Patient Satisfaction Survey With AI: Questions and Comment Summaries

A small clinic can ask AI to draft a seven to ten question patient satisfaction survey, export only the anonymous answers, have AI sort the free-text comments into themes, and then record each decision in an action log. Keep identifiable data out of any tool that has no data agreement for patient information.

Most clinic surveys fail in one of two places: they are too long to finish, or they produce a pile of comments that nobody reads. AI helps with both, and only if the survey is designed to produce answers you can act on and the comments reach the AI stripped of identity.

Follow me on Instagram@sagnikteaches

Decide what you want to change before you write a question

Start with the decisions the survey could inform. For a small clinic these are usually few: opening hours, phone access, waiting on the day, how well staff explain things, and how easy booking is. Write down three decisions you might make differently depending on the answers ("extend Thursday evening hours", "add a call-back option", "change the waiting-room sign"). Any question that cannot affect a decision is out.

Connect on LinkedInSagnik Bhattacharya

Then set a rhythm. A short survey sent continuously and reviewed monthly beats a big annual survey, because the numbers are fresh and you can see whether a change worked. The routine below assumes monthly.

Subscribe on YouTube@sagnikteaches

Draft the questions with AI, then cut

I run a small private clinic with [number] staff. Draft a patient
satisfaction survey of 8 questions that takes under two minutes on a phone.
Mix: 5 rating questions on a 1 to 5 scale (booking, waiting, staff listening,
clarity of explanation, overall), 1 yes/no question (would you book again?),
1 optional free-text question ("What is one thing we could do better?"), and 1
optional free-text question ("What went well?").
Rules: one idea per question, no leading wording, no jargon, no questions about
symptoms, diagnoses or treatments, no request for the patient's name.
Add a one-sentence introduction that says answers are anonymous and asks
people not to include names or health details in comments.

An illustrative result is a tidy eight-question set. Two things usually need editing. First, AI loves "double-barrelled" questions such as "How satisfied were you with the booking process and the reception staff?", which mixes two things. Split or cut them. Second, it often adds a ninth or tenth "just in case" question. Cut anything that fails the decision test above.

A filled-in example survey

1. How easy was it to book your appointment? (1 very difficult to 5 very easy)
2. How long did you wait on the day compared with what you expected? (much longer, a little longer, about right, shorter)
3. How well did the person you saw listen to you? (1 to 5)
4. How clear was the explanation of what happens next? (1 to 5)
5. How would you rate the reception team? (1 to 5)
6. Overall, how was your experience of our clinic? (very good to very poor)
7. Would you book with us again? (yes, not sure, no)
8. What is one thing we could do better? (optional, please do not include names or health details)

Eight questions, two minutes, one free-text box. The second free-text box ("what went well?") is optional and some clinics keep it for staff morale; if completion drops, remove it first.

Test the survey before it reaches a patient

Ask five people, a mix of staff and friendly outsiders, to complete it on their phones while you watch. Look for hesitation. If anyone asks "what does that mean?", the question changes. Time it with a stopwatch; over two minutes means cut a question.

Then use AI as a second reader. Paste the survey and ask: "Which questions are leading, ambiguous, or likely to be answered the same way by everyone? Which could be misread by a patient with limited English or low vision?" Treat its answer as a prompt for your own judgement. It often catches a lopsided scale (more positive than negative options), which quietly inflates your results.

Choose a form tool that keeps data where you can see it

The form tool collects the answers; AI only sees what you choose to export. Common choices are a free form builder such as Google Forms, Microsoft Forms if you already have Microsoft 365, or Jotform. Jotform's free Starter plan allows 5 forms, 100 submissions a month and 500 stored submissions; the clinic in the worked sum below, at 154 to 198 responses a month, would outgrow it in the first month, and the next tier (Bronze) lists at $34 a month billed annually ($39 monthly) for 1,000 a month. Jotform's pricing page also marks its healthcare-compliance features as unavailable on Starter, Bronze and Silver, appearing from Gold at $99 a month billed annually ($129 monthly), which matters only if identifiable patient data will ever sit in the form. Check the current limits before you commit. Things to check in whichever you pick:

• Can you turn off collection of email addresses and names? Do not collect them.
• Does it record IP addresses or location? Switch that off if you can.
• Can you export responses as a spreadsheet, so you control what is pasted into AI?
• Where is the data stored, and does the vendor offer a data processing agreement or business associate agreement if you decide to link survey data with patient records later?

