Global People Analytics, HR Technology and AI

Introduction

International organizations increasingly manage work through interconnected human resource information systems (HRIS), global payroll platforms, talent marketplaces, collaboration tools, learning systems and artificial-intelligence applications. These technologies create the possibility of faster, more evidence-based workforce decisions. They also create significant risks. Employee data may cross jurisdictions, payroll errors may affect legal entitlements, predictive models may reproduce historical discrimination, and generative AI may expose confidential information or produce misleading recommendations.

Global people analytics is therefore not simply the production of dashboards. It is the disciplined use of workforce data, statistical reasoning, technology and organizational judgment to improve decisions while respecting human rights, privacy, fairness and security. In international business, the challenge is intensified by differences in labor law, data-protection regimes, culture, language, employment practices, currencies and data quality.

A mature approach combines four capabilities:

1. Reliable infrastructure: integrated HRIS, payroll, identity, recruiting, performance and learning data.
2. Analytical competence: descriptive, diagnostic, predictive and prescriptive analysis.
3. Responsible governance: validity, fairness, explainability, privacy, accountability and cybersecurity.
4. Managerial translation: converting evidence into decisions, interventions and measurable outcomes.

The central principle of this chapter is that technology should augment responsible managerial judgment, not replace accountability.

Learning Objectives

By the end of this chapter, students should be able to:

• distinguish HRIS, global payroll, digital HR platforms and people analytics;
• design meaningful workforce metrics and interpret dashboards critically;
• calculate and explain common measures such as turnover, retention, absenteeism and selection validity;
• distinguish descriptive, diagnostic, predictive and causal analysis;
• evaluate experiments and predictive models used in HR;
• assess algorithmic recruitment and generative-AI applications for validity, bias, privacy, transparency and security;
• recommend governance controls for multinational organizations; and
• translate analytics into ethical and strategically relevant workforce decisions.

1. The Global HR Technology Ecosystem

1.1 HRIS and the system of record

An HRIS is a structured system used to store and manage core workforce information. Typical data include worker identity, employment status, job, location, manager, compensation, absence, training and organizational relationships. In a multinational enterprise, the HRIS must accommodate multiple legal entities, currencies, calendars, languages, tax regimes and definitions of employment.

The HRIS should normally be treated as a system of record, not automatically as a system of truth. A record can be complete yet incorrectly entered, or technically accurate but conceptually inconsistent across countries. For example, “headcount” might mean all workers in one country but only permanent employees in another. Analytics begins with definitions, data ownership and quality controls.

1.2 Global payroll

Global payroll systems calculate and administer pay, deductions, taxes, social contributions, benefits and statutory reporting. Payroll is operationally sensitive because errors can violate employment law, damage trust and create financial penalties. A global payroll architecture may include a central platform connected to local providers, or separate country systems integrated through a common data layer.

Important controls include:

• approval segregation between data entry, calculation and payment;
• reconciliation between payroll totals and the HRIS;
• validation of currency, exchange-rate and tax rules;
• change logs for bank accounts and compensation;
• restricted access to highly sensitive data; and
• contingency procedures for system outages.

Payroll data should not be repurposed casually for performance or behavioral surveillance. The fact that an organization can collect a variable does not establish that it should use it.

1.3 Digital HR platforms

Digital platforms support recruiting, onboarding, learning, performance management, scheduling, employee listening, collaboration and internal mobility. They can improve access and scale, but integration creates both opportunity and risk. Combining recruitment, productivity, communications and health-related data can generate intrusive employee profiles and may violate purpose limitations.

A sound architecture separates:

• identity and access management, which controls who can use systems;
• transactional applications, which execute HR processes;
• data warehouses or lakes, which support controlled analysis; and
• analytics and reporting layers, which present authorized information.

Data integration should be governed by documented purposes, retention rules and access roles. Data minimization is particularly important where information is transferred across borders.

2. From Metrics to People Analytics

2.1 Four levels of analysis

People analytics can be classified into four progressively demanding levels.

The levels should not be confused. A model that predicts resignation accurately may use variables that managers cannot ethically or practically change. Conversely, a variable may be causally important but weak for prediction.

