Performance Management

OKR alignment at scale is a graph problem: 1) Data Model: OKR Hierarchy: Company → Business Unit → Department → Team → Individual. Objective = {id, title, description, ownerId, parentObjectiveId, level, quarter, status}. KeyResult = {id, objectiveId, title, targetValue, currentValue, unit, score}. Alignment = {childObjId, parentObjId, contributionWeight}. An individual KR can align to multiple parent objectives (many-to-many). 2) Alignment Visualization: Tree view: company objective at top, drill down to contributing team and individual OKRs. DAG (Directed Acyclic Graph): because one objective can contribute to multiple parents. Color-coded: green (on track), yellow (at risk), red (behind). Aggregate progress: parent objective's progress = weighted average of child KR scores. Example: Company O: 'Become market leader in India' → BU O: 'Achieve ₹100Cr revenue' → Team O: 'Close 50 enterprise deals' → Individual KR: 'Close 10 deals worth ₹2Cr each'. 3) Check-in Workflow: Weekly: individual updates KR progress (numeric value or percentage). System automatically rolls up to team and department levels. Manager reviews team's aggregate progress. Monthly: team review meeting with OKR dashboard. Quarterly: company-wide OKR review, scoring, new cycle. Slack/Teams bot: weekly reminder on Friday — 'How's KR progress? Update in 2 clicks'. 4) Scale Challenges: 10,000 employees × 3-5 OKRs each = 30-50K objectives. Alignment tree: 5-6 levels deep with cross-functional links. Query: 'Show me all OKRs contributing to Company Objective #1' — recursive graph traversal. Performance: materialize aggregate scores, update on check-in (not compute on read). Cache alignment tree (changes infrequently — once per quarter). Search: find OKRs by keyword, owner, department, status (Elasticsearch). 5) Transparency vs Privacy: By default: all OKRs visible to all employees (radical transparency). Some companies: individual OKRs visible only to manager and skip-level. Configurable: per organization's culture preference. OKR scores should NOT directly equal performance rating (OKR is aspirational — 0.7 is good). 6) Integration: Jira/Asana: link tasks to key results (auto-update KR progress based on ticket completion). Salesforce: link deals to revenue KRs (auto-update from CRM pipeline). GitHub: link PRs/deployments to engineering KRs. Slack: OKR check-in bot, progress celebrations, at-risk alerts.

The bell curve (forced ranking/vitality curve) is one of the most debated practices in Indian HR: 1) What It Is: Forced distribution: manager must rate team members into predefined buckets. Typical distribution: Top 10-15% (Exceeds Expectations), 60-70% (Meets Expectations), 15-20% (Below Expectations), 5% (Does Not Meet — potential exit). Popularized by Jack Welch at GE ('rank and yank'). Indian IT (TCS, Infosys, Wipro) widely adopted it. 2) Why It's Controversial: Mathematical fallacy: assumes every team has poor performers (what if you have an A-team?). Team of 5: manager forced to give 1 person low rating even if all performed well. Creates toxic competition instead of collaboration. Managers game the system: rotate who gets low rating each year. High performers leave — they know they may get unlucky. Many companies (Microsoft, Adobe, Accenture) have moved away from it. 3) Calibration Technology: Calibration session: leaders from peer teams meet to normalize ratings. Purpose: ensure 'Exceeds' in Team A is equivalent to 'Exceeds' in Team B. Technology support: a) Calibration Dashboard: Grid view: all employees in department, current manager-proposed ratings. Drag-and-drop: move employees between rating buckets. Distribution chart: show current distribution vs target. Side panel: employee profile, goal scores, peer feedback, prior years' ratings. b) Guided Distribution: System shows: 'Current distribution: 25% Exceeds, 55% Meets, 20% Below'. Target: 15/70/15. Highlights: 10% excess in 'Exceeds' — which employees should be reconsidered? Not strict enforcement — guidance with manager override and justification. c) Bias Detection: Flag: manager gives all team members same rating (leniency bias). Flag: rating difference between male and female employees in same role (gender bias). Flag: new joiners consistently rated lower regardless of performance (recency bias). Compare: rating distribution by gender, ethnicity, tenure, work location. d) Data Points in Calibration: Goal achievement scores. 360/peer feedback themes. Recognition received. Attendance/productivity metrics. Client feedback (for client-facing roles). Prior year trajectory (improving, stable, declining). 4) Modern Alternatives: Continuous feedback model: no annual rating, ongoing development conversations. Strengths-based: focus on strengths, not categorizing underperformers. Relative ranking without forced distribution: rank top-to-bottom but no forced buckets. OKR-based: evaluate based on objective outcome achievement. Check-in model (Adobe 'Check-in'): quarterly conversations, no ratings, managers make comp decisions. 5) Technical Implementation: ReviewCycle → CalibrationSession → [CalibrationAction]. CalibrationAction = {empId, originalRating, finalRating, justification, calibratorId}. Audit trail: who changed which rating, when, why. Distribution enforcement: configurable (strict, guided, or none). Analytics: pre-calibration vs post-calibration distribution. Year-over-year: individual rating trajectory for consistency check.

