Workplace Survey Comment Analysis

 How ORRO used Natural Language Processing (NLP) and machine learning (ML) to make sense of a decade of employee feedback for a global design firm

The Challenge

Gensler, one of the world's largest architecture and design firms, runs the Workplace Performance Index — an online diagnostic tool measuring how well physical work environments support the people in them. Over a decade of deployments, the WPIx had accumulated an enormous archive of employee responses to two open-ended questions:

What do you like most about your workplace? What do you like least about your workplace?

By 2020, that archive represented hundreds of surveys and more than 174,000 individual responses. Each comment contained genuine, unprompted employee perspective — the kind of qualitative insight that no multiple choice question can fully capture.

The problem: at that scale, no human team could read and categorize the responses meaningfully. A word cloud could tell you that "noise," "desk," and "privacy" appeared frequently. It couldn't tell you how employees felt about those things, whether those feelings were changing over time, or what they actually predicted about workplace satisfaction.

Gensler needed a scalable, rigorous way to turn a decade of open-ended text into actionable design intelligence.

What We Did

We developed a multi-method NLP pipeline combining three complementary approaches:

1. Unsupervised Learning to Discover Topics We used BERTopic, a state-of-the-art topic modeling library, to let the data surface its own themes without imposing categories in advance. This was essential for catching topics no one would have thought to look for — and it worked. Alongside expected themes like noise, privacy, and meeting rooms, the model surfaced genuinely unexpected topics that human coders would never have created as categories.

2. Silver Labeling to Scale Human Expertise Gold labeled data — comments manually classified by experts — is the bottleneck in any NLP project. We developed a programmatic labeling approach that allowed subject matter experts at Gensler to encode their knowledge as labeling functions, generating "silver" labeled training data at scale. This let us train a classifier without requiring thousands of hours of manual annotation.

We organized the topic hierarchy around Gensler's Workplace Ecosystem framework: Space, Culture, Technology, and Policy — with subtopics nested beneath each.

3. GenAI to Supplement Training Data For topic categories where real examples were sparse, we used generative AI to create synthetic training text — prompting models to write workplace complaints and compliments from specific perspectives, then culling the output for quality. This "Create, Constrain, Cull" approach meaningfully expanded our training data for underrepresented categories.

What We Found

Noise and Privacy Dominate — But the Story Is More Nuanced Noise was the most frequently mentioned negative topic, and privacy was close behind. But the unsupervised modeling revealed that employees weren't just complaining about sound — they were describing a deeper need for control over their work environment. Comments about distraction, lack of focus space, and open plan layouts clustered together in ways that pointed toward variety and choice as the real underlying need, not just acoustic treatment.

Unexpected Topics Emerged That Would Never Appear on a Survey Among the topics the model surfaced that no human researcher would have created as a category: pests. Ants, rats, and spiders appeared with enough frequency across certain building surveys to constitute a meaningful signal about facilities management and employee experience. This is exactly the kind of insight that only emerges when you let the data speak for itself.

The Model Informs Real Design Decisions The topic classifier became a practical tool for Gensler's workplace strategy practice — allowing teams to rapidly profile any client's survey results against the broader benchmark, identify which workplace ecosystem dimensions were driving satisfaction or dissatisfaction, and make the case for specific design interventions with quantitative backing.

Privacy emerged as a critical design outcome — not just a nice-to-have. Variety of space types was consistently associated with positive sentiment across multiple topic categories. The data gave designers a richer, evidence-based brief to work from.

What This Means for Organizations Running Surveys

Most organizations treat open-ended survey responses as supplementary color — a few quotes to illustrate the quantitative findings. ORRO's approach inverts that relationship: the text becomes the primary analytical object, and the numeric data provides context and validation.

The implications are significant:

For HR and People Operations leaders: the comments your employees leave in open-ended survey fields contain more actionable intelligence than the Likert scale responses above them. You just need the right tools to access it.

For workplace strategists and architects: employee language about space, culture, technology, and policy can be systematically analyzed, tracked over time, and benchmarked — giving design decisions a quantitative foundation they rarely have today.

For anyone running surveys at scale: the bottleneck is not data collection. It is analysis. Text analytics closes that gap.

Interested in what your survey data could tell you?

ORRO builds bespoke text analytics solutions for organizations with complex, unstructured data challenges. Every engagement is custom — we work with your data, your questions, and your domain.

Contact us to start a conversation.

© 2026 ORRO AI Genius, LLC. All rights reserved.