2.2 Metric design

A useful metric has a clear numerator, denominator, population, time period, grain and owner. It should also specify exclusions and data sources. Common measures include:

Turnover rate

The average headcount is often calculated as opening headcount plus closing headcount divided by two:

This approximation is suitable when workforce changes are reasonably distributed. For highly seasonal or volatile workforces, the average of monthly or daily headcounts is preferable. The rate must state whether it includes voluntary, involuntary, internal transfers, temporary workers or only employees.

Retention rate for a defined cohort can be expressed as:

Retention and turnover are not always exact complements because their populations and treatment of internal movement may differ.

Absence rate is commonly calculated as:

The definition of absence must identify whether it includes sickness, authorized leave, unauthorized absence, occupational injury or statutory leave. Cross-country comparisons require caution because working-time regulations and recording practices differ.

Selection ratio is:

It describes hiring intensity but does not establish selection quality.

Worked Example: Interpreting turnover

A regional business begins a quarter with 1,200 employees and ends with 1,260. During the quarter, ninety employees leave. Average headcount is (1,200 + 1,260) ÷ 2 = 1,230. Quarterly turnover is 90 ÷ 1,230 × 100 ≈ 7.32%.

This result does not by itself indicate a problem. Analysts should segment it by country, job family, tenure, critical skill, voluntary status and reason for leaving. A 7.32% rate concentrated among newly hired employees may require a different intervention from the same rate concentrated among scarce technical specialists.

2.3 Dashboard interpretation

A dashboard is a decision interface, not evidence of causation. Effective dashboards provide:

• a clearly named population and period;
• definitions and data refresh dates;
• trends rather than isolated values;
• relevant comparison groups or benchmarks;
• confidence intervals or uncertainty where appropriate;
• drill-down controls that respect access rights; and
• a stated decision or action associated with each metric.

Dashboards can mislead through denominator changes, survivorship bias, aggregation and visual emphasis. A global average may conceal substantial country variation. A fall in absence may reflect improved attendance, under-reporting or a change in coding. Managers should ask: Compared with what? Measured how? For whom? With what uncertainty?

3. Analytical Methods for Workforce Decisions

3.1 Segmentation and association

Segmentation compares outcomes across groups such as country, function, gender, tenure, contract type or manager. It is useful for identifying patterns and prioritizing investigation. However, group differences can arise from confounding variables. For example, turnover differences between countries may reflect job mix, labor-market conditions, economic restructuring or differences in recording practices.

Correlation or association does not establish causation. A useful conceptual distinction is:

• Predictor: a variable associated with an outcome;
• Treatment or intervention: an action assigned or experienced by units;
• Outcome: the measured result; and
• Confounder: a variable related to both treatment and outcome that can distort the estimated effect.

3.2 Predictive modeling

A predictive model estimates the probability or expected value of an outcome. For a binary outcome such as resignation, logistic regression is commonly used:

Here, Y = 1 indicates the event of interest, such as resignation within six months; X₁ … Xₖ are predictors; β₀ is the intercept; β₁ … βₖ are estimated coefficients; and e is the base of the natural logarithm. The model assumes a specified relationship between predictors and the log-odds of the outcome, independent observations or an appropriately modeled dependence structure, and adequate data quality. It produces an estimate, not a certainty and not proof that a predictor causes resignation.

Model evaluation should use data not used to fit the model. Relevant measures include:

• Discrimination, such as area under the receiver operating characteristic curve;
• Calibration, the agreement between predicted probabilities and observed frequencies;
• Precision and recall, especially where positive cases are rare;
• False-positive and false-negative rates across relevant groups; and
• Operational utility, including whether managers can act appropriately on the output.

Accuracy alone is insufficient. A model can be accurate overall while performing poorly for a small group. Fairness analysis should compare error rates, selection rates and calibration where appropriate, recognizing that some fairness criteria cannot all be satisfied simultaneously when base rates differ.

3.3 Causal analysis and experiments

When an organization wants to know whether an intervention caused an outcome, it needs a credible counterfactual: what would have happened to the same or comparable workers without the intervention? Randomized experiments are often the strongest practical design. For example, eligible teams may be randomly assigned to receive a new manager-coaching program or business-as-usual support.

A simple average treatment effect can be represented as:

Y(1) is the potential outcome under treatment, Y(0) is the potential outcome without treatment, and E denotes the expected value. The fundamental problem is that for any one unit, only one potential outcome is observed. Randomization makes the groups comparable in expectation, subject to adequate implementation, compliance and sample size.

