AI 扩展计划 / Playbooks

AI Workforce / HR Decision / Employee Monitoring Governance Playbook

Workforce AI 不是普通企业效率工具, 它可能影响候选人和员工的机会、评价、排班、晋升、培训、纪律、调查和工作体验。治理重点是把 AI 输出放回 employment decision chain, 明确目的、数据、模型、人工判断、沟通、申诉、监控和证据的边界。

807 行AI_WORKFORCE_HR_DECISION_EMPLOYEE_MONITORING_GOVERNANCE_PLAYBOOK.md

AI Workforce / HR Decision / Employee Monitoring Governance Playbook

配对阅读：本手册的原理/架构解读版是 docs/ai-foundations/papers/125-ai-workforce-hr-decision-employee-monitoring-governance-architecture.md。先读 paper 建立机制与取舍，再用本手册落地为模板、RACI 与门禁，两者不需要重复精读。

版本: v1.0 日期: 2026-06-30

核心导读: Workforce AI 不是普通企业效率工具, 它可能影响候选人和员工的机会、评价、排班、晋升、培训、纪律、调查和工作体验。治理重点是把 AI 输出放回 employment decision chain, 明确目的、数据、模型、人工判断、沟通、申诉、监控和证据的边界。

控制链路: 从 use-case inventory、AI authority、impact tiering 和 employee data minimization 开始, 再进入 hiring / scheduling / QA / monitoring / mobility 等场景, 最后用 human review、notice、explanation、adverse impact testing、purpose separation、evidence ledger 和 governance cadence 形成闭环。

系统边界: 本文关注金融零售 workforce AI 的治理架构, 不把算法公平、员工监控、劳动关系、隐私和模型风险拆成孤立话题; 也不把 AI 工具能力等同于可上线的雇佣影响系统。

运行证据: 成熟治理应形成 workforce AI inventory、impact tier map、data minimization map、decision authority matrix、review protocol、notice / explanation / appeal flow、adverse impact test plan、monitoring purpose boundary、evidence ledger、KRI dashboard 和 limited-release decision memo。

重要说明: 本文是架构与治理参考, 不是法律意见、HR 合规建议、劳动关系建议、雇佣决策建议、监管解释或正式政策文本。真实项目必须由 Legal、HR、Employee Relations、Compliance、Privacy、Security、Works Council / Labor Relations where applicable、Model Risk、Internal Audit、业务 owner 和相关员工代表机制按 jurisdiction、role、employment law、union / works council context、data type、decision impact、vendor contract 和 internal policy 确认。

1. Executive Framing

1.1 核心观点

Workforce AI 的治理目标不是阻止 AI 提升 HR、运营和管理效率, 而是确保 AI 不会在不透明、不可复核、不可申诉、不可解释的情况下影响候选人和员工的工作机会、排班、绩效、培训、晋升、薪酬、纪律或离职。

1.1.1 Governance Formula

Workforce AI governance =
make every AI-influenced worker decision bounded, necessary, reviewed,
explainable, fair-tested, privacy-aware, challengeable and evidence-backed.

1.2 金融零售为什么特殊

金融零售企业通常同时拥有大规模 frontline workforce、高密度 customer / employee interaction data、高监管敏感流程、强绩效激励、多 jurisdiction 和劳动关系场景。

Workforce AI 会跨越:

ATS / HRIS / WFM / QA / LMS / CRM / SIEM / case management /
data lake / model gateway / vendor platform / manager dashboard / employee portal

系统设计任务不是采购一个 HR AI 工具, 而是设计一条可治理的 decision chain。

1.3 高管应该问的 10 个问题

Question	强答案应该包含
哪些 AI 工具影响候选人或员工?	workforce AI inventory, impact tier, owner
AI 是推荐还是决定?	decision authority matrix, final decision owner
使用哪些员工数据?	field-level data minimization map
候选人或员工是否被告知?	notice plan, communication evidence
人工复核是否真实有效?	review protocol, override log, calibration
是否测试 adverse impact?	testing plan, slice metrics, remediation record
是否考虑 accessibility?	inclusive design review, accommodation path
监控数据是否用途受限?	purpose separation, access controls
供应商声明是否被验证?	independent eval, contract-control evidence
出现争议如何解释和纠正?	explanation packet, appeal workflow, evidence ledger

Workforce AI 在一线运营中通常嵌入招聘、排班、质检、培训、监控和内部流动。治理重点是让 AI 影响链可见、可质疑、可纠正:

