Keep the survey short (10 to 15 questions, mostly 1-to-5 scales plus two or three open ones), switch off name collection, and ask every leaver the same exit questions. Remove names and identifying details, then have AI sort comments into themes with counts and quotes. Check a sample of its coding yourself, and report only groups of five or more.
Two things go wrong in small businesses, and neither is about the AI. First, anonymity breaks easily in a team of twenty: "customer service, joined this year" may describe one person, and staff know it. Second, open comments are where the useful information is, and they are also where an AI summary can quietly invent a theme or smooth an angry remark into a mild one. Protect anonymity in the design, and check the AI's coding against the raw comments, and the rest is quick.
A survey short enough that people finish it
Every question should link to something you could change. If you would not act on the answer, cut the question. Here is a 12-question set that works for most small teams. Scale questions use 1 (strongly disagree) to 5 (strongly agree).
- I know what is expected of me in my role.
- I have the tools and equipment I need to do my job well.
- I get enough notice of my shifts and any changes to them.
- My workload is manageable most weeks.
- I was trained properly for the tasks I do.
- My manager gives me useful feedback.
- I feel able to raise a problem without it being held against me.
- I think my pay is fair for the work I do.
- I can see how I could progress here.
- I would recommend this as a place to work to a friend. (Or use a 0 to 10 scale for this one.)
- Open: What is one thing we should keep doing?
- Open: What is one thing that would make your job better?
Add a third open question only if you have a specific topic, such as "What made last November's peak harder or easier than it needed to be?" Specific questions get specific answers. Keep it under ten minutes to complete, and say so in the invitation.
Keeping answers anonymous in a team of twenty
Anonymity is a promise you have to design in, not just state.
- Do not collect names or emails. In Google Forms, the Collect email addresses setting under Responses should be on Do not collect. In Microsoft Forms, the Record name option is only offered when the form is limited to people in your organisation; make sure it is unticked. Test by submitting a response yourself and checking what the results show.
- Ask about team only if every group has five or more people. Merge small teams into broader groups ("warehouse and packing", "everyone else") and use wide tenure bands ("under two years", "two years or more"), or leave both out.
- Warn about free text. Say in the survey: "Please avoid names or details that identify you or others." People will still write them, which is why the redaction step below matters.
- Decide who sees raw comments. Ideally one person outside line management, who prepares the anonymised version everyone else sees.
- Never use survey answers to judge an individual. Say so in the invitation, and mean it. One broken promise ends honest surveys.
When to send it, and what the invitation should say
Timing changes answers. Avoid the busiest fortnight of the year (for an online clothing shop, the run-up to the big sale weekends) and the week after a difficult announcement. Give people ten days, on work time, with one reminder. For staff who do not sit at a computer, a QR code on the break-room noticeboard and ten minutes set aside in a shift make the difference between half the warehouse answering and all of it.
An invitation you can adapt:
Subject: 10 minutes, anonymous: how's work going?
We'd like your honest view on how things are working here: rotas, training, workload,
pay and what we should keep doing. It's 12 questions and takes about 10 minutes.
It's anonymous. The form doesn't record names or emails, and we'll only report results
for groups of five or more. Please don't put names in the comment boxes.
[Name], who doesn't manage anyone, will see the raw comments and remove anything
that could identify someone before the rest of us see them.
Nothing you say will be used to judge anyone individually. We'll share what we heard
and what we're changing by [date]. The survey closes on [date].
[Link or QR code]Exit interviews that produce comparable answers
An exit interview is useful when you compare it with others. That only works if the questions are the same every time. A script for a 20-minute conversation, or a written form for people who prefer it:
1. What made you start looking for another job, or say yes to one?
2. When did you first think about leaving? What was happening then?
3. What did you enjoy most about working here?
4. What would you change about your day-to-day work?
5. How well did your manager support you? What would have helped?
6. Did you get the training you needed? Where were the gaps?
7. How does the new role compare on pay, hours and flexibility?
8. Is there anything that would have kept you here?
9. Would you consider coming back in future? Why or why not?
10. Anything else we should know?Practical points. Hold it in the last week or, for more candour, send the written form a few weeks after someone has left. Use someone outside the leaver's reporting line where you can. Write notes in the leaver's words rather than your summary of them. If you record the conversation to transcribe it, ask permission first and tell them how long the recording is kept. And log every interview in one sheet (date, role group, length of service, notes), so patterns can be found later.
