A Privacy-Preserving Architecture for AI-Driven People Intelligence: Design Science Research on Proactive Human Capital Management Systems

A Privacy-Preserving Architecture for AI-Driven People Intelligence: Design Science Research on Proactive Human Capital Management Systems Sebastian Kirsch Independent Researcher Technical Report November 2025 Abstract This technical report presents a comprehensive architectural framework for implementing artificial intelligence and machine learning systems in organizational people analytics. American organizations lose approximately $1 trillion annually to voluntary employee turnover, yet traditional HR systems operate reactively with backward-looking data. Drawing on design science research methodology, I propose a novel six-layer architecture integrating real-time behavioral signals while implementing strict privacy controls. The framework contributes: (1) production-ready system architecture, (2) taxonomy of 127 behavioral signals, (3) mathematical formulations for predictive models, and (4) ethical governance framework with role-based access control and bias mitigation. This work validates feasibility through architectural analysis, complexity assessment, and agentbased simulation. The framework provides actionable specifications for organizations seeking to implement AI-driven people analytics while maintaining ethical standards and employee privacy. Keywords: people analytics, machine learning systems, privacy-preserving analytics, HR technology, workplace AI, design science research, human capital management Contents 1 Introduction 4 1.1 The$1TrillionProblem .................................... 4 1.2 Research Approach: Design Science Methodology . . . . . . . . . . . . . . . . . . . . . . 4 1.3 Contributions .......................................... 4 1.4 NationalImportance ...................................... 5 1 AI-Driven People Intelligence Architecture 2 Literature Review 5 2.1 Machine Learning Systems in Production . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2.2 PeopleAnalyticsEvolution................................... 5 2.3 SignalDetectionTheory .................................... 5 2.4 AIEthicsandPrivacy...................................... 5 2.5 DesignScienceResearch.................................... 6 3 System Architecture 6 3.1 ConceptualFoundation..................................... 6 3.2 Six-LayerArchitecture ..................................... 6 3.2.1 Layer 1: Data Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 3.2.2 Layer 2: Signal Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 3.2.3 Layer 3: Feature Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 3.2.4 Layer4:Prediction................................... 7 3.2.5 Layer 5: Insight Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 3.2.6 Layer6:AccessControl ................................ 7 4 Behavioral Signal Taxonomy 8 4.1 Communication Signals (n=32) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 4.2 Temporal/Work Pattern Signals (n=18) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 4.3 Content/Sentiment Signals (n=24) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 4.4 ProductivitySignals(n=21)................................... 8 4.5 Collaboration Signals (n=14) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 4.6 Traditional HR Signals (n=12) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 4.7 External/Contextual Signals (n=6) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 5 Predictive Modeling Framework 9 5.1 ProblemFormulation...................................... 9 5.2 ModelArchitectures ...................................... 9 5.2.1 GradientBoosting ................................... 9 5.2.2 LSTMNetworks .................................... 9 5.3 HandlingClassImbalance ................................... 9 5.4 TemporalCross-Validation ................................... 9 5.5 InterpretabilityviaSHAP.................................... 9 6 Ethical Considerations and Governance 10 6.1 PrivacyArchitecture ...................................... 10 6.2 AlgorithmicFairness ...................................... 10 6.3 GovernanceMechanisms.................................... 10 2 AI-Driven People Intelligence Architecture 7 Framework Evaluation 11 7.1 EvaluationStrategy....................................... 11 7.2 AnalyticalEvaluation...................................... 11 7.3 ArchitecturalAssessment.................................... 11 7.4 ComplexityAnalysis ...................................... 