Control lens	Workforce AI design
风险链路	员工或候选人数据被采集 -> AI 生成 score/rank/flag/recommendation -> 管理者采纳或覆盖 -> 机会、负荷、评价或调查结果发生变化。
控制假设	AI 可以帮助总结证据、识别模式和提示复核, 但不能在不透明、不可申诉、无目的限制的情况下成为雇佣影响动作的实际决策者。
证据闭环	高影响场景需要 use-case card、数据最小化表、authority matrix、review log、override reason、notice/explanation、appeal result 和 adverse impact review。
例外处理	自动拒绝、纪律调查、监控二次用途、无障碍争议、群体影响信号、供应商模型变更或劳动关系敏感流程必须进入增强审查。
运营机制	use-case inventory 持续维护, 阶段性 fairness/accessibility review, manager calibration, notice/appeal workflow, KRI dashboard 和 stop trigger。
失败模式	把 manager review 做成橡皮图章、把安全监控挪作绩效排名、只信供应商公平声明、忽略无障碍路径或无法解释不利结果。

2. Source Anchors

以下锚点用于组织术语、治理问题和证据结构。它们不自动形成任何机构的法律或合规结论。

Anchor	Official source	本 playbook 使用方式
EEOC Artificial Intelligence and Algorithmic Fairness Initiative	https://www.eeoc.gov/ai	作为 employment AI fairness、algorithmic decision-making 和 hiring / employment decision 的主题入口。
EEOC technical assistance: software, algorithms, AI and Title VII	https://www.eeoc.gov/laws/guidance/select-issues-assessing-adverse-impact-software-algorithms-and-artificial	用于 selection procedures、adverse impact、algorithmic tools 和 employment selection evidence 的治理锚点。
U.S. Department of Labor AI Principles for Developers and Employers	https://www.dol.gov/general/AI-Principles	用 worker-centered design、transparency、meaningful human oversight、worker rights、AI training、worker data protection 组织原则。
DOL ODEP AI and inclusive hiring framework	https://www.dol.gov/agencies/odep/program-areas/employers/ai	用 inclusive hiring、accessibility、disability inclusion、AI hiring technology assessment 做补充锚点。
NIST AI RMF	https://www.nist.gov/itl/ai-risk-management-framework	用 Govern / Map / Measure / Manage 组织 workforce AI 风险分级、评测、控制、监控和持续治理。
FTC AI claims guidance	https://www.ftc.gov/business-guidance/blog/2023/02/keep-your-ai-claims-check	用于 vendor claims challenge: 不接受未经证据支持的 AI accuracy、fairness、bias-free、human-like 或 superiority 声明。

Source-to-artifact mapping:

Source lens	Artifact	高级表达
EEOC AI / adverse impact	adverse impact test plan, selection procedure inventory	AI screening 是 selection procedure risk, 不是普通排序功能。
DOL worker well-being	worker-centered governance checklist	worker voice、transparency、training、data protection 和 human oversight 必须进入 operating model。
ODEP inclusive hiring	accessibility and accommodation review	“招聘 AI 要同时看 fairness 和 accessibility, 不只看模型准确率。”
NIST AI RMF	workforce AI risk register and monitoring loop	“我用 Govern / Map / Measure / Manage 管 AI employment impact, 而不是一次性审批。”
FTC claims guidance	vendor claim substantiation matrix	“供应商说公平或准确不是证据, 我要求场景化评测和可审计材料。”

3. Workforce AI Use-Case Taxonomy

3.1 Lifecycle taxonomy

Workforce lifecycle	Use cases	典型 AI output	主要风险
Workforce planning	staffing forecast, branch coverage, attrition trend	demand forecast, capacity gap	聚合分析被误用为个人不利决定
Hiring / sourcing	candidate matching, job ad optimization, sourcing	match score, outreach list	proxy bias, accessibility barrier
Screening / assessment	resume parsing, skills test, interview analysis	rank, score, pass / review flag	adverse impact, opaque exclusion
Scheduling	shift allocation, overtime, workload balancing	schedule recommendation	unequal burden, accommodation conflict
Performance / productivity	KPI insight, anomaly, manager dashboard	score, ranking, trend	automation bias, context loss
QA / coaching	call / chat QA, script adherence, sales quality	defect flag, coaching topic	language / accent / customer mood bias
Monitoring / security	device, location, system access, data exfiltration	alert, risk score	surveillance creep, false accusation
Training / development	skill gap, learning path, certification	training recommendation	historical opportunity inequality
Promotion / mobility	internal candidate matching, succession	readiness signal	invisible gatekeeping
Discipline / investigation	misconduct signal, policy breach analytics	investigation flag	due process and evidence weakness
Exit / retention	attrition risk, retention campaign	risk segment, manager prompt	intrusive inference, retaliation perception

3.2 AI authority taxonomy

AI authority	Definition	Workforce examples	Governance expectation
AI-search	only retrieves workforce records or policies	HR policy search, candidate search	access, source, logging
AI-summarize	summarizes existing evidence	interview note summary, QA summary	source citation, human check
AI-score	assigns score or risk flag	candidate score, QA score, attrition risk	validation, fairness, explanation
AI-rank	orders people or cases	candidate ranking, promotion shortlist	adverse impact review, usage audit
AI-recommend	suggests action	coaching, training, shift changes	decision boundary, review
AI-monitor	observes behavior	screen, call, access, location	purpose limitation, notice
AI-trigger	starts workflow	investigation, manager review	threshold, false positive control
AI-decide	makes final employment-impacting decision	auto-reject, auto-discipline	high risk; requires strict review before use

Architecture principle:

The same model output becomes higher risk when it changes authority:
summary < score < rank < recommendation < trigger < final decision.

