Gains:
- Ability to analyze open-ended survey responses according to themes, frequency and mood
- Ability to interpret metrics such as eNPS and engagement and turn them into action recommendations
- Ability to establish a bias-free survey analysis framework that preserves anonymity and confidentiality
You conducted an employee engagement survey and received hundreds of free text comments from 400 people. It would take days to read them one by one and extract the main themes, and it would be subjective: from the eyes of a tired reader, some themes would stand out and some would disappear. Artificial intelligence (AI) groups hundreds of open-ended responses into themes, frequency, and mood in minutes; It creates a picture that will prompt management to take action. But this is where one of HR's most sensitive responsibilities comes into play: these comments belong to individuals, anonymity is promised, and this promise should never be broken. In this unit, we will learn how to conduct survey analysis end-to-end and preserve anonymity and fairness.
A few terms: eNPS (employee Net Promoter Score) is a metric that measures employees' propensity to recommend the company; It is found by subtracting the percentage of "those who recommend" from the percentage of "those who do not recommend". Participation/commitment refers to the emotional bond of the employee to his job and the company. Sentiment analysis is classifying the positive/negative/neutral tone of a text. Thematic coding is the process of dividing free texts into common topics (such as workload, management, career).
Theming Open-Ended Responses
AI turns free text into structural insight. The critical point: keeping the analysis anonymous and aggregated. Step by step process:
- Anonymize data. Remove or generalize name, unit, descriptive details.
- Extract the themes. Ask how many comments each theme appears in.
- Add the emotional state. Is each theme positive or negative?
- Ask for sample quotes — but safe ones. Representative quotes that will not give away the individual.
- Turn it into action. 1-2 concrete suggestions for each important theme.
- Interpret with humans. Numbers are read with context; The final comment is yours.
Your role: an employee experience analyst.Task: Analyze the following anonymous survey comments.1) Group comments into major themes (e.g. management, workload, career, pay, culture, tools). State how many comments you have for each theme.2) Summarize the general mood (positive/negative/neutral) of each theme.3) Identify the 3 most frequently mentioned positive aspects and 3 areas of development.Rule: Do not try to match comments with PEOPLE. If there are details (name, unit, specific event) that describe the individual, GENERAL them when summarizing. Do not quote a rare/specific comment alone that could reveal a person.
This prompt; It produces actionable output for management, such as "The workload theme appeared in 78 comments, predominantly negative; evening shift is the main reason" — without revealing anyone.
Tip: The biggest ethical risk in survey analysis is breaking anonymity. A comment such as "The only person in the accounting department whose child was born" can give a person away on its own. Explicitly tell the AI to generalize details that identify the individual and not highlight rare/specific comments.
Interpreting Metrics and Translating them into Action
The raw number alone is not meaningful; Is "eNPS 12" good or bad? AI helps contextualize metrics and translate them into action recommendations:
Interpret the survey metrics below and provide a 1-page summary to the manager.Data:- eNPS: 12 (25 last year) - Overall engagement: 3.6 out of 5- Lowest scoring area: "career development" (3.1)- Highest scoring area: "team relationships" (4.4)Required:1) Explain in plain language what is good and what is worrisome2) Possible hypotheses for eNPS decline (exact reason fabrication, hypothesis)3) Concrete action proposal for the 3 most critical areas
Divide the negative comments on the "Career development" theme into sub-themes (e.g. promotion uncertainty, lack of training, lack of feedback). For each sub-theme, write down the number of comments and an action idea. Do not expose the person; Keep quotes anonymous and representative.
Three Mini Cases
Case 1 — From days to minutes. In a company of 600 people, two HR specialists were hand-coding 1,400 free comments on an annual survey in one week. With AI, theming has been reduced to a few hours; The team devoted time to analysis, interpretation and action planning. Coding consistency also increased because the same criteria were applied to all comments.
Case 2 — Interception of secret signal. At one manufacturing company, overall scores were good, but AI highlighted the “shift schedule” theme with 46 negative comments; this was a pattern overlooked in hand reading. Management reviewed the shift system; In the following quarter, related complaints halved.
Case 3 — Preventing an anonymity crisis. In the first attempt, an expert also provided unit information to the AI; The printout contained the revealing quote "About the sole manager in unit X." The expert noticed this before publication, corrected the process: unit information was removed, a rule was introduced to "report separately any sub-group with less than 5 comments". Trust was preserved.
Weak Prompt / Strong Prompt
Weak prompt:Summarize these survey comments.
The result: superficial, numerous, with no guarantee of anonymity; a summary that could even give away an individual.
Powerful prompt:[anonymized data + theme and count + mood + safe/representative quote + "generalize individual, do not disclose rare comment" + action suggestion]
The result: an analysis that is numerical, themed, action-oriented and maintains anonymity.
Anonymity and Fairness Rules
Risk
precaution
Small group disclosure
Separate reporting of sub-groups with fewer than 5 respondents
descriptive detail
Generalize name/unit/event information
Quoting rare comment
Use only representative, majority quotes
biased comment
Question AI's systematic negation of a group (age, gender, etc.)
Exporting raw data to insecure tool
Use certified/institutional tools and anonymous data
Attention: Survey data is personal data; Pasting raw comments into a generic tool without corporate data assurance carries KVKK risk (we will dive into this in the next unit). AI's theme inference is also fallible; A human should review the numbers and examples and make the final interpretation. AI finds patterns; The human determines the meaning and action.
Common mistakes
- Breaking anonymity. Putting the unit/name/specific detail into analysis can give a person away.
- Reporting in small groups. The result of a 3-person unit is almost personal data.
- Blindly trusting the raw number. AI's thematization is fallible; Check by sampling.
- Highlighting the rare striking comment. It can be revealing rather than representative.
- Not connecting the comment to the action. If the analysis does not produce concrete steps, it remains on the shelf.
- Overlooking bias. Question if AI systematically negates a particular group.
In summary
- AI groups hundreds of open-ended responses by themes, frequency, and mood to extract actionable insight for management.
- Maintain anonymity: generalize details that identify the individual, do not disclose small groups and rare comments.
- Put metrics like eNPS and engagement in context; Don't make up a definitive reason, ask for a hypothesis.
- Tie each important theme to a concrete action; Analysis does not create value if it does not produce steps.
- Process raw data with secure tool; AI finds patterns, humans do the final interpretation and bias checking.
Application task
Prepare 30-40 survey comments, real or fictional (intentionally include a "telling" detail or two). Anonymize the data first. Have the AI generate thematics, sentiment analysis, and comment count for each theme; Request safe/representative quotes. Then check the output: do any quotes give away a person? Is a sub-group with less than 5 comments reported separately? Finally, write a concrete action for the 3 most critical themes.
checklist
- [ ] Has the data been anonymized before analysis (name/unit/detail)?
- [ ] Are the themes reported numerically (how many comments)?
- [ ] Has the mood been specified for each theme?
- [ ] Aren't subgroups with less than 5 responses reported separately?
- [ ] Are the quotes representative and non-disclosing?
- [ ] Has AI's output been questioned for bias?
- [ ] Is each critical theme tied to a concrete action?