11 7.5 Simulation-Based Feasibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 8 Implementation Insights from Industry Consultation 12 8.1 Practical Deployment Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 8.2 Common Technical Stack Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 8.3 Integration Challenges Observed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 8.4 Lessons Learned Across Implementations . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 8.5 PersistentOpenChallenges................................... 14 8.5.1 CausalInferenceGap.................................. 14 8.5.2 Feedback Loop Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 8.5.3 Cross-Organization Generalization . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 9 Conclusion 15 9.1 Summary ............................................ 15 9.2 ImplicationsforPractice .................................... 15 9.3 ImplicationsforResearch.................................... 15 9.4 LimitationsandFutureWork.................................. 15 9.5 NationalImpact......................................... 16 9.6 ClosingRemarks ........................................ 16 3 AI-Driven People Intelligence Architecture 1 Introduction 1.1 The $1 Trillion Problem American organizations lose $1 trillion annually to voluntary employee turnover (Work Institute, 2020). Replacement costs range from 50% to 200% of annual salary (Boushey & Glynn, 2012). Despite HR technology reaching $30 billion in 2023 (Bersin, 2023), employee engagement remains at 32% (Gallup, 2023). The root cause is architectural: traditional HR systems operate in batch mode with backward-looking data. Annual performance reviews capture information 6-12 months after issues emerge. Quarterly surveys identify problems weeks after mental disengagement. Exit interviews document reasons after decisions are irreversible. Modern organizations require continuous real-time intelligence. The technical challenge involves: (1) detecting early warning signals from multi-modal data, (2) predicting outcomes with intervention lead time, (3) protecting employee privacy, (4) scaling to enterprise complexity, and (5) maintaining interpretability for human decision-makers. 1.2 Research Approach: Design Science Methodology This work employs design science research methodology (Hevner et al., 2004), emphasizing creation and evaluation of innovative artifacts. Design science suits novel information systems architectures where contribution lies in the artifact itself rather than empirical human behavior observations. The methodology follows seven guidelines: (1) design as artifact, (2) problem relevance, (3) design evaluation, (4) research contributions, (5) research rigor, (6) design as search, and (7) communication to technical and management audiences. Scope Declaration: This paper presents technical architecture and framework design. It does not report empirical findings from deployed systems using real employee data, as such research requires IRB approval and sensitive information handling. I focus on technical design, implementation specifications, and feasibility demonstration. 1.3 Contributions Contribution 1: Production-Ready Architecture. Six-layer technical architecture for real-world deployment, addressing fault tolerance, real-time processing, consistency, and scalability. Contribution 2: Signal Taxonomy. Structured taxonomy of 127 behavioral signals across seven dimensions with technical specifications for collection and engineering. Contribution 3: Predictive Framework. Mathematical formulations for multi-horizon models with temporal dependencies, class imbalance handling, and interpretability. Contribution 4: Ethical Governance. Detailed mechanisms for privacy preservation, algorithmic fairness, and responsible deployment through role-based access control and bias mitigation. 4 AI-Driven People Intelligence Architecture 1.4 National Importance This research addresses substantial national importance. Voluntary turnover costs exceed $1 trillion annually in the U.S., with 47% average rates in 2023. Beyond direct costs, turnover disrupts knowledge transfer and productivity in healthcare, education, infrastructure, and manufacturing sectors critical to national interests. The framework provides practical implementation guidance that organizations can adapt for production deployment, bridging the gap between academic research and real-world application. 2 Literature Review 2.1 Machine Learning Systems in Production ML deployment presents challenges distinct from academic development. Sculley et al. (2015) observe that actual ML code comprises under 5% of production systems; majority involves data collection, monitoring, and serving infrastructure. For people analytics, additional constraints emerge: sensitive data requiring access controls, interpretable predictions for decision-making, class imbalance (10-15% turnover rates), and temporal dependencies violating i.i.d. assumptions. 