Source: [AI workplace training: check what people can do after assistance — Skills for Business]() [← All articles](https://skillsforbusiness.ai/blog/) HR & people · Guide # AI workplace training: check what people can do after assistance How HR and team leads can assess learning alongside AI-assisted output, using research on coding and tutoring to design a practical transfer exercise. [Skills for Business](https://skillsforbusiness.ai/blog/editorial/)30 October 20268 min read On this page 1. [What the coding study actually tested](https://skillsforbusiness.ai/blog/ai-workplace-training-measure-learning-after-assistance/#what-the-coding-study-actually-tested) 2. [A separate education experiment adds a design question](https://skillsforbusiness.ai/blog/ai-workplace-training-measure-learning-after-assistance/#a-separate-education-experiment-adds-a-design-question) 3. [Decide what the employee needs to retain](https://skillsforbusiness.ai/blog/ai-workplace-training-measure-learning-after-assistance/#decide-what-the-employee-needs-to-retain) 4. [Use a short cycle: attempt, assistance, explanation, transfer](https://skillsforbusiness.ai/blog/ai-workplace-training-measure-learning-after-assistance/#use-a-short-cycle-attempt-assistance-explanation-transfer) 5. [A hypothetical onboarding exercise](https://skillsforbusiness.ai/blog/ai-workplace-training-measure-learning-after-assistance/#a-hypothetical-onboarding-exercise) 6. [Give developers a debugging case, not a typing contest](https://skillsforbusiness.ai/blog/ai-workplace-training-measure-learning-after-assistance/#give-developers-a-debugging-case-not-a-typing-contest) 7. [Measure learning without turning it into surveillance](https://skillsforbusiness.ai/blog/ai-workplace-training-measure-learning-after-assistance/#measure-learning-without-turning-it-into-surveillance) 8. [Keep learning instructions separate from delivery instructions](https://skillsforbusiness.ai/blog/ai-workplace-training-measure-learning-after-assistance/#keep-learning-instructions-separate-from-delivery-instructions) 9. [Sources](https://skillsforbusiness.ai/blog/ai-workplace-training-measure-learning-after-assistance/#sources) What to take away - Assess the assisted output and retained understanding as separate outcomes. - Use a related case to check the judgment a human reviewer must retain. - Treat the proposed training cycle as a method to test, not a validated workplace intervention. When employees use AI to learn a new task, assess two outcomes: the work they finish with assistance and the judgment they can exercise afterward. A polished result can justify using a tool. It cannot establish that the person has learned enough to spot a mistake, handle an exception or explain the result to a colleague. For HR and team leads, the practical choice is to separate a production workflow from a learning exercise. In production, an approved assistant may complete a large share of routine work. In learning, deliberately leave some reasoning and diagnosis with the learner. Then check a small, related case without generated answers. This article proposes a training design for that purpose. It is informed by two studies, but it is not a validated training package or a claim that every use of AI weakens skills. ## What the coding study actually tested In [How AI Impacts Skill Formation](https://arxiv.org/html/2601.20245v2), Judy Hanwen Shen and Alex Tamkin studied developers learning Trio, an unfamiliar Python library. The main experiment included 52 participants, with 26 in each condition. The assistant used GPT-4o. Participants completed coding tasks and then took a quiz without AI. The researchers found lower quiz performance with AI assistance and no statistically significant overall improvement in completion time. This was a short experiment on a particular library and interface, not a longitudinal test of workplace competence or current coding agents. The study comes from researchers associated with Anthropic. Its [researcher-written explanation](https://www.anthropic.com/research/AI-assistance-coding-skills) highlights that some participants used assistance while remaining engaged in understanding. Those interaction patterns were observed within the experiment; they were not separately randomized treatments proving that a particular prompting style causes better learning. The paper’s version history records a February 2026 revision. We use that version rather than presenting the study as a new October model evaluation. ## A separate education experiment adds a design question A [2025 PNAS study by Hamsa Bastani and colleagues](https://pmc.ncbi.nlm.nih.gov/articles/PMC12232635/) examined nearly a thousand high-school mathematics students. It compared ordinary practice with two GPT-4-based aids: a general assistance interface and a tutor with safeguards, including teacher-designed hints and problem-specific guidance. Assisted practice improved, but students given the general interface subsequently performed worse on an unaided exam than the control group. The safeguarded tutor largely avoided that negative effect; it did not establish an unaided learning improvement over the control. These are school mathematics results, not employee training results. The paper’s June 2025 publication has an August correction concerning an author affiliation. The study measured short-term performance, not years of skill retention. The two studies do not estimate the same effect. Different populations, tasks and interventions prevent combining them into one workplace percentage. They do motivate a concrete question for a training owner: which part of the practice must the learner still do? ## Decide what the employee needs to retain Start by naming the capability the person will need when the assistant is unavailable, uncertain or wrong. Avoid a broad goal such as “become confident with AI.” It does not specify what successful learning looks like. For a developer, the capability might be identifying why a retry loop could duplicate a write. For an operations colleague, it might be explaining which source determines the deadline in an onboarding checklist. For a manager, it might be distinguishing an observed result from an assumption in a forecast. Choose something relevant to the person’s responsibilities. A learner does not need to memorize every command if reference material is normally available. They do need enough understanding to recognize when the suggested command is inappropriate. Write the learning goal as an observable action: explain, diagnose, compare or justify. Then identify a plausible wrong answer. If the training team cannot describe a meaningful failure, it will struggle to tell understanding from a fluent explanation. ## Use a short cycle: attempt, assistance, explanation, transfer The following cycle is our editorial proposal. The studies above inform its emphasis on retained understanding; they do not validate these exact steps or durations. First, give the learner a small attempt before asking an assistant for a complete