Send a short anonymous survey, 12 to 15 questions that take under ten minutes, a week or two before you choose any tools. Ask where people's time goes, what AI they already use, what worries them, how confident they feel and how they'd like to learn. Share the results within a fortnight, and act on at least one thing people said.
One caveat before you start: in a team under about fifteen people, anonymity is fragile, and a careless question can reveal who wrote what. The design choices below protect it. And only run the survey if you're prepared to change your plan because of the answers. Asking and then ignoring them does more harm than not asking.
Decide what the survey has to answer before writing a question
A pre-rollout survey should feed three decisions, and every question should serve at least one of them:
- Where to start. Which tasks eat time, happen often, and are ones staff would happily hand over.
- What rules and training come first. Who already uses AI, on which accounts, with what data, and how confident they are at checking its work.
- How to roll it out. Who wants to go first, what worries need answering, and how people prefer to learn.
If a question doesn't feed one of those, cut it. "How do you feel about AI in general?" is interesting but changes nothing you'll do next week.
Keeping answers anonymous in a small team
In a big company, a survey is anonymous by default. In a team of eleven, three answers are enough to identify someone. These rules close the usual gaps.
- The rule of five. Never report results for a group of fewer than five people. If you have two drivers and six counter staff, don't split answers by role at all.
- No combination questions. Role, length of service and full-time or part-time are each harmless alone. Together they point to one person. Ask none of them unless a decision depends on it.
- Free text gets paraphrased. People recognise each other's phrasing. When you share comments, reword them.
- Check the form settings. In Google Forms, leave "Collect email addresses" set to "Do not collect". In Microsoft Forms, either choose "Anyone can respond", or limit it to people in your organisation and untick "Record name". Microsoft's own help page warns that even then, a very small group or the response times can give people away, which is why the other rules here matter.
- Say exactly what you'll see. "I'll see the answers but not who gave them" is more believable than "this is completely anonymous".
Here's how easily it slips, in an illustrative ten-person car-repair garage. The owner's first results slide split the hours worry by team: "Workshop: 3 of 6 worried. Front desk: 2 of 2 worried." With two people on the front desk, both now knew the other had said yes, and so did everyone in the room. He also quoted a comment as written: "I'm the only one who does the warranty claims and nobody else knows how." That told the whole team who wrote it. The safer versions are dull, which is the point: "5 of 8 people are worried about their hours" with no split, and "One person said a task they handle alone would be hard to hand over."
Volunteers are the one exception. If you want names of people willing to test a tool first, collect them through a separate one-question form, so volunteering doesn't unmask the rest of someone's answers.
The question bank: 15 questions in five groups
Copy these as they are, or trim to twelve. Each has a note on why it's there, so you know what you lose if you cut it.
Where your time goes
- "Which three tasks take up the most of your week that you'd happily do less of?" (open text). Why: it gives you candidate tasks in staff's own words, which are often not the ones the owner would guess.
- "Roughly how often do you do each of these?" Rows such as answering the same question on the phone, retyping information between systems, writing routine emails or messages, looking up an order or stock status. Columns: never, 1-5 a day, 6-20 a day, more than 20. Why: rough volumes for scoring ideas later.
- "Which part of your job would you least want a computer to do, and why?" (open text). Why: it shows where the lines are, and often surfaces judgement calls the owner hadn't thought of as risky.
To see what questions 1 and 2 give you together, take an illustrative seven-person independent bookshop. Paraphrased answers to question 1 included "chasing special orders with the distributor", "the Friday stock spreadsheet", "the same 'do you have this in?' call again and again" and "writing the events newsletter". Question 2 put rough numbers on them: four of six people answered 6-20 stock-check calls a day, and two retyped special-order details between the till system and email "1-5 a day". That's two strong candidates (stock-check calls and special-order chasing) and one weak one (a monthly newsletter), before anyone has looked at a single AI tool.
