Imagine a new employee delivering working software sooner with AI. The team is pleased, until the employee is asked to explain why the same customer record can be imported twice. The code works on the demonstration data, but the understanding needed to review it has not yet been demonstrated.
For a Norwegian business, this is a capacity decision: who will be able to check the work, handle unfamiliar cases and train the next colleague? Buying more AI access does not answer it. Neither does requiring a human approval if that person has never learned what to look for.
My recommendation is to give each important workflow two explicit outcomes: an acceptable delivery today and a defined capability the team must retain or develop. The same AI tool can contribute to both, but the work needs to be organized accordingly.
Completion and learning are different results
A 2026 randomized study by Anthropic researchers examined 52 programmers learning an unfamiliar Python library. AI access reduced subsequent quiz performance; the average time advantage was not statistically significant. The study used a chat assistant and an immediate assessment, not months of workplace observation. It does not establish lasting skill loss or an effect of today's autonomous coding tools. The authors' analysis of different usage patterns was observational, so it cannot prove that a particular prompting style protects learning.
There is also evidence of benefit. Brynjolfsson, Li and Raymond's study, published in 2025, examined a phased rollout using data from 5,172 customer-support agents at one company. Productivity rose by 15% on average, with larger gains for less experienced workers. Performance during outages suggested learning, although those estimates were noisy and case mix could differ. This was a workplace rollout, not random assignment to the coding task above, and it does not establish the same effect in Norwegian companies.
These studies answer different questions. An assistant offering useful suggestions during repeated customer conversations is not the same learning environment as generating a solution to an unfamiliar programming problem. Treating either finding as a universal verdict on AI would hide the decision management actually controls: what employees practise, what help they receive, and how understanding is checked.
Keep the judgement, reconsider the routine
Consider a hypothetical Norwegian software consultancy introducing AI-assisted development for a customer integration. A developer new to this work can generate an import function quickly. The customer needs the integration to distinguish new records from records already processed, handle NOK amounts correctly and preserve the meaning of Norwegian field names.
The learning objective should not be to reproduce every line from memory. It could be to explain the duplicate-handling rule, identify a misleading test and diagnose a failed import using approved documentation. Those are capabilities the team expects the person to use when checking future work.
An experienced developer may already have those capabilities and benefit from delegating more routine implementation. A senior employee learning an unfamiliar system may need practice too. Set the learning requirement by the task and demonstrated knowledge, rather than by age, job title or confidence with the tool.
Some work should simply disappear. Do not preserve redundant copying or obsolete procedures as a training ritual. Identify the knowledge that still matters to delivery, review and continuity, then find the least wasteful way to develop it.
Change one training assignment
For the consultancy, a useful assignment could have three parts. This is a proposed management practice, not a method validated by the two studies.
First, require a small independent attempt. Before receiving a generated solution, the learner describes how duplicate records should be handled and proposes a test. Allow the ordinary source documents and reference material needed for the job. Keep the exercise in a safe test environment; customers should not carry the cost of a training mistake.
Then use AI as a source of feedback. Ask it to challenge the proposed test or explain an unfamiliar concept. The learner checks its explanation against the specification and discusses disagreements with a capable colleague. A plausible explanation from the same system that produced the code is not independent verification.
Finally, change the case. At a later session, present a different input or a deliberately flawed solution without an AI-generated answer alongside it. Ask the learner to locate the problem, explain the consequence and choose a next step. A fresh case gives better evidence of transfer than repeating a remembered answer. It still provides only a limited sample of competence.
Use the language and material people encounter at work. In this Norwegian consultancy that includes Bokmål customer requirements, English technical documentation and Norwegian business terminology. If a customer uses Nynorsk, include it. Fluent explanations in English alone do not show that the employee has understood a Norwegian requirement.
Budget for practice before spending the apparent saving
A small firm may have only one person qualified to supervise the integration. If every minute released by AI is immediately committed to more customer work, that person's coaching time has no place in the plan.
Allocate a modest, scheduled practice period within paid work and identify a supervisor and backup. Ask the team which routine assignments previously taught the essential judgement. Replace those learning opportunities deliberately when automation removes them. Employees and their representatives, where relevant, can help identify what is being lost or made harder.
Keep the assessment proportionate and transparent. Explain the learning purpose, what is recorded and how it will be used. Prefer a short demonstration and specific feedback to collecting every prompt or ranking people by AI usage. Do not quietly turn a training exercise into a performance-monitoring system.
Track delivery quality, review effort and demonstrated understanding separately. Faster completion with improving understanding supports expansion. Faster completion with repeated inability to explain critical decisions calls for different training or continued supervision. If practice consumes the entire benefit, narrow the requirement to the capabilities that truly matter before abandoning useful assistance.
What management should show the board
Management should own the training design and decide when someone is ready for independent review. Access to AI and authority to approve its work are separate decisions. A prompt-writing course alone should not confer that authority.
Where the dependency is material, the board should ask whether the company is developing enough people to check critical outputs and cover absences. Evidence might include a recent exercise on an unfamiliar case, remaining gaps and a funded plan to address them. The board does not need individual test scores or responsibility for running training sessions.
At the next management meeting, select one task being automated and name the judgement that must survive it. Assign a practice opportunity, a supervisor and a later demonstration. Connect this to the employee AI policy and the capacity decision about hiring, buying services or using AI. Better tools should make it easier to build a capable team, but that outcome needs its own plan.
Sources checked on 14 October 2026. This is an evergreen analysis, not a report of a new study released today. The consultancy and training assignment are hypothetical. The operating recommendations are my interpretation; the studies do not establish a universal training method or predict employment effects in Norway.