To analyse employee survey open-ended responses: export responses with metadata intact, anonymise the data before analysis begins, build a coding framework (inductive, deductive, or hybrid), apply codes consistently across all responses, quantify themes by segment (department, tenure, seniority), connect qualitative themes to your quantitative scores, and present findings with verbatim evidence. Tools like Skimle can process hundreds of open-text responses systematically, link every theme back to the source quote, and segment findings by any metadata variable, turning a task that takes a team days into something achievable before your next leadership meeting.
The open-text box is the part of your employee survey that actually explains what your numbers mean. Most organisations collect it and barely use it. This guide is for HR professionals and people analytics teams and organisational consultants who want to change that.
Why open-ended responses are valuable but chronically underused
Rating scales are good at measuring. An engagement score of 3.4 out of 5 tells you something is wrong. It does not tell you why, who is affected, or what would fix it.
The explanatory layer lives in the text.
When someone writes "I don't understand where we're going since the restructure and nobody seems willing to explain it," they are giving you the mechanism behind whatever score they submitted. When someone writes "the work is interesting but I feel invisible to senior leadership," they are describing an experience no Likert scale was designed to capture. Open-ended responses invite respondents to tell you what is actually salient to them, rather than to rate what you thought to ask about.
Gallup's research on employee engagement has consistently found that global engagement scores conceal enormous variation across teams, managers, and functions. Understanding that variation requires qualitative data. The score tells you there is a gap; the open text tells you where it is and what is causing it.
The reason open-ended responses are underused is not that HR teams fail to recognise their value. The reason is volume. A survey of 400 employees might produce 250 open-text responses. Reading all of them carefully, coding themes, and building a coherent picture takes a skilled analyst two to three full days. For most HR teams, that time does not exist in the window between survey close and leadership presentation.
The result is a well-documented pattern: the Likert scores go into a dashboard, the open-text responses go unread, and the most informative part of the survey is quietly discarded.
A second and less-discussed problem is depth. The standard open-text box produces short, often formulaic responses. Employees give you a sentence when they have three paragraphs of relevant experience to share — not because they lack things to say, but because the format signals brevity. If you want richer qualitative data from a survey-scale sample, Skimle Ask offers a different approach: AI-conducted follow-up conversations that probe employees' initial responses with context-aware questions, generating interview-quality qualitative data without requiring a human interviewer for each participant. The resulting responses are longer, more specific, and significantly easier to code into meaningful themes.
Common approaches and why most fall short
Before covering the method that works, it is worth being clear about why the common shortcuts do not.
Manual reading and highlighting is the default for smaller datasets. One person reads through the responses, highlights striking comments, and summarises impressions. The problem is subjectivity and inconsistency. You notice the vivid and strongly worded responses more than the quiet pattern that is actually more widespread. You code the first fifty responses with one implicit framework and the last fifty with a slightly different one without realising it. The findings are genuinely influenced by whoever happened to be doing the reading that day.
Word clouds have the attraction of being quick and visual. They are also almost useless for employee survey analysis. Word clouds count word frequency, not meaning. "Management" appearing large in a word cloud could mean people feel well-managed, poorly managed, or anything in between. The context that makes a word mean something is exactly what word clouds strip out. Presenting a word cloud as employee sentiment analysis is not analysis; it is pattern decoration.
General-purpose qualitative coding software (NVivo, ATLAS.ti, MAXQDA) can produce rigorous analysis, but these tools are designed for researchers who spend weeks or months on a single dataset. They have steep learning curves and are not built for the HR workflow of processing survey results quickly, segmenting by department, and producing a board-ready presentation within a fortnight. Our qualitative data analysis tools comparison covers the trade-offs across these tools in detail.
Generic AI tools like ChatGPT are increasingly used to summarise open-text survey responses, and they can produce a quick overview. The limitation is rigour. Pasting 200 responses into ChatGPT and asking for a summary produces a plausible-sounding synthesis with no audit trail, no link between themes and the specific responses that generated them, and no reliable mechanism for consistency across large datasets. There is no way to verify that the theme "people feel unrecognised" accurately represents what respondents said, or how many said it, or which teams it came from. Our post on why generic AI approaches fall short for structured analysis explains why the architecture of tools like ChatGPT creates these limitations specifically for systematic qualitative work. For a broader look at how AI can be used responsibly in research, see how to use AI in qualitative research.
The alternative is a structured, systematic method. It takes more discipline than any of the above, but it produces findings that are defensible and actionable.
Step-by-step: how to analyse employee survey open-ended responses
Step 1: Prepare, clean, and anonymise
Export your survey results with one row per respondent, keeping all metadata in separate columns: department, team, tenure band, seniority, location, manager. Remove duplicates and non-meaningful entries (test submissions, single characters), and standardise inconsistent metadata values — "Product" and "Product Management" need to resolve to one label before segmentation is possible. Keep metadata attached to each response throughout; a comment disconnected from its department has limited analytical value.
