AI Upskilling
When employees don't use AI after training, that alone does not prove the course failed or the person resists change. Often a specific barrier stands between the skill they learned in class and a real work task. Find that barrier, then help with exactly that point.
Why don't employees use AI after training?
The manager sees the team has completed the course but AI is nowhere in their work. The first impulse is to assign a second course or require a set number of weekly prompts. Both steps can miss the point if the specific barrier has not been removed first.
Humlum and Vestergaard surveyed about 18,000 workers across 11 occupations in Denmark. Some participants, picked at random, got a short message about AI. It changed what they believed about AI, but their use over the next two weeks stayed the same. The study was published in PNAS 122(1), 2025, DOI 10.1073/pnas.2414972121, CC BY 4.0. This was not a trial of a corporate course, and stated need does not measure actual skill. The practical takeaway: the belief that AI is useful does not by itself start use. Checking for a specific barrier is more productive than treating no use as proof of weak motivation.
The barrier framework and the order of help in this article are an original proposal for a conversation with the team. None of the studies listed tested this exact order, and it does not guarantee use or lasting mastery.
Start with the last work attempt
The first question to the employee is simple: when did you last try to use AI on a work task?
Ask them to walk through a specific case: what the assignment was, who was waiting for the result, what they typed in the prompt, what the tool returned, what edits they made, and how it ended.
If there was no attempt, don't invent one together. Choose a task the employee genuinely needs to complete in the coming days. That task becomes the material for the next check. No attempt after the course does not say anything by itself - there may have been no suitable task, no access, or no allocated work time.
Check access and task fitness
Two barriers come before any work with AI: no access to the tool and an unsuitable task. Check both before moving to prompt formulation.
Access. Check from the employee's own workstation: does the tool open, does the needed file load, are the permissions in place for this type of use. If the system returns a refusal or permissions are missing, that is a question for infrastructure, not a sign of weak skill. While the tool is inaccessible, no conclusion about skill is possible: a person cannot demonstrate an ability if they cannot start the work. The manager also allocates work time for the task - without it, even an open tab does not produce a result. If restoring access quickly is not possible, the simpler move is to choose a task where access already exists.
Readiness sign: the employee loads the file themselves and sees the content.
Task fitness. AI does not help on every assignment. Before starting, name the real recipient of the result and the minimum criterion: how will they know the work is done. Then assess whether the available material can support the conclusion needed.
Dell'Acqua and colleagues ran an experiment with 758 consultants at one firm. On tasks within the tool's capabilities, AI improved the result. On a task beyond those capabilities, it reduced accuracy. Results published in Organization Science, Copyright © 2026 The Author(s), DOI 10.1287/orsc.2025.21838, used under a Creative Commons Attribution License: https://creativecommons.org/licenses/by/4.0/. Limitations: quality and time were assessed separately, one version of the tool was used and one task was outside its scope, some ratings were subjective, and long-term use after the course was not measured. Practical takeaway: success on one type of assignment does not transfer automatically to another.
If you doubt fitness, ask an experienced colleague to try on the same permitted material. One colleague failure does not prove the task is absolutely unsuitable - it points to the need to examine the material, the tool, and the framing. If the colleague also struggles, change the task or narrow AI's role: for instance, let the tool prepare a draft while the employee makes the decision.

Help where the person is stuck
When access exists and the task fits, the employee will show you a specific difficulty. It might be formulating a prompt, passing material in the right amount, checking the answer, making edits, or deciding whether the result is good enough.
Sahni and Chilton interviewed ten experienced users of M365 Copilot, Microsoft's AI assistant for office apps, and recorded: independent trials, asking colleagues for help, lack of time, and failed attempts. This is a preprint, a paper not yet peer-reviewed - arXiv 2502.13281v1, February 2025, CC BY 4.0. The sample is small and may not include those who already quit the tool entirely; these are self-reports, not measured effectiveness. The description matches what a manager sees in a team: a person tries, gets stuck, asks a colleague, or gives up.
Help works like this: go through the exact action where the employee stops. Show it once if needed. Then the employee does it themselves. If the colleague finished the whole task, the skill cannot be counted as restored.
The next task of the same type - with a different recipient or a different amount of material - will show whether the person manages independently. One successful result is an observation about one task, not proof of lasting mastery.
What does a support plan look like?
