العربية Create account

Skills at work · 29 October 2026 · 10 min read

How do you use AI at work without losing the skill?

Imagine an employee hands you a tidy report ahead of schedule. Its recommendations are clear, and its language is persuasive. You ask him: "Why did you choose this solution? And what might make you change your mind?" He then returns to the conversation with artificial intelligence looking for the answer.

ZERAX Academy

Share

The short version in four numbers

  • 25.1% less time on average for a set of tasks suited to GPT-4's capabilities, in an experiment with Boston Consulting Group consultants.
  • 15% more issues resolved per hour on average, across more than five thousand customer service employees, according to the version published in 2025.
  • 17% lower test scores without help for high school students who practised with a ChatGPT-like assistant, against a group that did not use AI.
  • 52 developers in a small randomized experiment made available in 2026: assistant users scored lower on a later test without the tool.

What can fast delivery hide?

This hypothetical situation illustrates what delivery speed might hide: the quality of the document alone does not reveal the extent of its author's understanding. The employee can get things done with the help of the tool, while still needing to learn what enables him to review the result and defend his decision.

For the manager and the training manager, the artificial intelligence experience requires two parallel measurements: how do you improve work with its help? And what can the employee understand and execute when working alone?

The quality of work delivered with the tool does not show what the employee can do alone, so measure bothTwo parallel measurements. First, work with the tool: how do you improve work with its help? Measured by comparing time and quality with and without the tool, counting review and correction time. Second, what the employee can do alone: what can they understand and execute when working alone? Measured with a similar case with different details, solved and explained without the tool. The quality of work delivered with the tool does notshow what the employee can do alone, so measure bothThe two parallel measurements an AI trial needs1. Work with the toolHow do you improve work with its help?Measured by comparing time and quality with andwithout the tool, counting review and correctiontime.2. What the employee can do aloneWhat can the employee understand andexecute when working alone?Measured with a similar case with different details,which the employee solves and explains withoutthe tool.Source: ZERAX Academy, from the text of this article.
The quality of work delivered with the tool does not show what the employee can do alone, so measure bothTwo parallel measurements. First, work with the tool: how do you improve work with its help? Measured by comparing time and quality with and without the tool, counting review and correction time. Second, what the employee can do alone: what can they understand and execute when working alone? Measured with a similar case with different details, solved and explained without the tool. The quality of work deliveredwith the tool does not showwhat the employee can doalone, so measure bothThe two parallel measurements an AI trial needs1. Work with the toolHow do you improve work with its help?Measured by comparing time and quality with and withoutthe tool, counting review and correction time.2. What the employee can do aloneWhat can the employee understand and execute whenworking alone?Measured with a similar case with different details, whichthe employee solves and explains without the tool.Source: ZERAX Academy, from the text of this article.
The two measurements from the introduction's questions, and how each is taken.

Does artificial intelligence increase employee productivity?

It can, but the amount of benefit varies depending on the task, the employee's experience, and the way it is used.

In an experiment on Boston Consulting Group consultants, a set of tasks suited to GPT-4's capabilities was completed in 25.1% less time on average. In a separate study on more than five thousand customer service employees, the number of issues resolved per hour increased by 15% on average, according to the version published in 2025. [1, 2]

The gains of the less experienced and skilled customer service employees were greater, while the highly skilled benefited to a limited degree, with a slight decline in some quality metrics. So do not assume that the tool that helps the new employee will add the same value to the expert's work.

You can translate this into an experiment within your team: choose a repetitive task, set a clear acceptance criterion, then compare the time and quality with and without the tool. Include the review and correction time within the completion time; it is part of the work you need to deliver.

How does the answer look professional and still be wrong?

In the same consultants' experiment, the probability of reaching the correct solution decreased when participants used GPT-4 in a business case chosen by the researchers because it exceeded its capabilities at the time. Success in previous tasks was not a guarantee of success in this task. [1]

The researchers call this disparity "the jagged technological frontier": two tasks may be similar in difficulty for a human, and the tool succeeds in one and stumbles in the other.

Two tasks can be equally hard for a person, and the tool succeeds at one and stumbles on the otherAn illustration: a jagged line separates what lies inside the tool's capability from what lies outside it. Two tasks sit at the same height, the same difficulty for a person: the first under a peak of the line, where the tool succeeds; the second over a dip, where the tool stumbles. Two tasks can be equally hard for a person, andthe tool succeeds at one and stumbles on the other“The jagged technological frontier”: both tasks sitat the same height, the same difficulty for a personDifficulty for a personDifferent tasksInside the tool's capabilityOutside itThe tool succeedsThe tool stumblesSource: Dell’Acqua et al. (2026), Organization Science. An illustration, not data.
Two tasks can be equally hard for a person, and the tool succeeds at one and stumbles on the otherAn illustration: a jagged line separates what lies inside the tool's capability from what lies outside it. Two tasks sit at the same height, the same difficulty for a person: the first under a peak of the line, where the tool succeeds; the second over a dip, where the tool stumbles. Two tasks can be equally hard fora person, and the tool succeeds atone and stumbles on the other“The jagged technological frontier”: both tasks sitat the same height, the same difficulty for a personDifficulty for a personDifferent tasksInside the tool's capabilityOutside itThe tool succeedsThe tool stumblesSource: Dell’Acqua et al. (2026), Organization Science. Anillustration, not data.
An illustration of the idea, not the experiment's data.