3.3 Impact tiering

Tier	Description	Examples	Required controls
L1 Low impact	Personal productivity, no people ranking	HR draft email, policy summarizer	acceptable use, data protection
L2 Operational support	Helps process work, no direct adverse outcome	schedule forecast, training draft	review, logging, purpose control
L3 Employment influence	AI score/rank/recommendation can influence opportunity	candidate ranking, QA score	human review, adverse impact test, notice
L4 High-impact employment decision	AI materially affects hiring, pay, promotion, discipline, termination, schedule burden	auto-screening, bonus score, misconduct flag	formal governance gate, legal/HR review
L5 Not approved in current architecture	AI replaces accountable human decision or uses unacceptable data	emotion as sole criterion, undisclosed monitoring	block or redesign

4. Hiring / Screening Governance

4.1 Hiring AI workflow

job requisition
  -> job-relevant criteria definition
  -> AI tool intake and impact tiering
  -> data and accessibility review
  -> candidate interaction notice
  -> model / rule / vendor evaluation
  -> shortlist or score generation
  -> recruiter / hiring manager review
  -> final selection decision
  -> adverse impact monitoring
  -> appeal / accommodation / correction loop

4.2 Selection criteria architecture

Artifact	Good design	Weak design
Job criteria	tied to essential functions and documented competencies	generic “culture fit” or opaque score
Data fields	mapped to criterion and necessity	all resume text and public data ingested
Score explanation	shows evidence, uncertainty and missing information	single percentile without context
Human review	recruiter sees evidence and can override with reason	recruiter only sees green / red
Accessibility	alternative assessment path and accommodation workflow	one fixed AI assessment for everyone
Adverse impact	monitored by selection stage and group where lawful and available	only overall hiring rate tracked

4.3 Hiring controls

Control ID	Control objective	Evidence
HIR-01	AI screening is registered as workforce AI use case	use-case card, owner, tier
HIR-02	Selection criteria are job-related and approved by HR / Legal process	criteria matrix, review record
HIR-03	Data fields are necessary for the approved criteria	field map, denied fields
HIR-04	Candidate-facing workflow supports accessibility and accommodation	accessibility review, alternate path
HIR-05	AI score does not auto-reject unless explicitly approved under applicable law and policy	workflow config, sample trace
HIR-06	Recruiters review evidence, not only rank	review UI evidence, decision log
HIR-07	Adverse impact is monitored by stage	selection funnel report
HIR-08	Vendor claims are validated in institution context	eval report, substantiation file

4.4 Candidate explanation packet

Element	Purpose
tool purpose	what the tool supports
data categories	which data categories are used
decision role	whether AI screens, ranks, recommends or summarizes
human decision owner	who makes final selection decision
accommodation path	how accessibility needs are handled
correction route	how candidate can correct inaccurate data where applicable
evidence retained	what records are retained for review and audit

5. Scheduling Governance

5.1 Scheduling AI in financial retail

金融零售排班 AI 常见场景:

branch teller / banker shift optimization。
contact center staffing and schedule adherence。
fraud / AML queue capacity allocation。
overtime recommendation and peak coverage。
training slot scheduling。

5.2 Scheduling risk model

Risk	Example	Metric / evidence
Unequal undesirable shifts	late / weekend shifts disproportionately allocated	shift equity dashboard
Accommodation conflict	AI ignores approved accommodation or availability rule	blocked assignment log
Last-minute burden	certain employees receive more short-notice changes	change notice distribution
Hidden performance penalty	schedule adherence score penalizes system-driven changes	root-cause tagged adherence
Workload imbalance	high-risk calls or complex cases concentrated by employee	workload mix by skill / risk

5.3 Scheduling control design

forecast demand
  -> constraint library
  -> employee availability and accommodation guardrails
  -> AI schedule proposal
  -> fairness / burden check
  -> manager approval
  -> employee notice
  -> exception and appeal handling
  -> post-period equity review

Control	Evidence
Constraint library	approved rules, accommodation flags, labor rules where applicable
Schedule proposal trace	demand version, constraints, model version
Burden distribution check	late shift, weekend, overtime, split shift distribution
Manager approval	final approver, override reason
Employee challenge route	request, resolution, timing
Works council / labor review where applicable	consultation record, agreed conditions

6. Performance / QA / Employee Monitoring

6.1 Performance AI boundary

Performance AI is high risk when it affects bonus、performance rating、promotion readiness、disciplinary action、termination、schedule preference or opportunity。

设计原则:

AI performance analytics may inform coaching and quality improvement.
It should not silently become an automated performance management system.

