Start with a decision to investigate
Your leadership team asks, "If everyone’s using AI, why aren’t we seeing better work?" Interpreting survey results should help you choose what to investigate before you prescribe more training.
We think comparing dimension averages is a useful starting point, but a poor stopping point. Inspect who answered and how their answers fit together, then select one follow-up investigation.
The goal is a bounded interpretation and an evidence collection plan. This workflow is editorial advice supported by official survey guidance, not a validated diagnostic algorithm. Assemble these prerequisites before starting; they describe information you need, not promised product capabilities.
Interpreting survey results starts with people
Consider this interpreting survey results example: 150 account executives and 150 sales development representatives receive an email invitation and one reminder over two weeks. Participation is voluntary. All counts and categories below are hypothetical editorial illustrations, not validated questions or customer results.
Suppose 120 people respond. For Usage, 90 report frequent use and 30 report less frequent use. For Skills, 80 report lower skills and 40 report higher skills. For perceived Impact, 70 report little benefit, 40 report greater benefit and 10 leave the item unanswered. Keep those 10 missing answers separate from the 110 impact answers.
Now pair answers from the same people. Of 120 respondents answering both Usage and Skills, 65 report frequent use and lower skills. Of 110 answering both Usage and Impact, 60 report frequent use and little benefit. The overlap supports investigating those respondents’ tasks; separate averages alone wouldn’t show whether the same people reported both patterns.
Check composition before extending that interpretation. Account executives contribute 90 of 120 responses but comprise 150 of 300 employees; sales development representatives contribute only 30 responses despite making up half the workforce. Within those groups, frequent use with lower skills appears among 50 of 90 account executives and 15 of 30 sales development representatives. The pattern appears in both respondent groups, but participation is uneven.
AAPOR’s Best Practices, in its April 2024 version, recommends comparing sample composition with the population and describes cross-tabulation. Apply that guidance through the following suggested steps.
- Count each answer category. Keep missing answers separate and show the denominator.
- Pair Usage with Skills and perceived Impact for the same respondents, where available; repeat within relevant roles or departments.
- Compare respondent composition with workforce composition. Record which groups are missing or overrepresented.
- Identify uncertainty from recruitment, missing responses and answer accuracy separately.
- Choose one investigation using the decision table below.
When interpreting survey results across rounds, compare question wording and respondent composition before treating score movement as a change in practice.
Sources: Best Practices for Survey Research - AAPOR
Separate precision from missing voices
Understanding survey results means asking different questions about different errors. AAPOR’s Disclosure Standards, modified August 26, 2026, require defined precision measures and detailed model documentation for non-probability surveys. Don’t attach a conventional random-sample margin of error to voluntary employee responses without a defensible model.
AAPOR’s Standard Definitions, modified December 30, 2025, says response rates alone cannot establish whether nonresponse error exists or its size. ONS guidance, retrieved October 19, 2026, explains that inaccurate answers and other non-sampling errors can remain even when everyone is surveyed.
AI Adoption Score’s assessment documentation, retrieved October 19, 2026, describes its results as self-reported: they reflect perceived practice, not demonstrated capability or causal productivity effects. Its checks catch some overstatement, not all.
In our view, a precise-looking number shouldn’t outrank an unanswered question about who participated.
Sources: 360° AI Maturity Assessment: how it works · AI Adoption Score, Disclosure Standards - AAPOR, Standard Definitions - AAPOR, Uncertainty and how we measure it for our surveys - Office for National Statistics
Choose the check your pattern warrants
Use this editorial decision table to select the first investigation. Each row proposes a way to test an explanation; none diagnoses a cause from the scores alone.
| Observed pattern | First check | Suggested owner | Evidence to collect |
|---|---|---|---|
| Respondent mix differs from workforce | Review recruitment coverage | People analytics lead | Invitations and role counts |
| Same people report frequent use and low skills | Observe a bounded AI task | L&D lead | Work samples and task observation |
| Frequent use with low perceived impact | Trace one workflow | Operations manager | Outputs, review effort and rework |
| Opposing patterns across departments | Review groups separately | Head of AI | Paired answers and denominators |
Leave with one owner and one question
For the hypothetical sales team, start with a recruitment review because only 30 of 150 sales development representatives responded, compared with 90 of 150 account executives. Assign the People analytics lead to check invitation delivery and ask nonparticipants about access and participation obstacles. Suggested follow-up questions are editorial prompts, not validated instruments.
The paired answers still justify a later proposal-drafting observation with the L&D lead. First establish whether recruitment problems left particular employees unheard. Even if delivery was complete, you still wouldn’t know how nonrespondents would answer or whether question interpretation and task difficulty explain the reported pattern.
We’d finish with a named investigation, not a department ranking. Record the pattern, affected respondents, competing explanations, owner and evidence that would change your interpretation.
If groups are small, display answer counts and denominators: AI Adoption Score’s documentation warns that individual responses can move small-group averages substantially. If workforce composition is unavailable, record representativeness as unknown and obtain a comparison; if ambiguity persists, use focused interviews and task observation.
Your expected result is a clearer question and a collection plan. Choose the next check.
Sources: 360° AI Maturity Assessment: how it works · AI Adoption Score
Frequently asked questions
Who should coordinate the review?
Assign one coordinator who can bring the survey owner and relevant operational leads together. As editorial guidance, the Head of AI could hold the decision record while People analytics checks composition and a workflow owner arranges observation. Make access to individual answers a deliberate responsibility, rather than circulating them broadly.
Should we wait for another survey round?
You don’t need another round before investigating a specific ambiguity. Our editorial recommendation is to collect the missing evidence first, then decide whether repeating the survey would answer a different question. Set the repeat date around the intervention and the work cycle, rather than an arbitrary desire for another score.
Should we rewrite unclear questions immediately?
Clarify them for future rounds, but preserve the original wording alongside existing responses. As editorial guidance, ask a few respondents how they understood the item before changing it. Document any revision so a later comparison doesn’t quietly treat answers to different questions as movement on the same measure.
