A company can buy AI tools for everyone and still have no clear answer to a basic question: which tasks are actually becoming easier, and which employees are left dealing with extra checking or uncertainty?
For Norwegian managers, the next step is not to rank staff by how often they open a chatbot. It is to examine a few real tasks with the people doing them. That reveals where AI helps, where rules or poor information get in the way, and where using it would add risk without a worthwhile benefit.
The same job does not mean the same use
A US study by researchers at the St. Louis Fed, Vanderbilt and Harvard examined self-reported generative AI use in a national online survey. It linked occupations to detailed work tasks and asked respondents which of those tasks they performed and used AI to help complete. The analysis drew on almost 14,000 worker observations across four survey waves from August 2025 to May 2026.
The researchers found use across many occupations and tasks, but substantial differences among people doing similar work. In fewer than 3% of the studied tasks did at least half of the workers performing that task report using AI for it. Those figures describe the US survey and its task definitions. They are not Norwegian adoption rates or a measure of productivity.
There are important limits. The study used an online panel weighted to resemble the US population; the panel itself was not a random sample. People reported their own use, and the task questions were limited to the ten most important tasks associated with each occupation. Researchers found patterns consistent with experience helping adoption, but other explanations, including differences between early and later users, remain possible. The study does not prove that a particular training programme causes a gain.
A Harvard Business School AI Institute analysis published on 12 November revisited the findings. The underlying working paper dates from August 2026. It offers a useful question for managers, not a new measurement of Norwegian workers.
What the Norwegian figures do and do not say
Statistics Norway reported in September 2026 that almost five in ten enterprises with at least ten employees used one or more types of AI in 2026. The figure concerns enterprises, not the share of their employees using AI or the quality of their work. It also excludes smaller firms and does not tell a manager which tasks are suitable.
Together, the research points to a practical gap: a licence count or a company-wide adoption number cannot show whether a customer reply is better, whether staff have time to verify it, or whether one team is blocked by an unresolved data rule.
Map three tasks, not every prompt
Imagine a Norwegian service company where a team prepares customer proposals. One employee uses an approved assistant to produce a first outline; another avoids it because customer specifications are confidential; a third tried it on technical details and found the checking took too long. This is a hypothetical example, not a customer case.
Ask the team to select three recurring tasks and describe the process without collecting their private prompts or ranking individuals:
- Preparing a general outline: Which information is safe to use, and does the outline reduce the time needed to produce an acceptable proposal?
- Checking technical requirements: Which source is authoritative, and does the assistant save time after a qualified person verifies its answer?
- Writing a customer commitment: Which promises about price, delivery and product suitability must a named employee approve?
For each task, record the approved tool and data, the person responsible, the checking step and a simple measure of the completed result. A small business can do this in a short team meeting. A larger organisation may need input from IT, security, employee representatives and the owner of the customer process.
Remove the right obstacle
If people do not know what information they may use, management should clarify the rule with the relevant privacy and supplier experts. If a tool struggles with Norwegian terminology or produces unverifiable claims, test a different approach or retain the existing process. If the task is suitable but people lack a chance to practise, set aside paid time for a short guided exercise using safe examples.
A good outcome can also be a decision not to use AI for a task. More activity is not automatically more value. At the same time, leaving useful tools unused can mean slower service and missed opportunities to support staff. Compare completion time, review effort and error correction for similar cases before expanding. Do not treat employee confidence or the number of generated drafts as proof of a customer benefit.
Invite employees to report both useful examples and failures. Explain any usage data the company collects and involve the appropriate people before introducing monitoring. A learning discussion is more useful when staff can raise problems without feeling they are being scored.
The question for management and the board
Management should select the tasks, approve the boundaries, provide support and review outcomes. Where the investment is significant, the board should ask which business problem is being addressed, what evidence shows benefit and which risks or unresolved barriers remain. It does not need a list of individual employees' prompts.
The aim is not universal use. It is a team that knows when AI helps, how to check it and when another method is better. Connect the task review to your employee AI policy and your plan for measuring business results.
Sources checked on 13 November 2026. The company example and management approach are the author's analysis. The illustration is AI-generated.