Avoid single opaque productivity score. Use decomposed, contextual signals:

Signal type	Example	Context needed
Work volume	calls handled, cases closed	queue type, case complexity, staffing
Quality	QA defects, policy accuracy	sample method, reviewer calibration
Customer outcome	FCR, complaint, escalation	customer segment, issue type
Compliance	script adherence, required disclosures	channel, policy version
Learning	training completion, coaching response	access to training, manager support
Exceptions	outage, leave, accommodation	protected and sensitive handling

6.2 Contact center QA AI

voice / chat interaction
  -> transcription and redaction
  -> policy / script / complaint classifier
  -> AI QA scoring and issue tagging
  -> human QA sample and calibration
  -> coaching recommendation
  -> employee feedback
  -> performance boundary decision
  -> monitoring and appeal loop

Failure	Example	Control
Accent / language bias	non-native accent lowers sentiment score	slice testing and human review
Customer mood transfer	angry customer lowers agent score	separate customer sentiment from agent behavior
Script rigidity	AI penalizes correct deviation for vulnerable customer	exception taxonomy and review
Sample bias	only AI-flagged calls reviewed	mixed random + risk-based sample
Coaching-to-discipline creep	coaching flags feed performance file automatically	purpose boundary and approval

6.3 Employee monitoring taxonomy

Monitoring type	Examples	Primary governance concern
Security monitoring	data exfiltration, privileged access, abnormal downloads	protect systems while avoiding false accusation
Compliance monitoring	required disclosures, sales practice, policy adherence	evidence and proportionality
Safety monitoring	branch security, lone worker safety	purpose and retention
Productivity monitoring	activity, idle time, app usage	overreach and context loss
Location monitoring	device / branch / field location	necessity and notice
Sentiment / emotion analytics	voice tone, facial expression, emotion inference	high sensitivity and reliability concern
Biometric / wearable data	health, fatigue, physical movement	heightened data sensitivity

6.4 Purpose separation

Monitoring systems must distinguish purpose:

security incident detection
!= productivity measurement
!= performance evaluation
!= disciplinary evidence
!= training recommendation

Purpose separation controls:

separate data stores or access views。
separate role entitlements。
policy decision point before secondary use。
reason-coded access。
audit log review。
retention by purpose。
employee notice by purpose。

6.5 Monitoring evidence bundle

Field	Purpose
monitoring_purpose_id	proves approved purpose
data_source_id	source and sensitivity
collection_notice_version	communication evidence
model_or_rule_version	reproducibility
alert_threshold	threshold governance
employee_context	shift, role, access, system condition
human_triage_result	separates alert from finding
investigation_case_id	formal process trace
final_outcome	no action, coaching, security action, HR action
review_decision	appeal, correction, closure

7. Training / Promotion / Internal Mobility

7.1 Training recommendation

AI training recommendations can be positive, but they can also channel employees into different opportunity paths.

Controls:

recommendations should show skill evidence and optionality。
employees should see and correct skill profile data where applicable。
manager should not treat training recommendation as fixed potential score。
training access should be tracked to avoid unequal opportunity。

7.2 Promotion readiness

Promotion AI should avoid hidden ranking、historical bias replication、manager preference encoded as objective signal、over-reliance on visibility metrics, and penalizing leave, flexible work, accommodation or role assignment history without context。

Promotion evidence design:

Evidence	Why it matters
competency map	ties signal to role requirement
opportunity exposure	distinguishes performance from access
manager review notes	human context
employee-provided evidence	employee-submitted work evidence
calibration record	cross-team fairness
recommendation coverage	who is surfaced and who is missed

7.3 Internal mobility AI

skills profile
  -> employee consent / visibility settings
  -> role requirement matching
  -> gap explanation
  -> employee and manager review
  -> application / nomination workflow
  -> outcome and fairness monitoring

8. Human Review / Notice / Explanation

8.1 Meaningful human review

Human review is meaningful when reviewer has access to material evidence、knows AI limitations、can override without penalty、records reason、is calibrated、has feasible workload, and review occurs before employment-impacting action。

Human review is weak when reviewer only sees AI score、review is after action、override is discouraged、no reason code is captured、manager has no training, or queue pressure makes review automatic。

8.2 Review protocol

Review element	Design
Trigger	high-impact score, adverse action, low confidence, protected process, appeal
Reviewer	trained HR / manager / specialist depending on use case
Evidence	source data, score explanation, comparison, uncertainty, policy
Action	approve, modify, reject, escalate, request more evidence
Reason code	evidence insufficient, context missing, AI error, policy exception
Calibration	periodic sample review across reviewers
Quality metric	overturn rate, consistency, error found, review time

