On this page
TL;DR:
- AI for learning and development teams works best on first drafts: sorting needs-analysis notes, outlining a design, writing job aids, building practice scenarios and grouping evaluation comments. A person checks every output and owns the decision.
- Research on training transfer points to design, manager support, the chance to use the skill and follow-up. Spend the hours AI saves you on those.
- Plan backwards with the Kirkpatrick levels. Agree the result first, then the few job behaviours that drive it, then the learning, then the session.
- A happy feedback form tells you little about lasting learning. In one meta-analysis, reactions were only modestly related to immediate knowledge and unrelated to delayed knowledge. Measure behaviour on the job.
- Calculate ROI only for programmes that justify the effort, count staff time as a cost, isolate training's share, and expect modest numbers. Named staff survey answers and performance data stay out of personal AI accounts.
Who this guide is for
This guide is for learning and development managers, HR business partners, in-house trainers and training coordinators, in Kenya or anywhere else. If you are a one-person training function, we have written it with you in mind.
What does the research say makes training work?
Training works on average, and much of the outcome is decided outside the session. A meta-analysis by Arthur and colleagues found medium to large effects of organisational training, with average effect sizes of 0.60 to 0.63 across reaction, learning, behaviour and results measures.
These are the findings worth building around.
| What the research found | Source | What it means when you add AI |
|---|---|---|
| Behaviour modelling, error management and realistic practice are among the design factors most consistently linked to transfer | Grossman and Salas, 2011 | Use AI to produce more practice and better examples, and fewer slides |
| Transfer climate, supervisor and peer support, opportunity to perform and follow-up matter in the workplace | Same review | AI cannot create these. It can prepare managers to create them |
| Support from supervisors, peers and the organisation, coaching and feedback are linked to better transfer | QIC-WD research summary, drawing on Hughes and others, 2020 and Reinhold and others, 2018 | Brief the manager before you polish the deck |
| Post-training interventions such as goal setting showed no significant link to transfer in the Blume review. Grossman and Salas cite another review that found generally positive effects | Same summary, and Grossman and Salas | The evidence is mixed, so do not count on a follow-up email to rescue a weak programme |
| In leadership training, needs analysis, feedback, practice, spaced sessions, on-site location and face-to-face delivery were linked to better results | Lacerenza and others, 2017, 335 samples | Leadership is one kind of training, but this is a useful checklist for any design |
| Science-based training looks at what happens before, during and after, with jobs analysed first and supervisors aligned | Salas and colleagues, summarised by the Association for Psychological Science | Plan the before and after as carefully as the workshop |
You will often hear that only 10 per cent of training spending transfers to the job. Grossman and Salas cite that as an estimate from Georgenson in 1982. Treat it as an old estimate and measure your own programmes.
None of the studies in the table tested AI. The one AI study in this guide, further down, covers exam questions in medicine. So everything below applies well-tested training research to a new tool. Where we recommend a practice, that recommendation is our judgement.
Start with Kirkpatrick Level 4 and work backwards
The most useful planning move is to agree the business result before anyone opens a slide. Kirkpatrick Partners describe the New World Kirkpatrick Model as keeping the four original levels and beginning with Level 4, which they say is how the model was intended.
| Level | What it asks | Where AI can help | What stays human |
|---|---|---|---|
| 4 Results | Did the targeted outcomes occur, as a result of training and reinforcement? | Turning a vague sponsor request into leading indicators you could measure | Agreeing the result with the business owner |
| 3 Behaviour | Are people applying it on the job? | Drafting critical behaviours, manager checklists and observation forms | Deciding which behaviours matter and observing them |
| 2 Learning | Did people gain knowledge, skill, attitude, confidence and commitment? | Writing scenarios, quiz items and practice tasks | Judging whether the task shows real competence |
| 1 Reaction | Did they find it favourable, engaging and relevant? | Grouping open comments into themes | Reading the comments and acting on them |
Critical behaviours are the few specific actions that, done consistently, have the biggest impact on the result. If you cannot name three, you are not ready to design.
Required drivers are the processes and systems that reinforce, monitor, encourage and reward those behaviours: job aids, coaching, work review, recognition. Kirkpatrick Partners call them the key to on-the-job application.
Prompts work better when they follow a method
At FuKazee we teach the four Cs before any fancy technique:
- Clarity: say exactly what you want produced.
- Context: give the background the tool cannot guess.
