RoleSole Product ManagerWorkflow, requirements, controls, launch
WorkflowFeedback to actionSurvey comments through HR follow-up
PartnersEngineering + DataClassification, synthesis, product integration
StatusLive in productionAI-assisted, human-reviewed workflow
01Context
Open-text feedback was valuable but expensive
After each survey, HR had to read, group, summarise, and reply to hundreds of comments before they could act.
02Problem
Manual synthesis was slow and inconsistent
- Every admin grouped themes differently
- Effort ballooned with comment volume
- Employees waited too long for any follow-up
03Decision
Automate preparation, not accountability
Framework · Human-in-the-loop, escalate by confidence
AI classifies, clusters, and summarises the comments. HR keeps authority over interpretation and every employee-facing reply.
04Execution
Turn trust requirements into product behaviour
With Engineering and Data, I defined the taxonomy, confidence handling, review queues, anonymity rules, and the response workflow. In parallel, I owned a second AI surface built on the same taxonomy: an HR chatbot with hybrid lexical and semantic retrieval over a 3,500+ QA-pair knowledge base, where I structured the content and chunk boundaries myself.
Outcome
Faster synthesis, human judgement preserved
The efficiency figure is an estimate from a representative workflow. Controls are listed separately from measured outcomes.
Estimated~70% less synthesis timeA representative 500-comment cycle went from ~4 hours of manual reading to ~1 hour reviewing pre-grouped themes. Sentiment tagging scored >90% accuracy on a ~5,000-comment model-as-judge pass, with ~500 misclassified escalated to human review; sarcasm was the dominant failure mode, so low-confidence sarcastic comments route to neutral for human re-bucketing instead of forcing a wrong label, feeding each correction back as signal.
Designed controlHuman-approved repliesEvery employee-facing reply is reviewable and needs HR approval before sending.
FallbackManual review queueLow-confidence classifications are escalated, not forced into a theme.
Product lessonThe value did not come from displaying an AI summary. It came from redesigning what HR reviews, where judgement is required, and how weak evidence is handled.
What I would improve nextBuild a labelled eval set, track accuracy by theme and sentiment, and measure review time, reply acceptance, and time-to-action.