AI you already use
- "In the last month, how often have you used an AI tool such as ChatGPT, Gemini, Copilot or Claude for work?" Never, once or twice, weekly, daily. Why: this measures shadow AI, meaning tools staff use without the business having approved them.
- "What did you use it for?" Tick boxes: writing messages, summarising documents, looking things up, translating, checking spelling, other. Why: tells you what people already find useful.
- "Which account did you use?" Personal free account, personal paid account, a work account, not sure. Why: this is your data-risk question. Customer details in a personal free account is the first thing to fix.
Worries
- "How much do you agree with each statement?" Five-point scale from strongly disagree to strongly agree. "I'm worried AI will make mistakes I'll be blamed for." "I'm worried AI will reduce my hours or change my role." "I'm worried about customer information going into AI tools." "I think our customers would be uncomfortable if we used AI." Why: separates the four common worries, which need different answers.
- "What would you need to know or see to feel comfortable with AI at work?" (open text). Why: tells you what to say in the rollout announcement.
Confidence
- "How confident are you at checking whether an AI answer is correct?" 1 (not at all) to 5 (very). Why: checking is the core skill, and low scores here set your training priority.
- "Which of these could you do today without help?" Tick boxes: write a clear instruction for an AI tool, spot an invented fact, remove personal details before pasting text, explain to a customer how we use AI. Why: more honest than a general confidence rating.
- "How comfortable are you learning new software in general?" 1 to 5. Why: separates discomfort with AI from discomfort with change.
How to roll it out
- "How would you prefer to learn a new tool?" Short demo, written guide, sitting with a colleague, trying it alone. Why: shapes the training format.
- "Would you be willing to try a tool before everyone else?" Yes, maybe, no, with the separate volunteer form linked. Why: finds your first testers.
- "When in the week could you realistically spare 30 minutes to learn something?" Why: prevents training being booked into the busiest hour.
Last word
- "Is there anything else you want me to know before I decide anything?" (open text, optional). Why: the answer that matters most often turns up here.
Trimming works best when you cut for your team rather than for length. Take an illustrative domestic-cleaning company with eight cleaners who work alone in clients' homes and one office coordinator. The owner cut question 10 (the tick-box skills list assumed desk work), question 11 (everyone already used the rota app daily) and question 14 (every cleaner's free time was the same gap between morning and afternoon jobs). She rewrote question 2's rows around their actual week: "texting a client that you're running late", "reporting a breakage or missing supplies", "checking key or alarm instructions for a new client". And she sent it as a link by text message, since most of the team never opened a work email. Twelve questions took about six minutes on a phone, and eight of the nine staff replied within three days.
The invitation, word for word
The email matters as much as the questions. It has to make it safe to admit current AI use, or question 4 will come back as "never" from everyone.
Subject: 8 minutes of your view before we decide anything about AI
Hi all,
I'm looking at whether AI tools could take some of the routine work
off us. Before I choose anything, I want your view.
The survey has 15 questions and takes about 8 minutes:
[link]
- It's anonymous. I'll see the answers, not who gave them.
- If you already use ChatGPT or similar for work, that's fine and
useful to know. Nobody is in trouble. It helps me set sensible rules.
- Please do it in work time, by [date].
I'll share what came back, and what I'm going to do about it,
by [date two weeks later].
Thanks,
[name]Worked example: reading the results at an independent pharmacy
Picture an independent pharmacy with eleven staff that runs the survey and gets nine replies. The numbers are made up, but the pattern is a common one.
- Time: seven of nine name phone calls asking whether a prescription is ready as a top time sink. Several mention retyping delivery details.
- Current use: five of nine have used AI for work in the last month. Four of those used personal free accounts, mostly to draft customer messages and look up product information.
- Worries: six of nine agree they're worried about being blamed for AI mistakes. Only two are worried about their hours.
- Confidence: the middle answer on checking AI output is 2 out of 5.
- Learning: five prefer sitting with a colleague.
