How to Create a Training Program: 7 Steps, One Worked Example

Set up your branded LMS and deliver training with AI

Start Free

A manager gets told "we need training on the new refund policy" and is given two weeks. The usual next move is to open a slide deck and start writing about the policy. That is where many programs go wrong: nobody has yet said what should change on the job, or how anyone will know it did.

A training program is a planned set of learning activities for a defined group and purpose, with objectives, practice, assessment and support. Agree how to evaluate it before building; job improvement is an intended result, not part of the definition. I build every program backwards from that measure: the business number first, then the behaviour, then the assessment, and the lessons last.

I'm a cofounder of Coursebox, where we make an AI course builder for employee training, so weigh what I say about it with that in mind. Before Coursebox I was lead learning designer on three edX courses at the University of Adelaide. This guide walks through seven steps with one example carried all the way through: a contact centre rolling out a new refund policy. You finish with a filled one-page plan and a blank one to copy.

What is a training program?

A training program is a planned set of learning activities for a defined group and purpose, with objectives, practice, assessment and support. Agree how to evaluate it before building; job improvement is an intended result, not part of the definition.

Training is aimed at doing the current job better, now. Development is aimed at a future role, over months or years. Both can use the same methods, but a training program should be able to name the job behaviour it changes.

Does it work? The large review by Salas, Tannenbaum, Kraiger and Smith-Jentsch in Psychological Science in the Public Interest says the research shows two things: training works, and how it is designed, delivered and implemented affects how well it works. They split that into what happens before, during and after the training itself. The seven steps below follow the same order.

How to create a training program in 7 steps

1. Check training is the answer: find out whether people lack the knowledge or skill, or something else is in the way.
2. Pick the business number and how you will measure it: agree the baseline, the target and the evaluation plan now.
3. Write the objective as an observable behaviour: one sentence a manager could check by watching.
4. Write the assessment before the lessons: decision items, not recall items.
5. Build the lessons and practice: lesson, practice, assessment, with practice outweighing content.
6. Choose the format and schedule it for retention: spread it over weeks, not one afternoon.
7. Deliver it, support it on the job, and evaluate: pilot, brief the team leaders, then check against step 2.

The worked example is fictional: a contact centre has 40 agents handling voluntary change-of-mind returns. The following five company rules apply only to that process. Faulty goods and statutory remedies go to the approved consumer-rights process; the 30-day window does not remove those rights. The ACCC explains consumer remedies.

• Refunds are approved within 30 days of the purchase date in the CRM (the customer system), and day 30 is still eligible.
• After 30 days, decline the voluntary change-of-mind refund. A fault or another consumer-rights claim must be assessed separately under the approved remedies process.
• Any refund over $200 is escalated to a team leader before the agent promises money.
• The agent checks the purchase date in the CRM, not the customer's memory of it.
• The agent records the outcome and the reason.

Last quarter's quality assurance (QA) audit sampled 100 refund calls and found 18 that broke the policy. The source material is the policy document plus three anonymised QA call notes. All numbers in the example are illustrative.

Step 1: Check training is the answer

A needs analysis looks at three levels: the organisation (what result is off), the task (what the job requires) and the person (who cannot do it yet). Our training needs analysis questionnaire covers the questions for each level.

The point of the analysis is a decision, not a report. The CDC's guide to conducting a training needs analysis says a needs assessment helps you determine whether training is needed at all, and that other fixes, such as changes in policy, technology or organisation, might close the gap instead. For the refund calls, the branches look like this:

• If agents do not know the rule: train.
• If the CRM hides the purchase date: fix the screen.
• If the bonus rewards call speed over accuracy: change the incentive.

In the example, a 10-question check of 10 agents finds that 6 cannot state the $200 escalation rule. Knowledge is part of the gap, so training stays in. The team manager still reviews the call-speed bonus, because a knowledge fix will not hold against a bonus that pays people to skip checks. Our guide to performance gaps goes deeper on separating skill problems from system problems.

Output: a one-sentence needs statement. "Agents break the refund policy on 18 of 100 calls, partly because most cannot state the escalation rule."

Step 2: Pick the business number and how you will measure it

Decide how you will know the program worked before you write anything else. Kirkpatrick Partners, who maintain the Kirkpatrick Model, say it is critical to start with Level 4: Results. They define Level 3 as the degree to which people perform the critical behaviours on the job and are supported and held accountable for them.

For the refund program, the plan has four rows, each with a measure and a date:

• Level 4 (results): choose a business outcome the programme may contribute to, such as avoidable complaint cost, and agree its baseline, target and review period with the manager. The current 18 policy breaches per 100 audited calls is the operational behaviour measure below, not automatically a Level 4 result.
• Level 3 (behaviour): repeat a comparable QA sample against the baseline of 18 breaches per 100 calls, with an illustrative target of 5 or fewer. In week 2, team leaders also sample calls per agent using the checklist. Record the sampling limits and other changes that may affect the result.
• Level 2 (learning): the 10-scenario assessment score, from the course report, on day 2 (see the schedule in step 6).
• Level 1 (reaction): two questions at the end of day 2 on how useful and how relevant the training was.