If randomization is impossible, analysts may use matched comparisons, interrupted time series, difference-in-differences or regression adjustment. These designs require stronger assumptions and careful sensitivity analysis. A before-and-after change alone is weak evidence because other events may have occurred simultaneously.

4. Algorithmic Recruitment and Talent Decisions

Recruiting technologies may source candidates, parse résumés, rank applicants, schedule interviews, recommend assessments or generate communications. Their value depends on job relevance and measurement validity.

4.1 Validity

A selection procedure is valid when evidence supports the interpretation and use of its scores for a specified purpose. Relevant forms include:

• Content validity: the assessment reflects important job tasks;
• Construct validity: it measures the intended attribute rather than irrelevant characteristics; and
• Criterion-related validity: scores relate to a relevant job outcome, such as training performance or validated job performance.

Historical hiring decisions are not automatically a valid training target. If past decisions reflected unequal opportunity, biased interviews or restricted networks, a model trained on them may learn and scale those patterns. The target variable itself must be examined.

4.2 Bias and fairness

Bias may enter through data collection, labeling, feature selection, model design, threshold setting or human use. Proxy variables are especially important. Postal code, employment gaps, language patterns or school attended may correlate with protected characteristics or socioeconomic status.

A practical audit should compare:

1. representation of groups in the applicant and selected populations;
2. selection rates at each stage of the funnel;
3. false-positive and false-negative rates;
4. interview and assessment completion rates;
5. outcomes after hiring, not merely selection; and
6. performance across countries, languages, disability status and contract types where lawful and feasible.

The “four-fifths” or 80% rule is sometimes used as a screening flag: a group’s selection rate divided by the highest group selection rate below 0.80 may indicate potential adverse impact. It is not a universal legal test, does not prove discrimination, and should not replace jurisdiction-specific legal analysis or statistical assessment.

4.3 Human oversight

Human involvement is meaningful only when decision makers have sufficient information, authority and time to question the system. A person who merely approves an opaque recommendation is not effective oversight. Organizations should define when automated recommendations may be used, when human review is mandatory, how candidates can request correction or accommodation, and how decisions are documented.

5. Generative AI in Global HR

Generative AI can draft job descriptions, summarize interviews, translate communications, create learning content, answer employee questions and assist HR professionals with policy navigation. It can also hallucinate facts, reproduce stereotypes, infer sensitive attributes and reveal confidential information through inappropriate prompts or vendor retention practices.

Responsible use requires:

• approved tools and prohibited-use rules;
• data classification before prompting;
• no submission of unnecessary personal, health, disciplinary or payroll information;
• human verification of factual, legal and employment-related outputs;
• disclosure when employees or candidates interact with an AI system where appropriate;
• accessibility and multilingual testing;
• records of prompts, outputs and material decisions where legally permissible; and
• vendor controls concerning model training, retention, location, subcontractors and breach notification.

Generative AI should generally assist with low-risk drafting and retrieval before being used in high-impact decisions. It should not independently determine hiring, dismissal, promotion, compensation or legal eligibility. In high-impact applications, the organization must be able to explain the decision process, investigate errors and provide a meaningful route for review.

6. Privacy, Cybersecurity and International Governance

6.1 Privacy principles

Global HR data governance should reflect principles widely found in modern data-protection regimes:

• lawfulness and fairness: processing requires a legitimate legal basis and fair treatment;
• purpose limitation: data collected for payroll should not automatically be used for behavioral surveillance;
• data minimization: collect only what is necessary;
• accuracy: provide correction mechanisms;
• storage limitation: delete or anonymize information when no longer needed;
• security: protect confidentiality, integrity and availability; and
• accountability: document decisions and demonstrate compliance.

Cross-border transfers require attention to applicable transfer mechanisms, local restrictions, employee rights and vendor arrangements. Consent is not always a suitable basis in employment because of unequal bargaining power. Legal counsel and local specialists should be involved in high-risk processing.

6.2 Cybersecurity

HR systems are attractive targets because they contain identity, bank, compensation, health and immigration information. Controls should include multifactor authentication, least-privilege access, encryption in transit and at rest, network segmentation, secure application programming interfaces, vulnerability management, logging, backup testing and incident-response exercises.

