Training Needs Analysis: How to Find Out Whether Training Is the Answer

What Is a Training Needs Analysis?

A training needs analysis is the process of establishing what is going wrong, whether missing knowledge or skill is the cause, and what a person would have to be able to do differently for the problem to stop. It has three outputs and only one of them is a course.

The first output is a decision about whether training is the right intervention at all. The second, if it is, is a short list of the specific things a named group of people must be able to do afterwards. The third is the evidence you will look at later to find out whether it worked, agreed now rather than invented at the end.

Most of the discipline that surrounds this has grown a long way past what a working L&D team can run. You do not need a thirty-page document. You need to be able to answer six questions in writing, and the reason the analysis exists is that the answers are frequently embarrassing.

Why Do Most Training Requests Not Survive One?

Because a training request usually arrives as a solution rather than a problem. "We need a course on the new returns policy" is a solution. The problem underneath it might be that the policy is genuinely unknown, or that it is known and ignored, or that the system will not let anybody follow it, or that one store is responsible for most of the errors and the rest are fine.

Only the first of those four is solved by a course. The other three are solved by a manager, a software change and a conversation with one store, and a course aimed at any of them will be measured as a failure by whoever asked for it.

Six ways a request falls apart under a single round of questioning:

1. The problem is a symptom. Error rates are up, so training is requested. Nobody has looked at whether the errors cluster in one team, one shift or one product line, and clustering almost always means something other than knowledge.
2. The behaviour is understood and not performed. People can describe the correct procedure accurately when asked. They do not follow it because it takes longer, or because nobody checks, or because the fast way is rewarded.
3. The tool prevents the behaviour. The correct process requires a field that the system marks optional, a screen that times out, or a form that is only available on a desktop nobody on that shift has.
4. There is no shared definition of correct. Two supervisors would mark the same piece of work differently. No course can teach a standard that does not exist yet.
5. The audience is wrong. The request names everybody because naming everybody avoids an argument. The actual gap sits with new starters, or with one region, or with the people promoted in the last quarter.
6. Nobody has said what would change. If the answer to "what will be different in eight weeks" is "people will be more aware", there is no analysis, and there will be no way to defend the spend afterwards.

Is It a Can't-Do or a Won't-Do?

This is the single most useful test in the whole exercise, and it takes one question: if their job depended on doing it correctly today, could they?

If the answer is no, you have a can't-do. Something is genuinely missing: knowledge, a skill, practice, or a reference they can reach at the moment of the task. Training is a plausible answer, and often the right one.

If the answer is yes, you have a won't-do, and training will not fix it. The cause sits in one of five places instead, and each has a different owner:

1. Consequences. Doing it the wrong way is faster and nothing happens. Doing it the right way is slower and nothing happens either.
2. Feedback. Nobody finds out they got it wrong, so there is nothing to correct against.
3. Tools and process. The correct behaviour is harder than it needs to be, or occasionally impossible.
4. Priorities. Two instructions conflict and the other one is louder, usually because it is measured.
5. Expectations. The standard has never been stated plainly to the person expected to meet it.

Running a course against a won't-do is the most common expensive mistake in corporate learning, and it is expensive twice: once for the build, and once when the numbers do not move and the credibility of the next request is lower.

There is one honest exception worth naming. Sometimes a short piece of training is commissioned specifically so that the organisation can demonstrate it was communicated. That is a legitimate compliance purpose, but it is a documentation goal rather than a performance goal, and mixing the two is how a programme ends up being judged against an outcome it was never designed to produce.

The Four Levels Worth Looking At

Look at the problem from four distances. Most requests have been examined at only one of them, usually the third.

The job or task level is the one most often skipped and the one that most changes the content. A procedure document tells you what somebody wrote down once. Watching the task, or asking the person who does it best to talk through it, tells you where the real decisions are, and the real decisions are what the training has to be about. That is also why the best person to talk to is rarely the person who requested the course.

What Data Do You Already Have?

Teams commission surveys because a survey feels like analysis. Surveys are the weakest evidence in the list, because they ask people to self-report a gap they may not be able to see. Start with the records that already exist and cost nothing.

