If people have been through AI training and still do not use AI at work, look first at the workplace, not the trainees. In research on training transfer, what the workplace did before and after the course was more strongly linked to on-the-job use than what happened during it. A March 2026 report on AI adoption found that whether an employer encourages use was the strongest predictor of whether workers use AI at all. This guide goes through that evidence and proposes a check card that tests, two weeks after training, whether the workplace is giving trainees something to stand on.
This is a practical guide for people who run AI training, not a report on a customer engagement or a measured result. It rests on paper abstracts and NBER reports, all read on November 8, 2026. The studies cited are not all about AI training, and most rest on surveys and show associations rather than causes. The check card, the timing and the fictional sales-support team below are DMS.Labs design proposals and assumptions. They do not mean the approach worked in any organisation.
Why does AI training not turn into use at work?
It is hard to say the trainees simply lacked motivation. The studies report that on-the-job use moves together with conditions in the workplace. Which condition is missing differs from one organisation to the next, so you have to ask.
Satisfaction right after training cannot answer the question. Saying the lecture was good and actually using the tool next Monday are different things. A trainee may want to use AI and have no account. They may have an account and never have been told by their manager which material is allowed. Time to try it may appear nowhere on the work schedule. None of these gaps gets filled inside the classroom.
What does research say about workplace support?
The table below collects what I read. The right-hand column is DMS.Labs' interpretation.
Scroll horizontally to view a wide table.
| Source | What I confirmed | DMS.Labs' interpretation |
|---|---|---|
| Saks and Belcourt, 2006, Human Resource Management (abstract read) | Training professionals at 150 organisations reported that 62% of employees apply training material on the job immediately, 44% after six months and 34% after one year. The organisations used transfer-supporting activities more during training than before or after it. Yet activities in the work environment before and after training were more strongly related to transfer than activities during training | A hint that workplaces do relatively little before and after training, although that is where the stronger link to use sits. The figures are reports from training professionals, not measured employee behaviour, and the study is not about AI training |
| Blume, Ford, Baldwin and Huang, 2010, Journal of Management (abstract read) | A meta-analysis of 89 studies. Cognitive ability, conscientiousness, motivation and a supportive work environment were positively related to transfer. Most predictors were more strongly related to transfer for open skills such as leadership development than for closed skills such as computer software. When the same source reported the transfer outcome in the same measurement context, the relationships were inflated | Using AI involves judgement beyond a fixed procedure, so I see it as closer to an open skill. That classification is my interpretation; the study did not classify AI. It is also a reason not to judge on-the-job use from trainees' own satisfaction answers alone |
Both studies point the same way. The effect of training does not end in the training room; the environment trainees go back to shapes whether they apply it.
Does the same pattern show up for AI training?
I read the two most recent AI sources, and their results do not overlap completely. Both are NBER working papers, and their cover pages say they have not been peer reviewed.
The first is the Danish study by Anders Humlum and Emilie Vestergaard. It first appeared in May 2025 and was revised in March 2026; it originally circulated as "Large Language Models, Small Labor Market Effects". For the 2024 survey round it links about 25,000 valid responses to administrative records. Even in workplaces with no employer encouragement, no enterprise chatbot and no training, about 41% of workers had used an AI chatbot at work at least once, and 7% used one daily. Where a single support was in place, the share who had tried one was 76% with encouraged use, 57% with an enterprise chatbot and 69% with training. Where all three were in place, 93% had tried one and 28% used one daily. The authors note that these supports are workers' own reports of what they perceive, which may differ from what the employer intended. A footnote adds that training or tools without encouragement were associated with smaller reported gains.
The second is "Mind the Gap: AI Adoption in Europe and the U.S." (March 2026) by Alexander Bick, Adam Blandin, David Deming and co-authors. It includes a survey of workers in the United States and six European countries in January and February 2026. When encouragement, tool provision and training were put into one regression, encouragement was the strongest predictor of use. Among workers who received neither training nor tools, 47% of those whose employer encouraged AI use reported using it at work, against 10% of those who were not encouraged. Among workers who received neither encouragement nor training, 21% of those given tools used AI, against 10% of those not given tools. Once encouragement and tools were controlled for, training did not show up as a variable that raised use further.
A quiet office corner with a closed laptop, a paper-clipped printed training handout, a mug and an unused pad of sticky notes on a wooden desk by a window.View original
There are cautions in reading the two together. Both are observational, so neither can say that encouragement or training caused the use. The controls also differ. Humlum and Vestergaard compare workplaces where each support is in place on its own, while Bick and co-authors put all three in one equation to separate their shares. That does not support the conclusion that training is useless. What these analyses look at is whether people use AI, not the quality or safety of how they use it. What can be read from them is limited: training moves alongside encouragement and tools, and a workplace where those two are missing may not support the use of someone who has just finished a course.
The earlier PNAS paper by Humlum and Vestergaard, published at the end of 2024, points the same way. Its abstract says workers named employer restrictions on use and the need for training as the main barriers to adoption. I confirmed only the abstract and excerpts of that paper.
