Training Needs Analysis With AI: Survey to Client Report

A training provider can run a needs analysis with AI by writing a short survey and interview guide, anonymising the answers, asking the model to group comments into themes with quoted evidence, building a gap table from your own scoring, checking a sample against raw answers, and writing recommendations the client can act on.

The discipline that makes the difference is evidence. A client report built from smooth AI summaries that nobody traced back to what people actually said is a liability. The working rule is simple: each claim in the final report points to specific answers you have read, and anything you cannot trace comes out.

Follow me on Instagram@sagnikteaches

Decide what the client needs to decide

A needs analysis is only useful if it leads to a decision. Before writing a single question, ask the client what they will do with the result: choose between a half-day workshop and a programme, justify a budget, pick which team goes first, or settle an argument between two managers. Write that decision at the top of your project notes, because it decides what you collect.

Connect on LinkedInSagnik Bhattacharya

Then agree the scope in writing: which teams, which roles, what skills, what period. A needs analysis of "customer service training" for forty people across three shifts is a different job from one for the whole company. Also agree how anonymous answers will be, who sees raw responses (ideally only you), and what the client gets at the end. These agreements are the first lines of your privacy note to respondents, and they also protect you when someone asks later why a comment appears in a report.

Subscribe on YouTube@sagnikteaches

The data you can collect, and which AI step each needs

Most analyses blend four sources. The table shows what each gives you and where the model helps. Observation and performance data, such as error rates, are the client's evidence and you should ask for them even if they take time to find.

A survey that produces usable answers

Keep the survey to ten to fifteen minutes. Mix rating questions you can count with two or three open questions that capture detail. Ask the AI to draft, and then cut it hard, because models write surveys that are too long.

Draft a staff survey for a training needs analysis. Client: a wholesale distribution company, 45 warehouse and customer-service staff. Topic: handling difficult customer calls and order problems.
Include: 8 skill statements rated 1 (not confident) to 5 (very confident); 3 questions on how often each situation occurs (rarely/monthly/weekly/daily); 3 open questions (what makes calls hard, what help you have had, what would make the biggest difference); 2 questions on preferred learning format and time available.
Rules: plain English, no jargon, one idea per question, no leading questions, no questions about personal characteristics. Under 20 questions in total.

Edit the output in three ways. Remove double-barrelled statements, such as "I can calm angry customers and resolve complaints", which mix two skills and make a score meaningless. Remove any leading phrasing such as "How much has poor training affected you?". And make sure the scales are the same throughout, with labels at both ends.

Tell respondents how their answers will be used. A short intro such as "Answers are anonymous. Only [your name] sees individual responses. We will report themes and numbers for groups of five or more" does more for honest answers than any question wording. If you use a form tool, check where it stores responses and whether you can export them; the tutorial on Tally survey forms and AI summaries covers one simple option.

Manager interviews: a guide, and notes you can trust

Interviews tell you what the survey cannot: what a good performer actually does differently, and what the manager has already tried. Ask the AI for a guide of eight to ten open questions in sequence, from context to specifics: "Think of the last time this went well. What happened? What did the person do?" Keep the guide short enough to leave time for follow-ups.

An illustrative guide for the distributor's five managers, cut down from the fourteen questions the AI first proposed:

1. What does your team handle on a typical day, and which calls take longest?
2. Think of the last difficult call that went well. What did the person do?
3. And the last one that went badly: where did it turn?
4. What can a team member offer a customer without asking you, and do they know that?
5. How long does a new starter take to handle a refund or order problem alone?
6. What training or help have you already tried, and what happened?
7. Which problems come from systems, stock data or policy rather than skill?
8. If one thing changed in three months, what would you notice first?

Question 7 matters most for the report, because it lets managers name process problems before you mistake them for skill gaps. Question 8 hands you a success measure in the manager's own words, which is far easier to sell back to the client than one you invented.

Recording and transcribing an interview needs consent from the person. If you record, say so before you start, say where the file will be stored and for how long, and let them decline. Many consultants prefer to take their own notes and then use the AI to tidy them. In that case, paste only your notes, with names and company-specific details replaced by placeholders such as [manager A] and [Team 2]. Ask the model to restructure the notes under the guide's headings, and to mark anything it is unsure about instead of filling gaps.

