An AI workshop of two to three hours, or a one-day course, is enough when the goal is awareness: a shared vocabulary, basic safety rules and leadership agreement on direction. When the goal is that people use AI well in their daily work, a team needs several weeks of spaced practice on real tasks, with feedback and a way to see whether habits changed.
Key points
- A short workshop is good at three things: showing what the tools can do, setting the rules for data, and getting managers to agree on where AI should and should not be used.
- Memory research going back to Ebbinghaus shows the steepest forgetting happens in the first hours and days after learning, so a single session decays fast unless people use the skill again soon.
- A review of 317 experiments found that spacing practice over time beats cramming it into one session, and the best gap grows with how long you need to remember.
- A meta-analysis of 89 training studies links transfer to the job with motivation and a supportive work environment, which a programme can build in and a one-off event cannot.
- Free courses such as Elements of AI cover the basics well. Use them first, then pay for the part that free courses cannot do: practice on your own work with feedback.
What does a 2–3 hour AI workshop do well?
A two to three hour AI workshop does well at creating a common starting point: everyone sees the same demonstrations, hears the same rules and leaves with the same words for what the tools do. For a team where half the people have never opened ChatGPT, Copilot or Gemini, that is a real step.
Three jobs suit the short format.
Awareness. People see what a language model is good at (drafting, summarising, restructuring text, first-pass analysis) and where it fails (invented facts, outdated information, confident wrong answers). Seeing a model make a mistake live teaches more than a slide that says "check the output".
Basic rules. Which tools are approved, what may and may not be pasted into them, who to ask when in doubt. If the company has an AI policy, a workshop is a good place to walk through it with examples.
Leadership alignment. A half-day session for the management group can settle questions a programme depends on later: which tasks are candidates for AI, what the budget for licences is, what risk is acceptable. That conversation is often more valuable than the training itself.
A one-day course adds hands-on time, usually with generic exercises. It suits groups who need the same baseline, such as new hires or a company rolling out one tool to everyone at once.
Why does one-off AI training often fail to change how people work?
One-off training often fails to change behaviour because people forget most of what they practise once, and because the work environment they return to decides whether the new habit survives. Both findings are old, well replicated and apply to AI tools as much as to any other skill.
The forgetting curve
Hermann Ebbinghaus measured his own forgetting around 1880 by learning lists of nonsense syllables and relearning them after different intervals, and published the method and results in 1885. The loss was steep in the first hours and then levelled off. Jaap Murre and Joeri Dros repeated the experiment and published it in PLoS ONE in 2015; their results were similar to the original. Nonsense syllables are an extreme case, and work skills are unlikely to fade at exactly the same rate, but the shape is the useful lesson: what is not used soon after learning fades fastest.
Spaced practice beats one long session
Nicholas Cepeda and colleagues reviewed 317 experiments from 184 articles on distributed practice and published the synthesis in Psychological Bulletin in 2006. Spreading study over several sessions produced better retention than the same amount of study in one block, and the best gap between sessions grew with how long people needed to remember. In 2013 John Dunlosky and colleagues rated ten common learning techniques; only two, practice testing and distributed practice, earned their highest utility rating. Rereading and highlighting, the closest equivalents to watching a demo and keeping the slides, were rated low.
The job decides whether training transfers
Brian Blume and colleagues meta-analysed 89 studies of training transfer, meaning whether people apply what they learned back at work. Cognitive ability, conscientiousness, motivation and a supportive work environment all had positive relationships with transfer. Most of these relationships were stronger when the training targeted open skills such as leadership development, where people must judge how to apply what they learned, than closed skills such as using a specific piece of software.
Using generative AI well sits closer to the open end. The tool interface is easy; the judgement is hard: which task to hand over, how much context to give, when the answer is wrong, what must never be pasted in. That reading is ours, but it explains a familiar pattern. People leave a workshop impressed, try the tool once or twice, hit a bad answer on a real task and go back to the old way, because nobody is there to show them what went wrong.