A useful security model considers confidentiality, integrity and availability. A system that is private but frequently unavailable can disrupt payroll; a system that is available but permits unauthorized salary changes threatens integrity. Third-party providers must be assessed through contracts, audits, penetration-testing evidence and incident-notification obligations.

6.3 Governance structure

A global people-analytics governance committee may include HR, legal, privacy, information security, data science, employee relations and regional representatives. Its responsibilities should include:

• maintaining a register of analytical and AI use cases;
• classifying use cases by risk;
• approving data sources and purposes;
• requiring impact assessments for high-risk systems;
• reviewing validation and fairness evidence;
• monitoring drift and post-deployment outcomes;
• managing complaints, appeals and incidents; and
• retiring systems that no longer meet requirements.

A model card or use-case record should state the purpose, population, data, assumptions, performance, limitations, prohibited uses, responsible owner and review date.

7. Implementing a Responsible Analytics Capability

Implementation is best treated as an operating model rather than a software purchase.

Step 1: Define the decision

Begin with a decision and its owner: for example, reducing regrettable turnover among scarce engineers. Avoid starting with “What data do we have?” Clarify what action could follow from the analysis and what would make the action successful.

Step 2: Establish definitions and data lineage

Create a metric dictionary, identify source systems, document transformations and assign data owners. Reconcile headcount, worker identifiers, country codes, job families and dates. Test missingness, duplicates, inconsistent categories and changes in coding.

Step 3: Assess legal and ethical risk

Determine whether the use involves sensitive data, automated decisions, employee monitoring, vulnerable populations or cross-border transfer. Conduct a privacy and algorithmic-impact assessment before deployment.

Step 4: Build and validate

Use appropriate analytical methods, hold-out data or cross-validation, subgroup testing and sensitivity analysis. Compare the model with a simple baseline. A complex model is not automatically better if it is less stable or less explainable.

Step 5: Pilot and evaluate

Pilot with a limited population, define success measures in advance and monitor unintended consequences. Where feasible, use randomization or a credible quasi-experimental design. Collect feedback from HR professionals, managers and affected workers.

Step 6: Deploy with controls

Provide role-based access, explanations, escalation routes, human review and audit logs. Separate analytical recommendation from final employment decision. Communicate the purpose and limitations in accessible language.

Step 7: Monitor and improve

Monitor data drift, model drift, subgroup performance, complaints, security events and outcome quality. Revalidate after organizational restructures, labor-market changes, new countries, changes in law or major changes in data collection.

Illustrative Case: Meridian Foods’ Global Attrition Model

Background

Meridian Foods, a fictional multinational consumer-goods company, operates in 18 countries. Its HR leadership wants to reduce voluntary resignation among experienced supply-chain specialists. The company proposes an AI model that will assign each employee a resignation-risk score and display the score to line managers.

Problem

The initial model uses performance ratings, promotion history, absence, salary position, manager changes, collaboration metadata and commute distance. Training labels are based on past resignations. The data team reports high overall accuracy, but the works council in one country raises concerns about surveillance and protected personal information.

Analysis

The project team identifies several risks. First, resignation is not a uniform outcome: some departures are regrettable, some reflect promotions elsewhere, and some result from fixed-term contracts. Second, collaboration metadata are collected differently across countries and may disadvantage remote workers. Third, historic performance ratings contain manager-level inconsistencies. Fourth, a risk score could become self-fulfilling if managers reduce development opportunities for employees labeled high risk.

The team therefore changes the objective from “predict resignation” to “identify employees who may benefit from a voluntary career and workload conversation.” It removes unnecessary collaboration data, standardizes outcome definitions, tests performance across countries and job families, and compares the model with a transparent baseline. Employees are not shown to managers as a ranking. Instead, authorized HR partners receive a risk band together with uncertainty, contributing factors that are appropriate to disclose, and a requirement to offer the same development resources to comparable employees.

Approach and outcome

Meridian pilots the system in three countries. A randomized evaluation assigns some eligible teams to receive structured career-conversation training and others to existing practice. The evaluation measures retention, internal mobility, employee perceptions and adverse effects. After six months, the company finds modest improvement in internal mobility but no reliable reduction in resignation. It therefore continues the career-support intervention but does not expand the predictive model globally. The model is retained only as an internal analytical tool, subject to quarterly review.