Good presentation may be confused with the correctness of the conclusion. Therefore, when reviewing a recommendation, ask the employee to identify the data it is based on, the assumption that might change the decision, and the method by which they verified the result. Returning the question to the tool itself is not enough on its own for verification.

This experiment used an early version of GPT-4, and its results first appeared in 2023 before being published in a scientific journal in 2026. Its usefulness today is to remind you to test the limits of the tool you are actually using; it is not right to turn it into a fixed judgment that current models fail in every complex analytical task.

Do we lose our skills, or miss the opportunity to acquire them?

The risk of "skill erosion" or "skill atrophy" is raised when talking about reliance on automation. But there is a difference between a skill the employee mastered and then weakened, and a new skill they never acquired in the first place because the tool performed the learning steps for them.

The following research primarily addresses the second case, and measures results close to the training time. It does not prove that every artificial intelligence user loses their skills as years pass.

A skill the employee mastered and then lost is not the same as a new skill never gainedTwo illustrations of skill over time. First, losing a skill: a line rises to mastery, then falls with reliance on automation. Second, never gaining it: a line stays low because the tool performed the learning steps. The research in the article mainly addresses the second case. A skill the employee mastered and then lostis not the same as a new skill never gainedThe research in this article mainly addresses the secondcase, and measures results close to the time of trainingLosing a skillMastered, then weakened with reliance onautomationSkillTimeNever gaining itNever acquired, because the tool performed thelearning stepsSkillTimeSource: ZERAX Academy, from the text of this article. An illustration, not data.
A skill the employee mastered and then lost is not the same as a new skill never gainedTwo illustrations of skill over time. First, losing a skill: a line rises to mastery, then falls with reliance on automation. Second, never gaining it: a line stays low because the tool performed the learning steps. The research in the article mainly addresses the second case. A skill the employee masteredand then lost is not the sameas a new skill never gainedThe research in this article mainly addresses the secondcase, and measures results close to the time of trainingLosing a skillMastered, then weakened with reliance on automationSkillTimeNever gaining itNever acquired, because the tool performed the learningstepsSkillTimeSource: ZERAX Academy, from the text of this article. Anillustration, not data.
The difference between losing a skill and never gaining it.

The ready-made answer may improve the exercise and weaken learning

In an experiment on about a thousand high school students in Turkey, students' performance improved during training using a ChatGPT-like assistant. But when the assistance was removed, their test scores were 17% lower than a group that did not use artificial intelligence. [3]

Ready-made answers improved practice, then test scores without help came in 17% lowerTest scores without help for about a thousand high-school students in Turkey, against a group that did not use AI: students who practised with a ChatGPT-like assistant that gave ready-made answers improved during practice, then scored 17% lower. Students who used an assistant giving hints prepared by teachers showed no statistically significant difference. Ready-made answers improved practice, thentest scores without help came in 17% lowerAbout a thousand high-school students in Turkey; test scoresonce the help was removed, against a group that did not use AIDid not use AIreferenceChatGPT-like assistant thatgives ready-made answersPerformance improved during practice−17%Assistant giving hintsprepared by teachersno statistically significant differenceSource: Bastani et al. (2025), PNAS.
Ready-made answers improved practice, then test scores without help came in 17% lowerTest scores without help for about a thousand high-school students in Turkey, against a group that did not use AI: students who practised with a ChatGPT-like assistant that gave ready-made answers improved during practice, then scored 17% lower. Students who used an assistant giving hints prepared by teachers showed no statistically significant difference. Ready-made answers improvedpractice, then test scoreswithout help came in 17% lowerAbout a thousand high-school students inTurkey; test scores once the help wasremoved, against a group that did not use AIDid not use AIreferenceChatGPT-like assistant that gives ready-madeanswersPerformance improved during practice−17%Assistant giving hints prepared by teachersno statistically significant differenceSource: Bastani et al. (2025), PNAS.
The experiment was in high school mathematics, not a measurement of corporate employees.

As for the group that used an assistant designed to provide hints, and supplied with solutions and information prepared by teachers, their test scores showed no statistically significant difference from the group that learned without the assistant. This design helped avoid the decline observed with the version that provides ready-made answers.

The experiment was in high school mathematics, not a measurement of corporate employees. What can be utilized in professional training is the design question: does the trainee need to attempt and explain, or can they reach the end by copying the answer?