1. Quality and error records. Rework, returns, failed inspections, complaints, reopened tickets. Look for clustering before you look for a total, because a cluster points at a cause and a total points at nothing.
2. Support and helpdesk tickets. The questions customers and staff keep asking are a direct readout of what is not understood. The repeat question is the syllabus.
3. Incident and near-miss reports. In safety-critical work these are the most honest data you will get, because somebody had a reason to write them accurately.
4. Existing assessment data. If any current training has questions in it, the per-question results tell you which specific points fail rather than which courses were completed. Completion tells you almost nothing.
5. Time to competence. How long a new starter takes to reach normal output, and how much of that time is spent waiting for an experienced colleague to be free.
6. Exit and probation conversations. People leaving say things people staying do not.
7. Observation. Half a day watching the work is worth more than a hundred survey responses, and it is the only method that reliably finds the step everybody skips.
8. Interviews with three people, not thirty. The person who does the job best, the person who joined most recently, and the supervisor who signs the work off. Ask each what goes wrong most often. Where the three agree, you have found something.

Add a survey last, if at all, and only to test a hypothesis you formed from the list above. A survey asking "what training do you need?" produces a shopping list. A survey asking "which of these five steps do you feel least confident doing unsupervised?" produces data.

How Do You Run One in a Day?

Nobody is going to give you six weeks. A day, run in this order, gets most of the value.

1. One hour with the requester. Ask what will be different afterwards, who exactly is affected, and what they are seeing that made them ask. Write the answers down in their words.
2. Thirty minutes on the numbers. Pull whatever quality, ticket or assessment data exists and look only for clustering by team, site, shift, tenure or product.
3. Ninety minutes of observation or three short interviews. Watch the task, or talk to the best performer, the newest joiner and the supervisor.
4. Ten minutes on the can't-do test. Ask it about the specific behaviour, out loud, of somebody who does the work. Record the answer honestly even when it is inconvenient.
5. Thirty minutes to write the brief. Six answers, one page, in the next section's shape.
6. Twenty minutes to send it back. The requester either agrees with the one page or corrects it, and either outcome is worth more than the day cost.

If step four returns a won't-do, stop and say so in writing, with the alternative named. That conversation is unpopular once and useful permanently.

What Should the Analysis Actually Produce?

One page, six answers. If you cannot fill this in, the analysis is not finished, and a course built now will be built on an assumption.

1. The problem, in observable terms. What is happening that should not be, or not happening that should, stated so that somebody else could go and see it.
2. The audience, named narrowly. The specific group, with a number. "All staff" is almost always wrong and always more expensive.
3. The verdict. Can't-do, won't-do, or both, with the reason. If both, say which part training is taking and who owns the rest.
4. What they must be able to do afterwards. Three to five behaviours, each starting with a verb somebody could watch. Not "understand the policy". Instead, "check the two fields the policy requires before approving a refund".
5. The evidence you will look at. Which existing number should move, by roughly how much, by when. Agreed with the requester now, in writing, while it is still cheap to disagree.
6. What is out of scope. The parts of the problem training is not addressing, and who has them. This single line is what protects the programme from being blamed for the tool and the incentives.

Keep it to one page deliberately. A long analysis document gets read once and filed. A one-page brief gets argued about, which is the point.

Where Does AI Help, and Where Does It Not?

AI is genuinely useful either side of the analysis and not much use inside it.

It does not run the analysis. A model asked to produce a needs analysis will produce a plausible one for a problem you did not describe, because it has no access to your error data, your two disagreeing supervisors or the field the system marks optional. The output reads convincingly and is about a generic organisation.

It is useful for two jobs at the edges. Before the analysis, it can turn a pile of support tickets or free-text survey responses into themes faster than you can read them, which is a summarising job rather than a judgement one. After the analysis, once you know which behaviours the training is for, it removes most of the production work between the brief and a working draft.

That division matters because the production stage is where training programmes used to die. When a build takes three weeks, the analysis gets skipped to protect the deadline. When the build takes an afternoon, there is no longer an excuse for skipping the half-day that decides whether the build should happen.

How This Fits With QuikAuthor

QuikAuthor sits after the analysis, not inside it, and that is worth stating plainly rather than implying otherwise. It will not run your needs analysis, it will not verify that your source document is current, and it will not design your evaluation.

What it changes is the cost of acting on the brief. Once you have the three to five behaviours from the one page, upload the procedure, the policy PDF or a recording of the person who does the job best, and the AI Asset Converter transcribes and structures it into lessons with knowledge checks drawn from what the source actually says. A short, narrow course for the narrow audience your analysis identified takes an afternoon instead of a quarter, which is what makes "train only the forty people who need it" a practical answer rather than a theoretical one.