Anonymise before anything reaches the model

Free-text answers often include names ("Dave in dispatch never tells us"), customers, places and details that identify people. Remove those before pasting. A quick approach is to export the answers to a spreadsheet, add a column of respondent numbers, and scan each comment, replacing names with roles in brackets. For a 45-person survey, that takes about thirty minutes and is worth doing. Then use an account that does not train on your content and check the client agreement allows it. OpenAI says it does not use ChatGPT Business or Enterprise data for training by default, and Anthropic says the same of its commercial products, which include Claude Team; a personal free account is a different matter, so check its data settings before anything goes in.

If the client's data is sensitive, such as a regulated environment or personal data about staff, ask the client in writing whether AI tools may be used and on what terms. Mention it in your proposal so it is never a surprise. Some clients will say no, and you simply do the coding by hand, using the same method with a spreadsheet. The method matters more than the tool.

Theme coding with evidence attached

Coding means reading every comment and tagging it with a theme. The model can do a first pass in minutes, but only if you demand evidence for each theme. Paste the anonymised comments with respondent numbers and use a prompt like this.

Below are 41 anonymised free-text answers from a staff survey, each with a respondent number. Question: "What makes difficult customer calls hard?"
Task: group the answers into 5 to 8 themes. For each theme give: a short label; a one-sentence description; the respondent numbers that fit; a count; and two verbatim quotes. A comment may fit more than one theme. List comments that fit no theme separately. Do not infer anything that is not in the text, and do not rank themes by importance, only by count.
[paste numbered answers]

A realistic reply gives you something like: "Not knowing what I am allowed to offer (14 respondents)", "Customers who have already chased several times (11)", "Unclear stock information (9)", "No time between calls (6)", with quotes and numbers. Now check it. Open the spreadsheet, filter on the respondent numbers for the biggest theme and read the comments. If you find ones that do not fit, or fit a different theme better, tell the model and re-run, or correct the counts yourself.

Two failures are common. The first is merged themes: "communication issues" absorbs five different problems and hides the one the client can fix. Ask for splits where a theme exceeds a third of responses. The second is false precision: a count of "14" from a model that never counted. Always recount the biggest themes yourself against the spreadsheet.

Building the gap table from scores, not from the AI

The gap table is the centre of the report. For each skill, it shows the level the job needs, the level people report, and the evidence for the gap. The AI should not decide the target level. You and the client do, from the job requirements and manager interviews. Then calculate the numbers in a spreadsheet, as the tutorial on scoring a team's skills against target levels shows for AI skills; the same structure works for any skill.

The figures above are illustrative. The valuable line is the second one: a gap that looks like a training need but is really a system problem. A good report says so plainly, because recommending a course for a problem a course cannot fix damages trust. This separation of training needs from process problems is what clients rarely get from a generic template.

Ask the AI to review the table, not build it: "Here is my gap table and the themes. Which gaps might be caused by something other than skills, and what evidence would show it?" Treat the answer as a list of questions to ask the client.

From raw ratings to a gap figure you can defend

Averages hide things, so work the numbers in a spreadsheet where the client can follow every step. Take the first skill in the table. Fourteen customer-service staff rated "I can explain what I am allowed to offer a dissatisfied customer" on the 1 to 5 scale, and their illustrative answers were 1, 2, 2, 2, 3, 3, 3, 3, 2, 4, 3, 2, 3 and 3. They add up to 36, so the average is 36 divided by 14, or 2.57, shown as 2.6. Against a target of 4 the gap is 1.4. In the sheet that is =AVERAGE(C2:C15) for the average and =COUNTIF(C2:C15,"<=2")/COUNT(C2:C15) for the share scoring 1 or 2, which here is 6 of 14, or 43 per cent.

Now split the same column by length of service, which the survey should ask. Say five of the fourteen joined in the last six months and scored 1, 2, 2, 2 and 2, an average of 1.8, while the other nine scored 27 between them, an average of 3.0. That one split changes the recommendation. A whole-team workshop would spend most of its time on people who are nearly there; a short onboarding module for new starters, plus a one-page refund-limits checklist, targets the real gap. It also matches the manager who said new starters take about six weeks to handle refunds alone.

To rank skills, weight each gap by how often the situation comes up, using the survey's own frequency question: daily 4, weekly 3, monthly 2, rarely 1. Multiply the gap by the weight and sort.

