AI 底层逻辑 / 经典论文

AI Workforce Capability Academy：角色技能转型架构

AI Workforce Capability Academy 的核心价值不在“教会大家使用 AI”，而在建立一套可验证的组织能力系统：谁可以负责 AI use case、谁可以签署 eval/release evidence、谁可以改动高风险流程、哪些岗位仍存在 skills debt、这些能力变化是否真的改善了 adoption、quality、risk 和 business outcome。

599 行ai-foundations/papers/154-ai-workforce-capability-academy-role-skill-transformation-architecture.md

AI Workforce Capability Academy：角色、技能与组织能力转型架构

配对阅读：本篇的操作手册版（模板/RACI/门禁/runbook）是 docs/AI_WORKFORCE_CAPABILITY_ACADEMY_ROLE_SKILL_TRANSFORMATION_PLAYBOOK.md。第一遍读本篇建立原理与架构判断；第二遍做案例时再用 playbook 查表落地，两者不需要重复精读。

AI 时代的能力建设不是多开几门 prompt 课程，也不是把全员培训完成率报到管理层。真正的问题是：企业的 AI portfolio 需要哪些角色承担责任，这些角色需要什么技能证据，哪些能力缺口会阻塞上线、放大风险或削弱 adoption，学习体系如何反过来影响真实流程、质量、成本和控制结果。

这篇笔记把 AI academy 当成一套 workforce capability platform：它连接业务能力地图、角色演进、技能本体、熟练度等级、情景评估、能力证据、生产采用、风险控制和季度投资决策。对金融零售场景来说，学院不是培训中心，而是让组织安全扩展 AI 工作方式的运行架构。

重要说明: 本文是学习、作品集和内部架构训练材料, 不构成法律意见、合规结论、监管解释、人力资源政策建议、劳动关系建议、绩效考核方案、认证声明、采购建议或审计意见。正式采用时必须由 HR、Legal、Compliance、Risk、Model Risk、Information Security、Enterprise Architecture、Business Owner、Works Council / Employee Relations 和授权管理层结合司法辖区、员工政策、岗位体系、监管关系、数据使用范围和内部治理确认。访问日期按 2026-06-30 记录。

Source Anchors

Source	Official link	本文使用方式
SFIA AI skills resources	https://sfia-online.org/en/tools-and-resources/ai-skills-framework	用 AI skills framework 作为 AI 时代技能分类、岗位能力表达和组织技能管理的锚点。
SFIA 9 Skills A-Z	https://sfia-online.org/en/sfia-9/skills/all-skills-a-z	用 SFIA 技能目录和责任等级思想组织 role-to-skill mapping、proficiency language 和职业路径。
NIST NICE Framework Resource Center	https://www.nist.gov/itl/applied-cybersecurity/nice/nice-framework-resource-center	借鉴 NICE 对 work roles、tasks、knowledge、skills、abilities 的结构化表达, 设计 AI workforce capability ontology。
NIST AI Risk Management Framework	https://www.nist.gov/itl/ai-risk-management-framework	用 Govern / Map / Measure / Manage 组织 AI workforce risk、training controls、human oversight、measurement 和持续改进。
ISO/IEC 42001 AI management systems	https://www.iso.org/standard/81230.html	用 AI 管理体系、角色责任、运行控制、绩效评价、内部审核和持续改进组织 academy governance。
ISO 30414 Human capital reporting	https://www.iso.org/standard/69338.html	用 human capital reporting 的思路组织 workforce metrics、skills visibility、leadership accountability 和能力投资证据; 正式引用前需核对 ISO 当前版本状态。

核心导读

如果只用 LMS 课程、认证徽章和学习时长衡量准备度，企业会得到一种虚假的 ready signal。成熟的 academy 必须把学习路径设计成产品平台：有角色画像、技能图谱、情景实验室、证据登记、评审工作流、社区实践和生产遥测。它不是 HR 的单点项目，也不是技术团队的内部培训，而是业务、架构、风险、运营和人力共同治理的能力基础设施。

1. Thesis: Academy 不是培训中心, 是能力产品平台

低成熟度做法:

AI strategy announced
  -> launch AI literacy training
  -> ask teams to complete courses
  -> collect attendance and certificates
  -> declare workforce ready

这类做法无法回答高级管理层真正关心的问题:

