# How to run engagement surveys with AI and turn them into changes > Guide to climate surveys with AI: questions that inform, open-answer analysis with real anonymity and the close that prevents cynicism. - Canonical: https://serchai.com/en/guides/ai-engagement-surveys/ - Site: Serchai (https://serchai.com) — AI tools comparator - Language: en - Updated: 2026-07-26 --- ## Tools you will use - [Claude](https://serchai.com/en/reviews/claude/) — Anthropic's AI assistant for writing, analysing and thinking through documents. - [NotebookLM](https://serchai.com/en/reviews/notebooklm/) — The research notebook that answers only from your own documents, with the citation attached. - [Quizgecko](https://serchai.com/en/reviews/quizgecko/) — Turns any material into quizzes, flashcards and study notes. ## The steps, in short 1. **Design a short survey people can answer honestly** — Few questions, credible anonymity and the open space where the real information lives. 2. **Analyze the open answers without losing the nuance** — AI groups hundreds of comments into patterns, with quotes as evidence. 3. **Verify the patterns before drawing conclusions** — Detected themes get checked against the real answers, not accepted at face value. 4. **Close the loop: communicate what changes and what does not** — A survey without a visible response teaches staff not to answer the next one. > **TLDR:** The climate survey dies two ways: nobody answers honestly (dubious anonymity) or nobody does anything with the answers. AI fixes the part that was unviable by hand, analyzing hundreds of open answers into patterns with quotes, and the rest is design and courage: a short survey with real anonymity, pattern verification and the public close of what changes and what does not. The survey that produces no visible change produces cynicism. This guide is for anyone measuring climate who wants it to matter: HR, management, leads of large teams. The new tool is not the form (that existed) but the analysis: the open answers, where the real information lives, are no longer unreadable at volume. A note on limits: surveys detect collective patterns. Serious individual situations (harassment, health) need their own confidential channels, not a quarterly survey. ## 1. Design a short survey people can answer honestly The long survey and honesty are enemies: by question forty, people score on autopilot. The design that informs: eight to twelve closed questions on what you genuinely want to track (load, recognition, communication, trust in leadership), two or three open ones where the substance lives ("what would you change first?", "what would make you leave?") and the same questions every quarter, because the trend informs more than the snapshot. [Claude](https://serchai.com/en/reviews/claude/) helps phrase without bias: ask it to flag which questions are leading and why before you accept them. From $17 a month, with a permanent free tier. To build the questionnaire, [Quizgecko](https://serchai.com/en/reviews/quizgecko/) generates forms from your list, though any form tool works: that is not the critical piece. The critical piece is credible anonymity: staff answer honestly when they believe it is genuinely anonymous. That demands a minimum response threshold per segment (no breakdowns of the three-person team), asking for no identifying data and stating explicitly how it is protected. ## 2. Analyze the open answers without losing the nuance Here is the real change: two hundred open answers used to be unreadable and are now an afternoon. The flow: paste each question's answers (stripped of identifying metadata) and request the pattern analysis: which themes repeat, at roughly what frequency, with the representative literal quotes for each. [NotebookLM](https://serchai.com/en/reviews/notebooklm/) (from $4.99 a month, with a free tier) is this particular task’s tool rather than a chat assistant: upload the anonymized batch as a source and every pattern it returns opens the verbatim quote holding it up, which is exactly the verification the next step demands. The instruction that improves results: also ask for the minority-but-intense, because the theme five people mention with anger matters differently from the one twenty mention in passing. The analysis's privacy rule: answers travel to the tool without names or small-team breakdowns, and your company's data policies govern which tool at all. ## 3. Verify the patterns before drawing conclusions The generated analysis is a draft of conclusions, not the conclusions. The short verification: for each detected pattern, read the quotes supporting it (do they really say that?), check the counter-pattern (are there answers saying the opposite, and how many?) and separate facts from interpretations ("communication is bad" being a pattern does not yet say what the concrete problem is). Crossing with the closed questions completes the analysis: the open-answer pattern that matches a low closed-question score is solid signal. The one appearing on only one side asks for more looking before anything moves. From that verification comes the honest document: three to five findings with their evidence, not the forty-page report nobody reads. ## 4. Close the loop: communicate what changes and what does not The part deciding whether the next survey matters: the public close. The structure that works: this is what you said (the findings, honestly), this is what we will change (two or three concrete commitments with dates, not ten vague ones), and this is what will not change and why (the brave part that builds credibility). That third block is the one almost nobody does and the one that generates the most trust: staff do not expect everything to change, they expect no pretending. The [internal communications](https://serchai.com/en/guides/ai-internal-comms/) guide covers that announcement's how. And the loop closes at the next survey: what happened to the previous commitments is the implicit first question, and answering it before it is asked is the difference between a system and a ritual. The whole sector lives in [AI for human resources](https://serchai.com/en/ai-for/human-resources/). ## Common mistakes A long survey with dubious anonymity. It produces autopilot data and zero honesty: exactly the opposite of the goal. Stopping at the closed questions. The 7.2 average does not say what to do. The open ones say what to do, and now they can be read. Accepting the AI's analysis unverified. Generated patterns sometimes inflate the striking and miss the intense minority. The quotes are the verification. Surveying without closing the loop. A survey followed by silence teaches that answering is useless, and the next participation rate confirms it. Better not to ask than to ask and go quiet. ## Frequently asked questions ### How often should we survey? Quarterly is the usual balance: enough for trend, not enough for fatigue. The non-negotiable is closing each loop before launching the next. ### Is it safe to analyze answers with external AI? With the data precautions: answers without identifiers, no small-team breakdowns and your company's data policy deciding the tool. For strict environments, analysis with a local model is the alternative. ### What participation rate is good? Above 70% the picture is representative. Low participation is itself climate data: people do not answer when they doubt it matters or distrust the anonymity. ### Can AI detect serious problems in the answers? It can flag mentions deserving attention, and its role ends there: serious individual situations demand their own confidential channels and direct human handling, never the survey's standard flow.