How to Use AI in Business Training: Quizzes That Show What Stuck

A small training provider can use AI to turn its own course material into quiz questions, read feedback forms for themes, and send a short retention quiz two weeks later. A quiz at the end of the course shows what learners remember that day. The two-week quiz shows what stuck, and that is the one worth building.

Most providers measure the wrong moment. An end-of-day quiz mostly tests short-term memory, and a happy-faces feedback form tests the mood of the room. Both look fine on a good day. The two-week check is where gaps show up, and AI makes it cheap enough to run for every cohort rather than once a year. Here is how to build the three pieces, with the numbers from one illustrative course.

Follow me on Instagram@sagnikteaches

The three pieces and what each one tells you

The method has three parts that feed each other. Quiz questions built from your own slides and handouts check recall and application. A feedback form captures how learners felt about the delivery, the pace and the materials. A retention check, run 14 days after the course with fresh questions on the same learning points, shows what remained once the room was a memory.

Connect on LinkedInSagnik Bhattacharya

Keep the three separate. When providers merge feedback and quiz into one form, learners rush the quiz or skip the comments. The quiz needs a quiet few minutes; the feedback needs a moment of honesty.

Subscribe on YouTube@sagnikteaches

A worked example: a two-day customer-service course, twelve learners

Say a small training provider runs a two-day customer-service course for staff in local businesses. Twelve learners attend the March cohort. The provider has a slide deck of 46 slides, a three-page handout, and six learning points written on the course outline: greet within ten seconds, restate the problem, offer two options, confirm the next step, log the contact, and follow up within two working days. Everything below uses those six learning points, and every number is illustrative.

The session quiz on day two scores an average of 82 per cent. The provider is pleased. Fourteen days later the retention quiz, on the same six points but with new wording, averages 60 per cent. Broken down by learning point, "greet within ten seconds" holds at 83 per cent, "restate the problem" at 75, "offer two options" at 58, "confirm the next step" at 54, "log the contact" at 50 and "follow up within two working days" at 42. A single average would have hidden the real story: the last two points, the ones that happen after the customer leaves, are what people forget.

That finding changes what the provider does. It adds a five-minute practice of the follow-up step at the end of day two and a day-three reminder email. The next cohort's retention check is the proof. The sections below build each piece so that the same sort of finding is available to you.

Build session quiz questions from your own material

Start with what you teach, not with a generic quiz generator. Questions that do not match your handout teach nothing and annoy learners. Paste or upload your slides' text (or the handout) into a chat assistant, and tell it the learning points. Use a business or team plan that does not train on your content by default, and keep learner names out of anything you paste.

Several tools can do this. ChatGPT, Claude and Gemini all take pasted text or a document and will draft questions. Gemini Notebook (formerly NotebookLM) lets you load your documents as sources and generate a quiz from them with the Quizzes option in its Studio panel; the pencil icon lets you set the difficulty and add a prompt about focus, which keeps the AI closer to your material. Google's help page offers a CSV download for flashcards but not for quizzes, so copy the questions you keep into your form tool by hand. Google Forms and Microsoft Forms hold the final quiz and score choice questions against the answer key you set.

A prompt that gets usable questions

You are helping a small training provider write a quiz.
Source: the handout text below (do not use outside knowledge).
Learning points: 1) greet within ten seconds 2) restate the problem
3) offer two options 4) confirm the next step 5) log the contact
6) follow up within two working days.

Write 12 multiple-choice questions, two per learning point.
- Four options each, one correct, with plausible wrong answers
  drawn from real mistakes people make at work.
- At least six questions must be short scenarios, not definitions.
- Give the correct letter and a one-sentence explanation.
- Tag each question with its learning point number.
Handout text:
[paste text]

The tagging is what makes the retention analysis possible later. Without it you only get one average.

What the output looks like, and what you must fix

Here is an illustrative output for learning point 3:

Q7 (LP3). A customer says the delivery date you quoted is too late.
What should you do first?
A) Apologise and offer a refund
B) Offer two options, such as a faster service at extra cost or the
   original date with a goodwill discount
C) Explain why the date cannot change
D) Pass the call to your manager
Correct: B. Offering two options gives the customer a choice and
keeps the problem with you.