高级问题	培训完成率为什么不够
哪些岗位已经能安全负责 AI use case?	完成课程不等于能定义风险边界、设计 eval、管理 adoption。
哪些关键能力短缺会阻碍 AI portfolio?	课程目录无法显示 skill debt、capability bottleneck 和角色覆盖缺口。
谁能证明自己在真实场景中胜任?	证书不能替代情景评估、作品集证据、上线复盘和运营指标。
AI 投资是否转化为业务采用?	学习活动和 production adoption、quality、cost、risk、customer impact 需要连接。
哪些 workforce 风险正在积累?	过度依赖少数专家、角色边界不清、错误使用 AI、控制岗位缺口都不是普通 LMS 指标能发现的。

成熟做法:

business strategy and AI portfolio
  -> role taxonomy and role evolution
  -> skill ontology and proficiency levels
  -> skill gap and skills debt inventory
  -> learning path by role and capability
  -> scenario-based assessment
  -> portfolio artifact evidence
  -> supervised practice and community of practice
  -> adoption telemetry and operating outcomes
  -> workforce risk governance
  -> quarterly capability investment review

高级判断:

An AI academy succeeds only when it changes who can safely own AI work, not when it maximizes course throughput.

2. 学习主线：从岗位表到能力运行系统

读这类 workforce capability 架构，关键不是记住几个岗位名称，而是建立一条因果链：

AI strategy
  -> portfolio demand
  -> role responsibility
  -> skill ontology
  -> proficiency evidence
  -> authorization and release confidence
  -> production adoption and quality signal
  -> workforce risk and investment decision

这条链解决三个高级问题。

第一，岗位演进不能停留在“旧岗位加 AI 前缀”。产品、业务分析、架构、风险、数据、运营、一线主管都会发生责任重分配：有人要定义价值假设，有人要拆流程与例外，有人要设计 RAG/工具/权限/日志，有人要维护 eval 和 release gate，有人要负责 adoption、coaching 和 incident learning。

第二，技能必须被建模为可验证对象，而不是课程标题。一个技能节点至少要说明：它支持哪类业务能力，出现在哪些任务中，需要什么知识与实践，缺失会造成什么风险，达到某个熟练度需要什么 evidence。

第三，学院必须能反馈生产系统。真正成熟的指标不是学习完成率，而是高风险角色是否有有效能力证据，AI 产品是否更快通过 release readiness，真实工作流中的 error、override、incident、cycle time、rework 和 benefit 是否出现可解释变化。

3. Role Taxonomy: AI 时代岗位不是旧岗位加 Copilot

AI role taxonomy 应按业务能力、责任边界和风险影响定义, 不能只把现有岗位名称前面加 "AI"。

Role family	典型角色	核心责任	高级能力信号
Product leadership	Senior AI product lead, AI platform product lead, AI capability investment lead	定义 AI use case thesis、价值假设、adoption、roadmap、scale/stop gate	能把业务结果、eval、risk、cost、adoption 和 platform reuse 放进同一张产品决策表。
Business analysis	AI BA, Process Intelligence BA, CBAP-level AI Transformation BA	发现流程、规则、例外、stakeholder need、role redesign、acceptance evidence	能把 AS-IS/TO-BE、policy、control、human oversight 和 scenario pack 连接起来。
Architecture	AI Solution Architect, Enterprise AI Architect, Data/Integration Architect	设计 AI reference architecture、RAG、tool gateway、identity、logging、rollback、control plane	能证明架构支持可评估、可审计、可运营和可停用。
EvalOps and quality	EvalOps Lead, AI QA Lead, UAT Certification Lead	设计 eval sets、rubrics、regression gates、production sampling、quality review	能把 requirement、risk、test、model/prompt/RAG/tool version 和 release decision 连接。
Risk and governance	AI RiskOps Lead, Model Risk Partner, Compliance AI Control Owner	risk tiering、control design、incident taxonomy、monitoring、governance evidence	能把 NIST AI RMF / ISO 42001 类治理语言转化为可执行门禁。
Data and knowledge	Data Product Manager, Knowledge Owner, Data Steward	source of truth、metadata、lineage、freshness、permission、retention	能说明知识和数据是否足够支撑 AI 输出和审计重建。
Operations transformation	AML/KYC Ops Lead, Contact Center Transformation Lead, Branch Enablement Lead	真实流程 adoption、quality calibration、manager cadence、support model	能把 AI 工具嵌入排班、QA、升级、SOP、绩效和 coaching。
Workforce enablement	AI Academy Product Owner, Learning Architect, Community Lead	skill map、learning path、assessment、portfolio、community of practice	能把学习产品做成可度量、可迭代、可治理的 capability platform。