360 feedback collects performance input from multiple perspectives: 1) Feedback Sources: Self: employee's self-assessment. Manager: direct supervisor's evaluation. Peers: colleagues at same level who collaborate. Direct reports: for managers — upward feedback. Stakeholders: cross-functional partners, clients. Skip-level: manager's manager (for calibration). Each source evaluates different competencies from their unique perspective. 2) Technical Flow: a) Rater Selection: Employee nominates 3-5 peers and stakeholders. Manager approves/modifies rater list. System can suggest raters: based on collaboration data (shared projects, frequent email/Slack interaction). Minimum 3 raters per category for anonymity (if <3, responses combined with another category). b) Questionnaire Design: Competency-based questions: 'Rate [Name] on Communication: 1-5'. Behavioral indicators: 'Provides clear and timely updates to stakeholders'. Open-ended: 'What should [Name] continue doing?' 'What should they improve?'. Different questions for different rater types (direct reports rate leadership, peers rate collaboration). 10-15 questions (completion time: 10-15 minutes per person). c) Collection & Anonymity: Each rater receives unique, secure link. Responses stored with rater category but NOT individual identity. Aggregation: show average score per competency per rater category. Comments: presented verbatim but randomly ordered within category (prevent identification by writing style). Minimum respondents threshold: if only 2 peers responded, don't show peer category separately. d) Report Generation: Per competency: self-score vs manager vs peers vs reports (radar chart). Gap analysis: where self-perception differs from others' perception. Blind spot: self rates high, others rate low (development need). Hidden strength: self rates low, others rate high. Trend: comparison with previous 360 (if available). Comments: organized by theme (AI can cluster similar feedback). 3) Quality Controls: Social desirability: raters give inflated positive feedback. Mitigation: frame questions behaviorally ('How often does X demonstrate...'), forced distribution of ratings. Retaliation fear: direct reports afraid to give honest feedback about manager. Mitigation: strict anonymity, minimum respondents, aggregate only. Gaming: employees select only friendly raters as peers. Mitigation: manager approval of rater list, system-suggested raters, random assignment option. Low response rate: raters don't complete. Mitigation: automated reminders, manager nudges, keep survey short, deadline enforcement. 4) NLP for Comment Analysis: Sentiment analysis on open-ended comments. Theme extraction: cluster similar feedback (e.g., 5 people mention 'communication' → strong signal). Word cloud: most frequent themes. AI summary: 'Based on 8 rater comments, key strengths are technical expertise and mentoring. Key development area: delegation and time management.' 5) Data Model: FeedbackCycle → [FeedbackRequest]. FeedbackRequest = {subjectId, raterId, raterCategory, status, dueDate}. FeedbackResponse = {requestId, ratings: [{competencyId, score}], comments: [{questionId, text}]}. AggregatedReport = {subjectId, cycleId, competencyScores: [{competencyId, selfScore, managerScore, peerAvg, reportAvg}], themes: []}. Privacy: rater identity is NEVER stored in the report table. Even DBAs should not be able to link specific ratings to specific raters.