Anonymous collection is the simplest design. If you send each patient a unique link so you can follow up on low scores, you have created identifiable data, and the survey needs the same care as any patient record. Many small clinics keep the main survey anonymous and give a separate, optional "contact me about this" box that is not exported to AI.

Delivery: how to get enough responses

Three delivery routes, in the order I would try them:

1. A text with a link, sent a few hours after the visit. Short message, one tap, anonymous. The link is the same for everyone, so you cannot tell who answered.
2. A QR code at the front desk and on the receipt or appointment card. Cheap and good for people who prefer to answer on the spot.
3. A prompt from the clinician or reception at the end of the visit. "We have a two-minute survey, would you mind?" works better than a poster.

AI can draft the invitation wording and short variations to test. An example: "Thanks for visiting [clinic] today. Two minutes to tell us how it went? [link] It is anonymous, and we read every answer." Keep it under 160 characters if you can, using straight apostrophes, as a single curly apostrophe or emoji can cut a text's character allowance and add to the cost.

A worked sum, with invented numbers: a clinic sees about 1,400 appointments a month and has mobile numbers for 1,100 of them. It texts the link to those 1,100 and 154 people respond in the first month, a rate of 14 per cent. A second month with a better-timed text (sent the same afternoon) brings 198 responses from 1,100, or 18 per cent. That is an increase of 44 responses and four points of rate, which is the kind of change worth tracking more than the absolute figure.

Export the comments without identifiers

When the month closes, export to a spreadsheet and build a working copy for AI. In the working copy:

• Delete columns you do not need: timestamps to the minute, IP addresses, any email or name column. Keep an ID, the date reduced to the week, and the ratings.
• Read each free-text comment and redact names of staff, patients and family, conditions, treatments, dates and places. Replace with neutral tags such as [clinician], [family member], [condition]. For about 80 comments a month this is a 20-minute job.
• Remove comments that cannot be cleaned. If a comment is so specific that it identifies someone, leave it out of the AI batch and read it yourself. Record "1 comment withheld".
• Remove any comment that is a complaint or a safeguarding concern. Pass it to whoever handles that process today; it is not feedback to be themed.

The tutorial on anonymising client data before you paste it into AI goes through a reliable replace-and-restore routine. A rule that keeps you honest: you should be able to hand the AI's input to a stranger without them being able to name one patient.

Where AI is allowed to see the data

Be explicit about which tools may see which version of the data.

Check the plan, not the brand. A vendor that signs such an agreement often does so only for an enterprise or healthcare tier, so the team plan you use for drafting is usually not covered unless the vendor says in writing that it is. Free consumer plans are not in the table at all. If the tool and the agreement are not already in place, the redacted-only rule is what protects you, so apply it every month without exceptions. The tutorial on classifying data before using AI tools shows how to make those three tiers part of your clinic's habits.

Summarise the comments: prompt and sample output

Below are 70 anonymised patient comments from a small clinic survey. Each has
an ID and a rating from 1 to 5 for overall experience.
Task:
1. Group the comments into no more than 8 themes. Name each theme in 2 to 4 words.
2. For each theme give: number of comments, how many came from ratings 1-2,
   how many from ratings 4-5, and two short exact quotes.
3. List any comment that mentions a safety concern, a complaint about a named
   role, or something that looks identifying, with its ID only. Do not quote it.
4. Do not give advice and do not invent details.