3.1 产品、分析、架构与运营岗位的责任演进

Traditional role	AI-era evolution	不变的核心	新增能力
Product Manager	从功能 owner 变成 AI capability owner	用户问题、价值、优先级、路线图	eval literacy、AI cost economics、adoption telemetry、risk-aware release gate。
Business Analyst	从需求记录者变成 process and evidence architect	stakeholder alignment、需求质量、流程理解	AI-assisted requirements mining、scenario design、human oversight、role redesign。
Solution Architect	从系统集成设计者变成 AI control-plane architect	架构取舍、接口、NFR、韧性	model gateway、RAG governance、tool authorization、AI observability、eval hooks。
Operations Lead	从流程执行 owner 变成 human-AI operating system owner	SLA、队列、质量、产能、培训	adoption dashboard、override review、AI incident route、workforce confidence。
Risk / Compliance Partner	从评审者变成 continuous control co-designer	policy、control、issue management	AI risk tiering、monitoring evidence、human oversight control effectiveness。

4. Skill Ontology: 技能不是课程名称

Skill ontology 要能回答:

What capability does this skill support?
Which role needs it?
At what proficiency?
What evidence proves it?
Where does it show up in production outcomes?
What risk appears if it is missing?

4.1 Skill Object Model

Object	Minimum fields	为什么重要
Skill	skill_id, name, definition, domain, related capabilities, risk tier	防止把课程标题误当技能。
Task	task_id, skill_id, work activity, input, output, success criteria	对齐 NICE 式 work role / task 思路, 让技能落到工作任务。
Knowledge	concepts, policies, standards, system context, data context	说明知道什么。
Practice	repeated behaviors, decision routines, review habits, collaboration patterns	说明如何做。
Evidence	artifact type, scenario result, production signal, reviewer, validity period	说明如何证明。
Proficiency	level, autonomy, complexity, accountability, evidence threshold	区分能听懂、能执行、能负责、能设计体系。
Risk if absent	failure modes, customer impact, control impact, dependency exposure	把技能缺口转成 workforce risk。

4.2 AI Workforce Skill Domains

Domain	Example skills	代表性证据
AI literacy and responsible use	概率性输出、幻觉、数据边界、禁止用法、升级路径	情景题通过、错误输出识别记录、经理抽查。
Product and value	AI use case framing、benefit hypothesis、scale/stop decision、unit economics	AI product brief、benefits register、adoption-adjusted ROI。
Requirements and process	AI-assisted discovery、process mining、exception path design、role redesign	AS-IS/TO-BE、scenario pack、acceptance criteria、control-linked journey。
Architecture and integration	RAG、model gateway、tool gateway、identity、observability、fallback	reference architecture、ADR、integration control map、rollback drill。
EvalOps and quality	golden sets、rubric、regression、judge calibration、production sampling	eval report、defect taxonomy、release gate decision、quality dashboard。
Risk and governance	risk tiering、AI RMF mapping、control design、incident response	risk assessment、control evidence pack、incident postmortem。
Data and knowledge	source authority、metadata、lineage、freshness、entitlement、retention	source inventory、knowledge readiness scorecard、retrieval quality report。
Adoption and operations	training, support tier, adoption telemetry, coaching, community of practice	adoption dashboard、support runbook、manager cadence record、COP learning loop。

4.3 Role-to-Capability Map

Capability	AI PM	AI BA	AI Architect	EvalOps	RiskOps	Data Product	Ops Lead
AI use case thesis	A/R	R	C	C	C	C	C
Workflow and role redesign	C	A/R	C	C	C	I	R
AI reference architecture	C	C	A/R	C	C	C	I
Data and knowledge readiness	C	R	C	C	C	A/R	C
Evaluation and regression	A	R	C	A/R	C	C	C
Risk tier and controls	C	C	C	C	A/R	C	C
Adoption and benefit realization	A/R	R	I	C	C	C	A/R
Incident and corrective action	A	R	R	R	A/R	C	R
Academy evidence and skill review	C	C	C	C	C	C	C

说明: A/R 表示 accountability and execution responsibility, C 表示 consulted, I 表示 informed。学院体系本身应记录每个角色在哪些 capability 上达到何种 proficiency。

