A Norwegian business can demonstrate an impressive AI assistant with fictional customer cases and still be months away from using it on real work. The missing work may not be the model. It may be deciding which information can be used, preparing it, and giving the people responsible enough time to assess the risks.
That decision deserves a budget and a deadline of its own. A safe trial with invented material can teach a team quickly, but it cannot establish how well a service handles actual customer records or whether the organisation is ready to process them.
What a Norwegian public-sector trial found
The Norwegian government cloud marketplace's final report, published on 26 October 2026, describes a trial of Telenor AI Factory infrastructure and services. The active pilot ran from February to June. Participants included public bodies with different tasks, not a representative sample of Norwegian businesses.
The report says that, except for one organisation, participants did not complete a sufficiently thorough risk analysis to use real data in the trial. It also says the limited trial period contributed to an early choice by some participants to use test data in order to make progress. Organisations that wanted to use real data but did not do so cited capacity, expertise and data preparation among their difficulties.
This is not evidence that the participants' data were unsafe, that using test data was a mistake, or that one supplier is better or worse than another. The report draws on participants' and the supplier's reports, workshops, selected interviews and the marketplace's own assessments. The tested services were principally infrastructure and development environments, rather than a comparison of ready-made assistants. None of the participants reported that the solution failed their internal control and security requirements. Some reported missing documentation and support tools; the report did not establish how much came from the supplier and how much from participants' limited time or capacity.
The general lesson for a private business is an inference: make the work of approving and preparing information visible before judging when an AI pilot can expand.
Separate three questions
Imagine a Norwegian engineering company that wants an assistant to draft answers to maintenance enquiries. The company and examples below are hypothetical. The team can start with fictional equipment records and customer messages, without uploading actual client information. That can answer whether the workflow is worth pursuing, whether staff can spot errors and how much checking a draft needs.
It cannot, by itself, answer whether a real contract permits the proposed use, whether an unusual site description identifies a person, or whether the assistant finds the latest version of an actual maintenance procedure. Those questions need different evidence and may justify different decisions.
Management can organise the work in three stages:
- Prove the task is useful. Use invented or otherwise approved material. Define the result, the qualified reviewer and examples where the assistant should ask for help. Measure the whole task, including checking and corrections.
- Assess the real information. Identify the actual fields, customer commitments, sensitivity, people who need access and where information will be processed. Ask the customer, privacy and security specialists or supplier where necessary. Record unresolved questions and who will answer them.
- Decide on a narrow real-world test. Only after the relevant assessment and approvals, select a limited, appropriate set of real cases, name the reviewer, define stop conditions and compare results with the normal process. If the assessment does not support real data, continue with approved material, change the task or stop.
The point is not to push sensitive data into a tool to make a pilot look convincing. Nor should a test remain indefinitely disconnected from the work it is meant to improve. Each stage should state what it can prove and what remains open.
Removing names is not always anonymisation
A technician's account of a site may still identify someone through a location, time, role or unusual combination of details. Norway's Data Protection Authority explains that removing obvious identifiers alone is not enough to establish anonymity. Pseudonymised information can remain personal data. Fictional material can also disclose a real customer if it copies distinctive confidential facts.
Before using records, decide whether they contain personal or confidential information and involve the right expertise. Assess the applicable contractual and privacy requirements for the actual processing. A Norwegian hosting location can be relevant to that decision, but it is not a substitute for understanding access, processing and support arrangements. The public-sector trial does not establish legal obligations for every private company.
Give the assessment the same owner as the benefit
A pilot sponsor should bring the process owner, employees who know the work, IT or the supplier, and the appropriate privacy and security expertise together early. Assign each open decision an owner, an expected answer and a date. Fund the preparation and review work alongside the licence or development budget.
Ask for a one-page decision before expanding: Which task and information are proposed? What has been demonstrated with test material? What remains untested? Who can approve the next step, what controls and checking are needed, and what would stop it? Record the expected benefit as well as the cost of waiting or continuing manually.
Management owns the assessment and the operating decision. When the expenditure or exposure is material, the board should ask whether a credible path to useful, safe use exists and what resources or unresolved risks it entails. It does not need to approve every test record.
A good pilot can end with a decision to use less data, keep a human step or choose another process. It can also reveal a worthwhile opportunity that deserves the resources to move forward. Use the supplier evidence guide for the service review and the pilot-to-production guide for the wider operating decision.
Sources checked on 16 November 2026. The engineering company and staged decision are hypothetical and the author's analysis. The illustration is AI-generated.