Preparing comments for AI: strip out what identifies people
Before any comment goes into an AI tool, remove names, specific dates, rare roles and anything that points at one person. Do it in a copy of the data. A before-and-after, illustrative:
Before:
"Since Jo moved me to returns in March I've been doing two jobs, and when I told
Dan about my back he said everyone's tired in peak."
After:
"Since [MANAGER] moved me to [OTHER TASK] [DATE] I've been doing two jobs, and when
I told [MANAGER] about [HEALTH ISSUE] they said everyone's tired in peak."The meaning survives; the people do not. Health, family and other sensitive details deserve extra care: replace them with a neutral tag even if nobody could be identified. Use a business plan that does not train on your content (or switch the model-training setting off), and delete the uploaded files from the assistant afterwards. Redacting personal data from documents with AI covers faster ways to do this for larger batches; for a team of twenty, doing it by hand takes half an hour and is the safer choice.
Coding comments into themes: prompt, output and a check
"Coding" here means tagging each comment with one or more themes, so you can count how often each theme comes up. Doing it in two passes gives far better results than asking for "the main themes" in one go.
Pass one: draft a codebook
Below are 38 anonymised staff comments from a small online clothing business
(answers to "one thing to keep doing" and "one thing that would make your job better").
Propose a codebook of 6-10 themes. For each theme: a short name, a one-sentence
definition, and what does NOT belong in it. Do not count anything yet.Edit the result yourself. Merge overlapping themes, split vague ones ("communication" usually hides two or three different problems), and add any theme you know matters that the model missed.
Pass two: code every comment with the agreed codebook
Using ONLY the codebook below, tag each comment with up to 2 themes.
Output a table: comment number, themes, and the exact words that justify each tag.
Then give a count per theme and 2 example quotes per theme, copied word for word.
If a comment fits no theme, tag it OTHER. Do not paraphrase quotes.
[codebook]
[numbered comments]An illustrative extract of the output:
Theme counts (38 comments):
Shift notice and rota changes 11
Peak-season workload 9
Team and colleagues (positive) 9
Training for new tasks 6
Pay 5
Equipment and warehouse layout 4
OTHER 3
Shift notice, example quotes:
"we find out on Thursday what we're doing on Monday"
"rota changes by text the night before are the worst part"Then check it
Take 20 comments at random and code them yourself without looking at the AI's tags. Compare. If you agree on 16 or more, the coding is good enough to use. If not, tighten the codebook definitions and rerun. Also search the raw comments for each quoted phrase: in this illustration, one "quote" under Pay read "the pay doesn't reflect the heavy lifting", while the original said "honestly the pay doesn't reflect how heavy the lifting is in peak". A paraphrase presented as a quote is a small thing that makes staff distrust the whole report when they spot it.
Scores: what to calculate for a small team
With 20 responses, one person moves any percentage by 5 points. So keep the numbers simple and honest.
- Percentage favourable for each question (the share answering 4 or 5). Easier to read than an average, and it does not hide a split team behind a middling mean.
- The distribution, for any question where opinion is split. "Eight people said 5 and seven said 1 or 2" is a different story from "everyone said 3".
- Change since last time, only when the difference is large (three or more people) and the question wording has not changed.
- The recommend question. If you use a 0 to 10 scale, the usual score is the share answering 9 or 10 minus the share answering 0 to 6. With small numbers it swings wildly, so report the counts alongside it.
An assistant can calculate these from an exported file quickly. Ask it to show a table of counts per answer for each question; percentages you can check at a glance are better than a paragraph of interpretation.