5. Proficiency Levels: 从 Awareness 到 System Ownership

课程通过只能证明 exposure, 不能证明 proficiency。AI workforce capability 需要按责任、复杂度和证据定义等级。

Level	Name	能做什么	不应授权什么	Evidence threshold
0	No verified exposure	未证明具备基本 AI 使用边界	不能独立使用生产 AI 工具处理敏感任务	无正式证据。
1	Awareness	能解释适用场景、禁止用法、基本风险和升级路径	不能修改 prompt、审批输出或设计流程	role-based literacy scenario pass。
2	Guided practitioner	能在已定义 SOP 下使用 AI, 识别明显错误并反馈	不能独立设计 eval 或高风险流程	supervised task log、QA抽样、反馈质量。
3	Independent practitioner	能独立完成中等复杂任务, 处理例外, 生成可复核证据	不能拥有跨团队标准或高风险 release gate	scenario assessment、portfolio artifact、manager sign-off。
4	Lead / reviewer	能审查他人成果, 定义 standards, 处理复杂 trade-off	不能单独改变 enterprise policy 或 risk appetite	peer review record、release evidence、incident contribution。
5	System owner / strategist	能设计组织级能力系统、治理节奏、投资门槛和 workforce risk response	仍需与授权治理角色共同决策法律、合规、人事和监管事项	capability roadmap、governance pack、measured adoption outcomes。

高级原则:

Proficiency = autonomy + complexity + accountability + evidence.

5.1 AI Literacy Tiers

Tier	Audience	必备内容	评估方式
Tier 1 General AI awareness	全员和低风险使用者	基本概念、数据保护、禁止输入、错误反馈	短情景题和数据边界判断。
Tier 2 Role-based responsible use	一线、运营、PM、BA、技术、风控	角色任务中的 trust cues、human oversight、升级和记录	角色场景演练和主管抽检。
Tier 3 Builder and reviewer	PM、BA、Architect、EvalOps、Data、Risk	use case framing、eval、RAG、tool controls、release evidence	portfolio artifact review 和 scenario defense。
Tier 4 Accountable owner	Sponsor、业务 owner、平台 owner、风险 owner	risk appetite、benefit realization、operating model、incident accountability	decision memo review 和季度能力治理。

6. Academy Architecture

6.1 Reference Architecture

Enterprise strategy / AI portfolio / business capability map
        |
        v
Role taxonomy and workforce segmentation
  AI PM | AI BA | Architect | EvalOps | RiskOps | Data Product | Ops
        |
        v
Skill ontology and proficiency model
  skills | tasks | knowledge | evidence | risk if absent | validity
        |
        v
Capability gap and skills debt engine
  current evidence | target proficiency | portfolio demand | risk exposure
        |
        v
Learning path and practice studio
  modules | labs | scenario drills | supervised practice | COP
        |
        v
Assessment and evidence platform
  scenario exams | portfolio artifacts | reviewer workflow | certification ledger
        |
        v
Adoption telemetry and outcome analytics
  usage | quality | benefit | incidents | manager coaching | feedback
        |
        v
Governance and investment loop
  quarterly skill review | roadmap | funding | workforce risk treatment

6.2 Academy as Product Platform

Academy 应像产品平台一样运营:

Platform capability	产品化要求
Role profile registry	每个角色有目标能力、proficiency target、learning path、assessment pack 和 evidence validity。
Skill graph	技能可复用、可版本化, 与 capability map、use cases、risk controls 和 career path 连接。
Learning path composer	根据角色、当前证据、业务 portfolio demand 自动推荐路径, 不是所有人上同一门课。
Scenario assessment engine	生成和维护金融零售情景题、案例、数据包、评分 rubric 和 reviewer calibration。
Capability evidence registry	保存 PRD、architecture decision、eval report、process map、risk memo、adoption dashboard 等可复核能力证据。
Community of practice hub	支持案例复盘、peer review、office hour、模式库、反例库和实践标准沉淀。
Adoption telemetry bridge	把学习证据和 production adoption、quality、incident、benefit 关联, 避免学习系统自嗨。
Governance dashboard	给管理层看 capability readiness、skills debt、critical role coverage、risk trend 和投资优先级。

6.3 Key Data Objects

Object	Minimum fields
Role profile	role_id, role_family, business domain, target capabilities, required proficiency, risk-sensitive tasks。
Skill node	skill_id, definition, domain, related tasks, evidence types, adjacent skills, owner。
Evidence record	evidence_id, person/team, skill_id, artifact link, scenario id, reviewer, result, date, expiry, limitations。
Learning path	path_id, role, entry criteria, modules, labs, assessments, practice requirements, exit evidence。
Scenario pack	scenario_id, domain, task, inputs, constraints, expected artifacts, scoring rubric, critical failures。
Skills debt item	debt_id, capability, role, gap, portfolio impact, risk level, treatment owner, due date, progress evidence。
Adoption signal	use case, role, team, usage, quality, override, incident, benefit, friction, manager action。

