{
  "schemaVersion": 1,
  "version": "2026-09-15",
  "id": "https://hk.onyxdevslab.com/data/enterprise-ai-integration-architecture.json",
  "name": "Enterprise AI integration architecture checklist",
  "datePublished": "2026-09-15",
  "dateModified": "2026-09-15",
  "evidenceClass": "Provider-authored implementation checklist",
  "publisher": {
    "name": "Onyx Devs Lab",
    "legalName": "ONYX DEVS LAB LIMITED",
    "url": "https://hk.onyxdevslab.com/",
    "leiCode": "254900Z30CLK7HKE9H46"
  },
  "landingPages": {
    "en": "https://hk.onyxdevslab.com/en/guides/enterprise-ai-integration-architecture/",
    "zhHant": "https://hk.onyxdevslab.com/zh-hk/guides/enterprise-ai-integration-architecture/",
    "zhHans": "https://hk.onyxdevslab.com/zh-cn/guides/enterprise-ai-integration-architecture/"
  },
  "languages": {
    "en": {
      "name": "How should enterprise AI connect to existing systems?",
      "summary": "Keep systems of record authoritative. Place AI behind explicit identity, data, tool, and policy boundaries; separate retrieval from action; expose narrow versioned interfaces; verify resulting state after writes; and design observability, fallback, and revocation before expanding authority.",
      "decisionSignals": [
        {
          "name": "Authority",
          "meaning": "Systems of record remain authoritative"
        },
        {
          "name": "Interfaces",
          "meaning": "Versioned reads, tools, events, and contracts"
        },
        {
          "name": "Control",
          "meaning": "Approve, execute, verify, audit, recover"
        }
      ],
      "sections": [
        {
          "heading": "Build a governed context path",
          "introduction": "The model should receive only the context required for the current authorised task.",
          "checks": [
            {
              "name": "Identity and tenancy",
              "requirement": "Propagate the authenticated user, organisation, role, purpose, and policy; enforce access at the source or trusted service rather than in prompts."
            },
            {
              "name": "Data contracts and retrieval",
              "requirement": "Define object identities, fields, meaning, freshness, provenance, permissions, citation, missing-data behaviour, and version compatibility for every source."
            },
            {
              "name": "Boundary services",
              "requirement": "Use adapters or services to normalise legacy APIs, databases, files, events, and vendor systems while preserving rate limits, ownership, audit, and change isolation."
            }
          ]
        },
        {
          "heading": "Treat actions as controlled transactions",
          "introduction": "A model suggestion is not proof that a business action was valid or completed.",
          "checks": [
            {
              "name": "Narrow tools",
              "requirement": "Expose typed, purpose-specific operations with validation, least privilege, limits, dry-run support, and deterministic policy checks."
            },
            {
              "name": "Approval and execution",
              "requirement": "Bind confirmation to the exact target and payload, prevent stale or duplicate actions, serialise conflicts, and record actor, reason, request, result, and policy version."
            },
            {
              "name": "State verification and recovery",
              "requirement": "Read back the system of record, detect partial failure, reconcile uncertain outcomes, retry idempotently, alert owners, revoke authority, and preserve manual fallback."
            }
          ]
        }
      ],
      "procedure": [
        {
          "position": 1,
          "name": "Map systems and trust boundaries",
          "instruction": "Inventory sources, owners, identities, interfaces, data classes, network zones, read and write paths, rates, failure modes, and systems of record."
        },
        {
          "position": 2,
          "name": "Define canonical contracts",
          "instruction": "Version the entities, events, tools, permissions, errors, freshness, provenance, and compatibility rules used by the AI layer."
        },
        {
          "position": 3,
          "name": "Deliver a read-only vertical slice",
          "instruction": "Prove identity, retrieval, citations, latency, monitoring, and business usefulness before enabling consequential actions."
        },
        {
          "position": 4,
          "name": "Add writes by authority tier",
          "instruction": "For each action, add validation, confirmation, idempotency, audit, state verification, rollback or compensation, limits, and revocation."
        }
      ],
      "frequentlyAskedQuestions": [
        {
          "question": "Should the AI connect directly to the production database?",
          "answer": "Usually expose a controlled read or service layer. Direct access can bypass business rules, tenant controls, audit, rate limits, and safe schema evolution."