8.3 Override analysis

Override pattern	Interpretation
Very low override rate	possible automation bias or rubber-stamp review
Very high override rate	model / rule not useful or poor UX
Manager-specific override outlier	training, interpretation or incentive issue
Group-specific override pattern	potential fairness or process concern
Override reasons repeated	backlog for model / process improvement

8.4 Notice architecture

Notice is not one PDF. It is a lifecycle capability:

Moment	Example notice
before assessment	AI or automated tool may support screening
before monitoring	data categories, purpose, retention, access
before AI-assisted QA	calls may be scored for quality / coaching
before data reuse	workforce analytics purpose and limits
when decision is made	decision owner, review route, correction route

Notice quality: plain language、role-specific、channel appropriate、versioned、retained as evidence、aligned with internal policy and applicable law。

8.5 Explanation design

Explanation should answer:

What was the AI used for?
What data categories mattered?
What did AI recommend or flag?
Who reviewed it?
What final decision was made?
How can the person request review or correction?

Avoid explanation anti-patterns:

“The algorithm selected the most qualified candidates.”
“This is based on AI insights.”
“Your productivity risk is high.”
“The model is unbiased.”

9. Adverse Impact Testing / Data Minimization

9.1 Testing architecture

Adverse impact testing should be designed with legal and HR owners. Architecture must support measurement, evidence and remediation workflows.

population definition
  -> decision stage definition
  -> approved group data handling
  -> outcome and denominator construction
  -> metric calculation
  -> qualified review
  -> root-cause analysis
  -> remediation decision
  -> retest and monitoring

9.2 Testing scope

Testing can apply to:

candidate sourcing list。
resume screen pass rate。
assessment score distribution。
interview recommendation。
shortlist generation。
offer decision。
promotion recommendation。
schedule burden。
performance score distribution。
discipline trigger rate。

9.3 Metrics and slices

Metric / slice	Use
selection rate by stage	hiring funnel impact
score distribution	ranking / scoring behavior
pass rate ratio	screening disparity signal
false positive / false negative by group	QA / monitoring fairness
schedule burden by group / role	workforce scheduling fairness
promotion recommendation coverage	opportunity surfacing
appeal upheld by group	correction and process quality
manager / location slice	local process issues

Do not treat a single metric as legal conclusion. Use metrics as governance evidence for qualified review.

9.4 Remediation options

Finding	Possible response
Data source creates proxy risk	remove field, transform feature, add review
Score not job-related	redesign criteria, change rubric
Assessment inaccessible	provide alternative, redesign interaction
Ranking creates hidden cutoff	remove cutoff, add recruiter review
Manager usage creates disparity	training, dashboard redesign, audit
Model drift	retrain, restrict, rollback, pause
Vendor cannot explain	hold deployment or change vendor architecture

9.5 Employee data minimization

Data category	Examples	Governance concern
Identity and role	employee id, role, location, manager	access and role hierarchy
HR lifecycle	application, performance, promotion, discipline	employment impact
Work activity	cases, calls, tasks, schedule adherence	context and purpose
Communications	email, chat, call transcripts	privacy and labor expectations
Device / security	login, system access, file activity	secondary use risk
Location	branch, field, device location	necessity and retention
Biometric / health-adjacent	voiceprint, facial, fatigue, wearable	heightened sensitivity
Derived AI signals	score, risk flag, skill inference	explanation and correction

Minimization questions:

Is the field necessary for the approved workforce purpose?
Is there a less sensitive alternative?
Is raw data needed, or can an aggregate / manifest work?
Who can access it, and under which purpose?
How long is it retained?
Can it be reused for another purpose?
Is employee notice aligned?
Can material data be corrected?

10. Works Council / Legal / Labor Review Where Applicable

10.1 Applicability factors

Review depth depends on jurisdiction、worker status、union / works council context、monitoring type、data sensitivity、employment decision impact、whether tool changes work organization or performance management、internal policy and collective agreements。

10.2 Architecture implications

Review concern	Architecture artifact
what data is collected	data inventory and sample payload
why it is collected	purpose statement and necessity rationale
how it affects workers	decision boundary and impact assessment
whether workers are informed	notice and training plan
whether review is meaningful	human review protocol
whether data can be reused	purpose separation controls
whether monitoring is proportional	monitoring boundary and retention
how workers can challenge	appeal / correction workflow
how changes are governed	change notice and release gate

10.3 Consultation package

A pragmatic consultation package can include:

one-page executive framing。
current and target process map。
data field list。
AI authority taxonomy。
employee impact tier。
monitoring purpose boundary。
human review and appeal flow。
adverse impact / accessibility test plan。
retention and access rules。
pilot scope and stop triggers。