7. Evidence-of-Competence: 不是 Course Completion

能力证据需要从弱到强分层。

Evidence type	强度	适合证明	局限
Course completion	Low	接触过概念	不能证明能做任务或承担责任。
Knowledge quiz	Low-medium	概念和规则理解	不能证明复杂情境下的判断。
Scenario assessment	Medium-high	在受控案例中应用知识、处理 trade-off	需要持续更新场景和评分一致性。
Work artifact	High	能产出真实工作物, 如 PRD、eval plan、architecture decision、risk memo	需要 reviewer 和 evidence standard。
Supervised practice	High	在真实工作中被观察、反馈和改进	成本较高, 需要 manager cadence。
Production outcome	Very high	能力是否带来 adoption、quality、risk 或 benefit 改善	需避免把团队结果简单归因到个人。
Peer review and COP contribution	Medium-high	能审查、教学、沉淀标准、带动他人	需要防止只看活跃度不看质量。

7.1 Evidence Contract

每个高价值技能都应有 evidence contract:

skill: AI evaluation design for regulated retail banking workflows
target role: AI product lead / AI business analysis lead / EvalOps lead
proficiency target: Level 3 or Level 4
scenario: AI-assisted KYC document review release gate
required artifact:
  - business acceptance criteria
  - golden journey and edge case pack
  - eval rubric with critical failures
  - release decision rule
  - monitoring and feedback loop
critical failures:
  - no protected customer or vulnerable customer scenarios
  - no source authority or policy version control
  - no human override and escalation design
  - no re-certification trigger
reviewers:
  - EvalOps
  - Risk or Compliance
  - Business process owner
validity:
  - 12 months or until major framework / policy / platform change

8. Learning Path Design

8.1 Path Design Principles

Principle	说明
Role-first	先定义角色要承担的工作和风险, 再设计课程。
Evidence-first	每条路径以可审查 artifact 和情景表现作为出口, 不以课时作为出口。
Scenario-heavy	高级岗位必须在金融零售案例中练习, 例如 AML alert triage、KYC onboarding、contact center complaint。
Artifact-based	学习产出要能进入真实评审、上线和能力晋升证据, 而不是停留在课堂练习。
Practice plus community	学习路径必须包含 peer review、office hour、case clinic、COP pattern library。
Refreshable	AI 技术、政策、模型和工具变化快, path 必须有版本和复训触发。

8.2 Example Paths

Path	适用角色	核心模块	出口证据
AI BA Transformation Path	AI BA, CBAP, Process Owner	AI literacy, process mining, requirements-to-eval, role redesign, control-linked UAT	AS-IS/TO-BE, scenario pack, acceptance criteria, role change memo。
Senior AI PM Path	AI PM, Product Lead	use case thesis, value metrics, eval strategy, adoption, platform reuse, governance	AI PRD, benefit register, eval release gate, adoption dashboard。
AI Solution Architect Path	Architect, Tech Lead	AI reference architecture, RAG governance, identity, tool gateway, observability, rollback	architecture view set, ADR, control map, rollback runbook。
RiskOps / EvalOps Path	EvalOps, Risk, QA	risk tiering, golden sets, rubrics, regression, production sampling, incident learning	eval pack, defect taxonomy, risk control evidence, monitoring plan。
Contact Center AI Lead Path	Ops lead, QA, frontline manager	AI literacy, trust cues, knowledge governance, coaching, support model, adoption telemetry	team adoption plan, QA calibration pack, support runbook, benefit review。
AML/KYC Operations Lead Path	AML/KYC ops, compliance ops	case evidence, investigation copilot, escalation, audit trail, model limitations	alert triage scenario, SAR support evidence pack, override analysis。

9. Assessment Design

Assessment 要模拟真实决策压力, 而不是考定义。

9.1 Scenario-Based Assessment Pattern

business context
  -> messy artifacts
  -> role-specific objective
  -> constraints and policies
  -> expected artifact
  -> scoring rubric
  -> critical failure list
  -> reviewer calibration
  -> feedback and retake path