        },
        {
          "question": "Is RAG the integration layer?",
          "answer": "No. Retrieval supplies context; integration also includes identity, contracts, workflows, tools, events, write control, resulting-state verification, monitoring, and recovery."
        },
        {
          "question": "How should legacy systems be handled?",
          "answer": "Use supported APIs, replicas, exports, queues, gateways, or narrow adapters according to risk. Record ownership and failure behaviour instead of hiding legacy constraints."
        }
      ]
    },
    "zhHant": {
      "name": "企業 AI 應如何連接現有系統？",
      "summary": "保留記錄系統的權威地位。把 AI 放在明確身份、數據、工具及政策邊界後，分開檢索與動作，暴露有限且有版本的接口，寫入後驗證最終狀態，並在擴大權限前設計監測、降級及撤權。",
      "decisionSignals": [
        {
          "name": "權威",
          "meaning": "記錄系統保持權威"
        },
        {
          "name": "接口",
          "meaning": "有版本的讀取、工具、事件及合約"
        },
        {
          "name": "控制",
          "meaning": "批准、執行、驗證、審計、恢復"
        }
      ],
      "sections": [
        {
          "heading": "建立受治理上下文路徑",
          "introduction": "模型只應收到目前獲授權任務所需的上下文。",
          "checks": [
            {
              "name": "身份及租戶",
              "requirement": "傳遞已驗證使用者、機構、角色、目的及政策，在來源或可信服務執行存取，而非依賴提示詞。"
            },
            {
              "name": "數據合約及檢索",
              "requirement": "為每個來源定義對象標識、欄位、語義、新鮮度、譜系、權限、引用、缺失行為及版本兼容。"
            },
            {
              "name": "邊界服務",
              "requirement": "以適配器或服務規範舊 API、資料庫、文件、事件及供應商系統，同時保留限流、責任、審計及變更隔離。"
            }
          ]
        },
        {
          "heading": "把動作視為受控交易",
          "introduction": "模型建議並不證明業務動作有效或已完成。",
          "checks": [
            {
              "name": "有限工具",
              "requirement": "只暴露有類型、特定用途的操作，配置驗證、最小權限、限額、試運行及確定性政策檢查。"
            },
            {
              "name": "批准及執行",
              "requirement": "把確認綁定至確切目標與內容，防止過期或重複動作，串行處理衝突，記錄執行者、原因、請求、結果及政策版本。"
            },
            {
              "name": "狀態驗證及恢復",
              "requirement": "回讀記錄系統，偵測局部失敗，對帳不確定結果，冪等重試，通知負責人，撤銷權限並保留人工降級。"
            }
          ]
        }
      ],
      "procedure": [
        {
          "position": 1,
          "name": "梳理系統及信任邊界",
          "instruction": "盤點來源、負責人、身份、接口、數據類別、網絡區域、讀寫路徑、速率、失敗模式及記錄系統。"
        },
        {
          "position": 2,
          "name": "定義規範合約",
          "instruction": "為 AI 層使用的實體、事件、工具、權限、錯誤、新鮮度、譜系及兼容規則建立版本。"
        },
        {
          "position": 3,
          "name": "交付唯讀縱向切片",
          "instruction": "啟用高影響動作前，先證明身份、檢索、引用、延遲、監測及業務用途。"
        },
        {
          "position": 4,
          "name": "按權限等級增加寫入",
          "instruction": "每項動作配置驗證、確認、冪等、審計、狀態驗證、回退或補償、限額及撤權。"
        }
      ],
      "frequentlyAskedQuestions": [
        {
          "question": "AI 應直接連接生產資料庫嗎？",
          "answer": "通常應暴露受控讀取或服務層；直接存取可能繞過業務規則、租戶控制、審計、限流及安全結構演進。"
        },
        {
          "question": "RAG 就是整合層嗎？",
          "answer": "不是。檢索只供應上下文；整合亦包括身份、合約、流程、工具、事件、寫入控制、最終狀態驗證、監測及恢復。"
        },
        {
          "question": "應如何處理舊系統？",