11. Governance Workflow

11.1 End-to-end workflow

intake
  -> use-case taxonomy and impact tier
  -> data minimization and purpose review
  -> decision boundary design
  -> vendor / model / rule review
  -> adverse impact and accessibility plan
  -> notice / explanation / appeal design
  -> pilot release gate
  -> monitored pilot
  -> scale / hold / redesign / stop
  -> ongoing monitoring and annual review

11.2 Gate design

Gate	Decision	Evidence
Intake gate	accept / reject / narrow scope	use-case card, impact tier
Data gate	allow / minimize / block fields	data map, purpose review
Decision gate	approve AI role	authority matrix
Vendor / model gate	approve / require eval / block	vendor evidence, eval plan
Fairness gate	pilot allowed / restricted	adverse impact test design
Worker communication gate	launch ready / revise	notice, training, explanation
Pilot gate	limited go / no-go	controls, monitoring, stop triggers
Scale gate	scale / hold / stop	pilot metrics, KRIs, appeals
Change gate	release / rollback	regression and impact review

11.3 Stop triggers

Trigger	Response
confirmed critical adverse impact signal	pause affected workflow and review
unapproved data reuse	disable integration and investigate
human review capacity overwhelmed	restrict scope
high upheld appeal rate	root-cause and remediation
vendor model change without notice	freeze or rollback
monitoring purpose drift	access revoke and governance review
works council / legal condition not met where applicable	hold release
employee trust / complaint signal worsens materially	change management and investigation

12. Evidence Architecture

12.1 Evidence ledger

Every high-impact workforce AI run should be reconstructable:

Evidence field	Example
workforce_ai_use_case_id	`WAI-HIR-ResumeScreen-2026-01`
person_context	candidate / employee role category, not unnecessary sensitive detail
decision_stage	sourcing, screening, QA, schedule, performance
input_data_class	resume, call transcript, schedule, system access
model_or_rule_version	vendor model, internal rule, prompt version
AI output	score, rank, flag, summary, recommendation
confidence / uncertainty	score band or review flag
human_reviewer_id	trained reviewer or manager
review_outcome	accepted, modified, rejected, escalated
final_decision_owner	human accountable role
notice_version	communication proof
explanation_packet_id	worker / candidate explanation
appeal_case_id	challenge and correction trace

12.2 Evidence binder

Evidence artifact	What it proves
use-case inventory	no hidden workforce AI
impact tiering	governance depth matches decision impact
data minimization map	fields are necessary and purpose-bound
decision authority matrix	AI is not silently final authority
human review calibration	review quality is operational
adverse impact report	outcomes are monitored and reviewed
accessibility review	barriers are assessed
notice / explanation records	transparency is operationalized
appeal / correction log	challenges feed improvement
monitoring KRI dashboard	ongoing drift and misuse detection
vendor claim substantiation	claims are challenged with evidence
change gate record	model/rule/workflow changes are controlled

12.3 Audit replay

Audit replay should answer:

For this candidate / employee-impacting decision:
which AI tool was used,
which data categories were processed,
what score or recommendation was produced,
who reviewed it,
what final decision was made,
what explanation and appeal path existed,
and what monitoring evidence supports fairness and control operation?

13. Operating Model / Metrics / KRIs

13.1 RACI

Role	Responsibility
Business owner	owns workflow outcome and operational adoption
HR owner	owns employment process, policy alignment and employee experience
Product governance owner	owns product scope, user journey, release gates and metrics
Process / decision analyst	owns requirements, decision objects, data flows and exception paths
System architect	owns system design, integration, logging and evidence
Legal / Compliance	reviews applicable legal, regulatory and policy constraints
Privacy	owns employee data minimization, purpose, retention and rights workflow
Security	owns monitoring, access, incident and misuse controls
Employee Relations / Labor Relations	owns worker relations and consultation context
Works Council / union interface where applicable	provides required consultation / agreement path
Model Risk / Eval owner	owns validation, adverse impact testing and monitoring
Internal Audit	independently reviews control design and evidence quality
Vendor owner	owns vendor evidence, contract controls and change notices

13.2 Governance cadence

Cadence	Participants	Agenda
Weekly pilot review	product governance owner, HR owner, ops, architect, eval owner	defects, appeals, review load, adoption
Monthly workforce AI governance	HR, Legal, Privacy, Security, Risk, product governance owner	KRIs, adverse impact, monitoring boundary, incidents
Quarterly vendor review	vendor owner, HR tech, procurement, risk	model changes, evidence, SLA, claims, roadmap
Semiannual impact review	HR leadership, risk, audit, employee relations	outcome distribution, trust, policy updates
Annual inventory attestation	all owners	hidden tools, retired tools, data flows, access

13.3 Balanced metrics

Metric family	Examples
Efficiency	time-to-hire, scheduling cycle time, QA throughput
Quality	reviewer agreement, defect rate, rework, calibration score
Fairness	selection rate, score distribution, schedule burden, appeal upheld rate
Transparency	notice coverage, explanation packet completeness
Human oversight	review completion, override rate, review queue age
Privacy	minimized fields, secondary-use requests, access exceptions
Worker trust	survey, complaints, consultation issues, adoption sentiment
Risk	adverse impact signal, monitoring purpose drift, vendor change breach
Value	qualified decision support events, reduced backlog with stable risk