Newer evidence from learning programming

In a research paper made available in 2026, two researchers from Anthropic studied the impact of artificial intelligence assistance on learning a new programming library among 52 developers. The assistant users scored lower on a subsequent test without the tool, and no statistically significant improvement was proven in average completion time. [4]

This is a small randomized experiment, published as a preprint, in which participants used an assistant based on GPT-4o. The researchers measured understanding immediately after the task, so it adds evidence closer to professional work without settling the impact of long-term use or the performance of the latest programming assistants.

The researchers also noted that those who asked for explanations or asked conceptual questions achieved better results than some patterns of full delegation. This is an exploratory correlation in a small sample, but it helps formulate useful questions when reviewing how employees learn.

Seven rules to use the tool and maintain the employee's role in thinking

These results can be translated into the following practices in training and work management, while evaluating their usefulness within your team.

1. Determine the task's goal before choosing the assistance method

If the employee masters the work and needs to shorten repetitive steps, delegating a larger part to the tool with review might suit him. But if he is learning a skill he will need to judge the results later, leave him steps to perform himself. Clarify in the assignment whether you are measuring the quality of delivery, skill acquisition, or both.

2. Request a first attempt before showing the solution

In learning and decision-making tasks, ask the employee to record his understanding of the problem, the solution he favors, and what information he lacks. Then he uses the assistant to discuss this attempt. This way, the manager can see what changed in the employee's understanding, instead of settling for a polished final document.

3. Use hints to break through stumbling

The trainee can ask:

"This is my attempt. Identify the first step that needs review, and ask a question that helps me correct it. Wait for my answer before completing the solution."

This is a suggested request, not a guarantee of the tool's compliance or the correctness of its guidance. In structured training, review the hints and solutions with a content expert, as the teachers did in the educational experiment.

4. Review the evidence and logic before improving the phrasing

Separate content review from editing. Verify the numbers in their source, test the calculations, and discuss the assumptions. In important decisions, ask the employee to explain the reason for accepting the recommendation and what would make him reject it. Then improve the presentation and language after the content is settled.

5. Test understanding without the assistant

After training, present a similar case with different details, and ask the employee to solve it and explain his choice without using the tool. Use a specific evaluation criterion, and re-evaluate later if the skill is important for the role. The employee's satisfaction with the learning experience and his confidence in the answer do not substitute for this evidence.

6. Make the review an opportunity to transfer experience

Ask the new employee to state the suggestions he accepted and those he modified or rejected, with the reason. The manager discusses an example from his work with him instead of just asking "Did you use artificial intelligence?". For the expert, monitor the quality of the decision and the required corrections; increased speed alone may hide a decline in another aspect of performance.

7. Update usage rules whenever the tool changes

Record the tasks you tested, the tool or model used, the quality criterion, and the recurring errors. When changing the model or the working method, retest an appropriate sample. The tool's limits change, as does your team's experience in using it; do not make an old experiment a permanent policy.

Frequently Asked Questions

Should artificial intelligence be banned during training?

First determine what you want the employee to learn. You may allow assistance during practice, then require independent performance during evaluation. The employee can also be trained on the good use of the tool, provided that you measure this ability clearly instead of mixing it with his mastery of the task without assistance.

Are these results still useful after the evolution of the models?

Yes, to understand the difference between completing the task and learning it. But they do not give a fixed gain or loss percentage for today's tools. The customer service study relied on a tool based on GPT-3 and a broad deployment during 2020–2021, and the educational experiment was conducted in autumn 2023, while the newer developers' paper used GPT-4o. Test the results on your tools and tasks before generalizing them.

What is the first action a training manager can take?

Choose one skill the team needs to review artificial intelligence outputs. Design a training task for it with the help of the tool, then an independent case without it, and compare understanding and errors according to the same criteria. Use the results to adjust the assistance method in the next training.

Sources

  1. Dell’Acqua et al. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science, 37(2), 403–423. Publication record and official abstract at Harvard Business School. The paper first appeared in 2023.
  2. Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. Published text on the researcher's website. The numbers of the published version were adopted, not the estimates of the 2023 draft.
  3. Bastani et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122. Full text. The experiment was conducted in autumn 2023; the August 2025 correction relates to a researcher's institutional affiliation, not the results.
  4. Shen, J. H., & Tamkin, A. (2026). How AI Impacts Skill Formation. arXiv:2601.20245, second version. Preprint. It was first made available in January 2026, and the sample limits and short-term measurement were taken into account when presenting it.
  5. ZERAX Academy, Corporate training features.
Share

How does ZERAX help design training that asks the employee to think?

To apply these rules, training needs situations where the employee makes a decision, activities that reveal his understanding, and control over the assistance available to him. ZERAX provides training based on work scenarios and post-lesson activities, with the ability to make activities mandatory and determine whether the answer appears after three attempts or does not appear. [5]

The training assistant answers from the training content itself, mentioning the lesson it relied on, and it can be enabled or disabled for each training. These options allow the training manager to organize the use of assistance, and the manager complements it by evaluating the application at work.

We can build this for your team. Request an offer for your team: tell us about your organization and the number of your employees, and we will get back to you with an offer that suits you. Or email us at contact@zerax.online.

Browse courses

Read next