AI Follow-Ups generate a follow-up question in real time, based on what a participant wrote or selected. Unlike a question you write in advance, the follow-up is produced by an AI in the moment and is personalized to the individual response. Its purpose is to surface the "why" behind an answer without you having to anticipate it. You can enable AI Follow-Ups on rating questions and free-text questions, and the response is always qualitative (free text).
AI Follow-Ups and Rating Branching both collect follow-up insight, but they suit different situations. For help choosing between them, and for guidance on using both in the same survey, see People Science Guidance for Using Rating Branching.
The Science Behind AI Follow-Ups
Employee listening relies on two kinds of data working together. Your rating scores are quantitative: they tell you what is happening, and at what scale. Your comments and open text are qualitative: they tell you why. Neither is more valuable than the other, and strong insight comes from using them together. Rich qualitative feedback, of the kind a skilled interviewer draws out by probing and clarifying, has historically been hard to gather at scale because it depends on people's time. AI Follow-Ups applies this approach at the moment someone responds, which lets you pursue the "why" across a whole survey - rather than in a handful of interviews.
Four core principles informed the design of AI Follow-Ups: neutral, non-leading questions, psychological safety, transparency with user control, and manageable cognitive load.
1. Neutral, non-leading questions
Because the follow-up writes itself, its neutrality is deliberately built into the product rather than something you have to author question by question. AI Follow-Ups are designed to stay neutral and exploratory: they clarify what a person said, then ask for examples or detail, rather than implying a conclusion or amplifying a sentiment.
Two questioning methods shape this approach. Socratic questioning clarifies a response and then asks for the reasoning or evidence behind it, moving a general comment toward something specific. Appreciative inquiry invites people toward what would improve a situation, which encourages forward-looking, actionable feedback rather than dwelling on the problem. Together they keep the follow-up question from assigning blame, or presupposing an answer.
2. Psychological safety
Psychological safety is the sense that an employee can speak up honestly without fear of negative consequences. AI Follow-Ups are designed to support it in a few ways. The questions stay neutral and non-judgmental. Responses are always optional, so no one is pressed to say more than they want to. And for some people, a neutral AI feels less exposing to open up to than a manager or HR, which can make honest feedback easier to give.
If a conversation moves into more sensitive territory, or the participant feels comfortable to share specific details such as names or particular incidents that could make them identifiable, guardrails are in place. In these cases, AI Follow-Ups are designed to detect a sensitive topic or a potential identifier, and to stop the conversation rather than probing further. Guardrails do not replace your own review and redaction processes, and as with any comment, some sensitive content may still come through.
Alongside this, follow-up conversations carry the same confidentiality protections as any other survey response, with the same comment redaction process available if a name needs to be removed. Making these protections clear to participants up front reinforces the safety that prompts open feedback in the first place.
3. Transparency and user control
Participants should know they are engaging with AI, and always stay in control of the interaction. The follow-up is clearly identified as AI-generated in the survey, and is labeled as such in reporting. Responses are always optional, even when the parent question is required, a participant can stop at any point. Telling participants up front how many follow-ups they may receive supports both principles: it is honest about what to expect, and it leaves the participant in control of how far the conversation goes.
|
4. Manageable cognitive load
Cognitive load is the mental effort a task demands. A survey draws on limited working memory, and every question spends some of it. A survey that asks too much of an individual can produces rushed answers, or experience a drop-off.
AI Follow-Ups add to this cognitive load: a single question can lead to more than one follow-up in sequence, and that adds up for the participant. The design keeps this in check by capping how many follow-ups a question can ask, and by keeping each response optional.
When to Use AI Follow-Ups
AI Follow-Ups work best when you want the "why" behind a response but cannot predict it, because it varies from person to person. They are a good option when:
The reasons behind a response are likely to be broad. On a topic like work-life balance or psychological safety, no single pre-written follow-up would serve everyone.
A question scored low or had high variance in your last survey, and the root cause is unclear.
A question is a strong driver of a key outcome such as engagement or intent to stay. Understanding the "why" here has the highest return.
A free-text question is broad or high-level (for example, "Anything else you'd like to add?"), where a richer response would be more useful.
You want to understand what is working, not just what isn't. Enabling follow-ups on favorable responses surfaces concrete examples worth replicating.
They are likely not the best choice when:
You already know the specific probe you want to ask. In this case, another tool such as Rating Branching is a better choice, as you can control the follow up question you want to ask.
You need to compare responses across cohorts, or over time. Follow-up questions here adapt to each person, so the data is not directly comparable.
The follow-up needs to be a rating, select, or multi-select question. AI Follow-Ups only produce qualitative output.
The topic is highly sensitive and may surface content you cannot adequately act on.
Tip: AI Follow-Ups work best on top of a well-designed parent question and a survey of reasonable length, adding depth where the fundamentals are already sound. Keep them to subjective questions (i.e. the questions that ask how someone thinks or feels). Follow-ups aren't available on self-select demographic questions, because there is no subjective "why" to explore: the answer is a fact about the person. Probing one would also press for detail on sensitive personal characteristics, and because demographic fields are used to segment reports, that detail carries a higher risk of identifying the individual.
Designing Your AI Follow-Ups
1. Choose how many questions to enable.
We recommend follow-ups on no more than 3 questions per survey, or roughly 5% of your total. This mirrors our survey design guidance to use no more than three free-text questions, since each exchange asks for similar effort.
If you are also using Rating Branching, count both together: what matters is the participant's overall experience, not each feature in isolation.
2. Choose how many follow-ups per question.
You can set up to two follow-ups. We choose two because actionability rises sharply from one follow-up to two, then gains little from a third (while fatigue increases).