9.2 Financial Retail Scenario Examples

Scenario	Role assessed	Expected artifact	Critical failures
AI-assisted AML alert triage	AI BA / RiskOps / EvalOps	workflow map, human oversight design, eval scenario pack, incident route	No audit trail, no escalation for suspicious evidence, no QA sampling。
KYC document review copilot	AI PM / Architect / Ops	PRD, architecture control map, release gate, monitoring dashboard	AI allowed to reject customer without human control, no policy version evidence。
Contact center knowledge assistant	Product lead / Business analysis lead / Contact Center Lead	adoption canvas, training scenario pack, support model, quality scorecard	No citation requirement, no escalation for complaints, no frontline feedback loop。
Loan policy summarization assistant	Architect / RiskOps / Data Product	source authority model, RAG freshness design, access control, eval report	Retrieval after generation filtering, no policy effective date, no adverse action boundary。
Data product manager for AI personalization	Data Product / PM	data contract, consent and preference map, metric design, risk review	No retention boundary, no opt-out handling, no data lineage。

9.3 Rubric Dimensions

Dimension	What reviewers look for
Business problem clarity	Candidate ties AI to a measurable workflow problem and baseline。
Role and responsibility design	Human, AI, manager, risk, support and owner responsibilities are explicit。
Evidence quality	Claims link to artifacts, scenarios, policy, data or telemetry。
Risk and control thinking	Candidate identifies customer harm, privacy, fairness, operational, model and control risks。
Eval and monitoring	Candidate defines golden sets, critical failures, thresholds, recertification triggers and production sampling。
Adoption realism	Candidate explains behavior change, manager cadence, training, feedback and resistance。
Architecture feasibility	Candidate respects source authority, permissions, logging, fallback, integration and lifecycle controls。

10. Operating Model

10.1 Core Roles

Role	Responsibilities
AI Academy Product Owner	owns academy roadmap, role profiles, evidence platform, adoption outcomes and stakeholder backlog。
Workforce Capability Architect	owns role taxonomy, skill ontology, proficiency model and capability map alignment。
Learning Experience Lead	owns learning paths, labs, scenario sequencing and learner experience。
Assessment Board	owns scenario packs, rubrics, reviewer calibration, evidence validity and appeals。
Community of Practice Lead	owns case clinics, peer review, pattern library and expert network。
Business Capability Owners	define target proficiency for roles in their domain and accept production skill evidence。
HR / L&D Partner	integrates career paths, performance processes, learning systems and reporting。
Risk / Compliance / Model Risk	reviews risk-sensitive skills, control evidence and high-risk scenario assessments。
Enterprise Architecture	ensures academy skills align to AI reference architecture, platform standards and architecture governance。

10.2 Governance Cadence

Cadence	Forum	Inputs	Decisions
Monthly	Academy product review	path adoption, assessment results, learner friction, role demand	backlog, path improvement, scenario refresh。
Monthly	COP case clinic	real project artifacts, incidents, eval failures, adoption blockers	patterns, anti-patterns, reusable examples。
Quarterly	Workforce capability review	skills debt, portfolio demand, role coverage, risk exposure	investment priority, hiring vs training, critical path。
Quarterly	Assessment calibration board	reviewer variance, critical failure rates, appeal cases	rubric revision, reviewer training, evidence validity。
Semiannual	Role taxonomy review	strategy changes, platform changes, regulatory pressure, market role shifts	role profile updates and path changes。

11. Metrics and Adoption Telemetry

11.1 Balanced Metrics

Dimension	Metric	Bad interpretation	Better interpretation
Coverage	% target roles with role profile	Role profile exists, so role is ready	Check whether target proficiency and evidence thresholds are defined。
Learning	path completion	Completion means competence	Treat as entry signal, then require assessment and portfolio evidence。
Competence	assessment pass rate by role and scenario	High pass rate always good	Check reviewer calibration and scenario difficulty。
Evidence	% high-risk roles with current evidence	Evidence file exists	Check artifact quality, validity period and production relevance。
Adoption	target role active use in production workflows	More use is always better	Combine usage with quality, override, incident and outcome signals。
Quality	error, override and defect patterns	Lower override always better	High-quality override may indicate strong human oversight。
Business value	cycle time, rework, cost per case, STP, complaint trend	Training caused all improvement	Attribute carefully through baseline, cohort and workflow changes。
Workforce risk	skills debt age, critical role coverage, key-person dependency	Only HR concern	Treat as AI portfolio and operational resilience risk。

11.2 Adoption Telemetry Architecture

Learning and assessment system
  -> evidence records and role proficiency
Production AI systems
  -> usage, errors, overrides, feedback, incidents
Workflow systems
  -> cycle time, queue, SLA, rework, quality
Risk and control systems
  -> issues, audit findings, model monitoring, policy exceptions
HR and workforce systems
  -> role assignment, team structure, retention, career moves
        |
        v
Capability analytics
  role readiness | adoption outcome | skill debt | workforce risk | investment priority

12. Workforce Risk and Skills Debt

Skills debt 是一种真实的企业架构风险:

AI portfolio demand grows faster than verified workforce capability.