The recall row shows why the weighting matters. It has the biggest gap of all, but staff meet it a few times a year, so a one-page procedure kept by the phone may serve better than training time. Put the weights in the report's method note so the client can disagree with one and see what changes. If you involve the AI here, ask it to check your formulas and point out outliers, not to produce the averages: a spreadsheet leaves a trail the client can audit, and a chat answer does not.

Spot checks against the raw answers

Before any number or theme goes into the report, check it. Here is a routine that takes about forty minutes for a survey of around forty responses.

1. Recount the three largest themes by filtering the spreadsheet. Every count in the report must match.
2. Pick ten respondent numbers at random and read their full answers. Does each fit the theme the model gave it?
3. Read every comment the model put in "no theme". Real signals hide there, such as a single comment about a safety concern.
4. Check each verbatim quote by searching for it in the source. Models sometimes tidy quotes into words nobody said.
5. Recalculate each average in the gap table from the raw scores, and check the response rate.
6. Look for what is missing: a team with no responses, a role under-represented, a question skipped by many people.

Log what you found: how many corrections, and of what kind. If you correct more than one in ten items, tighten the prompt or code more by hand. Over time that log becomes a record of how reliable the process is, and a good thing to mention to a client who asks how you used AI.

Tidying interview notes: a before and after

Raw interview notes are fast to write and hard to use. Here is an illustrative before, then the kind of structure you ask the model to produce from your own placeholder-only notes.

Before (your notes): [Manager B] - new starters take ~6 wks to handle refunds alone. shadowing is ad hoc. best agent [Agent 3] uses a checklist she made herself. complaints spike after promo weeks. wants "scripts" but I think it's about knowing limits.

Prompt: Restructure these notes under the headings Context, What good looks like, Gaps seen, What has been tried, Manager's own view, and My interpretation. Keep my wording where possible. Mark anything you are unsure of with [?]. Do not add facts.

The result separates what the manager said (wants scripts) from your interpretation (the real gap is knowing limits), which is the habit that keeps a report honest. Check that the model did not move your interpretation into the manager's mouth; that is the commonest slip, and it quietly turns your opinion into a finding.

When the response rate is low or the groups are small

Sometimes only a third of staff reply, or a team has four people. Do not hide it. Report the response rate, say who is missing, and word findings carefully: "of the 14 who answered", not "staff think". For small groups, merge teams so nobody is identifiable, or report in words only. Ask the AI to flag any statement in your draft that generalises beyond the respondents, and soften each one.

A low rate is also a finding in itself. If night-shift staff did not answer, ask why: the survey was sent at the wrong time, or they do not trust that it is anonymous. You can add a short follow-up, such as a ten-minute group conversation with a few of them, to fill the gap before the report goes out. Then say in the method note what you did.

Recommendations a client can act on

Each recommendation needs four parts: what to do, for whom, how it will be known to have worked, and what it is likely to cost in time or money. Build recommendations from the gap table, in priority order, and separate the training from the non-training fixes.

A sample recommendation for the distribution company: "Run a half-day workshop for the 14 customer-service staff on refund and goodwill decisions, using the company's own refund policy and five real anonymised cases. Success looks like a fall in refund escalations from 11 a month to under 6 within eight weeks. Separately, the operations manager should review the stock-status screen, which 9 respondents described as unclear; this is a process fix, not training." Notice the specific numbers, the named audience and the split between training and process.

Ask the AI to draft the wording of each recommendation from your bullet points, and tell it to use only the figures you supply. Then check that every number in the text appears in your analysis. The same scoping rules apply if the recommendation becomes a proposal, a topic covered in the training proposal tutorial.

The client report: a structure that reads well

Most clients read the first page and the recommendations. Design for that. A structure that works is: a one-page summary with the decision the client needs to make; who took part and how; findings by theme with quotes and numbers; the gap table; training needs versus process fixes; recommendations with timing and success measures; and an appendix with the questions and method.

The summary should lead with the answer, such as "Of the eight skills assessed, three show gaps worth training; two further problems need process changes." Use plain language and avoid decoration. Quote respondents sparingly, choose quotes that show the range of views and not only the dramatic ones, and never use one that could identify a person. Where fewer than five people fall in a group, combine groups or report in words.