Comments:
[paste ID | rating | comment]

An illustrative result: "Phone access (14 comments; 12 from ratings 1-2; 'rang six times before anyone answered'). Friendly staff (19; 1 from ratings 1-2; 'everyone was so welcoming'). Waiting time (11; 9 from ratings 1-2). Clear explanations (9; 0 from ratings 1-2). Parking (6). Online booking (5). Other (6)." Counts like these should be checked in the spreadsheet, because chat assistants miscount; use a filter to confirm each number before it goes into a report.

Two checks make the summary trustworthy. First, ask for exact quotes, which exposes themes the AI invented. Second, read the ten lowest-rated comments yourself, whatever the AI says. A theme is not the same as an intensity: three comments that describe a real safety near-miss matter more than thirty about parking.

Reading the ratings with simple sums

Ratings need less AI than comments. A spreadsheet does the work, and three figures per question are enough: the number of answers, the share scoring 4 or 5, and the share scoring 1 or 2. Averages hide too much, because a 3.8 can be everyone saying 4 or half the room saying 5 and a fifth saying 1.

Take question 4, clarity of explanation, with 70 answers: 50 scored 4 or 5, 12 scored 3, 8 scored 1 or 2. That is 71 per cent high, 17 per cent middle and 11 per cent low (the three round to 99, which is fine). Now compare with question 1, booking, where 38 scored 4 or 5, 14 scored 3 and 18 scored 1 or 2: 54 per cent high and 26 per cent low. The booking question is where to look first, and the comments sort the "why" into phone access and online booking. With 70 answers a swing of a few people moves a percentage by several points, so avoid reading a change of three or four points as a trend. Wait for two or three months pointing the same way.

Ask AI to help with the arithmetic only as a check: paste the counts and ask it to compute the percentages, then compare with your sheet. If they differ, believe the sheet.

A realistic mistake and how it showed up

An invented case shows how easily this slips. A clinic exports its first month straight from the form, including a timestamp to the second and the free-text box, and pastes 60 comments into a free chat assistant to save time. Three comments name a clinician, one mentions a child's condition and a date. Nobody notices until a partner reads the chat history a week later.

The right recovery is ordinary and worth copying. Delete the conversation, check the tool's retention and training settings, write down what went in and when, tell the clinic's data-protection lead the same day, and decide together whether anyone needs to be informed. Then change the process: move the redaction step into the spreadsheet template as a separate tab called "AI-safe", and swap the chat tool to a business plan. The lesson: make the safe version the one with the fewest steps, because the shortcut is always tempting at month end.

Adjusting the survey for different kinds of clinic

The eight questions above suit a general outpatient clinic. A few swaps for other types:

• Physiotherapy and other repeat-visit services: ask after the third visit, not every visit, and add "How clear is your plan for the next few weeks?". Repeat surveys at every session produce fatigue and a stream of near-identical answers.
• Dental or procedure-based clinics: add one question on comfort and one on how clearly costs were explained before treatment, because those are the two places surprises arise. Do not ask about the clinical quality of the procedure, which patients cannot judge.
• Telephone or video appointments: replace the waiting-room question with "How easy was it to connect?" and add "Was the call close to the booked time?".
• Services for children or vulnerable adults: survey the parent or carer, keep it anonymous, and avoid any question that could draw out detail about the person's condition.

Keep a master version of the survey and a dated change list so that results before and after a wording change are never compared as if they were the same question.

The action log that turns comments into changes

A survey without a log is a hobby. After each monthly review, add rows to a simple table, shared by the team, with five columns: theme, evidence, decision, owner and review date.

Notice that one row has a decision of "no action". Recording a deliberate decision not to act, with a reason, is part of a credible log. At the next review, the first job is to open last month's log and ask whether each action happened and whether the number moved.

A simple measure keeps this honest. Take the phone row: if 14 of 70 comments mentioned phone access in March (20 per cent) and 7 of 74 in May (about 9 per cent), the change is real enough to mention to the team. If the proportion did not move, the call-back trial did not work, or did not run as planned, and you ask why. The sibling tutorial on summarising Friends and Family Test comments shows how a fixed code frame helps with this kind of comparison.