This is the step I see skipped most. We classified a sample of the requests people typed into the Coursebox course builder over 90 days, to see how often a training brief says how success will be judged.

A request is not a full plan, but a brief that lists topics and no measure is a brief for content. Our Kirkpatrick model guide explains each level in more detail.

Output: the four evaluation rows above, agreed with the contact-centre manager who owns the QA number.

Step 3: Write the objective as an observable behaviour

The objective is the behaviour from step 2, written so that someone could check it by watching or by giving a scenario. For the example:

"Given a change-of-mind refund call and the approved policy and CRM, the agent checks the facts, chooses the correct action and explains it without promising an unapproved refund." A 9-of-10 scenario score can prompt further checking; every essential rule must be demonstrated correctly before independent handling.

"Understand the refund policy" fails this test, because nobody can watch understanding. In their ten simple rules for developing a training program, Blevins and colleagues tell program designers to solidify, document and communicate the intended trainee outcomes before development starts. Our list of SMART learning objectives examples has more objectives written this way across job roles.

Step 4: Write the assessment before the lessons

If the objective says "decides approve, decline or escalate", the assessment has to make the agent decide. Here is one of the 10 items:

A customer bought a working blender 40 days ago for $89. They no longer want it and ask for a change-of-mind refund. Under the fictional policy, approve, decline or escalate?

Decline this voluntary change-of-mind refund because day 40 is outside the company's 30-day window. The blender is not reported faulty. If a fault is raised, use the approved remedies process instead.

Compare that with a recall item such as "What is the refund window?" An agent can answer "30 days" and still approve the blender on the call. Recall items test whether people remember the rule. Decision items test whether they can apply it.

Write the items around the boundaries, because that is where agents slip: day 30 against day 31, exactly $200 against $200.01, a customer who quotes a date that differs from the CRM. These items also teach: in Roediger and Karpicke's 2006 study, people who were tested on material remembered substantially more of it on delayed tests than people who studied it again. Short versions become the spaced quizzes in step 6.

Step 5: Build the lessons and practice (lesson, practice, assessment)

Build the content next. Every course I design comes down to three kinds of page: a lesson that gives people the information, a practice activity where they use it with feedback, and an assessment that counts. Most courses have far too much of the first, so the first question I ask a team turning a policy into a course is where the learner actually practises. For the refund program the outline looks like this:

That is three lessons, four practice activities and one assessment. The QA lead writes the scenarios from the three call notes, a better use of an expert than writing lesson text. Our guide to employee training materials covers the job aids and handouts that sit around a program like this.

Output: an outline with enough explanation and varied practice for the objective. Four practice rows and three lesson rows describe this example; counting rows does not establish the right balance or time allocation.

Step 6: Choose the format and schedule it for retention

Choose the format for each part of the outline, not for the program as a whole. The facts and the first practice go online, because every agent needs the same facts and can take them at their desk. The call goes to live pairs, because saying the escalation line out loud is the skill. The first week back happens on the job, with a team leader listening.

Plan revisits when retention is important. Cepeda and colleagues' review found benefits from distributed practice, with useful gaps depending on the retention interval and task. The following dates are an illustrative schedule, not an experimentally validated sequence for refund work.

• Day 1: online lessons, the branching call simulation and spot the error, about 30 minutes.
• Day 2: live paired calls, script rehearsal and the assessment, 45 minutes.
• Days 3, 10 and 24: a three-item scenario quiz, about 5 minutes each.

You need a tool that delivers the online part and tracks scenario scores per agent, which most learning management systems (LMSs) and course tools can do. Coursebox's AI course builder for employee training builds lessons, practice pages and a graded quiz from policy text you paste in; my test is below. Our guides to how long employee training should be and to employee training methods cover the format choices in more depth.

Output: the schedule, with dates.

Step 7: Deliver it, support it on the job, and evaluate

Pilot with one team of 10 agents, and fix the pages where they hesitate before the other 30 start. Then brief the team leaders, because the program does not end at the assessment. Blevins and colleagues make this a rule of its own: trainees need support incorporating what they learn into their work. Give each team leader a three-item listening checklist for week 2:

• Checked the purchase date in the CRM.
• Applied the 30-day rule.
• Escalated anything over $200 before promising money.

Then evaluate against the plan from step 2: the checklist results in week 2, the QA audit after launch against the 18 of 100 baseline, and a review with the contact-centre manager to change what did not work. Our guide on how to evaluate the effectiveness of a training program covers data collection and reporting.

A stronger evaluation compares trained and untrained groups on a number the business already records. One garment exporter's soft skills program shows what that looks like.