2.2 People Analytics Evolution People analytics has evolved from workforce planning (Huselid, 1995) through predictive turnover models (Maertz & Campion, 2004) to communication pattern analysis (Wu et al., 2008) and network structures (Kleinbaum et al., 2013). Recent AI applications target recruitment (Raghavan et al., 2020), performance prediction (Zhao et al., 2021), and engagement analysis (Green et al., 2022). However, most analyze historical data retrospectively; few address real-time systems. 2.3 Signal Detection Theory Signal detection from engineering and psychology (Green & Swets, 1966) applies across domains: financial early warnings (Borio & Drehmann, 2009), healthcare deterioration (Goldstein et al., 2016), ecosystem transitions (Scheffer et al., 2009). For organizations, behavioral indicators precede negative outcomes: decreased communication correlates with turnover (Feeley et al., 2010), work pattern changes indicate disengagement (Bakker & Bal, 2010), and linguistic markers predict decline (Holtom et al., 2008). 2.4 AI Ethics and Privacy AI employee monitoring raises ethical tensions: organizational intelligence needs versus privacy rights (Ball, 2010; Moore, 2018), algorithmic bias from historical discrimination (Dastin, 2018), and transparency versus proprietary methods. 5 AI-Driven People Intelligence Architecture Recent proposals emphasize oversight, audit trails, and human authority (Kellogg et al., 2020), though practical guidance remains limited. 2.5 Design Science Research Design science provides established methodology for artifact development (Hevner et al., 2004). Venable et al. (2016) distinguish artificial evaluation (analysis, simulation, experiments) from naturalistic evaluation (field studies). For technical architectures, artificial evaluation appropriately tests system properties when users are unavailable. Methods include complexity analysis, simulation, benchmarking, and expert evaluation. 3 System Architecture 3.1 Conceptual Foundation The Proactive People Intelligence (PPI) model rests on three pillars: Signal-Based Detection. Drawing from signal detection theory, employee outcomes manifest through detectable precursors. Continuous monitoring identifies at-risk situations enabling intervention. Multi-Source Fusion. No single source suffices. The framework integrates structured HR data, semistructured calendar/project data, and unstructured communications into unified analytical substrate. Hierarchical Privacy. Role-based access control ensures employees access own data, managers see team aggregates, executives receive organization-wide insights without individual identification. 3.2 Six-Layer Architecture 3.2.1 Layer 1: Data Integration Integration from HRIS (demographics, performance), communication platforms (email, Slack metadata), project management (tasks, collaboration), code repositories (commits, reviews), and calendars (meetings, schedules). API-based integration with OAuth 2.0. Collection focuses on metadata and patterns, not content. Example: from Slack, message frequency and response times—not message text. This balances utility with privacy. 3.2.2 Layer 2: Signal Detection Three detector categories process integrated streams: Frequency Analyzers: Monitor activity levels using statistical process control. For employee iand metric m: zi,m,t =xi,m,t −µi,m σi,m Alerts trigger when |zi,m,t|>2.5sustained. 6 AI-Driven People Intelligence Architecture Network Analyzers: Apply social network analysis quantifying position through degree, betweenness, and eigenvector centrality measures. Sentiment Analyzers: NLP assesses emotional tone using transformer-based models (BERT, RoBERTa) with transfer learning. 3.2.3 Layer 3: Feature Engineering Raw signals transform into predictive features: Temporal: Moving averages (7, 30, 90 days), rate of change, volatility. Comparative: Z-scores relative to peer groups, percentile rankings, deviation from baseline. Composite: Weighted combinations validated against outcomes. Feature selection via recursive elimination with cross-validation (RFECV) identifies predictive subsets avoiding overfitting. 3.2.4 Layer 4: Prediction Multiple models generate outcome predictions: Turnover Risk: Gradient boosting (XGBoost) predicts separation probability at 30, 60, 90 days. Engagement: Multi-class classifier categorizes: Highly Engaged, Moderate, At Risk, Disengaged. Performance Trajectory: LSTM neural networks forecast performance direction for next quarter. Team Health: Aggregates individual signals assessing team-level dynamics. Training uses 5-fold cross-validation with temporal stratification ensuring training precedes validation chronologically. 