How it started

The motivation for this project was to help HR teams quickly and accurately analyze and understand employee feedback from eNPS surveys.

In fast-paced workplaces, understanding employee feedback is essential for maintaining engagement and satisfaction. To address this, we developed an AI-powered tool to analyze eNPS (Employee Net Promoter Score) survey comments.

The tool summarizes key topics, and identifies sentiment, allowing HR teams to take meaningful action.

The Challenge: Making employee feedback actionable

HR departments often face challenges with the overwhelming volume of comments collected from employee surveys. For example, a company with 200 employees can receive thousands of comments, making it time-consuming to read through them all and delaying the identification of critical issues.

Our challenge was to simplify this process and deliver clear insights — removing the manual workload. This led us to ask: How might we simplify the analysis of survey comments data from eNPS surveys?

The process

1. Discover:
   Through user interviews and observations, we gained valuable insights into the challenges faced by HR teams, particularly the overwhelming task of analyzing large volumes of survey comments. These insights helped us understand the real pain points and prioritize the problem.

2. Define:
   We synthesized our findings to clearly define the problem: HR professionals needed a faster, more efficient way to analyze and interpret large sets of qualitative feedback while minimizing manual effort.

3. Ideate & Design:
   We brainstormed various solutions, exploring how AI could automate and smooth the analysis process. The focus was to create a tool that could quickly summarize key topics and analyze sentiments to surface insights. We also worked on designing a user-friendly interface to ensure the tool would be accessible and intuitive for HR teams.

4. Prototype & Testing:
   I designed initial prototypes of the AI-powered tool, prioritizing its ability to read and summarize comments. During this phase, I collaborated closely with developers to understand how Carol (our AI tool) works and what capabilities could be used or enhanced. This collaboration also marked the beginning of training Carol, teaching her how to read and analyze survey comments.

Collaboration is a keyword in my process, I involved the Product Manager and developers since the first wireframes — making sure everything we designed was technically viable but also supplied the users’ needs.

Image below description: Early wireframes shared with the team to openly discuss some options and if funcionalities were technically viable.

Research

Our goal was to understand how we could improve the qualitative data analysis for employee surveys. I worked together with the Product Manager to create a CSD Matrix (Certainties, Suppositions, Doubts) to guide our questions. This framework helped us define key areas of investigation.

Interviews approach

• Created an interview script with open-ended questions to explore users’ experiences with analyzing survey comments.

• Conducted 10 remote user interviews across different company sizes to understand how HR professionals interpret qualitiative feedback.

The solution:
AI-Driven Comment Analysis

To simplify the analysis of survey comments, we developed AI-driven topic detection and sentiment analysis tool.

The AI reads through all the comments, identifies sentiments and recurring topics. This automation significantly reduces the time required for HR to do the data analysis and helps surface trends that might otherwise go unnoticed.

Image below description:
Image 1: The tool analyzes survey comments, identifies sentiment, and summarizes key topics.
Image 2: It also highlights related topics, enabling HR to explore deeper insights.

Designing for Accessibility: Optimizing Charts for Colour Blindness

Ensuring that insights are clear and accessible to all users was a key consideration in our design process. The colours used in the sentiment graphs were carefully tested for different types of colour vision deficiencies using Stark, including:

• Deuteranopia: Red-green color blindness

• Protanomaly: Reduced sensitivity to red light

• Tritanomaly: Reduced sensitivity to blue light

• Protanopia: Complete red colour blindness

By testing with these conditions, we ensured that the data remained distinguishable and easy to interpret for everyone. This commitment to accessibility enhances the usability of our tool, allowing HR teams to make informed decisions without visual barriers.

How it works: AI Architecture & Data Processing

This project required close collaboration between product design, product management, and development to ensure AI-generated insights were both accurate and actionable. The engineering team played a crucial role in designing a scalable and efficient system while ensuring that HR teams could easily interpret and act on the insights.

AI Processing Flow

Carol is the AI platform that manages the entire data processing pipeline. Here's how it works:

1. Data Collection & Preparation

• Employee responses are stored in Carol’s database.

• Every 20 minutes, an automated process (cron job) identifies new, unprocessed responses and prepares them for AI analysis.
  

2. Topic Identification & Sentiment Analysis

• Carol sends each response to GPT in two steps:

  • Topic Detection: The AI identifies key topics mentioned in the feedback and aligns them with predefined categories.

  • Sentiment Analysis: A second AI request determines whether the sentiment for each detected topic is positive, neutral, or negative.

• The results are stored in the database.
  

3. Summarization & Insights Distribution

• Every hour, a cron job consolidates and structures the processed data.

• The final insights are sent via webhook to the Platform, allowing HR teams to access real-time, actionable feedback.
  

This structured flow ensures that AI-generated insights are accurate, scalable, and easy to interpret.

Project learnings

1. Collaboration drives success: Close teamwork with developers to understand how Carol and ChatGPT work ensured that the solution was technically feasible and met user needs.


2. Continuous feedback matters: Prototyping and testing early in the process highlighted usability challenges, such as the initial colour palette issues.

3. Constant feedback loops improved AI performance: Ongoing collaboration between me, engineers and developers helped refine the AI’s accuracy.

4. Accessibility Matters
Designing accessible visualisations required iterative testing with tools like Stark to ensure inclusivity for users with colour blindness. This improved the overall usability and equity of the platform.

Next steps

1. Enhance AI accuracy: Continue training the AI tool using real-world data and edge cases, and integrate additional human validation loops to improve context understanding and sentiment precision.

2. Optimize integration: Work on seamless integration with existing HR platforms, ensuring data flows smoothly and insights are delivered without disruption.

3. Explore new features: Identify opportunities for additional enhancements, such as deeper analytics on employee feedback trends and expanded reporting capabilities, to further support HR decision-making.