Here is an illustrative summary page for the distributor, filled in from the figures above. It is the part most clients will actually read, so every line carries a number or a decision:

• Decision needed: whether to fund a refund-decisions workshop for customer service this quarter, and who owns the stock-screen fix.
• Who took part: 41 of 45 staff answered the survey (91 per cent); five managers were interviewed; three months of refund-escalation figures were reviewed.
• Main finding: staff are least confident about what they may offer an unhappy customer (average 2.6 against a target of 4), and the gap sits mainly with people in their first six months (1.8).
• Not a training need: nine respondents described the stock-status screen as unclear; this is a system fix for operations.
• Recommendation: a half-day workshop on refund and goodwill decisions for the 14 customer-service staff, plus a one-page limits checklist added to onboarding.
• How we will know: refund escalations fall from about 11 a month to under 6 within eight weeks, reviewed together at the follow-up.

Read it back against your spot-check log before sending. Every figure on that page should appear in the spreadsheet, and every finding should trace to quotes you have read yourself.

Be honest about AI in the method note: "Free-text comments were grouped into themes with help from an AI tool, and all themes, counts and quotes were checked against the original responses." One sentence is enough. It is accurate, and it is far easier to say up front than to explain after a client finds out.

A worked example, with time and numbers

An illustration for a training provider serving a 45-person wholesale distributor. Forty-one staff answer the survey (a response rate of about 91 per cent). The provider interviews five managers for thirty minutes each. Time spent: three hours designing the survey and guide, one hour anonymising answers, one hour coding with the AI and a further forty minutes of spot checks, two hours building the gap table and reading performance data, three hours writing and editing the report. About eleven hours of desk time in all, plus two and a half hours of interviews. Coding the same 41 answers by hand and writing up without AI help might take this provider sixteen to twenty hours; all of these figures are illustrative, so time your own first project and price the next one from it.

The spot checks catch two problems: the model had assigned seven comments about stock data to the theme "customer anger", and a quote had been tidied from "I just guess at what I'm allowed to give" to "I am unsure what I am allowed to offer". Both are corrected. If those checks had been skipped, the report would have overstated one theme and altered a respondent's words, and the client's operations manager would have noticed.

Reusing the method across clients

After the first client, turn the process into a kit: a survey template with swappable skill statements, an interview guide, a prompt for theme coding, a gap-table spreadsheet, a spot-check log and a report skeleton. Keep it in a Claude Project or a ChatGPT Project so the instructions persist, but never store client answers there unless the plan and client agreement allow it. Name files by client code, not by name.

The tutorial on summarising training feedback forms for a client report covers the end-of-course version of this same discipline. Together they give you an evidence trail from the first discovery call to the follow-up measure, which is the evidence a client needs when someone asks whether the training budget worked.

Questions about AI in a training needs analysis

Can I upload survey responses to ChatGPT or Claude?

Only if your client agreed and the tool's terms suit the data. Strip names and identifying detail first, use a business plan such as ChatGPT Business or Claude Team, which their makers say are not used for training by default, and tell respondents in the survey how answers will be processed.

How many responses do I need for a credible report?

There is no magic number. Report the response rate and who answered, be cautious with groups under about ten, and do not present themes from three people as a company-wide finding. Interviews can add depth where the survey is thin.

How do I stop AI inventing themes that were not there?

Ask it to quote at least two raw answers for every theme, count how many responses fit each, and then check a random sample yourself. Drop any theme you cannot trace back to specific answers.

Should the client see the AI-generated analysis?

Show the findings, the evidence and your recommendations. You need not show the prompts, but say honestly that AI helped to group comments and that you checked the results against the raw answers. Saying so up front is far easier than explaining it later.

What is the difference between a training need and a want?

A need is a gap between what the job requires and what people can do now, backed by evidence such as errors, rework or observed performance. A want is a preference, such as a favourite topic. Your report should separate the two.

Further reads

• How to Analyse Customer Feedback Surveys With AI — General survey-analysis method with AI.
• How to Use AI in Business Training: Quizzes That Show What Stuck — Quizzes that test what stuck after the training.
• Best CRM for a Training Company: Bookings and AI Follow-Ups — A CRM to track client discovery and follow-ups.
• AI for Corporate Training Companies: Proposals to Certificates — Admin jobs across a training company worth handing to AI.
• How to Choose an AI Training Provider for Your Team — How clients judge a training provider, from their side.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI's business data page and Anthropic's privacy centre (business-plan data not used for training by default), and OpenAI's and Anthropic's Projects help articles, all read 4 November 2026.