Close the loop with patients

People answer again when they see an effect. Once a quarter, put two or three changes on a poster or in a short text or newsletter, in the form "You said, we did", with real actions. AI can draft the wording from your log, but each line must match an entry in it. Do not claim an improvement you have not measured; "we are trying a call-back option from June and will report in September" is better than "phone access is now better".

Four other jobs for the same set-up

1. A separate short survey for new patients. Three questions on booking and first impressions, since they notice things regular patients no longer see.
2. A staff version. Ask the team the same questions from their side; gaps between the two sets of answers are informative. The method in surveying your staff before an AI rollout adapts directly.
3. A "why did you not come back" message. A single-question text to patients who cancelled and did not rebook, anonymous reply, same redaction.
4. A tidy monthly one-pager. Ratings table, the top three themes, the action log changes, and one quote from a good week for the staff room.

Where patient surveys go wrong

• Leading questions. "How excellent was our service?" is not a question. Ask AI to flag leading wording and have a colleague read it cold.
• Only happy and angry people answer. Say so in the report: "These are the views of the 18 per cent who replied."
• Identifiers inside the comment box. Patients write names and conditions whatever the intro says. Redact every time and do not trust a form instruction alone.
• Pasting the raw export into AI to save ten minutes. The saving is real, so the rule needs a person who owns it, and a monthly reminder.
• A summary nobody acts on. If the action log has no owners and dates, the survey ends as a pleasant PDF.

Questions clinic managers ask about patient surveys and AI

How long should a patient satisfaction survey be?

Seven to ten questions that take under two minutes on a phone. Each extra question lowers completion, and the people who finish a long survey are often the keenest or angriest. One free-text question is worth more than five more tick boxes.

Can AI write the questions for us?

Yes, and it is good at first drafts and at spotting leading or double-barrelled wording. You still decide what you need to learn, test the survey on five colleagues or friendly patients, and make sure nothing asks for clinical detail you do not need.

Is it safe to paste survey comments into an AI tool?

Only after removing names, dates and health details, and only on a plan whose terms exclude training. If a comment could identify someone, leave it out or rewrite it. With a tool and agreement that cover patient data, the position changes, but check that first.

What response rate should we expect?

It varies widely with how and when you ask, so use your own first month as the baseline. A short link sent by text soon after a visit usually does better than a leaflet. Aim to improve on your own number, not to match a figure from elsewhere.

Further reads

• How a Small Clinic Can Use AI at the Front Desk Safely — Where survey invitations can sit in the front desk routine.
• Clinical Audit Template: A Small-Practice Cycle Summarised With AI — A structured audit cycle that complements patient comments.
• How a Clinic Can Use AI to Cut Missed Appointments — Missed appointments often show up in survey comments first.
• Patient Intake Automation in a Small Practice: Forms, AI Summaries — Another form your clinic could draft and summarise with AI.
• AI-Powered Chatbot Best Practices: A Weekly Review Routine for Clinics — A weekly review routine to adapt for survey results.
• Hotel Guest Feedback: Turn Post-Stay Surveys Into Fixes With AI — A post-visit survey turned into a fix list in another sector.
• AI Marketing Automation for Doctors: Recall, Reminder and Review Flows — Build recall, reminder and review-request flows in the practice software you already own, with consent, opt-outs and a worked month of numbers.
• AI Marketing for Medical Practices Without Making Health Claims — Service pages, review replies, referrer updates and local search for a medical practice, drafted with AI and kept free of health claims.
• AI Agents for Healthcare: Six Safe First Jobs for a Small Practice — Six admin-only first jobs for an AI agent in a small practice, the clinical work that stays human, and the contract checks each tool needs.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Jotform pricing page; ChatGPT Business and Claude pricing pages; Twilio's SMS character limit documentation; survey wording and figures in the examples are illustrative.