Employee Survey Theme Analyst

What it does

Turns open-ended survey answers (engagement / exit / pulse / 360 comments) into a systematic theme map, sentiment profile, and action-priority matrix while preserving anonymity. The core method is Braun & Clarke reflexive thematic analysis (6 phases: familiarization → initial codes → searching for themes → reviewing → naming → reporting), run in a hybrid mode: first inductive open codes at clause level (e.g. workload, no_recognition, manager_comms), then deductive grouping via HR frameworks. Every theme is anchored to two canons: the JD-R Model (Bakker & Demerouti) — Demands (burnout path) vs Resources (engagement path) — and Herzberg's Two-Factor Theory — Hygiene (pay, conditions, policy) vs Motivator (achievement, recognition, growth), yielding the "plug hygiene first, then build motivators" sequence. Themes are cross-tabulated with key-driver / importance-performance logic (frequency × sentiment intensity) and prioritized via Priority = Freq_norm×0.4 + Sentiment_intensity×0.35 + Controllability×0.25. Where available, eNPS (Promoter%−Detractor%) is reported and free text is split into Detractor/Passive/Promoter segments — because detractor themes ≠ promoter themes, the most action-relevant cut.

When to use it

When an open-ended survey item has hundreds-to-thousands of free-text answers and someone asks "what are the common themes, and which is urgent?" Ideal for exit-interview comments, eNPS explanations, pulse "why this score" fields, suggestion boxes, and 360 comments — especially when a Likert score dropped but the why is unclear and free text is the only context.

Method / frameworks

Braun & Clarke (2006/2019) reflexive TA; Krippendorff's α / Cohen's κ for inter-rater reliability (target κ≥0.61; single-pass flagged); JD-R Model; Herzberg Two-Factor; key-driver / Kano-like importance-performance mapping; eNPS segment distribution. Operational thresholds: 5% theme threshold, saturation check, clause-level coding for ambivalence. GDPR/KVKK + k-anonymity (k≥5): sub-cuts with n<5 are suppressed; re-identification is forbidden. Benchmarks are never fabricated — sector norms are requested from the user or stated as "typical range."