3. Choose which responses trigger a follow-up (rating questions).
On a rating question, you decide which responses prompt a follow-up:
Unfavorable responses (for example, Strongly Disagree or Disagree) when the "why" behind a low score is what you most want to understand. This produces data reflecting only that subgroup.
Favorable responses when understanding a strength is a priority, such as recognition or manager support. This captures what "good" looks like so it can be repeated, though it tends to take more effort to draw out detail, since people tend to comment more readily when they disagree than when they agree.
All responses - when you want the broadest signal. This is useful in a few cases: after a deliberate change, when you want to hear the full range of reactions. On topics where sentiment is genuinely mixed. On wellbeing questions, where a neutral answer often carries the most useful detail. And on the first run of a new question, when you don't yet know which responses will tell you the most.
Reading and Interpreting Your Results
Follow-up conversations appear in your comments report alongside regular comments. They are at their most useful next to your quantitative data, where the scores tell you what is happening and the conversations tell you why.
We recommend the following flow:
1. Start with your scores. Use driver analysis to find the questions with the biggest impact on your outcomes, and decide which ones warrant a closer look.
2. Read the conversations for those questions. Use the follow-up themes to understand what is behind a score: the specific barriers, examples, and context a number alone can't give you.
3. Cross-reference with your filters, carefully. Where the group is large enough to meet reporting thresholds, look at how themes vary across demographic groups. Remember that only some people who saw a follow-up chose to answer it, so a difference between groups is a lead to look into, not a conclusion on its own.
4. Use what you learn to focus action, not to prove it. Let the themes point you toward where to act and what to prioritize, alongside your quantitative evidence rather than in place of it.
Some important principles to consider when you are reviewing your results:
AI Follow-Ups are analyzed as a full conversation, not as individual messages.
The AI's questions and the participant's answers are shown together, and sentiment and topic tags apply at the conversation level.
Only the participant's answers inform those tags, not the wording of the question.
On a rating question where question level comments are enabled for a survey, a participant can leave both a regular comment and a follow-up conversation. The two arrive together as a single entry in your comments report, with the conversation shown beneath the comment text and the purple Follow-up response tag on that entry.
On a free-text question, the participant's answer is the comment, and the conversation sits beneath it in the same way.
Note: AI Follow-Up conversations aren't included in AI Comment Summaries, which cover standard comments only. You'll still find every follow-up conversation in the comments report.
Importantly, because each participant receives a different question based on their own response, this data is qualitative and directional, not a measure of your whole population.
It works best as a layer beneath your quantitative scores, adding the themes, examples, and stories that bring a number to life, which is especially useful when presenting results to leaders.
It can tell you more about the recurring themes across conversations, examples that illustrate a score, which topics people most want to talk about, and new issues worth adding to a future survey.
What it can't tell you is how common something is.
A theme coming up often doesn't make it the majority view, as only the people who engaged with the AI Follow-up are represented.
A topic not coming up doesn't mean it doesn't matter, someone simply may not have raised it.
Additionally, the sentiment tags here aren't directly comparable to those on your standard comments. A standard comment is tagged on its own, as a single piece of text. An AI Follow-Up is tagged across the whole exchange, the participant's answers together as one conversation, so the tag is summarizing something different in each case.
Follow-up conversations can surface sensitive content, as any open-text comment can. The same care, judgment, and comment-redaction processes apply. |
FAQs
When should I use AI Follow-Ups rather than Rating Branching?
When should I use AI Follow-Ups rather than Rating Branching?
Use AI Follow-Ups when you want the "why" behind a response but can't predict it, because it varies person to person. Use Rating Branching when you already know the question you want to ask, when you need comparable data across groups or time, or when the follow-up needs to be a rating, select, or multi-select question. See People Science Guidance for Using Rating Branching for a fuller comparison.
Which question types can have AI Follow-Ups enabled?
Which question types can have AI Follow-Ups enabled?
Rating questions and free-text questions. The response is always qualitative. Follow-ups are not available on self-select demographic questions, which are factual and carry reporting and sensitivity implications, or on eNPS questions using the 11-point scale.
How many questions should I enable them on?
How many questions should I enable them on?
No more than 3 per survey, or roughly 5% of your questions, as a ceiling rather than a target. Each exchange asks for effort comparable to a free-text question, so enabling it too widely lengthens the survey and increases drop-off. If you also use Rating Branching, count both together.
How many follow-ups will each person get?
How many follow-ups will each person get?
Up to two per question during the Early Access Program; we recommend two. The value of a response peaks at two follow-ups, and a third adds little while increasing fatigue. Tell participants up front how many they may receive.
Could AI Follow-Ups change how people answer?
Could AI Follow-Ups change how people answer?
They can. Some participants may keep their initial response vaguer if they expect a follow-up. Keeping follow-ups to a few well-chosen questions, and being clear about how they work, helps.
Can I treat these responses as representative of everyone?
Can I treat these responses as representative of everyone?
No. Only people who engaged with the follow-up are represented, so the data reflects that group rather than your whole population. Read it as qualitative and directional: strong for themes and examples, but not for measuring how common an issue is.
The AI asked about something I didn't expect. Is that a problem?
The AI asked about something I didn't expect. Is that a problem?
Follow-ups respond to what the participant wrote rather than a script, so the questions vary. Review the full conversation for context. Guardrails are designed to catch sensitive content and stop the conversation, but as with any comment, some sensitive content may still come through. Use the same review and redaction process you would for any comment.
Why do some comments show AI questions in them?
Why do some comments show AI questions in them?
Follow-up conversations display the AI's questions alongside the participant's answers so you can read them in context. Only the participant's answers inform sentiment and topic tags.
💬 Need help? Open a Support conversation to speak with our Product Support team.