Before coding begins, anonymise. Survey responses regularly contain identifying details: named colleagues, role descriptions, references to specific incidents. In small teams, even an unnamed comment can identify the writer. Skimle's Anonymise feature handles this systematically at scale, replacing identifying details with consistent placeholders while preserving the meaning needed for analysis — particularly useful when data will be shared with external analysts or third-party tools.
Step 2: Build your coding framework
Using Skimle: Upload your responses and let Skimle process them systematically. It discovers the theme structure from the data and builds an initial category hierarchy with every response already grouped under supporting quotes. You then take control: merge categories that overlap, split any that are too broad, rename themes to fit your organisation's language, and add codes for anything not captured. The framework emerges from the data rather than being imposed on it — and the entire dataset is coded by the time you start reviewing.
Manual approach: Decide upfront whether to use inductive coding (themes emerge from reading), deductive coding (you apply a predefined framework such as an engagement model), or a hybrid of both. Inductive is better for open-ended discovery; deductive is faster and produces year-on-year comparability. Start with ten to fifteen broad codes and refine as you go. Our guide on demystifying thematic analysis covers these choices in detail.
Step 3: Code your themes systematically
If you used Skimle in Step 3, your responses are already coded. Step 3 becomes a review: read through the grouped passages for each theme, confirm the groupings make sense, adjust anything miscategorised, and add codes for anything not surfaced. The dataset is fully coded before you start — your job is refinement, not construction.
If you are working manually, tag each response with the appropriate codes. The essential disciplines: distinguish positive from negative valence within the same theme (praise and criticism of a manager both code as "management" but mean opposite things), keep verbatim quotes linked to each code so findings are auditable, and revise earlier coding when your framework evolves mid-way. For deeper reference, see how to code qualitative data and how to summarise and synthesise qualitative data.
Step 4: Quantify themes by segment
Once coding is complete, the analytical work begins in earnest. The question is not only which themes exist in the data, but where they are concentrated.
For each significant theme, ask:
- Which departments or teams does it come from? Is it spread evenly or concentrated in one area? A theme mentioned by 40% of respondents in one division and 5% elsewhere is not a company-wide finding. It is a signal about that division.
- How does it vary by tenure? New joiners and long-tenured employees often experience the same organisation very differently. A concern about career development raised predominantly by people in their first two years is a different problem than the same concern raised by people with seven or more years of tenure.
- Does it vary by seniority? Individual contributors and managers often perceive the same situations differently. A theme about lack of strategic clarity might be concentrated at middle-management level, where people are expected to translate strategy but are not receiving enough to translate.
- How does it correlate with quantitative scores? Do respondents who mentioned a particular theme also tend to have given lower scores on a specific quantitative question? If so, the qualitative theme may be the explanation for a pattern already visible in your scores.
This segmentation is what turns an observation into a finding. "People mentioned unclear direction" is an observation. "Unclear direction was mentioned by 38% of respondents in the commercial team, predominantly those with three to six years of tenure, and this group gave the lowest scores on 'confidence in leadership'" is something you can act on.
Skimle is built specifically for this workflow. You upload responses with metadata columns intact and can immediately filter any theme by department, tenure, seniority, or any other variable you collected. What would otherwise require many manual cross-tabulation passes happens in a single view. The features overview explains how this works in practice.
Step 5: Connect qualitative themes to quantitative scores
The findings from open-ended analysis are most powerful when they bridge the qualitative and the quantitative.
For each significant theme, ask whether it maps to one or more of your quantitative questions. If "lack of recognition" is a recurring theme in your text responses, does it correlate with lower scores on whatever recognition question your survey included? If "unclear strategy" is prevalent, is it associated with low scores on "confidence in leadership" or "sense of purpose"?
When you can show that a qualitative theme aligns with a specific quantitative pattern, you are providing an explanation for the score rather than just an observation alongside it. This is where the two modes of survey analysis reinforce each other.
The reverse is equally informative. If you have a low quantitative score that no open-text theme clearly explains, that is a signal worth investigating. Either the open-text question was not designed to elicit explanation for that measure, or the concern is one people are reluctant to express in writing. Both are meaningful findings.
Step 6: Validate before reporting
Before you present anything, run a validation pass on your findings.
Check for coding inconsistency. Take a sample of twenty or thirty responses and recode them independently. If your new coding differs significantly from the original, resolve the inconsistency before reporting.
Test findings against the data. For each headline finding you plan to present, pull the verbatim quotes that support it and read them again. Does the finding accurately represent what people said, or have you overstated the pattern? The test is whether a sceptical colleague reading the same responses would reach the same conclusion.
Look for disconfirming evidence. When you have established a pattern, actively look for responses that complicate or contradict it. A finding that holds up under scrutiny is a finding you can present with genuine confidence.
Check your sample coverage. If responses from a particular team are underrepresented, your theme distribution for that segment may not be reliable. Flag the limitation in your reporting. The qualitative research sample size guide provides practical guidance on when you have enough responses to draw conclusions, which applies to survey data as much as interview data.
Practical examples by survey type
The method above applies across all types of employee surveys, but the emphasis shifts depending on what you are measuring.