13.4 KRIs

KRI	Why it matters	Response
high-impact AI not in inventory	hidden governance gap	freeze deployment and register
unexplained score drift	model or data change	investigate, regression test
low human override rate	possible rubber stamp	review training and UI
high review queue age	oversight not feasible	reduce scope or add capacity
adverse impact signal	possible fairness issue	qualified review and remediation
appeal upheld spike	data / model / process defect	root-cause and correction
monitoring data reused	purpose creep	revoke access and review
vendor unnotified change	evidence gap	freeze or retest
notice coverage gap	transparency weakness	communication remediation
manager outlier pattern	local process risk	calibration and audit

14. Financial Retail Case Patterns

14.1 Branch staffing optimizer

Area	Design
AI role	forecast demand and propose schedule
Human role	branch manager approves and handles exceptions
Data	traffic, appointments, skills, availability, approved constraints
Controls	accommodation guardrail, shift equity dashboard, override log
KRI	last-minute change burden, undesirable shift concentration
Governance evidence	balances customer service efficiency with worker fairness

14.2 Contact center QA copilot

Area	Design
AI role	tag calls, draft QA observations, suggest coaching
Human role	QA reviewer validates and separates coaching from discipline
Data	transcripts, policy, call metadata, complaint link
Controls	language slice test, random sample, challenge workflow
KRI	AI-human disagreement, appeal upheld, score drift by channel
Governance evidence	improves service quality without black-box performance management

14.3 Retail banking hiring screen

Area	Design
AI role	parse resume and surface job-relevant evidence
Human role	recruiter makes screening decision
Data	resume, application, assessment, job criteria
Controls	accessibility review, adverse impact by stage, no auto-reject
KRI	selection rate change, accommodation issues, recruiter override
Governance evidence	demonstrates EEOC-aware selection architecture

14.4 Insider-risk monitoring

Area	Design
AI role	detect anomalous access or data movement
Human role	security triage and formal HR process if needed
Data	access logs, customer data access, device telemetry
Controls	purpose separation, false-positive review, investigation evidence
KRI	false positive, monitoring purpose drift, access review exceptions
Governance evidence	separates security control from generalized employee surveillance

15. Reusable Evidence Patterns

Pattern	Completed example
Use-case card	`WAI-QA-CC-2026-01`: Contact Center QA Copilot; AI tags call topics, policy risk and coaching opportunities; QA reviewer validates before record finalization; L3 employment influence; excludes biometric identity, off-work behavior and unrelated HR records.
Decision authority matrix	Candidate shortlist: AI surfaces job-relevant evidence, recruiter decides; Branch schedule: AI proposes, branch manager approves; QA coaching note: AI drafts, QA reviewer validates; Security investigation: AI flags anomaly, security triage validates.
Employee data minimization	Call transcript allowed for QA coaching; customer sentiment limited to service context; private social media blocked; privileged access logs restricted to insider-risk detection; after-shift location blocked for scheduling.
Adverse impact test plan	Retail banking candidate screen; decision stage = resume screen to recruiter review; metrics = selection rate, score distribution, pass-rate ratio, recruiter override; escalation = material signal triggers root-cause review.
Monitoring purpose boundary	Security anomaly data may detect unusual customer record access, but is blocked from daily productivity ranking; QA transcript supports coaching, but not automatic discipline.

16. 30-Day Controlled Build-Out

目标: 30 天内完成一套 Workforce AI Governance Architecture Pack, 用来验证雇佣影响 AI 的系统边界、控制取舍和证据要求。

Day range	Theme	Outputs
1-5	Scenario, process, AI authority, impact tier, data inventory	use-case card, BPMN current-state, AI authority map, impact matrix, employee data map
6-10	Data minimization, decision boundary, human review, notice	minimization decision table, authority matrix, review protocol, notice lifecycle map
11-15	Explanation, appeal, accessibility, adverse impact	explanation packet, challenge workflow, accessibility checklist, test plan and metrics
16-20	QA / performance boundary, monitoring purpose, vendor claims, evidence	coaching-vs-discipline control, purpose matrix, substantiation matrix, evidence ledger
21-25	Operating model, KRI dashboard, legal / labor package, pilot gate	RACI, cadence, KRI thresholds, consultation pack, limited-go memo
26-30	Stop triggers, decision rationale, evidence pack, design review, operating principle	rollback checklist, decision rationale, evidence pack, gap fixes, 5-minute system explanation

Completion bar: workforce AI use-case card, data minimization map, decision authority matrix, human review protocol, notice / explanation / appeal flow, adverse impact and accessibility test plan, monitoring purpose boundary, evidence ledger schema, operating model, KRI dashboard, decision rationale set。

