AI Shadow AI / Citizen Development Governance Architecture 解读

面向对象: AI Governance PM / Senior BA / Enterprise Architect / AI Platform Architect / Security Architect / Financial Retail AI Owner。 核心问题: 员工已经在用公共 AI、浏览器插件、个人账号、低代码自动化和部门自建 Bot 解决真实工作流痛点。组织需要发现、分类、治理和迁移这些 uncontrolled AI uses, 而不是用一纸禁令把有效需求逼进更深的地下。 学习目标: 建立 shadow AI discovery、risk tiering、approved pathway、citizen developer guardrails、DLP、review gates、monitoring、exception 和 platform golden path migration 的治理架构。

---

Source Anchors

一句话:

Shadow AI governance architecture 是把“员工绕开正式路径使用 AI”转成可发现、可分级、可引导、可治理、可迁移和可审计的 enterprise control system。

---

1. Thesis: Shadow AI 是需求信号, 也是风险暴露

Shadow AI 不只是违规行为。

它通常说明正式平台没有满足真实工作流:

• 客服主管想快速总结投诉。
• BA 想把会议纪要转成流程和需求。
• 分行经理想批量生成客户沟通草稿。
• 风控分析师想用 LLM 辅助写 suspicious activity narrative。
• 营销团队想用外部插件生成 campaign copy。
• 开发者想用代码助手、agent、no-code builder 提高速度。

但在金融服务中, shadow AI 同时制造高风险暴露:

• data leakage: PII、账户、交易、投诉、信用、商户和员工数据外流。
• conduct risk: 对客户做出未经批准的承诺、建议或解释。
• records risk: AI 生成内容进入业务流程但没有留痕和保留。
• third-party risk: 未审查供应商、个人账号、数据训练条款和跨境处理。
• model risk: 输出被当成分析、评分或决策支持但没有验证。
• audit risk: 无 inventory、owner、approval、monitoring、exception 和 evidence。
• customer harm: 错误解释费用、贷款、理财、争议权利或投诉路径。

成熟治理的目标不是消灭所有探索, 而是把高价值需求导入受控路径。

---

2. Why It Matters

---

3. Core Concepts

---

4. Architecture Diagram

Shadow AI signals
  network egress | browser extension | expense | SaaS logs | DLP | interviews | surveys
        |
        v
Discovery and signal registry
  user | team | tool | data type | workflow | vendor | channel | frequency
        |
        v
Classification engine
  sanctioned | tolerated | migrate | restrict | block | investigate
        |
        v
Risk tiering
  data sensitivity | customer impact | automation | external sharing | vendor | records | tool authority
        |
        v
Approved pathway router
  approved AI catalog | employee copilot | RAG golden path | agent workflow | citizen builder sandbox
        |
        v
Review and control layer
  data policy | DLP | prompt policy | tool permission | HITL | evidence | exception
        |
        v
Monitoring and learning loop
  usage dashboard | KRI | incident | training gaps | platform roadmap | migration progress

Design principle:

Every discovered shadow AI use should resolve into:
approve, migrate, restrict, block, exception, training action, platform backlog item, or no-impact rationale.

---

5. Financial Retail Case

Scenario: a regional bank discovers that relationship managers use personal AI accounts to summarize customer meetings, create follow-up emails, and draft loan exception narratives.

Findings:

---

6. PM / BA / Architect Checklist

---

7. Code-Lite Experiment

Build a small classifier for discovered AI uses:

signal_id: SHAI-2026-0017
source: browser_extension_log
team: retail_lending
tool: external_ai_summary_plugin
workflow: loan_exception_narrative_drafting
data_entered:
  - customer_income
  - credit_score
  - collateral_notes
output_destination: loan_workflow_attachment
customer_impact: decision_support
vendor_status: not_approved
records_required: true
automation_level: draft_only
classification:
  state: migrate
  risk_tier: Tier3
  rationale: sensitive credit data, decision-support use, unapproved vendor, record retention gap
route:
  approved_path: internal_copilot_with_credit_policy_rag
  gates: [data_review, model_risk_screen, records_review, manager_handoff]
  controls: [DLP, source_citation, prompt_retention, reviewer_attestation]

Pseudo-query:

SELECT team, workflow, COUNT(*) AS signals
FROM shadow_ai_signal
WHERE risk_tier IN ('Tier2', 'Tier3', 'Tier4')
  AND classification_state IN ('migrate', 'restrict', 'block')
GROUP BY team, workflow
ORDER BY signals DESC;

Pass condition: every high-risk signal has an owner, a decision, a route, a control set and a closure date.

---

8. Interview Questions

9. Common Pitfalls

---

10. Final Principle

Shadow AI governance 的成熟标志不是“没有人尝试 AI”, 而是:

有价值的非正式用法能被快速发现、正确分级、导入受控路径、沉淀为平台能力; 高风险和禁止用法能被阻断、留证、纠正并转化为训练和架构改进。