What goes on a workplace check card?
Make one card per team that attended the training. The person who runs the check is the team's manager, not the trainee. The layout is a DMS.Labs proposal; no study has tested these five rows directly.
Scroll horizontally to view a wide table.
| Row | Question to check | Linked research | If it is empty |
|---|---|---|---|
| Say it | Has the team's manager named one task and said "try it on this"? | Encouragement was the strongest predictor of use (Bick et al.) | The manager picks one task and sends a one-line note |
| Tools | Does each trainee actually have an account and access? | Tool provision came next after encouragement (Bick et al.) | Compare the attendee list with the account roll-out |
| Time | Is there time in the working day to try it? | Workplace activities before and after training were linked to transfer (Saks and Belcourt). I did not read a study that tested a time row itself | Put about 30 minutes a week on the schedule as trial time |
| Rules | Does the team know which material may be entered and which may not? | Employer restrictions were reported as a barrier to adoption (Humlum and Vestergaard, PNAS) | Write the rules on one page. For choosing material, see Choosing Data to Share with AI |
| Checking | Who checks use, when, and with what? | Answers from the same source at the same time inflated the relationships (Blume et al.) | Add usage records or the manager's own observation to the trainee's one-line report |
A word on the fifth row. If you only ask trainees "are you using it?", the wish to use and the fact of using blend into one answer. Adding a second source, such as account usage records or something the manager saw, shows the gap between them.
What do you check two weeks after training?
I will explain with a fictional scene. It is not a DMS.Labs customer case, and the numbers are assumptions for illustration.
Suppose eight people on a sales-support team attended a one-day course, and two weeks later their manager talked to each of them. Three had used AI in their work and five had not. Sorting the five reasons into the card's five rows, suppose two said the manager had never told them to use it, two said they still had no account and one said they had no time. In that case, running the same course again would leave four of the five blocked exactly as before. What they need is a one-line note from the manager and an account.
The order DMS.Labs proposes is this. Before the course ends, write the check date and the responsible person on the card. After two weeks, open the card and go through the five rows one at a time; for every empty row, write who will fix it and by when. After three months, look at usage records once more. The intervals of two weeks and three months are my proposal, not what research recommends. Saks and Belcourt asked at six months and one year, and a team with a long work cycle may fit one month better. A way for trainees to recheck the boundary of what they hand over is covered in AI training should start with 'Should I hand this task over?'. That guide is the trainee's check; this one is the workplace's.
The purpose is not to grade the trainees. If nobody is using AI after two weeks, that is not in itself a failure. The card is a way to sort out whether the reason lies in the workplace.
A wall-mounted wooden key rack in an office hallway with five keys hanging and one hook empty, and a blank clipboard hanging beside it.View original
Frequently asked questions
How is telling people to use AI different from forcing them?
The studies I read show that encouragement moves together with use. They do not cover what happens when a number of uses is set for people. DMS.Labs proposes naming one task and, where AI is better left out of that task, saying why and writing down that people need not use it. I have not tested whether this still feels like pressure.
If few people use AI after two weeks, has the training failed?
I do not read it that way. In the Humlum and Vestergaard data, about 41% of workers had tried an AI chatbot even where the employer gave no support at all. A high usage rate also does not guarantee that work is handed over well, and some people do not use AI because a task should not be handed over. What to check is whether the reason for not using it is a gap in the workplace or a judgement about the work. The pattern in Saks and Belcourt, where the share applying training falls over time, was reported for training in general, not AI.
Will holding the course again fix it?
If a gap in the workplace is the cause, it may not. In the Bick et al. analysis, training did not explain use further once encouragement and tools were controlled for. That does not mean training can be cut, because the analysis looked only at whether people use AI. Still, it makes sense to check the card's five rows before repeating the same course.
What to read next
You can see the training outline on the DMS.Labs generative AI training page, which is provided in Korean. If you are preparing a consultation, pick one team that recently finished AI training and write down just one empty row on the card; it makes the current state easier to describe. For conditions, see the contact page.
For how to decide which numbers to look at after training, see Measuring AI adoption: revenue, costs and review time.
Sources
- Saks, A. M., & Belcourt, M., An investigation of training activities and transfer of training in organizations, Human Resource Management 45(4) (2006) (abstract only)
- Blume, B. D., Ford, J. K., Baldwin, T. T., & Huang, J. L., Transfer of Training: A Meta-Analytic Review, Journal of Management 36(4) (2010) (abstract only)
- Humlum, A., & Vestergaard, E., Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI, NBER Working Paper 33777 (May 2025, revised March 2026) (the revised version was read in full; not peer reviewed)
- Bick, A., Blandin, A., Deming, D. J., Fuchs-Schündeln, N., & Jessen, J., Mind the Gap: AI Adoption in Europe and the U.S., NBER Working Paper 34995 (March 2026) (read in full; not peer reviewed)
- Humlum, A., & Vestergaard, E., The unequal adoption of ChatGPT exacerbates existing inequalities among workers, PNAS 122(1) (2025) (abstract and excerpts only)