Read it as a learner would. Question seven is decent, but option A is arguably also a reasonable first move, and a good learner could argue for it. Fix the stem to "What does the course recommend you do first?" so that it tests the course's rule, not general judgement. Other typical fixes: one option is far longer than the rest and gives the answer away, two wrong answers say the same thing, or the explanation quotes a policy your course does not mention. Delete any question that relies on facts outside your handout. A rule of thumb is that you will keep about 8 of every 12 drafted questions without editing, fix three, and bin one.

Make the quiz teach as well as test

In Google Forms, open Settings and turn on "Make this a quiz". On each question use the Answer key control to pick the right option and set the points, then "Add answer feedback" to paste the one-sentence explanation. Under Release grades in Settings, choose "Immediately after each submission" or "Later, after manual review", and use the respondent settings to decide whether people see their missed questions, correct answers and point values. Immediate feedback helps learning on the day; for the day-14 check you may prefer to release results later so that learners cannot forward the answers.

Turn feedback forms into themes you can act on

A feedback form should have six questions at most: three rated (pace, materials, trainer) from one to five, and three open-text. Open-text questions are where AI saves hours: "What was the most useful part?", "What was confusing or missing?" and "What one thing would you change?"

Export the responses to a sheet, delete the name and email columns, and paste the open-text answers into a chat assistant. For a larger cohort, the tutorial on analysing customer feedback surveys with AI covers sheets of hundreds of responses. For twelve learners you can paste them in one go.

A feedback analysis prompt, with illustrative input and output

Below are 12 anonymous answers to "What was confusing or missing?"
from a two-day customer-service course.
Group them into themes. For each theme give: a short label, a count,
and one quote. Then list comments that fit no theme.
Do not invent comments. If a count is uncertain, say so.
[paste 12 answers]

An illustrative result:

Theme 1: Not enough time on follow-up and logging (5 of 12)
  "We rushed the last hour; I still don't know what a good log looks like."
Theme 2: Handout too dense (3 of 12)
  "Page two is a wall of text."
Theme 3: Wanted real examples from my own industry (3 of 12)
  "Examples were all retail."
Unthemed: 1 comment about the room temperature.

Notice how theme one lines up with the retention check: five of twelve said they did not understand logging and follow-up, and those were the two lowest-scoring learning points. Feedback comments and retention scores tell the same story from different angles, and a provider who reads both together can be confident about where to spend improvement time. Check the counts yourself: tally by hand against the raw sheet before you act, because a chat assistant will sometimes round a count of four up to five or merge two themes.

Two cautions about feedback

First, learners write in confidence about trainers and colleagues. Remove names, and do not paste free-text that identifies a person into a consumer-plan chatbot. Second, a rating scale rewards a pleasant day. Do not use the average rating as your quality figure, because a well-liked course can leave people unable to do the job. Treat the ratings as a mood gauge and the open-text themes as the diagnosis.

Run the two-week retention check

The day-14 check has one purpose: it measures what remains once the course is a memory. Choose day 14 because it falls after the first practice in real work, and far enough out that cramming has faded. Send it from a calendar reminder on the day the course ends, so you do not forget.

Write parallel questions, not the same ones

If you reuse the session quiz, learners may recall the answer letters instead of the point. Ask AI for parallel questions: same learning point, different scenario and different wording. Use this prompt, with the learning points tagged as before:

Here are the 12 questions from our session quiz and the six learning
points. Write 8 new multiple-choice questions, covering all six points
(at least one each, extra on points 4-6). Use different scenarios and
different wording from the originals, and do not reuse any option text.
Same format: four options, correct letter, one-sentence explanation,
learning point tag.

Run an extra review for fairness: could a learner who never went on the course guess the answer from common sense? If so, the question tests nothing. Ask a colleague who has not attended to take the quiz cold; if they score above 50 per cent on four-option questions, rewrite the easy ones. Pure chance on four options is 25 per cent, so that is the floor your numbers have to beat.

Send it, collect it, summarise it

Create the quiz in Google Forms (or Microsoft Forms), add a first question asking for a learner code, not a name (for example the cohort month and a number), and email the link with a two-line message: "Eight questions, about eight minutes, no pass mark. It helps me improve the course." Learners are more honest when it is clearly low-stakes. A reminder after three days lifts response rates; accept that you will not get everyone, and note the response count next to every figure you report.