3.2.5 Layer 5: Insight Generation Raw outputs undergo interpretation: Risk Scoring: Convert probabilities to discrete categories (Low, Medium, High, Critical) using calibrated thresholds. Explainability: SHAP (Lundberg & Lee, 2017) identifies contributing factors enabling targeted interventions. Natural Language: Automated systems convert statistics into accessible summaries. 3.2.6 Layer 6: Access Control Five-tier permissions ensure appropriate access: Individual Employee: Own data only, full detail. Cannot view peers or predictions. Manager: Direct reports aggregated, team analytics. Individual predictions anonymized. HR Partner: Department aggregates, anonymized patterns. Individual access requires escalation. Executive: Organization-wide patterns, benchmarks. No individual identification; minimum n≥10. Administrator: Full system access for maintenance. All access logged; no analytical use. This implements ”minimum necessary” access (Cavoukian, 2009). 7 AI-Driven People Intelligence Architecture 4 Behavioral Signal Taxonomy Through literature review and implementation, I identified 127 signals across seven categories: 4.1 Communication Signals (n=32) Volume metrics (messages sent/received, response ratio), temporal patterns (latency, consistency, after-hours activity), network position (centrality measures, clustering), interaction quality (reciprocity, thread depth, breadth). 4.2 Temporal/Work Pattern Signals (n=18) Start/end times (mean, variance), work hours (daily, weekly), consistency, weekend/late-night activity, meeting attendance/decline rates, calendar fragmentation. 4.3 Content/Sentiment Signals (n=24) Sentiment scores (positive, negative, neutral, volatility), linguistic features (length, questions, future-tense language, pronouns), topic modeling (work vs. social content), emotion detection (frustration, enthusiasm, stress). 4.4 Productivity Signals (n=21) Task completion/abandonment rates, duration vs. estimates, variety, project switching, code commits (technical roles), pull requests, document activity, milestone achievement. 4.5 Collaboration Signals (n=14) Cross-functional interaction, within-team density, external collaboration ratio, mentorship activity, meeting participation quality, pair programming, shared documents, knowledge contributions. 4.6 Traditional HR Signals (n=12) Tenure, role level, promotion history, performance scores and trends, compensation percentiles and growth, internal mobility, manager tenure, organizational changes experienced. 4.7 External/Contextual Signals (n=6) Industry hiring trends, local unemployment, competitor activity, organizational events (layoffs, acquisitions), team expansion/contraction, geographic factors. 8 AI-Driven People Intelligence Architecture 5 Predictive Modeling Framework 5.1 Problem Formulation Turnover as Binary Classification: Given employee iat time twith features xi,t, predict separation probability within horizon h∈ {30,60,90}days. Engagement as Multi-Class: Predict level yi∈ {1,2,3,4}(Highly Engaged, Moderate, At Risk, Disengaged). Performance as Regression: Predict performance change ∆pi,t+hover horizon h. 5.2 Model Architectures 5.2.1 Gradient Boosting XGBoost constructs ensemble of decision trees iteratively, minimizing regularized objective balancing prediction error with model complexity. Hyperparameter optimization via Bayesian search over learning rate, max depth, subsample ratios, and regularization. 5.2.2 LSTM Networks Long Short-Term Memory models temporal dependencies through gating mechanisms controlling information flow. Architecture: Two-layer LSTM, hidden size 128, dropout 0.3, trained with Adam optimizer and MSE loss. 5.3 Handling Class Imbalance Turnover affects 10-15% annually, creating severe imbalance. Strategies: class weighting inversely proportional to frequency, SMOTE synthetic minority oversampling, threshold optimization based on business costs. 5.4 Temporal Cross-Validation Forward chaining prevents information leakage: train on months 1-12, validate on 13, test on 14; then slide window forward. Ensures test data always future relative to training. 5.5 Interpretability via SHAP SHAP provides consistent feature attribution based on cooperative game theory, satisfying local accuracy, missingness, and consistency properties. TreeSHAP algorithm efficiently computes exact values for tree-based models in polynomial time. 9 AI-Driven People Intelligence Architecture Future Work: (1) Field deployment and validation with IRB approval, (2) RCTs comparing intervention strategies, (3) Cross-organizational benchmarking with privacy preservation, (4) Expansion to other outcomes beyond turnover, (5) Open-source implementation components. 