Can AI Summarise Training Feedback Forms for a Client Report?

Yes, with a split of jobs. Calculate scores in a spreadsheet, because arithmetic should never be guessed. Give AI only the anonymised comments to group into themes, then ask it to draft a one-page client summary around your numbers. Check the themes against ten comments you read yourself before you send anything.

The reason most trainers get lumpy results is that they hand a pile of forms to a chatbot and ask for "a summary". The output sounds fluent and misses the one comment that mattered, such as the delegate who said the room was too cold to concentrate or that the examples did not fit their sector. The process below stops that by separating numbers, themes and narrative.

Follow me on Instagram@sagnikteaches

Design the form so it can be summarised

Summarising starts before the course. A form that mixes six free-text boxes with vague scales gives you nothing to count. For a typical one-day course, ten questions is enough: seven rated, three written.

Connect on LinkedInSagnik Bhattacharya

Use one rating scale throughout. A five-point scale labelled at both ends ("1 = strongly disagree, 5 = strongly agree") is the easiest to total, and there is no need to remove the neutral midpoint. Mixing 1 to 5 with 1 to 10 makes every comparison awkward later.

Subscribe on YouTube@sagnikteaches

Keep the two "most useful" and "change" prompts separate. When delegates get one box asking "any comments", they write either praise or a complaint, and the mix makes theming harder. Add a delegate group code (for example "Group A") rather than a name field. Google Forms can save responses to a linked Google Sheet, Microsoft Forms opens them in Excel, and Tally, whose free plan allows unlimited forms and submissions, connects to Google Sheets; the Tally walkthrough for collecting feedback and summarising it covers a free route.

Score the ratings yourself, in a spreadsheet

Export responses to a sheet with one row per delegate and one column per question. Then compute, per question: the number of responses, the mean, and the percentage scoring 4 or 5 (often called "top-two box"). Those three numbers are what clients quote.

Take a customer-service workshop for a café group, run for 16 delegates of whom 14 returned the form, as the main worked example (all figures invented for illustration):

Note the 13 on the practice question: one delegate left it blank. Report "9 of 13", not "9 of 14", and never let a model recompute this for you. In a spreadsheet, the formulas are short: =AVERAGE(C2:C15) for the mean, =COUNT(C2:C15) for responses, and =COUNTIF(C2:C15,">=4")/COUNT(C2:C15) for the top-two share.

The weak spot is already visible: the pace question has the lowest mean (3.4) and only half the group rated it 4 or 5. That tells you where the comments should be read most carefully.

Anonymise the comments before AI sees them

Free-text comments leak. A delegate writes "[name] on my team found the role-play uncomfortable" or "as the only finance person in the room". You then paste that into a chatbot and put it in a client report. Before any theming, clean the comments in a separate column.

1. Copy the three comment columns into a new sheet with no delegate ID or timestamp.
2. Search for capital letters mid-sentence and replace names with "[colleague]" or "[manager]".
3. Replace job titles that identify one person ("the only HR manager") with a general role ("a delegate").
4. Delete anything that is a complaint about an individual rather than the course, and put it in a private note to yourself.
5. Shuffle the rows, so order does not reveal who submitted first.

Small groups are the real risk. In a group of eight from one department, a client reading "I felt unable to speak up with my line manager in the room" can guess who wrote it. Tell delegates on the form how comments will be used ("shared with your employer as themes, not word for word"), and honour that. If a business plan is available, use it: ChatGPT Business, Claude Team, Gemini in Google Workspace and the work version of Microsoft Copilot (formerly Microsoft 365 Copilot) do not train on your content by default, and consumer ChatGPT and Claude plans let you switch training off in privacy settings. See coding employee survey comments safely for a deeper treatment of the anonymity question.

Ask AI for themes, with counts and evidence

Now the AI step. The risk is a smooth list of themes with no sense of how common each is. Force it to count and to quote.

Below are anonymised comments from a one-day customer-service workshop
(14 completed forms; 31 comments in total). Group them into no more than six
themes.

For each theme give:
- a short label of 2 to 5 words
- the number of comments that mention it
- whether the comments are mostly positive, mostly negative or mixed
- two short verbatim quotes with names already removed

Also list any comment that does not fit a theme under "Outliers".
Do not invent comments. If two themes overlap, say so.

COMMENTS:
[paste 31 comments, one per line, numbered]

Numbering the comments lets you verify counts. Here is an illustrative output for the café workshop.

Now check it, because this is where errors hide. Count the numbered comments yourself for at least two themes. If the AI says seven mention pace and you find five, the theme is too broad or the model has counted comments twice. Also read the outliers, because the most valuable single comment is often there.

Three corrections most theme lists need

In practice, themes need three corrections.

Merged opposites. A model may bundle "pace too fast" and "pace too slow" into "pace". The client needs to know which. Ask: "Split any theme where comments point in opposite directions."

