7_AI 뱅크를 위한 에이전틱 AI 도입 전략과 기술적 접근_Jinsung Huh.pdf

AI 뱅크를 위한 에이전틱 AI 도입 전략과 기술적 접근

많은 금융사가 Gen AI 에이전트를 시범 운영에선 성공하지만, 데이터 통합과 인프라 구축의 복잡성으로 확장에 어려움을 겪고 있습니다. 본 세션에선 AWS SageMaker와 Bedrock을 활용한 데이터 통합과 대규모 에이전트 운영 방안을 소개합니다.

허진성 Solutions Architect, AWS

1. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.© 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Confidential and Trademark.
허진성
Solutions Architect
AWS
1
AI 뱅크를위한에이전틱AI
도입전략과기술적접근

2. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Agenda
4
Agentic AI의흐름과도입과제
주어진미션과해결과제
Amazon Bedrock AgentCore을통한Agent의빠른프로덕션화
결론

3. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.© 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Confidential and Trademark.
Agentic AI의흐름과도입과제
5

4. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.6
Gartner 2025 Top 10 Strategic Technology
Gartner, “Top strategic Technology Trends for 2025,” October 2024

5. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Agentic AI에대한엔터프라이즈기업들의투자
7
Gartner, “Top strategic Technology Trends for 2025,” October 2024 Gartner, “Top Strategic Technology Trends: Agentic AI—the
Evolution of Experience” February 2025
2028년까지기업용
소프트웨어에서33%가
AgenticAI를도입할것이다.
2028년까지일상업무결정의
대부분이에이전트형인공지능(AI)을
통해자율적으로이루어질것입니다.

6. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Agentic AI로의진화
8
인공지능(AI)
•분석,예측,자동화
수행
•인간의입력이나
감독이필요
생성형AI
•컨텐츠,텍스트,코드
생성에중심을둠
•출력의결과는사용자의
프롬프트에의존
Agentic AI
•자율적인의사결정이
가능한AI
•인간의개입없이의사
결정을내리고수행
수동적 자율적

7. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved. 9
인공지능(AI)을활용해‘인간이나시스템대신’추론하고,
계획하며, 작업을완료하는자율적소프트웨어시스템

8. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
금융사에서의Agentic AI 도입트렌드
10
6.60%
6.70%
7.10%
8.10%
9.60%
10.10%
제조업
헬스케어산업
도매업
정보산업
과학기술서비스산업
금융및보험업
0.00%2.00%4.00%6.00%8.00%10.00%12.00%
국내산업부문별생성형AI의활용전망
<출처:대한상공회의소>

9. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
무엇이Agentic AI의도입을어렵게하는가?
11
인프라관리Agent의
성능
보안외부연동
(MCP)

10. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.12
우리는많은Agentic AI PoC를진행합니다.
가능성
발굴
의미있는
비즈니스가치프로덕션과정에서의Challenge
성능 보안거버넌스POC AI production
agents
확장성

11. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.© 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Confidential and Trademark.
주어진미션과해결과제
13
“우리는AWS 뱅크의DX팀이고,
팀회의를시작합니다.”

12. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.14

13. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.15
우리에게주어진과제
“각개인의금융업무를도와줄
프라이빗뱅커Agent의도입해봅시다!”
재무분석가
Agent
리스크분석가
Agent
포트폴리오아키텍트
Agent

14. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.16
무엇을해결해가야하는가?
Agents
조직전체
Tools
“My Agent”Agents
내부부서
Tools
조직외부
Agents
Tools

15. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.17
Agent의Production화를위한필요요소
각사용자별격리된Agent 환경과확장가능한인프라가필요
내부및외부Tools과의유연한연동환경이필요
Agent 및Tools에대한안전한접근환경이필요
각사용자의대화및선호도등에대한개인화가필요
운영하는Agent에대한모니터링이필요

16. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.© 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Confidential and Trademark.
Amazon Bedrock AgentCore을통한
Agent의빠른프로덕션화
18

17. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Amazon Bedrock AgentCore
에이전트를안전하게배포하고운영하며, 어떤프레임워크와모델을
사용하더라도대규모로확장가능합니다.
어떠한모델및
프레임워크와도연동
신뢰할수있고안전한
에이전트배포
FLEXIBLETRUSTEDTIME TO VALUE
인프라운영에대한
번거로움을해결하여
AI 에이전트를구축

18. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
확장성,외부연동,모니터링
다양한Tools와연동
및Memory 기능
대규모의에이전트를
안전하게배포
Agent에대한
모니터링

19. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
다양한Tools와연동
및Memory 기능
대규모의에이전트를
안전하게배포
Agent에대한
모니터링
AgentCore Runtime
AgentCore Identity
대규모의에이전트를안전하게배포

20. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Agent에대한
모니터링
대규모의에이전트를
안전하게배포
AgentCore Browser
AgentCore Code Interpreter
AgentCore Memory
AgentCore Gateway
다양한Tools와연동
및Memory 기능
다양한Tools와연동및Memory 기능

21. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
대규모의에이전트를
안전하게배포
Agent에대한
모니터링
다양한Tools와연동
및Memory 기능
AgentCore Observability
Agent에대한모니터링

22. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.24
Agent의Production화를위한필요요소(1)
각사용자별격리된Agent 환경과확장가능한인프라가필요
내부및외부Tools과의유연한연동환경이필요
Agent 및Tools에대한안전한접근환경이필요
각사용자의대화및선호도등에대한개인화가필요
운영하는Agent에대한모니터링이필요

23. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
AgentCore Runtime
•최대8시간의VM유지및
시간과실시간확장지원
•다양한형식의페이로드지원
실시간확장지원
•오픈소스기반의프레임워크
지원
•몇줄의코드만으로
에이전트에대한인프라배포
빠른배포
•각세션별로분리된환경제공
•기존의IDP와연계된보안
지원
격리된환경

24. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Framework
Agent or tool code
Models
AgentCore Runtime decorator
+
configure
Docker file
AgentCore Runtime
AgentCore Runtime
Endpoint
Amazon ECR Repository
launch
Application
invoke
User
AgentCore Observability config
AgentCore Indentity config
AgentCore Runtime
Agent
AgentCore Runtime⭐

25. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
AgentCore Runtime
Session 1
각Session 별격리된환경제공
Session 2
Session 3
Session 4
Session 5
•고유한Session ID로Invoke시격리된
Runtime MicroVM을생성합니다.
•요청을처리하지않더라도Runtime 리소스를
유지하며 ColdStart를방지합니다.
•대화기록등을Runtime내임시저장소로활용
할수있습니다.
•최대8시간Runtime을지속할수있습니다.
Session 1
Session 2
Session 3
Session 4
Session 5

26. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.28
AgentCore Runtime - 재무분석가
•Strands기반의Agent 배포
•AmazonBedrock과의연동

27. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.29
Agent의Production화를위한필요요소(2)
각사용자별격리된Agent 환경과확장가능한인프라가필요
내부및외부Tools과의유연한연동환경이필요
Agent 및Tools에대한안전한접근환경이필요
각사용자의대화및선호도등에대한개인화가필요
운영하는Agent에대한모니터링이필요

28. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
•API, Lambda 함수및기존
서비스를MCP 호환도구로
변환
•표준화된인터페이스로다양한
Tools을접근
간소화된도구개발및통합
•단일하고안전한엔드포인트
•여러도구를하나의통합된
인터페이스로결합
안전하고통합된액세스
•에이전트가컨텍스트인식
탐색을통해적절한도구검색
지능형도구탐색
AgentCore Gateway

29. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Agent
MCP Client
AgentCore
Gateway
API Endpoint
Target
AWS Lambda
Target
RESTful services
withOpenAPI schema
Lambda functions
Tool 1
Tool 2
Tool 4
Tool 5
Tool 6
/mcp
Tool 3
APIs, tools, resources
AgentCore Gateway
List tools,
Invoke tool,
Search
•AgentCore Gateway는도구를자동으로색인화하며, 서버리스Semantic검색을
제공합니다.
•에이전트의LLM에전달되는컨텍스트를줄여정확성, 속도, 비용을개선
Search: “create a social media post”returns 4 most relevant tools

30. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved. 32
Agent의Production화를위한필요요소(3)
각사용자별격리된Agent 환경과확장가능한인프라가필요
내부및외부Tools과의유연한연동환경이필요
Agent 및Tools에대한안전한접근환경이필요
각사용자의대화및선호도등에대한개인화가필요
운영하는Agent에대한모니터링이필요

31. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Agent
(hosted/
self-hosted)
AgentCore
Gateway (tools)
User
AWS
resource/agents
(via IAM)
Agent
Inbound
Auth
App
Inbound Auth
(IAM/Oauth)
Your users need to
access the agent
through an application
External
resources/agents
(via Oauth, API
keys, etc.)
AgentCore Gateway
Inbound Auth
Outbound Auth
(IAM)
Outbound Auth
(OAuth)
Outbound Auth
(OAuth)
AgentCore Identity
Inbound AuthOutbound AuthAgent에접근하는보안제어Agent가외부Tools에접근하는제어

32. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.34
AgentCore Gateway/Identity – Lambda연동
•AgentCore Gateway로Lambda 기반Tools 사용
•AgentCore Identity로Outbound OAuth 접근제어

33. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.35
Agent의Production화를위한필요요소(4)
각사용자별격리된Agent 환경과확장가능한인프라가필요
내부및외부Tools과의유연한연동환경이필요
Agent 및Tools에대한안전한접근환경이필요
각사용자의대화및선호도등에대한개인화가필요
운영하는Agent에대한모니터링이필요

34. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
AgentCore Memory
Short term Memory
대화기록
세션상태
Long term Memory
Semantic
유저선호도
요약정보
Shor term
->
Long term
AgentCore Memory
async
Agent 실행
메세지
Agent 상태
List
Events
Events
Retrieve
Memory records
sync
sync
sync
async
Agent

35. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.37
AgentCore Memory – 리포트저장
•AgentCore Memory을
통한사용자의최종리포트저장
및해당리포트는향후활용

36. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.38
Agent의Production화를위한필요요소(5)
각사용자별격리된Agent 환경과확장가능한인프라가필요
내부및외부Tools과의유연한연동환경이필요
Agent 및Tools에대한안전한접근환경이필요
각사용자의대화및선호도등에대한개인화가필요
운영하는Agent에대한모니터링이필요

37. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Amazon Bedrock
AgentCoreObservability
AgentCore Runtime
AgentCore Memory
AgentCoreGateway
…
OTEL logs
AgentCore Observability dashboards
Third-party observability dashboards
AgentCore Observability

38. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.© 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Confidential and Trademark.
결론
40
“AWS 뱅크DX팀의프로젝트발표
시간입니다.”

39. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Bedrock AgentCore을활용한프라이빗뱅커

40. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.

41. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
묻고 Bedrock AgentCore로 가!

42. © 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Thank you!
© 2025, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Confidential and Trademark.44