Two experienced Icelandic teachers volunteered to help run an AI pilot. Their description of the job was disarmingly simple: “We became the students.”
That is a better starting point for an AI rollout than a licence count. Give the people who understand the work time to learn where AI helps, where it fails, and what has to change around it. Access is procurement. Adoption begins when the method of working changes.
The honest result is that learning takes time
Anthropic’s short film, How Icelanders are thinking about AI, follows teachers, students, and parents during an Icelandic education pilot. It is not an evaluation. Anthropic made the film and supplied one of the tools used in the pilot.
That limitation is also what makes one comment stand out. A teacher says AI is not saving her time yet. She has to put substantial effort into changing what she teaches.
This is more useful than another effortless productivity claim. A new tool may write a first draft in seconds while making the whole workflow slower. Someone still has to decide what a good draft contains, supply the right source material, check the result, and work out what happens when it is wrong. Early in a rollout, that learning cost is real.
Ignoring that cost makes the programme look better than it is. Declaring the pilot a failure after a week may be just as premature. Measure two things separately:
- Learning cost: time spent testing, preparing context, reviewing failures, and changing the workflow.
- Steady-state value: time, quality, capacity, or revenue after the new method has settled.
If the second number never improves, stop. But do not hide the first number to make an AI programme look successful. The distinction matters when you later measure AI ROI beyond the initial time estimate.
Put the people who know the work at the edge of the tool
The two pilot coordinators in the film had taught together for 17 years. They were not waiting for an AI specialist to explain teaching to them. They volunteered, learned the tool, and tried it against work they already understood.
That choice matters. A domain expert can notice failures that a fluent demo misses: a technically correct explanation pitched at the wrong level, a lesson that cannot be delivered in the time available, or an exercise that removes the thinking it was supposed to develop.
In a business, look for a respected operator rather than an “AI champion” whose main qualification is enthusiasm. The operator can answer three less glamorous questions:
- What arrives before this task begins?
- What must be true before the work can move on?
- Which mistakes are merely annoying, and which ones matter?
The Iceland pilot paired tool access with training material and a support network, according to the announcement from Anthropic and Iceland’s education ministry. That support is what turns account distribution into a rollout. Without it, use cases stay in private browser tabs.
The OECD’s 2025 review of AI adoption in education systems reaches a similar, broader conclusion: capacity, buy-in, continued professional development, and support influence whether teachers adopt AI in practice. A school system is not a sales team or a wholesaler. Still, the operating lesson transfers: the people responsible for the work need a way to learn together. Permission to use a tool is only the beginning.
A boundary should describe the work
The film contains a sharper AI policy than many company handbooks. One student distinguishes using AI as a teacher from using it to produce an assignment from beginning to end. Another says a good teacher inspires students in a way the tool cannot.
Both are describing a boundary around the work, not the brand of software.
“Do not put confidential information into public AI tools” is necessary. A team still needs an operating method for each approved use case:
| Decision | What to record |
|---|---|
| Purpose | The job AI is helping with and the person who receives the output |
| Inputs | The documents, systems, and data the tool may use |
| Output | The draft, classification, recommendation, or action it may produce |
| Review | Who checks it and what evidence they need to see |
| Boundary | The decision AI may not make and the data it may not receive |
| Failure | What happens when the output is uncertain, incomplete, or wrong |
| Measure | The observable change that would make the use case worth keeping |
This record can fit on one page. Its value is not governance theatre. It gives the team a shared object to correct when a test fails. A company-wide AI policy people can actually use should provide the common boundary; the one-page record makes it specific to this workflow.
Start with one workflow, then earn the right to expand
A useful first rollout is small enough to observe but frequent enough to teach you something. Pick one workflow that happens every week. Follow a real item through it: one lead, report, order, assessment, or support request.
If the first use case is still undecided, start with the queue rather than asking the team to rank AI ideas in the abstract.
Record the current time, delay, and common errors before adding AI. Then let a small group work with the tool for several cycles. Review examples together, including the embarrassing ones. Change the instructions, source material, review step, or even the task itself. At the end, decide whether to standardize the method, run another bounded test, or remove AI from that part of the workflow.
The artifact from the pilot should not be a presentation saying people were excited. It should be a working method:
- an example of acceptable output;
- a short review checklist;
- known failure cases;
- an owner; and
- a baseline and follow-up measure.
Only then does wider access have something to spread.
What Iceland does not prove
The film captures several thoughtful participants. It does not show that Icelandic teachers as a whole support AI, that the pilot improved learning, or that it saved time. The launch announcement describes intended uses and support, not measured outcomes. The project may eventually produce that evidence; these sources do not contain it.
Nor does it settle how AI should be used in companies. Schools protect learning and childhood; businesses optimize different outcomes and face different obligations. The parallel is narrower: both fail when access races ahead of a shared method.
The teachers’ posture is worth borrowing precisely because it is not triumphant. They are curious, practical, and still unsure. They use the tool, keep the human boundary visible, and accept that the first phase is learning.
That is what an AI transformation looks like before it earns the name.
Common questions
Before you decide.
What is the best way to start an AI rollout?
Choose one recurring workflow, involve respected people who already understand it, define what good output looks like, and record where the tool helps or fails before expanding access.
Should an AI pilot save time immediately?
Not always. Early use can add work while people redesign the workflow, prepare reliable context, and learn how to review output. Measure the learning cost separately from the steady-state result.
Who should lead an AI pilot?
A credible practitioner who knows the work and wants to experiment should lead alongside someone responsible for data, risk, and implementation. A detached innovation team should not define the workflow alone.
How do we prevent employees from using AI in the wrong way?
Set rules at the task level: what data may be supplied, what output AI may produce, who reviews it, which decisions remain human, and what happens when confidence is low.