Over-generic labels. "Course content" says nothing. Rename to "Wanted more on handling complaints". A specific label is something a client can act on.

Missing praise for the trainer, or blame. Models sometimes drop personal comments. If the comments name the trainer, decide how to handle them before they enter a client document; a mention of an individual trainer belongs in your own quality file, not necessarily the client summary.

Sample sizes and how to word small groups

A mean of 4.6 from 14 people is evidence. A mean of 4.6 from four is an anecdote. Set your own rules and apply them to every client, so you are not accused of favouring flattering numbers.

• Under 8 responses: no percentages. Give counts ("5 of 7 rated the pace 4 or 5") and treat comments as individual views.
• 8 to 14 responses: counts alongside percentages, themes labelled "mentioned by" rather than "most delegates".
• 15 or more: percentages stand on their own; compare against previous sessions if you have three or more.
• Response rate under 60%: say so. A client should know that eight of 20 delegates answered.

Add a line to every report: "Based on 14 completed forms from 16 delegates (88%)." It takes seconds and shows that you did not hide a poor response rate.

The one-page client summary

Clients want a page, not a spreadsheet. Give the AI the score table, the themes table and your own notes on what you changed in delivery, and use a prompt that fixes the layout.

Write a one-page client summary of a training evaluation. Use only the
two tables and my notes below. Do not add numbers or causes.

Layout:
1. Headline (one sentence: overall result and the main point to act on)
2. Scores at a glance (keep the table as given)
3. What delegates valued (top two themes, with counts)
4. What to improve (top two themes, with counts)
5. Our response (from my notes)
6. Suggested follow-up for the client in the next 30 days

Plain British English. Under 350 words excluding the table. No praise
of our company. State the number of responses beside every score.

SCORES: [paste table]
THEMES: [paste table]
MY NOTES: [paste notes]

An illustrative output: "Delegates rated the workshop 4.6 out of 5 for clarity of objectives and 4.8 for the trainer. The main point to act on is pace: only 7 of 14 rated it 4 or 5, and 7 of 31 comments said the afternoon felt rushed. Role-play was the most valued element (9 comments)..." Then the follow-up: "Ask each manager to observe one complaint-handling conversation within two weeks."

The "our response" section matters most to the client and is yours alone. If you plan to move the lunch break or cut the final module to allow more practice, say so. Clients remember trainers who act on feedback.

Spot checks that take ten minutes

Before sending, run four spot checks. They are quick and they catch most of what goes wrong.

1. Recompute two scores. Pick the highest and the lowest, and recalculate the mean from the raw sheet.
2. Read ten comments. Choose them at random and see whether the theme assigned fits. If three of ten feel wrong, rerun with tighter instructions.
3. Search for names. Use find to look for capital-letter words in the final page, and delete any person's name or job title.
4. Read the headline aloud. If the first sentence would make a client wince or shrug, rewrite it to reflect the actual finding.

Then compare with a human: ask a colleague who attended or observed the session whether the summary matches their memory. Where it does not, trust the colleague and find out why.

Turning vague comments into usable ones

Some comments are too thin to theme, and a client report full of "good course" does nothing. Two fixes help. First, improve the prompt on the form itself. "Any other comments?" produces "no". "What is one thing you will do differently on Monday?" produces something specific. Second, when you have a pile of thin comments, tell the AI to separate them rather than stretch them.

A before and after shows the idea (illustrative):

Raw comment: "Good trainer, a bit rushed, loved the role play, handout could be better."

The same comment split into four coded pieces:

• Trainer: positive.
• Pace: negative (rushed).
• Role-play: positive.
• Handout: negative (could be better, no detail).

One comment now counts in four themes. That is correct, and it is why the theme counts in your table can add up to more than the number of comments. Say so on the page, with a footnote such as "one comment can touch several themes", or a client will try to add the column and find the total does not match.

Comparing this course with earlier ones

A single score is hard to read; a trend is not. If you have run the same course before, add a column to the score table for the previous run, and compute the change in a spreadsheet. Say the pace question scored a mean of 3.4 this time and 4.0 on the last run with a different group of 12.

The difference is 0.6, which is large enough to look at, though with groups this small it could be a different mix of people rather than a real change. Say that in plain words: "Pace scored lower than the previous group (3.4 against 4.0). Two groups of this size are not enough to call a trend, so we will watch it over the next three sessions." That is more honest than "pace has declined", and clients trust a trainer who says what the data cannot show.

For three or more runs, a small table with one row per session and one column per question lets you spot a pattern at a glance. Ask the AI only to describe the table you have built, not to produce one. A prompt such as "Describe in three sentences which scores rose, which fell, and which stayed within 0.3 points, using only the table below" keeps it to what the numbers say.

When the venue theme outranked a missing briefing

Imagine a trainer pasted 40 raw forms, names included, into a chatbot and asked for "the top five themes". The reply listed "Excellent facilitation" as theme one and "Venue issues" as theme five. It looked tidy. But the venue theme had been created from just two comments, while a cluster of six comments about a missing pre-course briefing was folded into "Other".