Learning points

The case illustrates that a technically accurate model may not solve the managerial problem. Defining the decision, selecting a legitimate intervention, protecting autonomy, testing causal effects and monitoring unintended consequences are more important than maximizing a single predictive score. It also demonstrates that responsible organizations may narrow or discontinue an AI application when evidence does not justify its risks.

Challenges and Limitations

People analytics faces persistent limitations. Data are often incomplete, historically biased and generated for administration rather than research. International comparisons suffer from differences in law, culture, language, job architecture and reporting. Small groups create statistical uncertainty and heightened confidentiality risks. Models may drift when labor markets or organizational strategies change. Employees may alter behavior when they know they are monitored, and managers may game metrics.

Ethical limitations are equally important. A technically possible inference may be inappropriate. Predicting vulnerability, health status, union activity or personal circumstances can undermine dignity and trust. Transparency must be meaningful without exposing security-sensitive details or enabling gaming. Fairness metrics require normative choices, and no single metric can settle all ethical questions.

The most important safeguard is institutional humility: analytics should inform inquiry and support better decisions, while preserving the possibility that the data, model or objective is wrong.

Contemporary Developments and Best Practices

Current developments include skills ontologies for internal mobility, privacy-enhancing computation, federated analytics, synthetic data for testing, continuous listening and AI copilots for HR service delivery. These developments are promising but do not remove the need for governance. Synthetic data may still reproduce the structure of biased source data; federated systems may reduce data movement but do not guarantee lawful purpose; and AI copilots may amplify errors at scale.

Best practice is to:

• begin with a clearly defined business and human decision;
• use the least intrusive data that can answer the question;
• standardize definitions before comparing countries;
• distinguish prediction from causation;
• validate models against job-relevant outcomes;
• test subgroup performance and accessibility;
• document assumptions, uncertainty and prohibited uses;
• maintain meaningful human review and appeal;
• secure systems throughout their lifecycle; and
• involve employees and regional stakeholders in design and evaluation.

Chapter Summary

Global people analytics connects HRIS, payroll, digital platforms and analytical methods to workforce decision-making. Reliable analysis depends on clear definitions, data lineage, valid denominators and attention to international comparability. Descriptive dashboards show what happened; diagnostic analysis explores associations; predictive models estimate likely outcomes; and causal designs investigate whether interventions produce effects.

Algorithmic recruitment and generative AI require more than technical accuracy. Organizations must assess validity, bias, explainability, privacy, security, accessibility and human accountability. Predictive models should not be treated as causal explanations, and AI recommendations should not replace responsible employment decisions. A mature governance model includes risk classification, impact assessment, access controls, monitoring, documentation, appeal and system retirement. The strategic value of HR technology is realized when trustworthy evidence supports humane, lawful and effective action.

Key Terms

Algorithmic accountability: Responsibility for the design, use, effects and review of an algorithmic system.

Calibration: The extent to which predicted probabilities correspond to observed outcome frequencies.

Causal inference: Methods used to estimate the effect of an intervention relative to a counterfactual.

Data minimization: Collecting and using no more personal data than necessary for a defined purpose.

Data lineage: Documentation of where data originate and how they are transformed.

Descriptive analytics: Analysis that summarizes what has happened.

Digital HR platform: Technology supporting activities such as recruiting, learning, performance or employee service delivery.

Global payroll: Systems and processes administering pay, deductions, taxes and statutory contributions across jurisdictions.

HRIS: Human resource information system used to store and manage workforce records.

Human oversight: Meaningful human ability and responsibility to review, challenge and act on automated outputs.

Model drift: Decline or change in model performance as data, behavior or conditions change.

People analytics: Disciplined use of workforce data and analytical methods to improve people-related decisions.

Predictive validity: Evidence that a measure relates to a relevant future job outcome.

Purpose limitation: Using data only for specified and legitimate purposes.

Selection ratio: The number hired divided by the number of applicants.

Turnover rate: Separations during a period divided by average headcount, expressed as a percentage.

Validity: Evidence supporting the interpretation and use of a measure for a specified purpose.

Generative AI: AI that produces text, images, code or other content in response to prompts.

Review Questions

5 Short Answer Questions

1. What is the difference between an HRIS and a people-analytics platform?
2. Why can a global average turnover rate be misleading?
3. Distinguish predictive analysis from causal analysis.
4. Identify three risks associated with generative AI in HR.
5. What is meant by meaningful human oversight?