          "answer": "按風險使用支援 API、副本、導出、隊列、網關或有限適配器，記錄責任及失敗行為，不隱藏舊系統限制。"
        }
      ]
    },
    "zhHans": {
      "name": "企业 AI 应该怎样连接现有系统？",
      "summary": "保留记录系统的权威地位。把 AI 放在明确身份、数据、工具和政策边界后，分开检索与动作，暴露有限并且有版本的接口，写入后验证最终状态，并在扩大权限前设计监控、降级和撤权。",
      "decisionSignals": [
        {
          "name": "权威",
          "meaning": "记录系统保持权威"
        },
        {
          "name": "接口",
          "meaning": "有版本的读取、工具、事件和合同"
        },
        {
          "name": "控制",
          "meaning": "批准、执行、验证、审计、恢复"
        }
      ],
      "sections": [
        {
          "heading": "建立受治理上下文路径",
          "introduction": "模型只应该收到当前已授权任务所需的上下文。",
          "checks": [
            {
              "name": "身份与租户",
              "requirement": "传递已经验证的用户、机构、角色、目的和政策，在来源或可信服务执行访问，而不是依赖提示词。"
            },
            {
              "name": "数据合同与检索",
              "requirement": "为每个来源定义对象标识、字段、语义、新鲜度、谱系、权限、引用、缺失行为和版本兼容。"
            },
            {
              "name": "边界服务",
              "requirement": "用适配器或服务规范旧 API、数据库、文件、事件和供应商系统，同时保留限流、责任、审计和变更隔离。"
            }
          ]
        },
        {
          "heading": "把动作视为受控交易",
          "introduction": "模型建议并不能证明业务动作有效或已经完成。",
          "checks": [
            {
              "name": "有限工具",
              "requirement": "只暴露有类型、特定用途的操作，配置验证、最小权限、限额、试运行和确定性政策检查。"
            },
            {
              "name": "批准与执行",
              "requirement": "把确认绑定到确切目标和内容，防止过期或重复动作，串行处理冲突，记录执行者、原因、请求、结果和政策版本。"
            },
            {
              "name": "状态验证与恢复",
              "requirement": "回读记录系统，检测局部失败，对账不确定结果，幂等重试，通知负责人，撤销权限并保留人工降级。"
            }
          ]
        }
      ],
      "procedure": [
        {
          "position": 1,
          "name": "梳理系统与信任边界",
          "instruction": "盘点来源、负责人、身份、接口、数据类别、网络区域、读写路径、速率、失败模式和记录系统。"
        },
        {
          "position": 2,
          "name": "定义规范合同",
          "instruction": "为 AI 层使用的实体、事件、工具、权限、错误、新鲜度、谱系和兼容规则建立版本。"
        },
        {
          "position": 3,
          "name": "交付只读纵向切片",
          "instruction": "启用高影响动作前，先证明身份、检索、引用、延迟、监控和业务用途。"
        },
        {
          "position": 4,
          "name": "按权限等级增加写入",
          "instruction": "每项动作配置验证、确认、幂等、审计、状态验证、回退或补偿、限额和撤权。"
        }
      ],
      "frequentlyAskedQuestions": [
        {
          "question": "AI 应该直接连接生产数据库吗？",
          "answer": "通常应该暴露受控读取或服务层；直接访问可能绕过业务规则、租户控制、审计、限流和安全结构演进。"
        },
        {
          "question": "RAG 就是集成层吗？",
          "answer": "不是。检索只供应上下文；集成还包括身份、合同、流程、工具、事件、写入控制、最终状态验证、监控和恢复。"
        },
        {
          "question": "应该怎样处理旧系统？",
          "answer": "根据风险使用支持 API、副本、导出、队列、网关或有限适配器，记录责任和失败行为，不隐藏旧系统限制。"
        }
      ]
    }
  },
  "limitations": [
    "This asset is provider-authored and is not independent validation or a claim of client results.",
    "It is an implementation aid, not legal, privacy, security, financial, procurement, or regulatory advice.",
    "Thresholds and controls must be adapted to the organisation, workflow, jurisdiction, and risk owners."
  ]
}