Amazon Web Services (AWS) is gaining attention for its fully managed generative AI service, Bedrock, as a technical solution to the issue of AI silos within organizations. The platform allows departments to freely use their preferred models while maintaining unified security and data standards across the enterprise.
On August 19, AWS presented Bedrock as a way to address internal AI silos in companies. Bedrock is an integrated platform that provides access to major foundation models from industry leaders such as Amazon, Anthropic, OpenAI, Meta, Mistral, Cohere, and AI21 Labs through a single API. It operates serverless, requiring only usage-based billing without the need for infrastructure management.
Choi Young-jun, head of AWS AI technology, explained, "Bedrock provides access to various foundation models through a single API, integrates scattered internal knowledge, and applies a common responsible AI policy across the enterprise using guardrails and agent cores."
The core of Bedrock is to ensure both departmental autonomy and enterprise governance. The 'Managed Knowledge Base' feature connects data scattered across internal documents, databases, and storage, offering fully managed retrieval-augmented generation (RAG). This automation eliminates information silos between departments, allowing the entire organization to share the same knowledge source.
Security and responsible AI policies are also centrally applied. The 'guardrail' feature implements content filtering, personally identifiable information (PII) masking, verification to prevent hallucinations, and defenses against prompt attacks across all AI applications based on common standards. Additionally, AWS's Identity and Access Management (IAM) system allows for detailed control over who can use which models and data based on roles.
Solutions for the agent domain have also been established. Bedrock's 'Agent Core' is an infrastructure platform that accommodates agents developed using any framework, including CrewAI, LangGraph, LlamaIndex, and Google ADK.
It supports fully isolated execution environments for sessions of up to eight hours, a gateway that transforms API, Lambda, and MCP servers into agent-compatible tools, policies that define agent behavior boundaries in natural language, and evaluation features that monitor quality based on 13 pre-built criteria.
Even if departments independently develop agents, they can operate on a common security, monitoring, and quality management infrastructure, allowing for seamless enterprise-wide expansion without governance gaps.
Choi emphasized, "If companies hesitate to adopt AI, productivity gaps will widen, and if each department independently implements AI without a system, models, data, and security standards will become fragmented, making it even more difficult to implement AI agents that cross departmental boundaries. A structure that allows for autonomous innovation in each department while maintaining enterprise governance is the technical solution to AI silos."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.