Two features earn their place specifically because of how an analysis ends. Assessment results in QuikAuthor go down to individual question responses, not just completions, so the evidence you agreed in answer five is available per behaviour rather than per course. And the knowledge checks generated from a document are traceable to passages in it, which is what stops a question being true in general and absent from your actual procedure.

If the verdict was a won't-do, none of that helps, and building the course anyway is the mistake this guide exists to prevent. The training ROI guide covers the measurement end once you have a genuine can't-do to act on.

Frequently Asked Questions

What is a training needs analysis?

A training needs analysis is the process of establishing what is going wrong, whether missing knowledge or skill is the cause, and what a named group of people must be able to do differently for the problem to stop. It produces three things: a decision about whether training is the right intervention at all, a short list of behaviours the training is for, and the evidence you will look at afterwards to judge whether it worked.

How do you know whether a problem is a training problem?

Ask one question about the specific behaviour: if their job depended on doing it correctly today, could they? If the answer is no, something is genuinely missing and training is a plausible answer. If the answer is yes, it is a won't-do, and the cause sits in consequences, feedback, tools, conflicting priorities or unstated expectations. Training will not move any of those five.

What are the four levels of a training needs analysis?

Organisation, job or task, person, and cause. The organisation level asks whether the problem matters to something the business already measures. The job or task level establishes what doing the work well actually involves. The person level identifies which specific people cannot do it. The cause level asks why the gap exists and whether knowledge is one of the reasons. The job or task level is the one most often skipped and the one that most changes the content.

What data should a training needs analysis use?

Start with records that already exist: quality and error data, support tickets, incident reports, per-question results from existing assessments, and time to competence for new starters. Add observation of the work and short interviews with three people, the best performer, the newest joiner and the supervisor who signs the work off. Put a survey last, and only to test a hypothesis you already formed, because a survey asks people to self-report a gap they may not be able to see.

How long should a training needs analysis take?

Half a day to a day for most requests. An hour with the requester, half an hour looking for clustering in the data you already hold, ninety minutes observing the work or interviewing three people, ten minutes on the can't-do test, and half an hour writing a one-page brief. A longer document is usually read once and filed, while a one-page brief gets argued about, which is what makes it useful.

What should a training needs analysis document contain?

Six answers on one page: the problem in observable terms, the audience named narrowly with a number, the verdict of can't-do or won't-do with the reason, three to five behaviours the learner must be able to perform afterwards, the existing number you expect to move and by when, and an explicit statement of what is out of scope and who owns it. The last line is what stops the training being blamed for the tool and the incentives.

Can AI run a training needs analysis?

No. A model asked to produce a needs analysis will produce a plausible analysis of a problem you did not describe, because it has no access to your error data, your disagreeing supervisors or your systems. AI is useful either side of the analysis rather than inside it: summarising support tickets or free-text responses into themes beforehand, and removing the production work between the finished brief and a working course afterwards.

What is the difference between a training needs analysis and a skills gap analysis?

A skills gap analysis compares the skills a workforce holds against the skills it is expected to hold, usually across roles and usually for planning. A training needs analysis starts from a specific performance problem and asks whether knowledge is one of its causes. The first is inventory work aimed at a future capability, the second is diagnostic work aimed at something happening now, and a skills gap is only one of the possible findings.

Where This Guide Stops

Two limits are worth naming, because both are outside what any analysis method can fix.

An analysis only works if somebody is willing to report a won't-do verdict to the person who asked for the course. The method is straightforward and the conversation is not, and no template makes that easier.

And an analysis cannot tell you whether the behaviours you chose are the right ones if nobody in the room knows the work. Talking to the person who does the job best is not a nice addition to the process. When it is skipped, the training ends up teaching the documented procedure rather than the real one, and the people who already know the difference will notice in the first two minutes.

Test the Verdict on One Real Request

Take the next training request that lands and run the six questions against it before you build anything. If it comes back a genuine can't-do, QuikAuthor's free plan includes 25 courses, up to 50 active monthly learners and every AI feature, with no credit card and no expiry, which is enough to turn the brief into a working course the same afternoon.

Related reading: Measuring Training ROI · AI Tools for Instructional Designers · Rapid Authoring Tools