- Constraints: set limits on length, tone, format and what to leave out.
- Customisation: shape it for the reader. Ask where the answer will be used and who will read it.
Replace the square brackets. Remove names first.
Needs analysis: let AI sort, and keep the diagnosis
AI is good at organising what people told you and poor at knowing why performance is weak.
Kirkpatrick Partners warn against a common and costly mistake: treating poor performance as a lack of knowledge or skill when the cause is motivation or the environment. A chatbot asked to "design training for a team missing targets" will happily design training. It may never ask whether the system is down every Friday.
The evidence for needs analysis is thinner than textbooks suggest. Arthur and colleagues found only 6% of the data points in their meta-analysis came from studies reporting a needs assessment (others may have done one without saying so), and found no clear pattern, on very few data points. Lacerenza's leadership review did link it to better results. Our view is to keep it short and treat it as a test of whether training is the right answer at all.
A workable routine:
- Talk to the sponsor. What result is wrong, and how do they know?
- Talk to five to eight people who do the job, plus one or two who do it well.
- Look at the work itself: error logs, complaint categories, a sample of reports.
- Remove identifying details, then use AI to sort the notes.
- Write one page that says what training can fix and what it cannot.
Prompt 1: pressure-test the request
I am the learning and development lead at a [type of organisation] in [country]. A manager has asked for training on [topic] because [stated reason]. Give me 10 questions to ask the manager and 5 to ask staff. They should reveal whether the real cause is skill, motivation, tools, process, workload or unclear expectations. Keep each question under 20 words. Do not assume training is the answer. Finish with three signs that training would not fix this.
Prompt 2: sort anonymised interview notes
Below are anonymised notes from interviews with [number] staff in [role]. Group what was said into themes. For each theme give the number of people who raised it and one short quote. Label each theme as: skill gap, tool or system, process, workload, motivation, management, or unclear. Do not infer anything that is not in the notes. Mark uncertain items [CHECK]. Finish by listing what the notes do not tell us. Notes: [paste]
Read the notes yourself first, then compare.
Designing the programme: ask for options, then choose
Use AI to widen your options and test your logic. Your knowledge of the people decides.
Give it the result, the behaviours and the real constraints: shifts, branches, devices and systems. A design that needs everyone away from their desks for two days is useless to a team that cannot close the counter.
Practice appears repeatedly in the evidence above, and Lacerenza's review favoured spaced sessions. Where you can, swap one long day for shorter sessions.
Prompt 3: design from the behaviours
Result we want: [measurable result]. Critical behaviours: [list 3]. Learners: [number, role, prior experience]. Constraints: [time, locations, devices, connectivity, languages]. Propose three different programme designs. For each give the sessions, what learners practise in each, what happens between sessions, what the manager does, and what we would measure at 30 and 90 days. Say what each design sacrifices. Do not include icebreakers or content that does not serve the behaviours.
Then remove anything you cannot defend in front of the sponsor. If your provider or your own team proposes modules that map to no behaviour, those are the first to cut. Our guide to budgeting for AI training helps if money is part of the choice.
Materials: build fewer slides and better job aids
AI drafts materials quickly, so the risk moves from too little content to too much. Job aids are named by Kirkpatrick Partners as a required driver, and Grossman and Salas list them under follow-up, so a good one-pager may be worth more than twenty slides.
Three rules keep materials honest:
- Work from a source pack. Paste your policy, procedure or product sheet and tell the tool to use only that. Anything else it adds is a guess.
- Check every fact, figure and legal reference against the original. AI states wrong clauses and invented numbers in the same confident tone as correct ones.
- Have a fluent colleague read any Kiswahili and any text about local law or culture.
Prompt 4: one-page job aid
Using only the procedure below, write a one-page job aid for [role]. It will be printed and kept beside the workstation, or read on a phone. Include: when to use it, the steps as a numbered list, the three most common mistakes, and who to ask. Maximum 200 words. Where the procedure is unclear, write [ASK PROCESS OWNER] and do not fill the gap. Procedure: [paste]
Practice activities: where AI earns its keep
Practice is the part of training that was always expensive to build, and AI makes it cheap. This is where we would spend most of your new capacity.
