign-off, and then found that scaling agents meant changing who builds them and putting every agent behind one shared set of controls.
6. Implementation hypotheses (not in the source)
6.1 A complaint drafting pipeline keyed to type and to the 14-day window
Basis: The agent “summarizes” the complaint, “classifies its type,” searches for material, and drafts the reply; the employee confirms the facts and does the final review. Kim ties the work to replying “within 14 days.”
The classification step most likely decides where the agent searches. A complaint typed as, say, a card fee dispute or a loan interest question would pull from that product’s terms, the relevant internal rules and procedures, and past replies to complaints of the same type, all of which a bank would already hold as documents. The draft would probably come back with the passages it relied on, so the employee can check each claim against its source instead of searching again. The agent likely does not confirm account-level facts itself. The article assigns “confirming the facts of the complaint” to the employee, which suggests the employee checks transactions and account history in the bank’s own systems before approving. A work queue that shows days elapsed against the 14-day window would let reviewers take the oldest cases first.
6.2 Business-built agents registered once and routed through one gateway
Basis: Security and control were separated from business logic into “the platform’s single security system”; every AI call passes through the platform, the APIM common gateway, and Private Link with authentication, permission checks, PII masking, guardrails, and block logs; the agent management system runs “from pre-review of an agent through planning, development, operation, and monitoring.”
An agent built in the AI Lab would likely start as an entry in the agent management system. Pre-review would record what data the agent reads, which models it calls, and who may use it. On approval, the agent would get credentials for the APIM gateway scoped to that use rather than direct access to any model endpoint, so it cannot reach an external model except through the common checks. This is how the AI center can approve many agents without building them: it reviews the registration and the gateway policy, not the agent’s code. The same gateway is a natural place for the cost controls the bank lists. Per-agent usage can be metered at APIM, which feeds FinOps reporting and lets the platform team move a high-volume agent to a cheaper model without the agent’s owner changing anything.
6.3 Judge scores turned into three queues
Basis: LLM-as-a-Judge scores answers on accuracy, groundedness, regulatory compliance, and hallucination detection, and routes them to Pass, Review, or Block. The KB release describes a “real-time answer quality evaluation and control system” for the financial consultation agent, piloted in corporate-lending consultations.
In a consultation flow, each generated answer would probably be scored before it is shown. Pass would release it. Review would hold it in a queue for a staff member, and Block would suppress it and fall back to a safe response or a hand-off. The thresholds are likely set per criterion, with regulatory compliance and hallucination acting as hard gates and accuracy and groundedness as scores. Block events would feed the real-time violation monitoring and audit logs the guardrail layer keeps, so the platform team can see which agents and which question types fail most and adjust retrieval or prompts for those. If complaint drafts also pass through the judge, the Review route would matter less there, since every complaint reply already ends with an employee’s sign-off.
7. Sources
- Byline Network (Kim Woo-yong) — “플랫폼 2.0” KB국민은행이 그리는 금융 AI 확장 로드맵 (“Platform 2.0”: KB Kookmin Bank’s roadmap for scaling financial AI)
https://byline.network/2026/10/1-1496/
2026-10-01 (talk at Microsoft Industry Summit 2026, 2026-09-30); accessed 2026-11-17. Korean-language article; quotes translated from Korean. - KB Financial Group press release — KB금융, AI와 함께 일하는 문화로 AX 가속화 (KB AI Lab launch)
https://www.kbfg.com/kor/pr/press/view.htm?CONTENT=15095&B=8
2026-07-14; accessed 2026-11-17 - KB Financial Group press release — KB금융, 임직원 아이디어로 AI 업무 혁신 가속화 ‘2026 그룹 통합 AI 에이전트 경진대회’ 개최
https://www.kbfg.com/kor/pr/press/view.htm?CONTENT=15146&B=8
2026-09-07; accessed 2026-11-17 - KB Financial Group press release — KB금융, ‘KB 오픈이노베이션 프로그램’ 우수 협력 사례로 디캠프의 ‘스타트업 OI(오픈이노베이션) #금융권’에서 금융위원장상 수상
https://www.kbfg.com/kor/pr/press/view.htm?CONTENT=15081&B=8
2026-06-25; accessed 2026-11-17