The client's main fix was invisible. The error only showed up because the trainer counted comments for the last theme and saw the number was small. The repair was simple: ask for counts, ask for outliers, number the comments, and rerun. The lesson is that a summary is not wrong because it lists a false theme; it is wrong when it quietly ranks the themes by what reads best rather than by how often delegates said it.

Three other training formats that need small changes

A half-day webinar, a multi-day programme and a coaching series each need slightly different handling.

Multi-day programmes. Collect feedback after each day with a short three-question form, then a longer one at the end. Summarise each day separately and ask the AI to compare days ("Which theme appears in day 2 and not day 1?"). A one-day summary may hide the fact that the first morning was weak.

Online sessions. Add a question about audio, screen sharing and breakout rooms. Low scores here are a delivery problem, not a content problem, and the client will want them separated.

Dog-grooming or driving-school style trade courses. If you train staff in practical skills, add one rated question per skill, for example "I can now brush out a matted coat safely" or "I can demonstrate a hill start to a learner". Self-rated confidence is not competence, so label it as confidence and, where possible, pair it with an instructor assessment.

Following up after the summary goes out

Feedback is most useful two or three weeks later. Send delegates a three-line message asking what they have used. If you manage bookings and reminders in a CRM, record the summary against the client and schedule the follow-up there; choosing a CRM for a training company compares options. You can ask the AI to compare the two-week answers with the end-of-day comments and see whether the "Monday action" actually happened.

Keep a running sheet of scores by course, trainer and date, so you can spot drift. If three sessions in a row score the pace question below 3.5, change the course, not the summary wording. The summary is only as good as the form behind it, and a form built for counting makes every later step quicker. If you are still writing the brief that comes before all this, the needs analysis report tutorial helps you decide what the evaluation should measure in the first place.

Questions trainers ask about feedback summaries

How many forms do I need before themes mean anything?

Below about eight responses, report comments individually and avoid percentages. From around fifteen, themes start to hold. Always show the number of responses next to every score so the client can judge the weight for themselves.

Can I paste delegates' comments into ChatGPT?

Only after you remove names, job titles that identify one person, and anything naming a colleague. Use a business plan or confirm the tool's privacy settings first. Comments often contain personal detail that delegates did not expect a third party to see.

Should I send the client the raw comments?

Agree this in advance. Many clients are happy with themes plus three or four representative quotes, lightly redacted. Sending every comment verbatim can expose individuals, especially in a group of ten, and removes the reason for anonymity.

Can AI calculate the average scores?

Let the spreadsheet do it. Averages, percentages and counts are arithmetic, and a spreadsheet formula never guesses. Use AI for the words: grouping comments, naming themes and drafting the summary around numbers you have already worked out.

Further reads

• Training Course Cancellation Policy: Terms Drafted With AI — Another client-facing document worth drafting with AI.
• Training Proposal: Scope, Price and Outcomes Written With AI — Write outcomes you can later report against.
• AI for Corporate Training Companies: Proposals to Certificates — The other admin jobs a training firm can hand to AI.
• Joining Instructions Template: Details and Reminders With AI — Send the feedback link in the same reminder flow.
• AI Training for School Staff: Session Plan, Tasks and Handout — How another sector plans and reviews staff sessions.
• How to Use AI in Business Training: Quizzes That Show What Stuck — Add a knowledge check alongside the happy sheet.
• Event Feedback Survey: Write It and Summarise the Answers With AI — Ten event survey questions, a frozen theme codebook, counts you verify in the sheet and a one-page report, with the AI doing the reading.
• Event Planning Using AI After the Day: Feedback, Sponsor Reports, Debrief — Post-event only: survey themes with counts you check, a sponsor recap built on promised against delivered, and a debrief that ends in owners and dates.
• Museums and AI: Turn Visitor Feedback Into Next Season's Plan — How a small museum can sort a season of visitor comments into themes with AI, check the tags, and turn the counts into five fixes for next season.
• AI in Employee Training: Turn Your SOPs Into Short Modules and Quizzes — A practical SOP-to-lesson workflow, including quiz checks, manager approval and a small-team completion sheet.
• First Aid Training Certificate Template: Issue and Renew With AI — One sheet, one certificate template, one mail merge and a renewal reminder sequence: a first aid trainer's admin loop, with AI on the wording.
• How to Start a Training Company: Accreditation, Insurance and AI — First-year costs, how to pick an accreditation route, which insurance to price, and the proposal, booking and certificate admin AI can handle.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI enterprise privacy and data-controls pages, Anthropic privacy centre and pricing page, Microsoft Learn Copilot privacy page, Google Workspace generative AI privacy hub, Google Forms and Microsoft Forms help pages, Tally pricing page (all checked 4 November 2026). Survey design points are general practice, not vendor claims.