Satisfaction scores measure the wrong thing
The Kirkpatrick model, a common way to evaluate training, has four levels: reaction, learning, behaviour and results. A feedback form at the end of a workshop measures level 1, whether people found it engaging and relevant. A manager who paid for AI training wants to know about level 3, whether people perform the new behaviour in their job. A one-off event rarely produces any evidence at that level.
What does a multi-week AI programme add?
A multi-week AI programme adds the parts the research points to: practice spread over time, practice on the person's own tasks, feedback on each attempt, and a record a manager can read. Content matters less than structure here; the same prompting technique taught once and taught six times over six weeks gives different results.
| 2–3 hour workshop | One-day course | Multi-week programme | |
|---|---|---|---|
| Typical length | 2–3 hours, once | 6–7 hours, once | 6–8 weeks, one session and one assignment per week |
| Best for | Awareness, rules, leadership alignment | Common baseline for a large group | Changing how a team works day to day |
| Practice on own work | Rarely | Some, mostly generic exercises | Every week, assignments built from real tasks |
| Spacing | None | None | Built in, with a week of real use between sessions |
| Feedback | Questions in the room | Some, during exercises | Written feedback on each assignment |
| Measurement | Satisfaction form (Kirkpatrick level 1) | Sometimes a quiz (level 2) | Scored assignments and a baseline to compare with (towards level 3) |
| What the manager sees | Attendance | Attendance, maybe quiz results | Team report: progress, strengths, gaps, what to train next |
| Time per participant | 2–3 hours | 6–7 hours | Roughly 12–16 hours spread over the period (estimate: 1 hour session plus 1 hour assignment per week) |
Each element matches a finding above. The week between sessions is the spacing. The assignment from the person's own work is what makes the skill open-ended and real, so that the first bad answer happens with support available. Scoring each assignment is practice testing in Dunlosky's sense. The manager report creates the supportive environment Blume's review links to transfer: when a manager can see that someone in finance has started using AI to draft variance comments, they can ask about it, share it and give time for it.
The trade-off is time. A programme asks for 12 to 16 hours per person over six to eight weeks, and it needs a sponsor who keeps it on the calendar when a busy week comes. If nobody will protect that time, a programme becomes a series of workshops with homework nobody does.
Workshop or programme: how do you decide?
Choose a workshop when the goal is shared understanding and choose a programme when the goal is a measurable change in how people do their work. Most companies need both, in that order: a short session to align leaders and set rules, then a programme for the teams where AI touches the most work.
| Factor | A workshop is probably enough | A programme is probably needed |
|---|---|---|
| Goal | Awareness, a common vocabulary, a policy walkthrough | People using AI weekly on real tasks, with fewer errors |
| Who | Management group, the whole company at kick-off | A team with recurring writing, analysis, customer or reporting work |
| Team size | Any; one session scales to a large room | Groups of roughly 6–20 per track, so each person gets feedback |
| Similarity of tasks | Tasks differ widely, generic examples are fine | Tasks repeat, so one good workflow pays back every week |
| Risk | Low: public information, internal drafts | Higher: client data, personal data, numbers that go into decisions |
| Current level | Most people have not tried the tools | Some use AI already, unevenly and without shared rules |
| Evidence needed | Attendance is enough | Management wants to see progress and gaps per team |
| Time available | One block of 2–8 hours | 1–2 hours per person per week for 6–8 weeks |
Example
Example: a fictional 30-person logistics company in Rotterdam. The leadership group of 5 spends a 3-hour workshop agreeing which tasks are in scope and approving two tools. That costs 15 person-hours. The whole company then gets a 2-hour introduction with the AI policy: 60 person-hours. The customer service team of 8, who write around 40 replies each per day, joins a 6-week programme at about 2 hours per person per week, which is 96 person-hours. Before week 1 each person records how long a standard complaint reply takes them; in week 6 they record it again and submit a reply workflow as the final assignment. Total: about 171 person-hours, with more than half aimed at the one team where the work repeats most.