Those answers lead to four decisions. First, before any new tool, a short rule on accounts and customer data, because customer details were going into personal accounts; the tutorial on shadow AI covers how to do this without a crackdown. Second, the first project is prescription-status calls, the approach in how independent pharmacies shorten phone queues. Third, training focuses on checking outputs, not writing prompts. Fourth, the rollout note says plainly that the pharmacist remains responsible for clinical decisions and that anyone who flags an AI error is doing their job. That answers the blame worry directly, and it matches the lines drawn in pharmacy tasks you should never hand to AI.
Note what the pharmacy doesn't report: answers split by role. With two pharmacists and two drivers, any split would identify people.
Expect a few answers that contradict each other, and read them as information. Two of the pharmacy's nine replies said "never" to question 4, then ticked "writing messages" and "checking spelling" in question 5. The likeliest reading isn't dishonesty: people often don't count the AI features built into their phone keyboard or email app as "using AI". Count those two as occasional users, and add a line to the training about AI features inside everyday apps, because they follow the same data rules.
Grouping the free-text answers without losing the sharp ones
Nine replies to three open questions is about 27 comments, which you can read in ten minutes. AI can help group them, on a business account and with anything identifying removed first:
Below are anonymous staff comments from a survey about AI.
Group them into themes. For each theme give: a short name,
how many comments fit it, and one paraphrased example.
Then list separately any comment that raises a risk, a
complaint or something only one person mentioned.
Comments: [PASTE, with names and shift details removed]An illustrative extract of the reply:
"Theme 1: Phone workload (7 comments), e.g. 'the prescription calls never stop'. Theme 2: Concerns about accuracy and responsibility (5 comments), e.g. 'I don't want to be blamed if it gets something wrong'. Theme 3: Use of personal AI tools (4 comments), e.g. 'I sometimes use ChatGPT on my phone for customer messages'. Single mentions: none that raise significant risks."
The themes are fine; the last line is wrong. One of the four theme-3 comments said a customer's list of medicines had been pasted into a personal account to draft a message about interactions. The AI filed it as an ordinary example of personal-account use, even though it was the most serious thing in the survey. That's why the separate "risk" instruction isn't enough on its own, and why you still read every raw comment yourself.
Questions that backfire
- "Do you support the move to AI?" It's loaded and it announces a decision that's supposedly not made yet.
- "Would you take on new responsibilities if AI frees up your time?" Read in a small team, that sounds like "your job is changing whether you like it or not".
- "How many hours a week do you waste on admin?" "Waste" implies blame. Ask how often tasks happen instead.
- Double-barrelled questions such as "Is AI useful and safe?" Someone who thinks it's useful but unsafe can't answer.
- Compulsory open-text boxes. People type "n/a" to get past them, and the survey takes longer.
- Questions that invite politeness. How it shows up: an illustrative eight-person print shop asked "Do you think AI could help us?" and got eight yeses. The rollout stalled anyway. A re-run using the four worry statements in question 7 found five of eight agreed they were worried about being blamed for AI mistakes. The yeses had been polite rather than informative, and the statements gave people a way to disagree without saying no to the boss.
- A name field "for follow-up". One optional name field on the same form and nobody believes the rest is anonymous.
Closing the loop: the "you said, we're doing" note
Within two weeks, send a one-page reply. It is the part that earns you honest answers next time.
- How many replied, out of how many.
- The three things that came up most, in plain words.
- Three decisions you've made because of them.
- One thing people asked for that you're not doing yet, and why.
- When they'll hear more.
For the pharmacy above, the note might read (illustrative):
YOU SAID, WE'RE DOING: AI survey results
9 of 11 of you replied. Thank you.
What came up most
- Prescription-ready calls take up a big part of the day
- Several of you already use AI for messages, mostly on
personal accounts
- Most of you are worried about being blamed for AI mistakes
What we're doing
1. From Monday, no customer details in personal AI accounts.
We're setting up one work account for everyone instead.