The example below is fully constructed. No actions, meetings, or deadlines were observed. The situation, the prompt, and the plan are designed to illustrate the principle. We do not know how a real employee would manage - this example only shows the structure of fit work.
Situation. A project coordinator must prepare a draft working note from permitted public material for a manager. The file does not open in the work tool. The recipient has not yet named a criterion. The coordinator is unsure how to separate a source fact from their own recommendation. An experienced colleague exists, but their time has not been agreed.
Prompt to the AI:
I am a project manager. For an educational example, compose a finished support plan before the coordinator's next attempt.
Situation: the coordinator must prepare a draft working note from permitted public material. The file does not open in the work tool. The note's recipient has not yet named a criterion. The coordinator is unsure how to tell a source fact from their own recommendation. An experienced colleague is available, but their time and the way to approach them are not yet agreed. Time and quality are not measured.
Public material: Government Digital Service, Artificial Intelligence Playbook for the UK Government, https://www.gov.uk/government/publications/ai-playbook-for-the-uk-government/artificial-intelligence-playbook-for-the-uk-government-html. © Crown copyright 2025. Contains public sector information licensed under the Open Government Licence v3.0. https://www.nationalarchives.gov.uk/doc/open-government-licence/version/3/. Verified independent paraphrase of the needed part: adoption is chosen based on the need of the work and the people; training is based on learning needs; support is chosen based on the situation, using existing structures when they fit. Do not use other provisions of the document in this task. Purpose: own illustration, not GOV.UK endorsement.
Compose a plan that connects: access check, clarity of the note's goal, checking the conclusion, colleague help, and the next independent work. Specify who does what, how to agree on help availability without invented deadlines, and define the next check on new material right now. Do not judge skill before access is restored. Do not promise to resolve all barriers. This is a plan, not an action taken.
Finished constructed response:
Support plan for the coordinator before the next attempt
Paraphrase source: Government Digital Service, Artificial Intelligence Playbook for the UK Government, https://www.gov.uk/government/publications/ai-playbook-for-the-uk-government/artificial-intelligence-playbook-for-the-uk-government-html. © Crown copyright 2025. Contains public sector information licensed under the Open Government Licence v3.0. https://www.nationalarchives.gov.uk/doc/open-government-licence/version/3/. Purpose: own illustration, not GOV.UK endorsement.
Situation: the coordinator needs a draft note from permitted public material. The file does not open in the work tool. The recipient has not named a criterion. The coordinator is unsure how to separate fact from recommendation. A colleague exists, but there is no agreement yet.
Step 1. Access. Responsible: manager and infrastructure owner.
The manager checks whether the file opens from the coordinator's workstation. If the system returns a refusal or permissions are missing, the infrastructure owner restores access to the permitted material. Bypassing restrictions is not allowed. Until access is restored, the task does not run and skill cannot be assessed: a person cannot show a skill if they cannot start the work.
Readiness sign: the coordinator loads the file themselves and sees the content.
Step 2. Goal and criterion. Responsible: manager.
Before work begins, the manager names the note's recipient and the minimum criterion: the text contains a fact from the source, a separate conclusion, and a separate unresolved question; source links are verifiable. The recipient confirms this is sufficient to accept the draft. If the recipient cannot name a criterion, the task does not start: formulating the goal becomes the first work operation.
Step 3. Help checking the conclusion. Responsible: experienced colleague by agreement.
Once access is restored and the criterion is named, the manager asks the colleague to help the coordinator specifically with separating fact from recommendation. The coordinator shows the specific difficulty - for example, which line of the source they cannot distinguish from their own conclusion. The colleague explains the principle; the coordinator checks it themselves. The colleague does not write the note for them.
Availability agreement: the manager asks the colleague when they can spare one conversation on this specific topic. Time is chosen based on both people's actual workload, not on a norm from this article or a preset number of hours. If the colleague cannot in the coming week, the plan does not break: access and criterion are already agreed, and help shifts to the next available slot.
The existing task card in the tracker is the place to reach out. No new tools are needed.
Step 4. Next independent work. Defined now.
After the coordinator works through one material with the colleague's help, the next check is assigned: a new assignment of the same type with different permitted public material, with no ready template. The coordinator extracts the facts, writes the recommendation, and marks what is unknown. The manager accepts the text by the same criteria from step 2.
If an access problem remains on the new task, return to step 1. If the conclusion is not supported by the material, change the framing. If the action does not work out, adjust the help: perhaps the colleague needs to show one more example or narrow AI's role.