Three situations argue against a programme. A team that is about to change its core system should wait until the new tools are in place. A team whose work does not repeat, such as a small legal team with one-off matters, may get more from individual coaching. And a company without approved tools or data rules should fix that first, because a programme built on unapproved tools teaches habits that later have to be undone.
Which free AI courses are worth using first?
Free courses are worth using first for the basics, because they teach concepts well at no cost and let paid time go to practice on your own work. Assign one before a workshop or as week 0 of a programme, so the live time does not go on explaining what a language model is.
| Course | Made by | Length | Good for | Keep in mind |
|---|---|---|---|---|
| Elements of AI (Introduction to AI) | University of Helsinki and MinnaLearn; offered in Norwegian by NTNU | Self-paced; the Norwegian page gives 25 hours on average | Concepts: what AI is, how machine learning works, limits and societal questions | Not focused on generative AI tools at work; no exercises on your own tasks |
| AI Fluency: Framework & Foundations | Anthropic, with Prof. Joseph Feller (University College Cork) and Prof. Rick Dakan (Ringling College) | Self-paced, with a final assessment and certificate | A framework for working with AI: delegating, describing tasks, judging output, using it responsibly | Made by an AI vendor (the company behind Claude); the course page does not state a duration |
| Introduction to Generative AI | About 45 minutes | A short primer on what generative AI is and how it differs from traditional machine learning | Also covers Google tools for building your own generative AI apps, which most office teams do not need |
None of these courses look at your files, your customers or your rules for sensitive data. They cannot tell a project manager that her status-report prompt leaves out the one figure her director always asks for. That gap is what you pay a provider to close. If you are still choosing tools, our comparison of ChatGPT, Claude and Gemini and the guide to AI tools for business are useful reading before any training starts.
Does the EU AI Act require AI training for staff?
The EU AI Act requires providers and deployers of AI systems, a group that includes companies using AI tools at work, to take measures that support their staff's AI literacy, but it does not prescribe a course, a certificate or a set level. Article 4 has applied since 2 February 2025. The Digital Omnibus on AI, Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026, rewrote the article: the duty to take measures remains, it no longer requires any specific level of AI literacy for any individual, and the Commission and Member States must support companies, SMEs in particular, in meeting it. The European Commission's Q&A says there is no need for a certificate, that an internal record of training and other initiatives can be kept, that relying on the instructions for use is not sufficient, and that different levels of training can suit different staff. Deployers of high-risk AI systems also have a separate duty under Article 26 to make sure the people who oversee those systems are trained for it.
For Norwegian companies the regulation is not yet part of Norwegian law. In a press release on 4 August 2026 the government said it is working to bring the AI Act into the EEA Agreement, that the Norwegian AI law will go out for a new public consultation in autumn 2026, and that the aim is to present it to the Storting in spring 2027. A company working in EU markets, or planning for the Norwegian law, gets a head start from training that leaves a written record: who was trained, on what, and how they did.
What should you ask an AI training provider?
Ask any provider what participants will produce, how it is judged and what you will see at the end. A good answer is specific; a vague one tells you the training will be generic.
- What will each person work on? Ask whether exercises come from participants' own tasks or from a standard case library. Own tasks transfer better; standard cases are easier to deliver.
- How is the work judged, and by whom? Ask for the criteria in writing before the start: is the answer correct, was the right tool used, was sensitive data handled properly, can the method be reused.
- What does the manager receive? A team-level report on progress and gaps is useful. Ask whether individual results are shared, and whether participants consent to that before the start.
- How is data protection handled during training? Exercises on real work must follow your rules on client and personal data. Ask how the provider prevents confidential material from going into unapproved tools.
- Which tools will be taught? Training should match the licences you have or plan to buy. Teaching a tool the company has not approved creates habits you will have to reverse.