Annual engagement surveys are typically the largest datasets, with the broadest question coverage. The priority is segmentation: finding which themes are concentrated in which teams, functions, or tenure cohorts. The thematic analysis is most valuable when it can explain the divergence between high-scoring and low-scoring units. A function with an engagement score of 3.1 compared to a company average of 3.7 needs a qualitative explanation, not just a number.
Exit surveys tend to produce more direct language, because departing employees have less reason to be diplomatically vague. Themes to look for: management quality (often understated even here), growth and development, compensation relative to market, workload, and culture fit. The key analytical question is whether departure themes cluster around specific managers, teams, or tenure stages, because the answer determines whether intervention is targeted or systemic. Our dedicated guide on exit interview analysis covers this in detail.
Pulse surveys generate smaller datasets per cycle but accumulate over time. The value in pulse open-text analysis is trend detection: is a theme that was minor in Q1 becoming more prominent in Q2 and Q3? Are concerns that were concentrated in one team spreading? Tracking theme frequency across cycles requires consistent coding frameworks, which is an argument for establishing your theme structure early and maintaining it across survey rounds.
How to present findings to leadership
Analysis that does not reach the people who can act on it is analysis that changes nothing.
Lead with the finding, not the process. Your audience does not need a methodology section. They need the headline insight, two or three pieces of supporting evidence, and a clear statement of what you recommend.
Anchor qualitative claims in numbers. "A significant proportion of respondents in the commercial team described feeling unclear about strategic direction" is more credible than "people in commercial feel unclear about strategy." Frequencies matter. Even approximate ones ("roughly one in three respondents in this function") are more useful than none.
Use quotes as evidence, not performance. One or two verbatim quotes that are specific, articulate, and representative do more for a finding than a paragraph of paraphrase. Choose for representativeness and clarity. Quotes should illustrate the pattern; they should not be selected because they are the most emotionally vivid outliers.
Show the connection to quantitative data. Where a qualitative theme corresponds to a specific low score on a quantitative measure, make that connection explicit. It demonstrates that the qualitative findings explain the patterns already visible in your numbers, rather than sitting alongside them as a separate, harder-to-verify body of evidence.
Prepare for challenge. Leadership teams sometimes push back on qualitative findings with "but is that really representative?" The response is the methodology: you can point to how many responses the theme appeared in, which segments it came from, and which quantitative scores it corresponds to. Rigour in the analysis is preparation for scrutiny in the presentation. Our guide on presenting qualitative research findings to executives covers this in detail, including how to handle defensive reactions to management-related themes.
Where AI tools help and where they fall short
The step-by-step method above is sound at any scale. The limiting factor, without tool support, is time.
A dataset of 400 open-text responses takes a careful analyst two to three full days to code, cross-cut, and validate properly. Most HR teams do not have that capacity in the window between survey close and reporting.
Purpose-built AI analysis tools change the practical equation. A tool like Skimle can process all 400 responses, identify recurring themes, build a structured code hierarchy, link every theme to the verbatim responses it came from, and segment by department or tenure, in minutes rather than days.
What this does not do is replace analytical judgement. The theme structure an AI generates is a starting point, not a finished product. You review it: merging categories that are too similar, splitting those that are too broad, renaming themes to fit your organisation's language, and ensuring the resulting framework accurately represents what employees actually said. The AI removes the most laborious parts of the work; the expert judgement that makes analysis reliable and defensible remains yours.
The contrast with generic AI tools is worth making explicit. Asking ChatGPT to summarise 200 survey responses produces a plausible synthesis. It does not produce a traceable one. There is no audit trail from theme to source response. There is no segmentation by department or tenure. There is no systematic mechanism to ensure that every response was read with equal attention rather than the most recent ones dominating the context window. For casual reading of a small dataset, this may be acceptable. For analysis that will inform a board-level talent strategy, it is not. Quality in the AI era makes this case more broadly: the availability of fast AI-generated summaries raises the stakes for rigorous approaches, not lowers them.
Skimle is built specifically for structured qualitative analysis at the scale of employee surveys. Responses upload with metadata intact, themes are linked to source quotes throughout, and findings segment by any variable you collected. For teams who need to go from survey close to defensible, board-ready insights in a matter of hours rather than days, this is the practical alternative to choosing between "quick and unreliable" or "rigorous and takes two weeks." You can see how Skimle fits into HR and people analytics workflows on the HR and people teams use case page.
Ready to make sense of your open-ended survey responses? Try Skimle for free and turn hundreds of text responses into structured, traceable insights in a fraction of the time.
Want to learn more about qualitative analysis? Read our guides on thematic analysis methodology and how to summarise and synthesise qualitative data.
About the authors
Henri Schildt is a Professor of Strategy at Aalto University School of Business and co-founder of Skimle. He has published over a dozen peer-reviewed articles using qualitative methods, including work in Academy of Management Journal, Organisation Science, and Strategic Management Journal. His research focuses on organisational strategy, innovation, and qualitative methodology. Google Scholar profile
Olli Salo is a former Partner at McKinsey & Company where he spent 18 years helping clients understand the markets and themselves, develop winning strategies and improve their operating models. He has done over 1000 client interviews and published over 10 articles on McKinsey.com and beyond. LinkedIn profile