17. Architecture Decision Rationale

Design question	Control rationale	Evidence and architecture focus
Workforce AI governance 和普通 AI governance 有什么不同?	Workforce AI 直接影响机会、评价、排班、晋升、纪律和监控, 所以要先做 inventory、impact tiering、data minimization、decision boundary 和 evidence ledger。	说明 employment impact、worker power imbalance、accessibility、employee data sensitivity、monitoring purpose creep 和 labor relations; 用 AI authority 从 search 到 decide 分级。
如何设计 hiring AI 治理架构?	把 hiring AI 当成 selection procedure risk: job-related criteria、accessibility review、no silent auto-reject、stage-level adverse impact monitoring。	讲 job requisition -> criteria -> tool intake -> data review -> candidate notice -> score -> recruiter review -> final decision -> monitoring; vendor claim 必须场景化验证。
Meaningful human review 怎么判断?	Reviewer 必须看到证据、理解限制、能 override、记录原因, 且有时间和校准。	用 review before action、source evidence、real override、reason code、training、workload 六项判断; 极低或极高 override rate 都要调查。
如何做 adverse impact testing?	与 Legal、HR、Model Risk 定义 population、stage、outcome、denominator、group data handling、metric 和 remediation workflow。	强调按 source、screen、interview、offer 等阶段测试; 指标是治理证据, 不是单独法律结论。
员工监控 AI 最大风险是什么?	最大风险是 purpose creep: security detection 不能未经审查变成 productivity ranking。	用 monitoring purpose boundary、purpose-based access、reason-coded access、audit log 和 secondary-use review 说明架构。
如何处理 works council / union context?	不把劳动关系审查放到最后; discovery 阶段就准备数据、目的、影响、监控边界、人审、notice、appeal 和 pilot scope。	说明适用性由 Legal、HR、Labor Relations 确认; consultation package 要描述具体系统行为。
如何审查 vendor “bias-free AI”?	不接受绝对声明, 要训练范围、适用岗位、评测方法、adverse impact / accessibility evidence、change notice 和 audit log。	用 FTC-style claim discipline: fairness claim must map to population、job、outcome、time window and customer-side testing。
如何形成系统证据包?	选 contact center QA 或 branch scheduling, 做 inventory、impact tier、data map、authority matrix、review、notice、impact test、monitoring、evidence、KRI。	重点不是泛泛谈 HR 合规, 而是证明能把 employment-impacting AI 变成可运行 architecture, 并用 scale / stop memo 讲治理判断。

18. System Evidence Pack

Evidence artifact	内容	验证重点
System framing memo	business value, worker risk, governance thesis	说明为什么该系统需要 workforce-specific governance
Workforce AI inventory	use cases, owners, impact tiers	发现 hidden AI
AI authority map	search / score / rank / monitor / decide	决策边界
Data minimization map	employee data fields, purpose, retention	隐私和数据治理
Decision authority matrix	AI role, human role, final owner	accountability
Human review protocol	triggers, evidence, override, calibration	meaningful oversight
Notice / explanation pack	worker-facing transparency architecture	信任和申诉
Adverse impact test plan	population, stage, metric, remediation	fairness governance
Accessibility review	inclusive hiring / worker access path	inclusive design
Monitoring boundary	purpose separation and secondary use controls	employee monitoring governance
Evidence ledger schema	audit replay fields	可审计性
KRI dashboard	risk, fairness, review, privacy, trust	运营治理
Vendor claim matrix	AI claims and substantiation evidence	供应商治理
Pilot release memo	limited go, conditions, stop triggers	成熟上线判断
Decision rationale set	8 system design Q&A	能用证据解释架构取舍

5 分钟系统说明: problem = workforce AI can silently influence employment decisions; method = classify by lifecycle, authority and impact tier; architecture = minimization, authority, review, notice, appeal and evidence; controls = adverse impact, accessibility, purpose separation and KRIs; case = contact center QA or branch scheduling; close = worker-aware, evidence-backed decision support。

Operating control summary:

Concept	Control statement
Workforce AI	AI that affects candidate or employee opportunity, evaluation, monitoring or work experience
Impact tiering	governance depth follows employment decision impact
AI authority	search, summarize, score, rank, recommend, monitor, trigger, decide
Data minimization	employee data must be necessary, purpose-bound and retention-bound
Meaningful review	reviewer has evidence, authority, time, training and override path
Adverse impact testing	stage-level outcome testing with qualified review and remediation
Monitoring boundary	security, compliance, QA, productivity and discipline are separate purposes
Evidence ledger	every high-impact AI-influenced decision can be reconstructed

Operating principle:

Workforce AI governance protects the decision chain: purpose, data, model, score, human judgment, worker communication, challenge, monitoring and evidence.