The evidence favours practice that lets people get it wrong. Keith and Frese's meta-analysis of error management training, covering 24 studies and 2,183 people, found an overall effect of d = 0.44. The effect was larger for performance after training (d = 0.56) and larger still when the later task was different from the training task (d = 0.80). Grossman and Salas add that behaviour modelling works best with both good and bad examples plus a chance to practise, and that practice should resemble the workplace.
So build these four things:
| Activity | How AI helps | Your check |
|---|---|---|
| Good and bad examples | Writes a strong and a weak version of a customer reply, report paragraph or feedback conversation | Is the weak version realistic, or a cartoon? |
| Flawed drafts | Produces a document with planted errors for learners to find | Keep your own answer key, and confirm each error |
| Scenario role-play | Plays a difficult customer, nervous junior or sceptical board member | Run it yourself first, several times |
| Branching cases | Writes a situation with decisions and consequences | A subject expert confirms the consequences are true |
Prompt 5: flawed draft for practice
Write a [document type] of about 200 words for a [role] at a [type of organisation]. Plant exactly five errors: two factual errors contradicting the source below, one wrong figure, one tone problem and one missing required step. Then give a separate answer key listing each error and why it matters. Learners will work in pairs for ten minutes to find them. Source: [paste]
Prompt 6: live role-play
You are a [customer or colleague] who is [situation and emotion]. I am a [role] practising how to handle this. Stay in character. Reply in two or three sentences at a time and react realistically to what I say. If I handle it badly, become harder to deal with. When I type END, step out of character and give me two things I did well and two things to change, quoting my words.
If you test knowledge with quiz items, review each one. In a study published in PLOS ONE in 2023, ChatGPT drafted 50 medical exam questions in about 20 minutes, against about 211 minutes for two human examiners. Independent assessors scored the AI's questions lower on relevance, the only domain with a significant difference, and the AI's scores varied more. That was one study in medicine. An expert should still read every item.
Evaluation: ask better questions and read the comments yourself
Good evaluation starts before the workshop and carries on after it. Decide the measures when you write the design, and take a baseline while you still can.
Research by Sitzmann and colleagues, summarised by QIC-WD, reports that reactions are strongly connected to post-training motivation, modestly related to immediate knowledge and unrelated to delayed knowledge. Asking whether the training was useful does a little better than asking whether people liked it: in the Blume review, as summarised there, utility reactions have a modest link to transfer, and enjoyment-type reactions none.
Keep the form short and aim it at use.
| Level | A better question or measure | When |
|---|---|---|
| 1 | "Which part of this will you use in the next week, and on what?" | End of session |
| 2 | A task done on real work, marked against a checklist | During and end of session |
| 2 | "How confident are you that you can do this on the job? What might stop you?" | End of session |
| 3 | Manager or peer observes the critical behaviour on a sample of work | 30 and 90 days |
| 3 | Work sample: the same report or case, before and after | 60 to 90 days |
| 4 | The leading indicator agreed at the start, such as turnaround time or error rate | 90 days and later |
Self-reports are often chosen as the simplest way to measure transfer, the QIC-WD summary notes. Add something a manager or a work sample can confirm.
Keep the barriers people list in their comments separate. AI can group the themes, and you should still read the raw comments.
ROI: calculate it for the few programmes that deserve it
ROI is worth the effort when a programme is expensive, broad, or watched by senior people. The ROI Institute's own basics say that not every programme should be evaluated to impact and ROI.
The formula is simple: ROI (%) = (net programme benefits ÷ programme costs) × 100. The ROI Institute's guiding principles include using the most credible sources, choosing the most conservative option, using at least one method to isolate the programme's effect, adjusting estimates for error, assuming little or no improvement where you have no data, and counting only the first year of benefits for short-term programmes.
A worked example
Every number in this example is invented to show the method. It is not a client result and not a price.
A team of 20 officers each write 96 reports a year. Using a sample, you measure the whole task, including checking.
| Step | Calculation | Result |
|---|---|---|
| Time per report before training | Measured on a sample | 3.0 hours |
| Time per report after training | Measured on a sample, including checking | 2.4 hours |
| Gross time saved if everyone keeps it up | 0.6 hours × 96 reports × 20 officers | 1,152 hours |
| Adjust for adoption | 60% of officers still using it at 90 days | 691 hours |
| Isolate training's share | Managers estimate 70%, after adjusting for doubt | 484 hours |
| Convert to money | 484 hours × KES 800 loaded hourly cost | KES 387,200 |
| Fully loaded costs | Design, delivery, evaluation and 160 staff hours in the room | KES 300,000 |
| Net benefit | 387,200 − 300,000 | KES 87,200 |
| ROI | (87,200 ÷ 300,000) × 100 | 29% |
The headline saving looked like KES 921,600. After honest adjustments it is KES 387,200.