5 Long Answer Questions

1. Explain how an international organization should design governance for integrated HRIS, payroll and analytics systems.
2. Discuss the importance of validity and fairness in algorithmic recruitment.
3. Explain the assumptions and limitations of logistic regression when used to estimate employee attrition risk.
4. Compare descriptive, diagnostic, predictive and prescriptive analytics, using HR examples.
5. Evaluate the privacy and cybersecurity challenges of cross-border workforce data.

3 Analytical / Application Questions

1. A company reports falling absence after introducing an employee-monitoring system. Design an analysis to determine whether the fall represents improved attendance, changed reporting behavior or a data-quality problem.
2. An AI screening tool selects 36% of applicants from Group A and 27% from Group B. Calculate the selection-rate ratio for Group B relative to Group A and explain why the result is only a screening indicator, not a legal conclusion.
3. Design a pilot experiment to evaluate whether structured career conversations reduce regrettable turnover among international technical employees. Specify the treatment, comparison group, outcome, ethical safeguards and likely threats to validity.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes predictive analytics from causal analysis in HR?

A. Predictive analytics proves that a variable causes an outcome.

B. Predictive analytics estimates likely outcomes, while causal analysis investigates effects of interventions.

C. Predictive analytics is always descriptive, while causal analysis is always prescriptive.

D. Predictive analytics can be used only for payroll data.

Correct Answer: B. Predictive analytics estimates likely outcomes, while causal analysis investigates effects of interventions.

Question 2

An organization begins a year with 1,200 employees and ends with 1,260. It records 90 separations. Using average headcount, what is the quarterly turnover rate?

A. 6.67%

B. 7.32%

C. 7.50%

D. 8.00%

Correct Answer: B. 7.32%

Question 3

Which feature is most important for a meaningful HR dashboard?

A. The largest possible number of visualizations

B. A clear population, period, denominator and metric definition

C. Only global averages without segmentation

D. Removal of uncertainty information

Correct Answer: B. A clear population, period, denominator and metric definition

Question 4

Why may training an applicant-screening model on historical hiring decisions be problematic?

A. Historical data are always too small to analyze.

B. The model cannot process structured data.

C. Past decisions may contain and reproduce discrimination or restricted opportunity.

D. Models cannot be validated against job outcomes.

Correct Answer: C. Past decisions may contain and reproduce discrimination or restricted opportunity.

Question 5

Which control most directly reduces unauthorized changes to payroll bank-account information?

A. A colorful dashboard

B. Segregation of duties, multifactor authentication and change logging

C. Increasing the number of job advertisements

D. Using only annual headcount reports

Correct Answer: B. Segregation of duties, multifactor authentication and change logging

Question 6

In a logistic attrition model, what does the predicted probability represent?

A. A guaranteed resignation decision

B. The estimated probability of the specified outcome given the predictors and model

C. The causal effect of every predictor

D. The employee’s legal entitlement to leave

Correct Answer: B. The estimated probability of the specified outcome given the predictors and model

Question 7

What is the strongest general design for estimating whether an HR intervention caused an outcome?

A. A single post-intervention opinion survey

B. A comparison of two unrelated countries

C. Random assignment to treatment and comparison conditions, when feasible

D. A dashboard showing before-and-after values only

Correct Answer: C. Random assignment to treatment and comparison conditions, when feasible

Question 8

Which principle requires that data collected for payroll not automatically be reused for behavioral surveillance?

A. Purpose limitation

B. Model drift

C. Selection ratio

D. Calibration

Correct Answer: A. Purpose limitation

Question 9

A model has high overall accuracy but much higher false-negative rates for one employee group. What should the organization do?

A. Deploy it immediately because overall accuracy is sufficient.

B. Ignore group-level results if the model is complex.

C. Investigate subgroup performance and reassess the model’s fairness and fitness for use.

D. Remove all human review.

Correct Answer: C. Investigate subgroup performance and reassess the model’s fairness and fitness for use.

Question 10

Which is an appropriate low-risk use of generative AI in HR, subject to verification?

A. Automatically dismissing employees

B. Independently deciding compensation increases

C. Drafting a general learning announcement without confidential employee data

D. Inferring an applicant’s health condition

Correct Answer: C. Drafting a general learning announcement without confidential employee data