2. Our first project is prescription-status calls. Nothing
else changes until that's working.
3. Training will focus on checking AI answers, in pairs,
because most of you said that's how you'd like to learn.
Not doing yet: AI help with delivery routes. Our current
delivery software may already do it, so I'm checking first.
On blame: the pharmacist stays responsible for every clinical
decision. Flagging an AI mistake is doing your job well.
More on [date].If worries about jobs came up strongly, deal with them face to face as well; the tutorial on talking to staff who fear AI has the conversation. And if you have a lot of free text, the approach in running staff surveys with AI analysis shows how to group comments without losing the sharp ones.
Staff AI survey questions owners ask
Should the survey be compulsory?
No. Make it clearly optional and anonymous, and ask for it during paid time, for example the first ten minutes of a shift. Compulsory surveys in a small team feel like a test, and people answer the way they think the owner wants. You get fewer but more honest answers from a voluntary survey.
Can I use AI to analyse the free-text answers?
Yes, on a business plan that doesn't train on your content, and after removing anything that identifies a person, such as names or references to a particular shift. Ask it to group comments into themes and count them, then read the raw answers yourself. AI summaries tend to smooth over the one sharp comment you most need to hear.
What response rate is good enough?
In a team of ten, aim for eight or more. Below about half, the results mostly reflect the keenest and the most worried people, and the middle is missing. If replies are low after a week, the owner asking in person, with a reminder that it's anonymous, works better than a second email.
Should I survey staff again after the rollout?
Yes, around 90 days in. Reuse five questions word for word: current AI use, which account people use, the blame worry, confidence at checking answers, and the open question. Identical wording is what lets you compare before and after. Change the questions and you can't tell whether attitudes moved or the wording did.
Further reads
- How to Get Your Staff to Actually Use AI Tools — What to do with low-appetite results.
- How to Train Staff to Use AI in a Small Business — Build training around the confidence gaps the survey finds.
- How to Write an AI Usage Policy for Your Small Business — Turn the account and data answers into rules.
- How to Choose and Support an AI Champion in a Small Team — Pick from the volunteers the survey turns up.
- AI Literacy Requirements: What Your Staff Need to Know — What every employee should know, whatever the survey says.
- AI Change Management for Small Teams: A Practical Plan — Plan the rollout around the concerns people raised.
- How Much Time and Money Does AI Staff Training Take? — Training hours by role, free and paid course options with current prices, a costed plan for an eleven-person decorating firm, and how to check it paid off.
- Barriers to AI Adoption in Small Businesses and How to Clear Them — Eight practical barriers that keep small firms from adopting AI, the cheapest way past each, and the few barriers you should respect rather than clear.
- How to Build a Restaurant Staff Rota With AI Demand Forecasts — Turn a covers forecast into an hourly staffing plan and a fair rota, with a worked Saturday that shows the labour-cost difference in dollars and hours.
- AI Rollout Plan for a Ten-Person Marketing Agency — A ten-person agency's 12-week AI rollout, week by week: platform choice, client rules, two pilots, the numbers after a quarter and what went wrong.
- How to Measure Time Saved After Rolling Out AI in a Small Firm — A before-and-after timing method for five named tasks, with the net-saving formula, a worked calculation and the checks that keep the number honest.
- How to Roll Out Microsoft 365 Copilot to a Small Team — Start everyone on free Copilot Chat, license three heavy users for four weeks, then let the usage report decide who gets a seat. A café group shows how.
- How to Get Staff Buy-In When You Introduce AI — A veterinary practice introduces AI with its team rather than to it: one-to-ones, a team-chosen first job, a sceptic's veto and a written time-dividend deal.
- Small Business AI Statistics 2026: Adoption, Spend, and Results — 2026 figures on small-business AI adoption, spending and results from eight named reports, with what each means for a five-person business.
- 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 on collecting email addresses; Microsoft Support, 'Set up your survey so names aren't recorded when collecting responses'.