Two thousand responses and a deadline on Friday
As part of a culture diagnostic, your firm ran an employee survey for the client. The export from Qualtrics, SurveyMonkey or Microsoft Forms arrives with two thousand responses, forty questions and three open text boxes. The analysts clean it, recode scales, build pivots by function, level and location, remove groups too small to report, read every comment and tag it, then build chart after chart for the deck.
The analysis is repeated almost identically on the next engagement, from scratch.
Why survey analysis is so manual
- Each survey tool exports differently, and the cleaning steps are redone every time.
- Standard cuts are rebuilt as new pivots on each engagement.
- Small-group suppression, to protect anonymity, is applied by hand and easy to miss.
- Free-text comments are read and coded one by one.
- Charts are made in Excel and pasted into PowerPoint.
Survey tools are part of the reason. Each has its own export layout, and a question added or reordered changes the columns. Analysts write formulas that break silently when the structure shifts, and errors are found when a chart looks odd rather than when the data arrives.
What that costs
| Issue | Effect |
|---|---|
| Analyst days per survey | Budget spent on mechanics, not insight |
| Missed suppression | Individual respondents potentially identifiable |
| Inconsistent coding | Themes depend on who read the comments |
| Late results | Less time to discuss findings with the client |
Survey respondents were promised anonymity. Protecting it is not optional, and doing it by hand is where mistakes happen.
The repetition is the real waste. The cleaning steps, cuts and charts for an engagement survey are nearly identical from one client to the next, yet each time they are rebuilt by whichever analyst is free, with small differences that make results harder to compare across clients.
How we build survey analysis
- Exports from your survey tools are read and cleaned by a set of steps agreed once and reused: recoding, missing values, scale direction.
- Your standard cuts, by function, level, location, tenure, are produced automatically, with groups below your threshold suppressed.
- Comparisons with benchmarks or previous waves are included where you have them.
- A language model codes free-text comments against your theme framework and suggests new themes, and analysts review a sample and any low-confidence codes.
- Representative comments are selected with identifying details removed, for analysts to check.
- Charts are generated in your PowerPoint template, ready for the team to add commentary.
- The whole analysis can be rerun if the client sends more responses or a correction.
We usually build the pipeline from your most common survey type first, using a past dataset, so the team can compare the output with what they produced by hand. Once the standard cuts match, the pipeline is used on live work.
Survey week, afterwards
The analysts receive a clean dataset, standard cuts, draft comment themes and charts, and spend their time on what the results mean for the client.
Anonymity rules are applied every time, by design, rather than depending on someone remembering.
Is this your survey process?
- Every survey export is cleaned by hand.
- Pivots are rebuilt on each engagement.
- Small-group suppression is manual.
- Comments are coded one by one.
- Charts are pasted into slides individually.