Time saved only becomes money if the hours go to something the organisation values, so say what happened to them. Report the lower levels alongside the ROI figure, as the ROI Institute advises, and be suspicious of any provider quoting a huge return without saying how training's share was isolated.
Following up after the workshop
The evidence on follow-up is mixed, as the table above shows, so our advice is to tie it to real tasks and involve a manager. Grossman and Salas describe useful follow-up as further learning opportunities such as after-action reviews, feedback and job aids.
If-then plans. Before people leave the room, each writes one: "If I get the monthly report on Friday, then I will draft it with the template before lunch." Gollwitzer and Sheeran's meta-analysis of 94 tests found a medium-to-large effect of such plans on reaching goals (d = 0.65). That research covers goals in general, so treat it as a promising habit.
Short structured debriefs. A meta-analysis by Tannenbaum and Cerasoli covering 46 samples found that debriefs improved effectiveness by about 25% over control groups. Facilitation and structure may help, though the authors had few studies to compare.
| When | What happens | How AI helps |
|---|---|---|
| Before the workshop | Manager and learner agree one task to apply it to | Drafts the short briefing for the manager |
| Last 15 minutes | Each person writes an if-then plan on real work | Nothing. People write these themselves |
| Day 7 | Short message with one tip and one question, on the channel people actually read, such as WhatsApp or Teams | Drafts a series of messages from your content |
| Day 30 | Manager check-in on the critical behaviours | Drafts the conversation guide |
| Day 60 | Group debrief: what worked, what failed, what to change | Groups the anonymised notes into themes |
| Day 90 | Work sample and leading indicator | Helps format the report for the sponsor |
Prompt 7: manager check-in guide
My team completed training on [topic]. The critical behaviours are [list]. Write a 10-minute check-in guide for a busy line manager who did not attend. Include: what the training asked people to do, five questions that invite real examples, how to respond if the person has not tried it yet without blaming them, and one way the manager can remove a barrier. One page, plain English.
Our piece on getting middle managers to use AI covers their role, and why corporate AI training doesn't stick covers the common failures.
Which AI tools, and what must stay out of them
Start with what your organisation already licenses. We do not recommend one tool over another.
As of November 2026, Microsoft says Copilot Chat comes with eligible Microsoft 365 licences and has enterprise data protection when people sign in with a work account. It works from the web, files you upload and content you have open in Outlook or Teams. Reaching your emails and files automatically needs the paid add-on. Other assistants such as Gemini, ChatGPT and Claude can do this drafting too. Check the data settings on any free plan.
Training records that identify a person are personal data under Kenya's Data Protection Act. That covers attendance, assessment scores, manager comments and named survey answers. Health status and family details count as sensitive personal data under section 2.
| Safe in tools your organisation allows | Only after removing identifiers, in an approved account | Keep out |
|---|---|---|
| Public information, blank templates, generic scenarios | Interview notes, anonymous comments, work samples | Named appraisals, disciplinary notes, health or family details, payroll, exported HR records |
In a team of six, "the supervisor in Mombasa" identifies someone, so anonymising takes more than deleting the name.
Section 35 gives people a right not to be subject to a decision based solely on automated processing that significantly affects them, with limited exceptions. Promotion, disciplinary or pass-or-fail decisions should rest with a person. Our guide to writing an AI usage policy covers the wider rules.
What goes wrong, and how to avoid it
| Problem | What it looks like | What to do |
|---|---|---|
| More content, same result | A longer deck, built faster | Cap slides. Put the time into practice |
| Training the wrong problem | Sessions on a skill people already have | Run Prompt 1 before you agree to design anything |
| Invented facts in materials | A wrong clause or made-up statistic on a slide | Work from a source pack and check everything |
| Vanity evaluation | High ratings reported as impact | Report Levels 3 and 4, even when the news is mixed |
| Inflated ROI | Gross time saved reported as money | Adjust for adoption and isolate training's share |
| One workshop and done | Nothing planned after the session | Fix the follow-up calendar before the workshop date |
Some problems sit beyond the reach of training or AI. If the tools are blocked, the workload is impossible or managers punish experimentation, tell the sponsor.