- What is measured before and after? Without a baseline at the start, nobody can say whether anything changed. Ask what the first-week survey or test looks like.
- What stays with the team afterwards? Templates, prompt libraries and the final workflows should belong to you. Ask who owns them and in what format they are handed over.
Two follow-up reads help you hold a provider to these answers: the guide to prompt engineering shows what good practice looks like in detail, and the guide to calculating AI ROI shows how to put hours saved against hours spent.
How does LeapJet run its AI training programme?
LeapJet's AI training for teams is a multi-week programme, typically 6 to 8 weeks, with one live session and one assignment per week, online or on site, in English or Norwegian. Each assignment is built from the participant's own work, reviewed with written feedback and scored on criteria agreed at the start: correct result, right tool, sensitive information handled properly, and a method that can be reused. At the end the manager receives a team-level report on progress, strengths, gaps and what to train next; individual scores are shared only if participants agree before the start. Programmes are quoted after a conversation about group size and goals. If your team is ready to move from trying AI to using it in recurring work, the AI automation guide shows where those workflows usually lead next, and the next step is to list the five tasks your team repeats most and bring that list to the first conversation.
Frequently asked questions
How long should AI training for employees be?
It depends on the goal. Two to three hours is enough for awareness and rules. For people to change how they work, plan for several weeks with practice between sessions, because spaced practice is retained better than one long block.
Is a one-day AI course enough for a team?
A one-day course gives a team a shared baseline and some hands-on time. It rarely changes daily habits on its own, because the exercises are generic and there is no feedback once people are back at their desks. Follow it with practice on real tasks if behaviour change is the aim.
Are free AI courses good enough?
Free courses such as Elements of AI, Anthropic's AI Fluency course and Google's Introduction to Generative AI teach the basics well. They cannot give feedback on your own work or check how your team handles sensitive data. Use them as preparation, then spend paid time on practice.
How do you measure whether AI training worked?
Measure what people do at work after the training. Record a baseline before training, such as time spent on a standard task or the quality of a typical output, and measure the same thing at the end. Scored assignments and a team report give a manager evidence at Kirkpatrick level 3.
Do employees need a certificate under the EU AI Act?
No. The European Commission says Article 4 does not require a certificate. Companies should take measures suited to their staff and the AI systems they use, and can keep an internal record of training as documentation. Since the July 2026 amendment, Article 4 also does not require any specific level of AI literacy.
Sources
- Blume, Ford, Baldwin and Huang: Transfer of training, a meta-analytic review. Journal of Management 36(4) (2010)
- Cepeda, Pashler, Vul, Wixted and Rohrer: Distributed practice in verbal recall tasks, a review and quantitative synthesis. Psychological Bulletin 132(3) (2006)
- Dunlosky, Rawson, Marsh, Nathan and Willingham: Improving students' learning with effective learning techniques. Psychological Science in the Public Interest 14(1) (2013)
- Association for Psychological Science: Which study strategies make the grade? (2013)
- Ebbinghaus: Memory, a contribution to experimental psychology (1885; English translation 1913)
- Murre and Dros: Replication and analysis of Ebbinghaus' forgetting curve. PLoS ONE (2015)
- Kirkpatrick Partners: The Kirkpatrick Model
- EUR-Lex: Regulation (EU) 2024/1689 (AI Act), Article 4 (2024)
- EUR-Lex: Regulation (EU) 2026/1744, Digital Omnibus on AI (2026)
- European Commission: AI literacy, questions and answers (updated 27 July 2026)
- Regjeringen: Tung vil sende KI-loven med endringer på høring (4 August 2026)
- University of Helsinki and MinnaLearn: Elements of AI
- Elements of AI, Norwegian version offered by NTNU
- Anthropic Academy: AI Fluency, Framework & Foundations
- Google Skills: Introduction to Generative AI
- Google Cloud blog: Generative AI on Google Cloud, new training content (2023)