Skills debt pattern	Signal	Risk	Treatment
Expert bottleneck	少数人审所有 eval、architecture、risk memo	release delay, burnout, key-person risk	create reviewer guild, train Level 4 reviewers, define reusable rubrics。
Role ambiguity	PM、BA、Risk、Architect 都以为别人负责 human oversight	control gap, incident confusion	update RACI, scenario drills, release gate ownership。
Tool enthusiasm without controls	团队会用 AI, 不会设计权限、日志、eval	data leakage, hallucination, customer harm	builder path requires architecture and risk evidence。
Course-rich, evidence-poor	学习平台数据好看, production adoption 差	false readiness	link path exits to portfolio artifacts and adoption signals。
Manager capability gap	一线学了工具, 经理不会 coach 和 review	behavior drift, inconsistent use	manager path, QA calibration, adoption huddle。
Data/knowledge owner shortage	RAG 项目多, source authority 和 freshness 无 owner	wrong answers, audit gaps	data product and knowledge owner capability path。

13. Financial Retail Examples

Role	Target capability	Skill evidence	Adoption outcome
AI BA	Contact center complaint triage workflow redesign	TO-BE flow, exception path, vulnerable customer scenario pack, acceptance criteria	First-assignment cycle time down, rework stable or lower, complaint escalation quality improved。
AI Product Manager	KYC onboarding document review copilot	PRD, benefit hypothesis, release gate, adoption dashboard, scale/stop memo	Manual review queue within threshold, false unsupported rejection in test pack at zero, reviewer adoption stable。
AI Solution Architect	Loan policy RAG assistant	source authority model, entitlement-aware retrieval design, ADR, monitoring plan	Citation accuracy and policy freshness monitored, no unauthorized knowledge exposure。
RiskOps / EvalOps Lead	AML alert investigation assistant	golden set, rubric, critical failure list, production sampling plan	QA issue rate decreases, hallucinated rationale incidents tracked and remediated。
Data Product Manager	Customer 360 AI context product	data contract, lineage, consent/preference handling, freshness dashboard	AI use cases reuse governed context, fewer duplicate data extracts。
AML/KYC Operations Lead	Human-AI case review operating model	reviewer SOP, override taxonomy, manager coaching routine, incident route	consistent override reasons, better calibration, reduced unsupported escalations。
Contact Center Transformation Lead	Knowledge assistant adoption system	role-based training, support tiers, feedback loop, COP clinic	target teams repeatedly use the assistant, feedback becomes eval backlog, AHT improves without quality drop。

14. Governance

Governance should avoid two extremes:

HR owns training but cannot connect to AI portfolio and risk.
Technology owns AI skills but cannot change roles, incentives and frontline behavior.

14.1 Governance Principles

Principle	Practice
Business capability owns demand	Capability owners define which roles must reach which proficiency for roadmap execution。
Academy owns platform, not all decisions	Academy runs role/skill/evidence system; business, risk, architecture and HR co-own decisions。
Evidence before authorization	High-risk AI responsibilities require current competence evidence before access or approval authority。
Continuous refresh	Role profiles and scenarios refresh when models, policies, tools, regulations or incidents change。
Workforce risk is enterprise risk	Skills debt and critical role gaps are reported in portfolio and operating risk reviews。

14.2 Control Points

Control point	Trigger	Required evidence
AI tool access	User enters risk-sensitive workflow	role-based literacy and responsible-use evidence。
Builder authorization	Team designs AI use case	Level 3 evidence for product, analysis or architecture path, or explicitly supervised assignment。
Release reviewer	Person signs eval/risk/adoption evidence	Level 4 reviewer evidence and reviewer calibration record。
Major role change	AI changes job tasks or authority	role impact assessment, training path, employee communication and manager enablement。
Incident response	AI misuse or quality issue	incident learning loop updates scenario pack and evidence requirements。