Then export and ask AI for a per-point summary. This is a calculation job, so prefer a tool that can run the numbers, and check the totals by hand for at least one learning point.

Attached is a sheet: one row per learner, one column per question,
1 = correct, 0 = wrong, plus a row with the learning point tag.
Calculate, per learning point, the percentage correct. Compare with
the session quiz figures I pasted below. List the three points with
the biggest drop. Show the working so I can check it.

"Show the working" matters. If a tool reports 58 per cent for "offer two options" and the sheet has 7 of 12 correct, that is 58 per cent, and you can see it. If it reports 67 per cent, its working will show 8 of 12, and a recount of the sheet's 7 tells you one answer was misread, often a blank cell treated as correct. For a deeper look at checking numbers a chat assistant produces, see the ChatGPT and Excel guide.

Reading the day-14 results as a table

Lay the two sets of figures side by side before you decide anything. Here are the illustrative numbers from the twelve-learner cohort, with the feedback themes alongside. All 12 learners took both quizzes. The session quiz had two questions per point, so each point's figure is out of 24 answers; the day-14 check had eight questions, one each on points 1, 2, 3 and 6 and two each on points 4 and 5, so those figures are out of 12 or 24 answers.

Two readings come out of it. Points one and two fall by single digits, so the course already teaches them well enough to keep. Points five and six fall by 25 and 33 points and are also the ones learners said they did not understand, so they are the priority. Point three drops sharply but nobody complained, which suggests learners felt fine in the room and lost it afterwards: a spaced reminder is the likely fix, not a rewrite of the slides.

With twelve learners, one person moves a one-question figure by about eight points, so do not chase differences of one or two answers. Treat anything under ten points as noise, and make big decisions only from patterns that repeat across two cohorts.

Check that the check worked

Before you trust the figures, run three sanity checks. Confirm the response count (all 12 replied in this cohort; in most cohorts a few will not) and note it beside every percentage. Re-add one learning point by hand from the raw sheet. And look for learners who scored 100 per cent in under two minutes, which usually means they guessed fast or looked at the handout; that is fine for a low-stakes check, but do not count it as proof of strong retention.

Other places this approach pays off

The same three-piece pattern changes shape with the kind of course. Here are five short cases, all illustrative.

• A software skills course (Excel basics). Replace multiple choice with three practical tasks, such as "Write a formula that totals column C only for rows marked Paid". Day-14 check: the same style of task on a different sheet. AI can write the practice sheet and the answer key; you verify the key by solving it yourself.
• A compliance refresher for care staff. Scenario questions only ("A client's family member asks for a copy of notes. What do you do?"). Add a confidence rating after each answer, from one to five. Wrong answers with high confidence are your priority for retraining.
• A short sales course. Use a role-play recording or written reply as the day-14 task; ask AI to score it against a four-line rubric, then spot-check three scores yourself. The tutorials on building a staff training programme with AI and induction packs and refresher quizzes show similar layouts for in-house teams.
• A hands-on trade course. A written quiz does not show skill. Use it for knowledge (safety limits, sequences) and add a one-minute photo or video task at day 14, judged by a person.
• An online self-paced course. Set the day-14 email to send automatically from your email tool 14 days after enrolment, with the quiz link inside. The tutorial on AI answering student questions for an online course shows how the same material can support a help desk afterwards.

Where AI-built quizzes mislead you

These are the failures that cost providers credibility. Learn the signs before they show up in a cohort.

• Questions that reward test-taking tricks. If the correct answer is always the longest or always option B or C, learners notice. In Google Forms, turn on Shuffle option order from each question's More menu, and check the answer pattern across the quiz.
• Answers drawn from the model's memory, not your handout. The AI invents a rule that your course never taught. Add "use only the text below" to every prompt and read the explanations for anything unfamiliar.
• A false sense of progress. Quiz scores can rise because the questions got easier. Keep a master set of three or four anchor questions that stay unchanged each cohort so you can compare like with like.
• Unfair marking of open text. If you use AI for rubric scoring, spot-check ten answers per cohort and keep the rubric in the prompt each time. Mark certification yourself.
• Pasting learner details. The retention sheet should contain a code, not a name. Keep the key linking codes to names in a separate file.

