How do we handle small groups?

deepsight works with fixed base limits: below a base of 10 responses, values in group comparison and dashboards are flagged or hidden, and significance indicators only appear from 30 responses per group. These limits can’t be adjusted. They are a statistical indicator for small base sizes, not a protection rule: quotes and exports remain readable. The minimum group size for your reports is set in your reporting rules, and you apply it before sharing.

Who can see responses and results?

Admins manage the users and roles of your organization. Roles apply to the whole organization, not per project: anyone with access sees all of the organization’s projects. If departments must not see each other, we set up separate organizations in the Enterprise plan. Support access to your content is only possible with time-limited approval from your admin. Sign-in via SSO (OpenID Connect, tested with Microsoft Entra ID) is included in the Enterprise plan; we set it up together with your IT team, and it can be made mandatory. You share individual results via a link, with a recipient list, password, expiry date and revocation. Guests have read-only access and don’t need an account of their own – so share deliberately.

How long are the comments stored?

For uploaded datasets, you set a retention period per project or for the whole organization. deepsight warns you by email before deletion, and every deletion is recorded in the deletion log. If you wish, only the original data is deleted; key figures, dashboards and reports remain, including the quotes in them. The automatic retention period does not yet apply to responses from surveys in deepsight cloud.

How do we analyze exit surveys with few responses?

Read the responses in full and use the topics to organize them, not as statistics. For comparisons, combine longer periods, such as six months instead of one. Shares then rest on a broader base, and conclusions about individuals become harder to draw.

Can we run the survey in deepsight too?

Yes, with surveys in deepsight cloud, which we activate for your organization as part of onboarding. You share the participation link via your intranet or email list; participants don’t need an account, and participation is currently available in German only. An AI follow-up question for brief answers is optional and can be skipped; it’s best to agree its use in employee surveys with your employee representatives. Important for that discussion: the analysis started directly from the survey and the AI follow-up question work without anonymization. You get masked results by also analyzing the survey dataset as a separate analysis with anonymization; it counts toward your text allowance again.

HR & People Analytics: employee feedback

Common hurdles

Why only three out of five for “would recommend”? The reasons are in the comments.

Engagement, pulse and exit surveys produce open-ended responses by the hundreds. Before team leads can work with them, it takes a lot of manual work and coordination.

The challenge: Comments nobody reads in full
The scale values are analyzed quickly. The comments go to HR as a long list or only get spot-checked.
With deepsight: All responses, not a sample
Topic analysis assigns the open-ended responses to topics – based on suggestions from deepsight or on your category system. You check the result against the original instead of sorting comment by comment.
Go to topic analysis
The challenge: Names in the comments
Employees mention managers, colleagues or details about themselves. Before comments are passed on, someone redacts them by hand.
With deepsight: Details masked before the analyses run
Our own anonymization works without a language model. The “names” category is always on; you choose others such as email addresses or phone numbers. If translation is switched on, texts are translated before masking. Not every detail is detected – check results before sharing them.
Go to anonymization
The challenge: Every wave coded from scratch
Annual survey or pulse: the categories are rebuilt every time. Whether a topic is mentioned less often or was just coded differently is hard to tell.
With deepsight: Your category system in every wave
Apply your own codebook, including parent and child categories, for example along the lines of your survey model. You save adjustments as a new topic version.
Apply your codebook
The challenge: Open questions before the start
The works council and data protection want to know where responses are processed and what the AI does with them. Without documentation, the approval process drags on.
With deepsight: Documentation you can put on the table
deepsight cloud runs on Microsoft Azure in the EU, including the AI features via Azure OpenAI; Microsoft does not use your data to train AI models. Translation via DeepL or Google is optional; the list of service providers shows who is involved. It is available on request, together with the DPA and TOMs.
Go to documentation

Example: employee survey

A rating and the response behind it

Open-ended response · What should we change?

“My team is great, but decisions from the top reach us late and without any explanation.”

Would recommend: 3 of 5
Department: Production

Select a category to see the matching passage.

In the analysis

You can look at topics from open-ended responses alongside scale values, for example by department or location. Which breakdowns make sense and are permissible should be settled before the survey.

Multiple mentions: One response can be assigned to several topics – here teamwork (positive) and leadership communication (negative).

Illustrative example.