Step 5. Profitability check. Unknown for now.
If the question of whether working with AI pays off in this process is decided later, record the full cycle: preparing the material for the prompt, waiting for the response, corrections, checking, transfer to the work document, handoff to the recipient. Separately record one-time instruction preparation and recurring costs per run. Without measurement the balance is unknown. One success on one task does not establish a causal effect of the course.
The plan does not promise to resolve all barriers. The chosen colleague has not agreed yet. Real availability, quality, and time remain unknown until actual work happens.

When is the full AI cycle not worth it?
The fourth barrier concerns balance rather than skill or the task: is it worth using AI in this process at all?
Compare the full path, rather than the tool's response speed alone, to an accepted result at the same quality. This includes preparing the material for the prompt, waiting for the response, edits, checking, transferring to the work document, and handing off to the recipient. Count one-time instruction preparation separately from recurring costs per run: an instruction can pay off with enough repetitions. Payback depends on measured one-time and recurring costs, the same accepted quality, and the number of tasks needed.
For a fair comparison, use new comparable tasks rather than re-running the same assignment. Re-running a familiar task is faster simply because the person already knows it and types the same text again. That does not reflect a real difference in the process. Even such a local measurement does not establish a causal effect of the course - it shows that in this process, this employee, at this moment, managed.
If there is no data for comparison, call the balance unknown. If AI consistently falls short on comparable new tasks, return to the ordinary method in this process. That does not mean AI is unsuitable for other team tasks. Frequency of AI use is not by itself a success criterion: an employee can open the tool daily for minor things and still not apply it where it would bring real benefit.
Frequently asked questions
When is another course needed?
Retraining makes sense when a check of the last attempt showed the employee does not remember basic actions: does not know where to go, cannot load a file, cannot formulate a first prompt. In that case, retraining on the specific tool and the specific team task makes sense. If the person remembers the steps but is not applying them, a second course will not remove an access barrier, will not make the task suitable, and will not allocate work time. First find out what exactly is in the way.
What if the employee never tried to use AI after the course?
That does not necessarily mean the training failed. There may have been no suitable task, no access, or no allocated time. There is no need to invent an attempt together. Choose a real work task the employee must complete in the coming days, and agree that the next check will be built on it. This gives a concrete starting point and lets you check whether there is a barrier without creating an artificial situation.
How do you agree on colleague help without pressure?
Name the specific topic where help is needed and ask when the colleague can spare one conversation on it. Do not set a deadline in advance and do not set a standard number of hours - time is chosen based on both people's actual workload. If the colleague cannot in the coming days, the plan does not break: access and criterion are already agreed, and help shifts. The task card in the tracker remains the place to reach out and save the result.
When is it reasonable to stop using AI in this process?
If the full cycle with AI consistently loses to the ordinary method on comparable new tasks at the same accepted quality, and one-time instruction costs do not pay off at the expected number of repetitions, return to the ordinary method in this specific process. That does not mean AI is useless for other team tasks or that the course was wasted. One failure is not grounds for that conclusion - you need comparable new tasks, not re-runs of a familiar assignment.
Can you use AI query counts as a training success metric?
No. How often an employee opens the tool says nothing about the quality of results or whether the tool is being used on suitable tasks. An employee can open AI daily for minor things and still not apply it where it would bring real benefit. A useful measure is an accepted result on a work task at comparable quality and time. One such result is an observation about one task, not proof of lasting mastery.
What does "skill restored" mean after a course?
It is an observation, not a permanent status. One accepted result says only that, at this point, on this task, the person managed independently. It does not explain why they were not using AI before, and it does not guarantee the next task will go the same way. For a more reliable observation, you need the next task of the same type, without a template and without help, accepted by the same criteria as the previous one.
How do you know the access barrier is gone?
One sign: the employee opens the tool from their own workstation, loads the needed file, and sees its content. Until that moment, no conclusion about skill is possible - positive or negative. A specific system refusal or missing permissions is an observation about infrastructure, not about the employee's motivation or ability. Removing the barrier is the infrastructure owner's job, not the employee's.
What next
If the team shows no AI use after the course, start with one employee's last attempt. Check access to the tool, name the task criterion, agree on help where the person is stuck, and assign the next task right now. This plan does not guarantee a result, but it makes the result checkable.