HR Analytics, HR Tech and AI: Revision Notes

HR metrics measure activity and outcomes; HR analytics uses data and statistics to improve people decisions. Types: descriptive (what happened), diagnostic (why), predictive (what will happen) and prescriptive (what to do). Bersin’s maturity levels: operational reporting, advanced reporting, strategic analytics, predictive analytics; most firms stay at levels 1–2.

Process: define the business question and decision, form hypotheses, gather and clean data, analyse simply first, interpret for practical meaning and causality, act, then measure against a comparison group. LAMP (Boudreau and Ramstad): Logic, Analytics, Measures and Process must all be present.

Data comes from the HRIS, payroll, attendance, performance, ATS, LMS, surveys, exits and business systems; common problems are inconsistent definitions and titles, missing data, mismatched IDs and small groups. Use the least data needed and aggregate.

Use medians for skewed pay. Correlation does not prove causation. Linear regression models continuous outcomes such as pay; logistic regression models yes/no outcomes such as leaving. Distinguish statistical from practical significance. Confounding: remote workers may seem to leave more because they are newer. Simpson’s paradox: women can look less promoted overall yet equal within every function. Multicollinearity (age and tenure) is checked with VIF. Experiments with control groups establish cause.

Attrition: segment leavers versus stayers, then use logistic regression for drivers and risk and survival analysis for timing; act on drivers (e.g., pay below 90% of midpoint plus no progression doubling risk) and use risk scores only for support. Pay equity: regress log pay on grade, family, experience, rating and location plus gender; the gender coefficient is the adjusted gap. Selection validity correlates assessment scores with later performance. Bradford factor = spells² × days: ten one-day absences score 1,000 versus 10 for one ten-day absence; use for conversations, not penalties. ONA maps collaboration but needs privacy safeguards.

Core metrics: attrition = separations ÷ average headcount; regretted attrition; compa-ratio; personnel cost ratio; revenue per employee; HCROI = (revenue − (opex − employee cost)) ÷ employee cost; absenteeism = days absent ÷ scheduled days. Dashboards: decision-led, 6–10 metrics, trends, targets and segments, shared definitions, colour only for exceptions.

Google Project Oxygen (2008–09) proved managers matter and defined eight (later ten) good-manager behaviours used in feedback and training; Google also found four interviews predicted as well as more. IBM claimed (2019) about 95% accuracy predicting quits and about US$300 million saved. Salesforce has audited pay annually since 2015. (The compendium’s Under Armour, AON and similar examples are unverified.)

HRIS holds records, HRMS adds processes, HCM adds strategic talent management; plus ATS, LMS/LXP, engagement platforms, self-service and chatbots. Indian vendors include Darwinbox (unicorn since 2022), Keka, greytHR, Zoho People and PeopleStrong. Implementation: select, plan, redesign processes, configure and test with real users, train managers first, go live with support. Failures come from dirty data, automating bad processes, over-customisation, weak adoption and no owner.

AI in HR includes predictive models (attrition, matching), generative AI (JDs, policies, summaries, answers) and agents (multi-step tasks). Risks: bias (Amazon’s CV tool, 2018), opacity, hallucinated policy answers, privacy leaks into public tools and surveillance. Govern by keeping humans accountable for people decisions, bias testing, transparency, data minimisation, vendor checks and an inventory of uses. The EU AI Act treats employment AI as high-risk (obligations phasing in from 2026–27); NYC Local Law 144 requires bias audits; India relies on MeitY’s AI Governance Guidelines (2025) and the DPDP Act.

The employer is the data fiduciary, employees and candidates are data principals, and payroll, HRMS and verification vendors are processors. Processing for employment purposes is a legitimate use under section 7, so payroll and attendance data need no consent; non-employment uses (e.g., photos in marketing) do. Duties: notice where consent applies, accuracy, reasonable security, erasure when the purpose ends, grievance redressal, processor contracts and breach reporting (inform individuals and the Board, detailed report within 72 hours). Significant Data Fiduciaries need an India-based DPO, audits and impact assessments. Transfers abroad are allowed except to restricted countries. Penalties reach ₹250 crore for security failures. Rules were notified 13 November 2025, with most duties applying around May 2027.

Versus GDPR: GDPR covers paper records, applies since 25 May 2018, also requires 72-hour notification and fines up to €20 million or 4% of turnover. The compendium wrongly says DPDP needs employee consent and has no breach deadline. Monitoring should follow purpose, minimisation, fairness and awareness.