Case study: Soft skills training in five Indian garment factories

India's largest ready-made garment exporter ran P.A.C.E. (Personal Advancement and Career Enhancement), an 80-hour workplace program designed by Gap Inc., for female garment workers in five factories in Bengaluru. Workers sew in production lines where each person does one operation, so a line's output depends on people communicating about bottlenecks and problems.

• What they did: Workers attended two-hour weekly sessions during 2013 and 2014, on modules such as communication, problem solving and time management, ending with a team simulation of a garment order.
• Evidence and result: Adhvaryu, Kala and Nyshadham's randomised evaluation assigned production lines to treatment or control, then selected trainees within treatment lines, and measured productivity from the factory's own production data. Trained workers were 13.5% more productive than controls, untrained workers on trained lines also improved, and the authors report a net return to the firm of 256% eight months after the program ended. There was no effect on wages or retention.
• Limitation: This is one firm, one country and one workforce, and the return is the authors' calculation from program costs and productivity gains. The result belongs to the whole 80-hour program, not to any one module.
• Lesson: The program was judged on a number the factory already tracked every day, against a control group. If your program changes a number you already record, such as QA breaches per 100 calls, compare a trained team with one that has not started yet.

What changes for other types of training program

The seven steps stay the same across program types. What changes is the answer at three of them.

Our guide to types of employee training programs describes each type in full, the guide on how to train new hires in 10 simple steps covers onboarding, and our professional training examples show programs from other industries.

What makes a training program effective?

For a workplace programme, I look for these four features:

• It is tied to one business number, agreed before the build.
• It gives practice more time than content, with assessment items that ask for decisions.
• It is assessed on the job behaviour, not only on a course score.
• It is supported after the session, by someone who watches the work.

One feature it does not need is matching content to learning styles. Preferences exist, but Pashler, McDaniel, Rohrer and Bjork's review found no adequate evidence base for building learning-styles assessments into teaching, so choose the format by the task instead.

Building the example in Coursebox (my test)

I wanted to see whether an AI course builder would start at the content or at the decision, so I ran the refund brief through the Coursebox course agent in a test workspace. I pasted the five-rule policy and three anonymised QA breaches into the prompt. In breach B, a contact-centre agent told a $340 customer "the refund will be with you in 3 days" before escalating.

What I asked for. "Build a short course for contact-centre agents on our new refund policy. Objective: 'Given a refund call, the agent decides approve, decline or escalate under the policy, without notes, in 9 of 10 scenarios.' Build it as lessons, practice and a 10-scenario assessment."

What it produced. Seven pages at about 50 minutes: three lessons, three practice pages (a guided call, boundary and amount drills, a whole-call rehearsal) and one assessment, so the outline was not content-heavy. The assessment followed the pasted policy on the items I inspected: 10 graded items, a 90% pass grade, and the boundaries I would have written myself, including exactly 30 days, exactly $200, $200.01 and a customer who says "five weeks ago" when the CRM shows 26 days.

What was wrong. Every item asked only for the outcome, with Approve, Decline or Escalate as the options. Breach B was not a wrong decision: the agent escalated correctly but promised the money first. Someone could score 10 out of 10 and repeat breach B on the next call. Most item stems also began "The CRM confirms that the purchase age is...", which does the check-the-CRM step for the learner.

The pasted test policy itself had a more serious flaw: it treated faulty goods after 30 days as automatic declines. That is not a safe consumer-remedies rule. The worked teaching example above now covers change of mind only; the generated test course was not rebuilt or retested with that correction. Its 90% overall grade also did not ensure every essential rule was correct. Both need repair before real use.

The fix, and how I checked it. I asked the AI editor to replace the repeated decline item with a "what do you say" item about breach B, retaining nine questions and the 90% pass grade. After reload, question 10 asked about a $340 coffee machine bought 12 days ago. Its correct option required team-lead approval before confirming the refund. There were still 10 items and the pass grade was 90%. I did not test a learner attempt or reporting, and would add further wording choices before use.

At a much larger scale, Secure Meters, with more than 7,200 staff in 10 countries, has published 40 AI-built courses and trained more than 500 employees with Coursebox, working to a one-month target window for new-joiner onboarding.

Template: one-page training program plan

The plan puts all seven steps on one page, in build order, with the refund program filled in and a blank copy for yours.

Putting the plan to work

Run all seven steps whenever a business number is at stake, and agree that number before anyone writes a lesson. For a one-off briefing with nothing to measure, steps 1, 3 and 4 are enough: check training is the answer, write the behaviour, and test it with a decision item. If the needs check shows a system or incentive problem, fix that instead, because no course will.

Your next action: take the one training request on your desk right now, fill in rows 1 and 2 of the plan with the manager who asked for it, and do not open a slide deck until row 5 has a real assessment item in it.