9.5 National Impact This research addresses substantial national importance. The $1 trillion annual turnover cost represents massive economic inefficiency. Reducing turnover 10-20% through early intervention could save $100-200 billion annually, improving U.S. competitiveness and productivity. Benefits include workforce development through targeted career interventions, organizational knowledge preservation in specialized fields, and ethical AI deployment models balancing intelligence needs with privacy rights. The framework provides organizations with actionable specifications for implementing these systems responsibly and effectively, advancing both technical capabilities and ethical standards in AI deployment for human capital management. 9.6 Closing Remarks Transition from reactive batch-mode HR to proactive real-time people intelligence represents fundamental architectural shift. This paper provides comprehensive blueprint for organizations undertaking transformation, addressing technical, ethical, and practical considerations. The framework balances competing demands: accuracy vs. privacy, comprehensive collection vs. minimization, automated insights vs. human judgment, organizational intelligence vs. individual autonomy. Through careful design, these tensions can be managed. Ultimately, the goal is not replacing human decision-making with algorithms, but augmenting human judgment with timely, accurate, interpretable insights. Successful implementations combine advanced AI with thoughtful change management, ethical governance, and genuine employee wellbeing commitment. This architecture provides foundation. Continued research, responsible deployment, and ongoing refinement will determine whether AI-driven people intelligence fulfills its promise of creating more effective, more humane organizations. References [1] Bakker, A. B., & Bal, P. M. (2010). Weekly work engagement and performance. Journal of Occupational and Organizational Psychology, 83(1), 189–206. [2] Ball, K. (2010). Workplace surveillance: An overview. Labor History, 51(1), 87–106. [3] Bersin, J. (2023). HR technology market: 2023 analysis. The Josh Bersin Company. [4] Borio, C., & Drehmann, M. (2009). Assessing banking crises risk. BIS Quarterly Review, March. 16 AI-Driven People Intelligence Architecture [5] Boushey, H., & Glynn, S. J. (2012). Business costs of replacing employees. Center for American Progress. [6] Cavoukian, A. (2009). Privacy by design: 7 foundational principles. Information and Privacy Commissioner of Ontario. [7] Dastin, J. (2018). Amazon scraps AI recruiting tool showing bias. Reuters, October 10. [8] Feeley, T. H., Hwang, J., & Barnett, G. A. (2010). Predicting turnover from friendship networks. Journal of Applied Communication Research, 38(2), 115–134. [9] Gallup. (2023). State of the global workplace: 2023 report. Gallup, Inc. [10] Goldstein, B. A., et al. (2016). Opportunities in risk prediction models. JAMA Cardiology, 1(9), 1064– 1070. [11] Green, D. P., Gino, F., & Staats, B. R. (2022). Learning while doing. Organization Science, 33(2), 414–434. [12] Green, D. M., & Swets, J. A. (1966). Signal Detection Theory. Wiley. [13] Hevner, A. R., et al. (2004). Design science in IS research. MIS Quarterly, 28(1), 75–105. [14] Holtom, B. C., et al. (2008). Turnover and retention research. Academy of Management Annals, 2(1), 231–274. [15] Huselid, M. A. (1995). Impact of HRM practices. Academy of Management Journal, 38(3), 635–672. [16] Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work. Academy of Management Annals, 14(1), 366–410. [17] Kleinbaum, A. M., Stuart, T. E., & Tushman, M. L. (2013). Discretion within constraint. Organization Science, 24(5), 1316–1336. [18] Lundberg, S. M., & Lee, S. I. (2017). Unified approach to interpreting model predictions. NIPS, 30. [19] Maertz, C. P., & Campion, M. A. (2004). Profiles in quitting. Academy of Management Journal, 47(4), 566–582. [20] Moore, P. V. (2018). The Quantified Self in Precarity. Routledge. [21] Raghavan, M., et al. (2020). Mitigating bias in algorithmic hiring. FAT* 2020, 469–481. [22] Scheffer, M., et al. (2009). Early-warning signals for critical transitions. Nature, 461, 53–59. [23] Sculley, D., et al. (2015). Hidden technical debt in ML systems. NIPS, 28. [24] Venable, J., Pries-Heje, J., & Baskerville, R. (2016). FEDS framework for evaluation. European Journal of Information Systems, 25, 77–89. 17 AI-Driven People Intelligence Architecture [25] Work Institute. (2020). 2020 Retention report. Franklin, TN. [26] Wu, L., et al. (2008). Mining face-to-face interaction networks. ICIS 2008. [27] Zhao, Y., et al. (2021). Employee turnover prediction with ML. SAI Conference, 737–758. 18