A cost and time sum for one cohort

For a twelve-learner course, estimate the work like this, again illustratively. Drafting and editing twelve session questions: about 40 minutes. Writing the feedback form: 15 minutes, once, then reused. Analysing feedback with AI plus the hand tally: 25 minutes. Writing eight parallel retention questions and building the form: 35 minutes. Day-14 summary and your notes: 30 minutes. That is about two hours and 25 minutes the first time, and roughly 80 minutes for later cohorts because the forms already exist. At $40 an hour of your time, the first cohort's added cost is under $100, plus whatever you already pay for a chat assistant and a form tool (Google Forms is free). A business-tier chat plan that does not train on your content is the main extra expense, and many providers already have one.

Using what you find

A retention check is only worth the effort if it changes the next course. After each cohort, write down three lines: the lowest learning point, what you changed because of it, and the target for next time. Share the lowest point with the trainer, not as a verdict but as an agenda item. A drop in "log the contact" may mean the handout hides it on page three, or that the practice exercise had no logging step. Try the cheapest fix first, which is usually a short practice and a reminder email, and see what the next cohort shows.

If you also build SOP-style modules for in-house teams, the tutorial on turning SOPs into short modules and quizzes with AI shows the same quiz-building step on process documents. The retention check applies to those just as well.

Quiz and feedback questions from training providers

How many quiz questions are enough for a retention check?

Eight to twelve questions, taking under ten minutes, is enough for a short course. Fewer than eight makes a single question swing the average. More than twelve and learners start skipping or guessing, which tells you about their patience, not their memory.

Can AI mark open-text answers?

It can give a first-pass score against a rubric you supply, but treat that as a draft. Check a sample of ten answers by hand before trusting the rest. Keep high-stakes marking, such as certification, with a person, and tell learners how the first pass works.

Should learners know the quiz is AI-generated?

Say so in one line if you use it for anything that counts towards a result. For a low-stakes retention check there is no harm in telling them anyway. What matters more is that you read every question and answer key yourself before it goes out.

Is it safe to paste learner feedback into a chatbot?

Remove names, emails and any detail that identifies a person before you paste, and use a business plan that does not train on your content by default. Feedback often contains personal remarks, so the same care applies as with customer data.

What if the two-week scores are low?

Look at which questions dropped, not the average. If the drop is in one topic, the teaching or the material for that topic needs work. If it is across the board, the course may be too long for one sitting, or the follow-up reminders too thin.

Further reads

• How Much Does It Cost to Create an Online Course With AI? — What it costs to build a course around the material.
• Turning One Online Course Into Emails and Posts With AI — Reuse course content for the follow-up messages in week two.
• Bubble Tea Staff Training: Drink Recipe Cards and Quizzes With AI — Quizzes for hands-on recipes, a different kind of retention.
• Fast Food Employee Training: Reference Cards and Quizzes With AI — Reference cards that support what the quiz checks.
• Pharmacy Staff Training: Induction Packs for Counter Staff With AI — Build induction sheets and quizzes for counter and dispensary-support staff from your own SOPs with AI, then have the pharmacist approve every version.
• AI Tools for Learning and Development: A Small-Company Shortlist — A job-by-job shortlist for whoever runs training in a small firm: skills matrix, training plan, onboarding and checks, using tools you may already have.
• Can AI Summarise Training Feedback Forms for a Client Report? — Score ratings in a spreadsheet, theme the comments with AI, and produce a one-page client summary that protects delegates' anonymity.
• Joining Instructions Template: Details and Reminders With AI — What goes in course joining instructions, a filled-in template, and an AI-drafted reminder sequence merged from the booking sheet.
• Training Course Cancellation Policy: Terms Drafted With AI — How a training company drafts cancellation terms with AI: notice tiers, substitutes, venue costs, postponement and refund wording for review.
• Training Needs Analysis With AI: Survey to Client Report — How a training provider runs a needs analysis with AI, from survey and interviews to a gap table and a client-ready report with spot checks.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Google Docs Editors Help (Forms quiz settings), Google Workspace user help (shuffle option order), Gemini Notebook Help (flashcards and quizzes), Microsoft Forms support (create a quiz), OpenAI and Anthropic business privacy pages, checked 4 November 2026.