solution. Ask them to state the problem and predict one possible failure. This establishes what they currently understand and gives the discussion something concrete to address. Second, allow assistance with a defined role. The assistant might explain an unfamiliar term, offer a hint or question the proposed approach. Make the permitted help clear. Do not quietly expect a learner to resist a full answer when the exercise rewards only fast completion. Third, ask for an explanation tied to the actual work. “Why is this check here?” is more useful than “Do you understand?” Ask the learner to point to a source, trace a dependency or explain what would change under a different condition. Finally, use a related transfer case. Change one important detail and ask for an independent response. Keep approved reference documents available if those are normally part of the job. Temporarily withholding generated answers is an assessment design choice, not a demand that people work from memory. ## A hypothetical onboarding exercise Imagine an HR operations team learning to prepare an onboarding checklist from an approved policy and a role description. The learning goal is source judgment: identify mandatory steps, preserve uncertainty and escalate missing information. The exercise concerns process preparation, not an employment decision about an individual. | Stage | Learner’s task | What the reviewer checks | | --- | --- | --- | | Initial attempt | Mark mandatory steps in a short policy extract | Required actions are distinguished from suggestions | | Assisted practice | Ask an approved assistant for hints about omissions | Hints are checked against the supplied sources | | Explanation | Explain why each proposed deadline applies | Dates have a source and assumptions remain visible | | Transfer | Handle a new role with one missing input | The learner identifies the gap instead of inventing a requirement | Use fictional records or approved training material. Do not bring employee data into an unapproved tool to make the exercise feel realistic. A successful checklist and a successful learning result should have separate records. The checklist may be correct because the assistant supplied it. The transfer case shows whether the learner can identify the relevant rule in a new situation. If the learner misses the exception, revise the exercise before drawing a conclusion about their overall ability. The source may be unclear, the example may introduce an unfamiliar policy issue, or the review criteria may be inconsistent. ## Give developers a debugging case, not a typing contest For developer onboarding, select a small piece of code using an unfamiliar internal abstraction. Ask what it does, how it fails and which test would expose the failure. Keep the scope small enough that understanding can be discussed directly. After assisted practice, provide a second snippet with a changed assumption. For example, a hypothetical queue consumer may now retry after a timeout. Ask whether the write is safe to repeat and which evidence is needed before deciding. The point is not to prove that unaided coding is morally superior. It is to establish whether the developer can review the kind of output the production workflow will ask them to accept. Separate knowledge gaps from environment failures. Missing documentation, unavailable test services and confusing error messages need different remedies from a misunderstanding of concurrency. Record the obstacle before adding another general instruction to a training prompt. ## Measure learning without turning it into surveillance Collect the minimum evidence needed to improve the exercise: the selected case, the explanation, the transfer response and the correction required. Tell participants what is assessed and who sees it. Prompt counts and time spent chatting are poor substitutes for a defined learning outcome. For a small business, a shared exercise and a short reviewer note may be enough. A larger organization can calibrate reviewers with the same sample answers and check whether different teams interpret the criteria consistently. Do not rank employees by one small quiz. A short assessment samples a narrow capability. Offer feedback, another practice opportunity and a later related case before using the result to make broader judgments. Track production performance separately. Our [guide to measuring completed work](https://skillsforbusiness.ai/blog/workplace-ai-adoption-measure-completed-work/) explains how to include review and repair effort. Training adds a different question: can the person still exercise the judgment the workflow assigns to them? ## Keep learning instructions separate from delivery instructions A [reusable agent skill](https://skillsforbusiness.ai/blog/skill-md-files-as-team-work-instructions/) can preserve the learning procedure: ask for an initial attempt, offer bounded hints, request an explanation and provide a new case. A delivery procedure may instead authorize drafting the complete output immediately. Label the mode clearly and test whether the intended assistant follows it. A file describing a tutoring approach does not establish that a product enforces it, and an instruction alone cannot replace access controls or an accountable reviewer. Do not withhold help during urgent production work just to create a learning opportunity. Set aside a bounded practice session with safe inputs and explain its purpose. People should know when the objective is delivery and when it is building capability. The next step is small: choose one task your team expects AI to support, name the judgment its human reviewer must retain, and prepare one changed case that tests it. Keep the assisted result and the independent explanation side by side. That comparison gives a training owner a much better basis for the next lesson than a completion badge. ## Sources & further reading Sources checked for this article. Research findings and vendor claims are distinguished in the text. 1. [How AI Impacts Skill Formation (v2)](https://arxiv.org/html/2601.20245v2)Shen and Tamkin; arXiv · Accessed 30 October 2026 2. [How AI assistance impacts the formation of coding skills](https://www.anthropic.com/research/AI-assistance-coding-skills)Anthropic · Accessed 30 October 2026 3. [Generative AI without guardrails can harm learning: Evidence from high school mathematics](https://pmc.ncbi.nlm.nih.gov/articles/PMC12232635/)Bastani et al.; PNAS · Accessed 30 October 2026 Produced with AI-assisted research and writing. Examples and recommendations are our editorial synthesis unless attributed. This article does not claim independent product testing or human expert review. [Read our editorial policy.](https://skillsforbusiness.ai/blog/editorial/) ## Keep reading [Give coding agents a reproducible bug and a clear acceptance test](https://skillsforbusiness.ai/blog/coding-agent-bug-report-acceptance-tests/)[Before installing an agent skill, review its access and updates](https://skillsforbusiness.ai/blog/third-party-agent-skills-installation-review/)