An online clothing shop's first survey and six exit interviews
An illustrative online clothing shop has 22 staff: nine in warehouse and packing, four in customer service, three in photography and content, three in buying and merchandising, two in the office and the owner. Warehouse turnover had been high, and the owner wanted to know why.
The survey went to 21 people (everyone but the owner), with team asked only as "warehouse and packing" or "everyone else", since both groups had nine or more. Nineteen responded. Results for the lowest-scoring questions:
| Question | Warehouse and packing (9) | Everyone else (10) |
|---|---|---|
| Enough notice of shifts and changes | 2 of 9 favourable | 7 of 10 favourable |
| Workload manageable most weeks | 3 of 9 | 6 of 10 |
| Trained properly for my tasks | 4 of 9 | 8 of 10 |
| Pay is fair | 3 of 9 | 5 of 10 |
| Would recommend as a place to work | 4 of 9 | 8 of 10 |
The six exit interview notes from the previous year, all from warehouse leavers, were anonymised and coded with the same codebook. Four of six mentioned short notice of overtime and rota changes in peak season; three mentioned pay compared with other warehouse work; three said they were moved onto returns or new tasks without training. Only one mentioned their manager as a reason for leaving.
Read together, the survey and the exit interviews told one story: the problem was planning, not people. Shift notice scored lowest, it was the biggest comment theme, and it appeared in most exit interviews. The actions the owner committed to:
- Publish warehouse rotas two weeks ahead, with peak-season overtime offered by the start of each month.
- A one-page task guide and a buddy shift before anyone is moved onto returns or goods-in.
- Review warehouse pay against a few current job adverts for similar roles before the next peak.
Adjusting the approach to your team's size and shape
- Under about eight people, such as a handmade jewellery workshop of five, an anonymous survey cannot really be anonymous. Use the same questions as a structured one-to-one instead, or ask someone outside the business to collect answers and report themes only.
- Seasonal teams, such as a farm shop's summer and autumn staff, give better answers if you run a short end-of-season form for everyone leaving at once, rather than individual exit interviews. Code the batch together and compare it with next year's.
- Shift-based production teams, such as a craft brewery's brewhouse and packing crew, answer more readily on paper or a tablet at the start of a shift. Add one question on safety ("I can stop a job if I think it's unsafe"), because it is the answer you most need to hear.
Reading exit interviews over time, not one at a time
A single exit interview is an anecdote, often coloured by how someone feels on their last day. The value comes from reading them together every quarter or six months. With an anonymised log sheet, this prompt does the job:
Attached: exit_log.csv (Quarter, RoleGroup, ServiceBand, Q1-Q10 answers, anonymised).
Using the codebook below, count themes by quarter and by role group.
Which themes appear in at least a third of interviews in any role group?
Has any theme appeared for the first time in the latest quarter?
Quote one supporting answer per finding, copied exactly. Don't speculate on causes.Watch for small numbers here too. Two leavers in a quarter cannot show a trend. Keep the log going for a year before drawing firm conclusions, and look at service band: people leaving in their first three months usually point to recruitment or onboarding, while people leaving after two years usually point to pay or progression. If early leavers dominate, planning a new starter's first two weeks is the fix to look at first.
Where AI analysis of staff feedback misleads
- Leading prompts. "Why are staff unhappy with management?" produces management themes whether or not they are there. Ask neutral questions and let the codebook decide.
- Long comments dominating. One person's 400-word comment can generate three themes and several quotes. Count themes per person, not per sentence.
- Sarcasm read straight. "Love finding out my shifts by text at 10 p.m." coded as positive about communication. The 20-comment check catches this.
- Invented or merged themes. A summary that mentions "concerns about remote working" in a warehouse team is a sign the model is drawing on generic patterns. Every theme needs quotes that exist.
- Sentiment scores on tiny samples. "Overall sentiment: 62% positive" from 19 people means very little. Skip it.
