Find insights from survey responses
Turn survey responses into clear themes and actionable findings.
Steps
1. Describe the task
Sifting open-text survey responses for the themes that matter is slow work, and AI can be a good thinking partner for it - spotting themes, summarising what respondents are telling you, and surfacing patterns and surprises. You’ll still apply your own judgement and decide what to act on: the goal of using AI for this task is to create a strong first pass, not a finished analysis.
Be clear about who you are, the role you want it to play, what you’re looking for, and who the findings are for.
Copy and paste this prompt into the chat box, then send it as is - there's nothing to edit first. The AI will respond with questions rather than a draft. That's by design as it needs context before it can produce something useful. Answer as if you're briefing a colleague, and if you're unsure of an answer, just say so - working it out together is part of the process.
Using Worthwhile Chat?
This prompt is ready-made: just select it in your chat box and it will self-populate.
Act as a research analyst with experience in the charity sector.
Read through the survey responses I'll share with you and identify the key themes - what's coming through strongly, what's mixed and anything unexpected. Produce a summary with the main themes, how strongly each one comes through, and a few representative quotes from respondents. Use clear, accessible language.
Start by asking me questions about: my organisation, where we’re based and who we support; the specific service I want to focus on; our main goal for this service; and if I have any drafts or notes to upload. Then draft the improvement ideas.
As you draft, watch out for and avoid: changing or altering respondents' words when quoting them; and including personal or identifying details in the summary - flag these instead.
2. Provide context
As you answer the AI’s questions, you can upload documents or paste in information to help it understand your work more deeply. The more complete and relevant the data, the better the output will be.
Important context (add as much of this as you can)
- Open-text responses: These are where the richest insights tend to be. Free-text answers to questions like “What did you find most helpful?” or “Is there anything we could improve?”
- Ratings or scores: Satisfaction scores, Net Promoter Scores or any scaled responses (e.g. 1-5 ratings)
- Any context about the survey; When it was run, how many people responded, what programme or service it relates to, and what questions were asked
You don’t need perfect data. A CSV export from Google Forms, a spreadsheet of responses, or even copied-and-pasted text from a Word document all work fine. Share what you’ve got - it doesn’t need to be too tidy.
Before you upload files: read your data first
Before asking the AI to analyse your survey responses, read through them yourself. You don’t need to do a full analysis - but you should have a feel for what’s in there. This means you’ll be better placed to spot if the AI misses something important, overweights a minor theme or gets something wrong.
Check for sensitive data
Review your documents for personal or sensitive information about beneficiaries. Anonymise names and identifying details before uploading. Even though Worthwhile Chat is privacy-first, good data practice starts with you. If you're using other AI tools, check their data settings too - many train on your data by default. How much you share depends on your existing privacy and data policies, and how much control you have over the tools you use.
If you’re unsure, ask yourself: “If this document were seen by someone outside our organisation, could anyone be identified or harmed?” If yes, anonymise first.
This is an anonymised CSV export of 85 responses to our end-of-programme survey from March 2026. Please read through all the responses and then summarise the key themes.
3. Review the output
Review the initial summary carefully. AI can be confident and articulate even when it’s wrong, so your judgement matters here. Read through what it has produced and ask yourself:
Does it feel right? Does anything jump out as missing? Think of this as a working check, not a formal process. If something doesn’t feel right, that’s useful information - use it to refine in Step 4.
Review checklist
- Does this match what I know?
- Is it accurate?
- Is the sample size big enough to draw conclusions?
- Are the people we support described in a way they’d be comfortable reading about themselves?
4. Follow up prompts
Getting a useful output often takes more than one go. Think of this as a conversation - you’re directing the analysis, not just accepting the first output. The more specific your questions, the more useful the results.
Ask for revisions - see some examples below:
To explore themes further:
- “Tell me more about the responses related to [theme]. What are respondents actually saying?”
- “Are there any differences between people who rated the service highly and those who didn’t?”
To find what’s missing or surprising:
- “What’s not being said? Are there any gaps or topics I’d expect to see that aren’t coming through?”
- “Were there any responses that didn’t fit the main themes - any outliers?”
To make it actionable:
- “Based on these themes, what are the two or three things we should prioritise improving?”
- “If I had to present three key findings to my programme team in two minutes, what should they be?”
Representation:
- “Check the quotes you’ve pulled represent the range of respondents, not just the most articulate or most critical voices.”
- “Make sure no single group’s experience is being presented as everyone’s - flag where a theme only reflects part of who responded.”
If anything needs correcting or rebalancing, ask the AI to revise. For example: "The negative feedback is overrepresented - can you rebalance the summary to reflect the overall picture more fairly?”
And if a conversation starts to feel stuck, you can always start a fresh one with slightly different instructions - sometimes a clean start gets you closer to what you need.
Add what only you know
The AI works with what you give it - it can’t know everything about your context. Once you have a solid analysis, fill in what it can’t provide:
- What you’ve been hearing informally - patterns from sessions, conversations with staff, things that don’t always make it into a survey
- Any important context about who responded - and who didn’t
- Recent changes or decisions that would affect how the findings should be interpreted
5. Tips and good practice
Share with a colleague
Don’t treat AI-produced analysis as the final word. Share the findings with colleagues who know the programme or service - they’ll spot things you missed and challenge interpretations that don’t hold up. A short conversation with a colleague is worth more than another round of prompts.
Be honest about what the data tells you
Every dataset has limitations. When sharing findings, be upfront about:
- How many people responded (and - if calculable - how many didn’t)
- Whether the respondents are representative of the people you work with
- Whether the findings are themes worth exploring further, or evidence strong enough to act on
- Whether the specific figures or counts in the summary match your original data - check these before you rely on them
Be transparent about AI use
If you’re including these findings in a report, board paper or funding application, consider noting that AI tools were used to support the analysis. More organisations are being open about this, and it’s in line with our Worthwhile AI values.
Protect your respondents
Before sharing any output externally, double-check that no identifying information has slipped through - in quotes, in combinations of details or in how individuals have been described
Save time next time
Once you’ve been through this process:
- Save your prompts so you can reuse or adapt them for future surveys
- Save the structure that worked well for your summary
- Note which follow-up questions produced the most useful insights
- Compare findings over time - running the same analysis on successive surveys can reveal trends that a single snapshot misses
This will make future analysis faster and more consistent.