15. Anti-Patterns

Anti-pattern	Why it fails	Better design
"Everyone gets the same AI course"	不同角色风险、任务和责任完全不同	role-based paths with scenario assessment。
"Prompt engineering is the academy"	AI 能力远超过 prompt, 包括 eval、data、risk、architecture、adoption	skill ontology across product, BA, architecture, risk, data and ops。
"Badge equals capability"	badge 常证明参与, 不证明工作能力	evidence contract and portfolio review。
"Academy owned only by HR"	HR 能运营学习, 但无法独立定义 AI control and business capability demand	joint ownership with business, architecture, risk and HR。
"Assessment is multiple choice only"	高级 AI 岗位需要情景判断和 artifact production	scenario-based exam and portfolio defense。
"Adoption is a change-management afterthought"	不改变工作流和经理节奏, 学习不会转化为结果	adoption telemetry and manager routines embedded in path。
"Skills debt ignored until hiring crisis"	AI portfolio 会被关键岗位短缺阻塞	quarterly skills debt review and capability investment。
"Community is chat group activity"	活跃不等于标准沉淀	COP must produce reusable patterns, reviewed artifacts and corrected anti-patterns。

16. 零售银行能力学院蓝图

一个可落地的 AI Workforce Capability Academy 可以从 contact center 和 KYC operations 两个高频、高风险、强运营的场景切入。它们足够接近客户，能暴露 frontline adoption、quality review、policy boundary、exception handling 和 manager coaching 的真实问题；同时又足够结构化，适合沉淀角色画像、技能证据和治理节奏。

16.1 角色与能力蓝图

Capability area	需要覆盖的角色	能力证据
Workflow redesign	business analysis lead, process owner, operations lead	AS-IS/TO-BE flow, exception path, human oversight design, role impact memo。
Product value and adoption	product lead, business owner, frontline manager	benefit hypothesis, adoption dashboard, scale/stop memo, support model。
AI architecture and controls	solution architect, data product owner, platform engineer	architecture view set, ADR, entitlement-aware retrieval, logging, fallback and rollback design。
Eval and release confidence	EvalOps, QA, risk partner, business reviewer	golden set, critical failure list, release threshold, production sampling plan。
Workforce risk management	academy owner, HR/L&D, risk, capability owner	skills debt register, reviewer coverage, recertification trigger, critical-role gap treatment。

16.2 核心交付物

Artifact	内容
Role taxonomy	定义 product, analysis, architecture, EvalOps, RiskOps, Data Product, AML/KYC Ops, Contact Center Lead 的责任边界。
Skill ontology	8-12 个 skill domains, 每个 domain 关联 task、knowledge、practice、evidence、proficiency 和 risk if absent。
Role-to-capability map	显示哪些角色对 use case framing、workflow redesign、RAG governance、eval、risk control、adoption 和 incident learning 负责。
Proficiency model	用 Level 1-5 定义授权边界和 evidence threshold, 避免“学过课程就能签字”。
Scenario assessment	设计 KYC、AML、contact center 三类 case, 包含输入材料、约束、预期 artifact 和 critical failures。
Evidence registry	定义 evidence object 字段、reviewer、validity period、limitations 和 recertification trigger。
Adoption telemetry	连接 learning path、production usage、quality、override、incident、benefit 和 manager action。
Governance pack	支撑月度 academy product review、COP case clinic、季度 workforce capability review。

16.3 评审清单

Criterion	高质量信号
System design	Academy 被建模成有数据对象、反馈回路和治理节奏的能力系统, 不是培训列表。
Product operating logic	学习路径按角色、风险和 adoption 结果设计, 有明确 backlog 和版本管理。
Business analysis maturity	流程、规则、例外、stakeholder accountability、acceptance evidence 和 scenario assessment 能互相追溯。
Architecture maturity	source authority、permission、observability、fallback、rollback、release control 都能进入能力证据。
Risk maturity	skills debt、critical role coverage、high-risk authorization 和 control evidence 能进入治理讨论。
Outcome linkage	能解释哪些能力变化影响了 cycle time、quality、incident、override、benefit 和 customer experience。

17. Final Mental Model

AI workforce capability is not learned once.
It is architected, evidenced, practiced, governed and measured against real adoption.

The senior-level move is to stop asking:

How many people finished AI training?

and start asking:

Which roles can now safely own which AI capabilities,
what evidence proves it,
where does production adoption confirm it,
and which skills debt still threatens our AI portfolio?

SOTA 状态标注 (2026-07-01)

本篇属于第二、三遍深读池（参考架构/深读笔记），未列入 12 周主线必读。时效基线为写作时点；引用前请按 CLAUDE.md 全局时效性硬规则复查最新进展。模块级 SOTA 对照见 docs/AI_SYSTEMATIC_LEARNING_ROADMAP_2026.md 各周「2026 SOTA 对照」行与文末「SOTA 检查」。