- Softened wording. Summaries tend to make angry comments polite. Read the raw comments yourself, all of them; with a small team it takes twenty minutes and it is the most useful twenty minutes in the process.
Telling staff what you heard and what changes
The next survey's response rate depends on what happens after this one. Share results within a month, in a short "you said, we're doing" note. The clothing shop's, adapted:
Thank you: 19 of 21 of you answered the survey.
You said: rota changes come too late, especially in peak.
We're doing: warehouse rotas published two weeks ahead from next month, and peak
overtime offered by the 1st of each month.
You said: people get moved onto new tasks without training.
We're doing: a one-page guide and a buddy shift before anyone starts a new task.
You said: pay doesn't feel right for warehouse work.
We're doing: reviewing warehouse pay against similar jobs before October, and we'll
tell you the outcome either way.
What went well: you rated your colleagues and team spirit highly. Keep it up.
Next survey: March. Questions to [name] any time.Include at least one thing you are not changing and why. It makes the rest more believable.
The legal and HR side
Staff survey responses and exit interview notes are personal data, and data-protection law such as the GDPR sets rules on why you collect them, how long you keep them and who can see them. Keep exit interview notes only as long as you need them for analysis, store them away from personnel files, and do not use survey data for decisions about individuals. If an exit interview raises a complaint about harassment, discrimination or safety, that is no longer feedback to analyse: it needs to be handled under your grievance or safety procedure, and it may be worth speaking to an HR or employment adviser. An AI helpdesk for staff HR questions can take some routine queries off your plate, but not these.
For the survey itself, the methods in analysing customer feedback surveys with AI transfer well, and if your next survey is about a specific change, surveying staff before an AI rollout has a ready question set. Wider HR tasks you can hand to AI safely are covered in AI for HR in small businesses, and the themes you find often feed straight into better conversations at review time.
Staff surveys and exit interviews: common questions
How often should a small business survey its staff?
Once or twice a year for the full survey is plenty for most teams under 50, with a two- or three-question pulse in between if something big changes, such as a move or a new manager. Surveying more often than you can act on the results teaches people that the survey changes nothing, and response rates fall quickly after that.
Should the owner run exit interviews personally?
In a very small business there may be nobody else, but people are more candid with someone outside their reporting line. If you can, have a manager from another area or a trusted outside adviser hold the conversation, and offer a written form as an alternative. Whoever runs it should use the same questions every time and write notes in the leaver's words.
What response rate is good enough?
In a small team, aim for most people to respond, because each missing voice changes the picture. Around three-quarters or more is a reasonable target when the survey is short and people trust the anonymity. A low rate is itself a finding: it usually means the survey is too long, the timing is bad, or people doubt it is anonymous or that anything will change.
Is it safe to paste staff comments into ChatGPT or Claude?
Only after removing names and identifying details, and only on a business plan or with the model-training setting switched off. Staff comments are personal data, and exit interview notes can include health or family details. Check with your data-protection adviser before uploading, keep the files you upload to the minimum needed, and delete them from the assistant when the analysis is done.
Further reads
- How to Write an Employee Handbook With AI and What to Check — Turn what staff told you into clearer written policies.
- How to Write Fairer Performance Reviews With AI — Carry survey themes into fairer one-to-one review conversations.
- How to Build a Staff Training Matrix With AI — Act on 'nobody trained me' comments with a proper matrix.
- AI Shift Scheduling vs a Rota Spreadsheet for Small Teams — Rota complaints are common; plan shifts with better notice.
- GDPR and AI Tools: What a Small Business Must Do — Data-protection basics before staff data goes into AI tools.
- Staff Offboarding Checklist for AI Tools and Shared Accounts — The practical side of someone leaving: accounts and access.
- How HR Consultants Use AI for Policies, Letters and Cases — Three workflows for HR consultants, with a client fact sheet, letter before-and-after, a grievance chronology prompt and the lines AI must never cross.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Google Forms help and settings (Collect email addresses options); Microsoft Support on setting up a survey so names aren't recorded in Microsoft Forms.