Questions people ask
Can AI do a training needs analysis?
It can help you prepare questions and sort anonymised notes. It cannot find out why people are underperforming, because it has not met them or seen the work. Treat its themes as a first sort and check them against the source.
Will AI replace learning and development teams?
We do not expect it to replace the judgement, relationships and credibility the work depends on. It will take over a lot of drafting. Teams that spend that time on practice design, manager support and measurement should be more valuable to the business.
Is the Kirkpatrick model still useful?
Yes, as a planning tool. Its current form, the New World Kirkpatrick Model, starts from results and adds critical behaviours and required drivers. It will not tell you how to separate training's effect from other causes, so pair it with a method for that when money is at stake.
Can we use AI to mark assessments after a workshop?
Use it to suggest feedback on anonymised work against your checklist. A person should make any pass, fail or certification decision. Section 35 of the Data Protection Act protects people from solely automated decisions with significant effects.
Where should a one-person L&D team begin?
Pick one programme you already run. Name its three critical behaviours, replace one slide section with a practice activity, and book a 30-day manager check-in.
Sources
Training transfer and evaluation research
- Arthur, Bennett, Edens and Bell, Effectiveness of Training in Organizations, Journal of Applied Psychology, 2003, last checked 8 November 2026
- Grossman and Salas, The transfer of training: what really matters, International Journal of Training and Development, 2011, last checked 8 November 2026
- QIC-WD, Training Transfer: a summary of workforce research evidence, including Blume and others, 2010 (December 2022), last checked 8 November 2026
- QIC-WD, Trainee Reactions: a summary of workforce research evidence (February 2022), last checked 8 November 2026
- Sitzmann and others, A review and meta-analysis of the nomological network of trainee reactions, Journal of Applied Psychology, 2008 (abstract only; the knowledge findings above come from the QIC-WD summary), last checked 8 November 2026
- Lacerenza and others, Leadership Training Design, Delivery, and Implementation: A Meta-Analysis, Journal of Applied Psychology, 2017 (abstract), last checked 8 November 2026
- Association for Psychological Science, summary of Salas and others, The Science of Training and Development in Organizations, 2012, last checked 8 November 2026
- Keith and Frese, Effectiveness of error management training: a meta-analysis, post-review version, Journal of Applied Psychology, 2008, last checked 8 November 2026
- Tannenbaum and Cerasoli, Do Team and Individual Debriefs Enhance Performance? A Meta-Analysis, Human Factors, 2013, last checked 8 November 2026
- Gollwitzer and Sheeran, Implementation intentions and goal achievement: a meta-analysis, 2006, last checked 8 November 2026
Evaluation frameworks
- Kirkpatrick Partners, An Introduction to The New World Kirkpatrick Model (October 2022), last checked 8 November 2026
- ROI Institute, ROI Basics, chapter one, last checked 8 November 2026
AI and tools
- PLOS ONE, ChatGPT versus human in generating medical graduate exam multiple choice questions, 2023 (abstract record, Kabale University library), last checked 8 November 2026
- Microsoft Learn, What is Microsoft Copilot?, page dated 5 November 2026, last checked 8 November 2026
Law
- Data Protection Act, No. 24 of 2019 (ODPC copy), last checked 8 November 2026
Where FuKazee fits in
FuKazee trains people in tech, the future of work and the future of careers. We have trained more than 3,000 people, and we design corporate sessions around each team's real tasks.
We assume nobody in the room knows AI yet, use no jargon, show the free version of every feature and never push one tool. Microsoft 365 and Copilot are part of the work because that is where many teams live. We teach the four Cs before anything else, and we will tell you plainly that nobody becomes an expert in a few days.
If you run learning and development, you can use this guide without us. If you would like help building AI into your training function, or training your staff, see how our corporate AI training in Kenya works, or weigh it up with our guide to in-house versus external AI training.
Tell us your team size, the result you are after and what has not worked so far. Email lead@fukazee.com
Want guided support instead?
Plan AI training around your team's real work. Answer a few quick questions about your sector, team size and goals, and we will recommend the right programme and format. Sessions are built on your workflows and include a usage policy.
Rodney has run FuKazee's training since 2022 and still teaches inside every programme line: corporate sessions, the postgraduate cohort and the KCSE examiner webinars. First Class Honours in Journalism and Mass Communication, University of Nairobi.