Attrition prediction is one of the highest-impact applications of people analytics: 1) Feature Engineering: a) Performance Data: Current and historical performance ratings (trend: improving or declining). Goal achievement scores. PIP history (strong predictor). Peer feedback sentiment. Skill gap (role requirements vs current skills). b) Engagement Data: Pulse survey scores and trends. eNPS (Employee Net Promoter Score). Feedback frequency (giving and receiving). Recognition received (quantity and recency). c) Compensation: Compa-ratio: salary vs market midpoint (below 0.85 = high risk). Time since last raise or promotion. Stock vesting cliff (common to leave after cliff). Total comp vs external offers (market data). d) Career: Tenure in current role (>3 years without promotion = risk). Manager quality (employees leave managers, not companies). Internal mobility: applied for internal roles? (signal of restlessness). Career development plan: does one exist? Is it followed? e) Work Patterns: Work hours trend (burnout indicator). PTO usage (not taking leave = burnout, excessive = disengagement). Overtime frequency. Remote vs office pattern changes. f) External: Job market conditions (unemployment rate, tech hiring trends). LinkedIn profile updates (changed photo, added skills, open-to-work flag). Glassdoor review activity from company. 2) Model Building: Algorithm: gradient boosted trees (XGBoost, LightGBM) — handles mixed features, interpretable. Training data: historical exits (voluntary only — exclude layoffs, retirements). Target: left within next 6 months (binary classification). Class imbalance: typically 85% stay, 15% leave — use SMOTE or class weights. Feature importance: identify top predictors (usually: tenure, comp ratio, manager rating, recent rating decline). Cross-validation: time-based split (train on past years, validate on recent). 3) Output & Action: Risk score: 0-100 for each employee. Risk categories: Low (<30), Medium (30-60), High (60-80), Critical (>80). Explanation: top 3 factors driving risk for each employee. Example: 'Priya (Risk: 78) — Drivers: compa-ratio 0.82, no promotion in 3.5 years, recent pulse survey score decline'. 4) Action Framework: For high-risk critical talent: Manager alert (confidential): 'Consider having a career conversation with Priya'. HR Business Partner: review compensation, discuss retention actions. Retention toolkit: compensation adjustment, role expansion, learning opportunity, flexible work. NOT automated: never auto-email employee ('We think you might leave') — that would be terrible. Track intervention effectiveness: of employees flagged high-risk who received intervention, what % stayed? 5) Ethical Considerations: Privacy: employees don't know they're being scored (transparent about data collection, but not individual predictions). Bias: if model penalizes employees who take maternity leave → discriminatory. Regularly audit model for demographic bias. Self-fulfilling prophecy: if manager treats high-risk employee differently, they may actually leave. Use: aggregate insights ('Engineering team has 30% high-risk') more than individual targeting. Legal: in some jurisdictions, predictive HR analytics may face regulatory scrutiny. 6) Accuracy: Typical model performance: AUC 0.75-0.85. Not perfect — some leavers won't be predicted (sudden life events). False positives: some flagged employees won't actually leave. Focus on: correctly identifying the top 10-20% highest risk (precision at top). ROI: if model identifies 50 potential leavers and retention actions save 20, and replacement cost is ₹10-15L each, ROI = ₹2-3 Cr.

Performance-linked compensation (merit increases, bonuses, equity) is the most impactful output of the review cycle: 1) Compensation Components Linked to Performance: Base salary increment: typically 5-15% annually (India IT average: 8-10% for 'Meets', 12-15% for 'Exceeds'). Variable pay (bonus): percentage of target payout based on individual + company performance. Equity (ESOP/RSU): grants based on rating and strategic importance. Promotion: performance is a prerequisite, combined with readiness assessment. 2) Merit Matrix: 2D matrix: Performance Rating × Position in Pay Band = Increment %. Example: Exceeds + Below Midpoint = 15% raise. Exceeds + Above Midpoint = 10% raise. Meets + At Midpoint = 8% raise. Below + Above Midpoint = 0% raise. Position in band (compa-ratio) prevents runaway salaries for long-tenured employees. 3) Variable Pay: Company performance factor (0-1.5×): based on revenue, profit targets. Business unit factor (0-1.5×): based on BU-specific goals. Individual performance factor (0-1.5×): based on rating. Payout = Target Bonus × Company Factor × BU Factor × Individual Factor. Example: Target bonus = ₹2L. Company achieved 110% target = 1.1×. BU achieved 90% = 0.9×. Individual 'Exceeds' = 1.3×. Payout = ₹2L × 1.1 × 0.9 × 1.3 = ₹2.57L. 4) Budget Allocation: Total compensation budget: typically 3-5% of payroll for merit increases. Distribution: HR allocates budget by department based on headcount and rating distribution. Manager's budget: based on team size and ratings. Manager decides individual allocations within budget constraints. System enforces: total allocations ≤ department budget. 5) Technical Implementation: CompPlanningCycle = {year, totalBudget, status, meritMatrix, variablePayFormula}. EmployeeCompPlan = {empId, currentSalary, compaRatio, rating, proposedIncrement, proposedBonus, proposedEquity, managerNotes}. BudgetAllocation = {departmentId, allocatedBudget, usedBudget, remainingBudget}. Workflow: Manager proposes → HR reviews → Department head approves → CHRO/CFO final approval. Scenario modeling: 'If we give Exceeds 15% instead of 12%, how much additional budget needed?' Parity analysis: are we paying fairly across gender, ethnicity, location for same role + performance? 6) Communication: Increment letter: generated from comp planning system, sent through employee portal. Total rewards statement: shows salary + bonus + equity + benefits = total value. Market benchmarking: show employee their compensation vs market percentile. Timing: announce within 2-4 weeks of performance review completion. 7) India-Specific: CTC restructuring: increment may change CTC components (Basic goes up → PF increases → take-home may not increase proportionally). Tax impact: higher salary → higher tax → show employees net impact. Notice period: high performers in notice period — accelerate counter-offer from comp planning data.