Before the analysis

What you decide in advance

HR, data protection and employee representatives decide these points together. We discuss with you which of them deepsight covers and which you set down in your reporting rules.

Recipients
Who receives which results: management, HR, team leads, teams?
Minimum group size
From how many responses do you report results for a group?
Wording
Which responses may be passed on verbatim, and which only in summarized form?
Involvement
How are the works council or staff council involved?
Retention
How long are raw data and results stored?

Security & data protectionHosting, AI processing, access and deletion.View security MethodologyHow to check assignments against the original responses.View methodology

Research questions

What your employee survey can tell you

Which topics do employees with a low “would recommend” score mention, and which do those with a high score mention?
Which responses lie behind a notable score in a department or at a location?
Which topics are new since the last survey, and which are mentioned less often?
What reasons do departing employees give for leaving?
What concerns new employees in their first months?
How do employees comment on a change, such as new working-time rules or new software?
Which suggestions for improvement come up again and again across departments?

Good to know: You can also ask the assistant questions like these directly. Depending on the question, its answers include charts and original responses as evidence.

Data and sources

What you start with

Usually the Excel or CSV export from your survey tool: one row per respondent, with open-ended responses, scale values and attributes in separate columns.

Employee surveysScale questions on topics such as satisfaction, leadership or recommendation, open-ended questions like “What should we change?” and attributes such as department or location.

Pulse surveysShort, recurring surveys with a few open-ended questions, often monthly or quarterly.

Exit and onboarding surveysReasons for leaving or impressions from the first weeks, often collected on a rolling basis with few responses per month.

Responses in other languagesFor workforces in several countries: integrated translation (AI-based, or DeepL or Google if you prefer) supports the analysis. If translation is switched on, texts are translated before masking.Go to translation

Process

From export to report for team leads

Prepare the data
Upload the file and add the research question and internal terms, such as department names or abbreviations, to the study context. Filter out empty or content-free responses like “n/a” up front.
Data quality check
Check the masking
Read a sample of masked responses, especially for indirect references such as job role and location together. Anything the detection missed, you mask manually – versioned and reversible.
Anonymization
Check topics against the original
Read the responses for each topic verbatim, correct assignments, and rename, merge or split topics.
Topic refinement
Compare groups
Put topics and sentiment side by side by department, location or rating level, with significance indicators if you wish.
Group comparison
Check before sharing
Check quotes, filters and small groups against your reporting rules, then share the results as a dashboard, report or PowerPoint deck.
Dashboards

Results

What HR and team leads receive

Topics by department, with evidenceShare of analyzed responses and sentiment for each topic, by department or location side by side and with the verbatim responses; multiple mentions possible; anything the data quality check filters out does not count toward the base.

Results for each group of recipientsFor management, HR or individual departments, you put findings together as a dashboard or report. You share reports via link; recipients don’t need their own account.Go to reporting

Trends across surveysYou add follow-up waves and pulse surveys to a time series and match topics across the waves.

Data for your HR reportingResponses with their topics as Excel or CSV, charts and quotes in PowerPoint.

Good to know: Wave comparison is included in every paid plan – see the plan comparison. If your result tables already exist in Excel, the slide factory turns them into slides based on your template, for committee meetings, for example.

Methodology

What matters with employee feedback

Choose attributes sparingly

Upload only the attributes you need for comparisons: every additional attribute splits the workforce into smaller groups.

Choose quotes deliberately

Wording and details can point to a person even when names are masked. Select quotes for reports carefully, or summarize what was said.

Put differences in context

Significance indicators show whether a department differs noticeably from the rest of the respondents in statistical terms. They don’t show why – but they do show where a conversation with the team is worthwhile.

Rare but serious comments

A topic with few mentions can matter more than the most frequent one. Read such responses in full, and decide in advance how you will handle reports of serious incidents, such as harassment.

FAQ

Questions about analyzing employee surveys

Related to HR & people analytics

Anonymizing text data: methods, examples and limitsWhy removing names isn’t always enough – with a review process for free text.Read the article Enterprise requirementsSSO, users and roles, separate organizations, and documentation for your review.Go to enterprise requirements Custom projectsWe assess a connection to your HR system or survey tool as a custom project.Go to custom projects

Discuss your employee survey

We will work through the analysis, recipients and data protection for your project with you.

Request a demo Get in touch

Solutions · HR & People Analytics Understand employee feedback better.