The rollout starts when the experts become students

We watched Anthropic’s short film about Iceland’s education pilot expecting to hear the familiar story about teachers getting hours back, but instead the line that stayed with us came from two teachers who had worked together for 17 years and volunteered to help run the project: “We became the students.”

A little later, another teacher admits that AI isn’t saving her time yet. Changing what she teaches has taken substantial effort, and it’s an awkward thing to include in a film made by Anthropic, which also supplied one of the tools, and it is probably the most useful moment in the whole eight minutes.

The teacher isn’t describing a failed button. She’s changing a lesson, trying the output, checking whether it helps students, and deciding what still belongs to her.

That looks more like an AI rollout than an afternoon spent buying licences.

And it makes the early mess much easier to explain.

The first result may be more work

A tool can produce a draft in seconds and still make the week slower, because somebody has to find the right source material, explain what a good draft contains, catch the failures (including the plausible ones), and change the surrounding workflow. Those jobs are easy to leave out of a pilot report because they make the early numbers look worse, but they are the work of learning what was actually bought.

We’d therefore separate the hours spent learning from what happens after the method settles, since the first number includes testing, preparing context, reviewing odd results, and changing the way the task moves. The second is the change in time, quality, capacity, or revenue once people have a method they can repeat. Hiding the first number flatters the rollout, while expecting the second number in the first week can kill a useful experiment before it has taught anyone much. Nevertheless, if steady-state value never appears, of course, we should stop. The distinction simply helps us say what stage we’re in when we measure the result.

The two teachers in the film are interesting because they already know the work. They can spot a lesson pitched at the wrong level, an exercise that won’t fit the available time, or an activity that quietly removes the thinking it was meant to develop (the part a polished demo won’t show). A general AI team might see fluent text, the teacher sees the class on Tuesday morning.

For a company, the equivalent isn’t necessarily the most enthusiastic person in the room, because it is a respected operator who knows what arrives before the task, what has to be true before it can move on, and which mistakes are inconvenient versus dangerous. We still need somebody to own data, risk, and implementation, but a detached innovation team can’t define good work on behalf of the people who do it every day.

That practical knowledge is the reason for the pairing.

So the group can notice mistakes that outsiders miss.

According to the announcement from Anthropic and Iceland’s education ministry, participating teachers were offered training material and a support network as well as access to the tool. The OECD’s 2025 review of AI adoption in education systems also points to capacity, buy-in, continued professional development, and support as influences on adoption. Neither source proves that the Iceland pilot improved teaching (the film certainly doesn’t), and a school is obviously not a wholesaler or a sales team. The narrower business analogy is still worth keeping: private browser tabs don’t add up to an operating method, even when lots of people have access.

Give the group something it can correct

One student in the film draws a useful line between treating AI as a teacher and asking it to produce an assignment from beginning to end. Another says a good teacher inspires students in a way the tool can’t. They are talking about the work that should remain visible, instead of arguing about whether a particular software brand is good or bad.

Company guidance often stops one step earlier, since “don’t put confidential information into public AI tools” is necessary, but it doesn’t tell an approved user what a good use looks like. For one recurring workflow, we’d write down the purpose, the information the tool may receive, the output it may produce, the person who reviews it, and the decision it may not make. We’d also note what happens when the answer is uncertain and what observable change would make the experiment worth keeping.

Otherwise, permission is still too vague to use.

The page turns private guesses into something discussable.

That can fit on a page (alongside an example of acceptable work and a few failures from real cases). It gives the group a shared object to amend after a poor result instead of relying on everybody to remember a different lesson. The company-wide AI policy supplies the common limits, and this smaller record explains how those limits apply to the report, lead, order, or support request in front of us.

If we don’t yet know which workflow deserves that page, we’d look for something that recently got stuck. Follow one real item from arrival to outcome, record its delay and common errors, and let a small group try several cycles. Looking at the strange examples together matters more than collecting a satisfaction score. The group may change the instructions, the source documents, the review step, or decide that AI doesn’t belong in this part of the work at all.

Perhaps the ordinary rule was enough after all.

That is a perfectly good result from a pilot.

At the end, there should be more to share than a presentation saying people were excited, and consequently there should be a method that another colleague can try, with an owner, a known boundary, examples, and a before-and-after measure. Wider access then has something to spread.

The uncertainty is part of the result

Anthropic’s film isn’t research, and the people who appear in it don’t speak for every Icelandic teacher, so it doesn’t establish that the pilot saved time, improved learning, or won national support. The launch announcement describes the access, training, and intended uses, but it doesn’t report those outcomes. We shouldn’t fill in the missing evidence because the participants are thoughtful or the film is persuasive.

That gap matters (even if the eventual results are excellent).

For now, uncertainty is one of the findings.

It also can’t tell a company how to use AI. Schools protect learning and childhood, while businesses have different aims and obligations. However, we can borrow the posture of those two experienced teachers: they know their field, they are willing to become beginners again, and they haven’t confused early curiosity with a finished result. For now, that may be the honest shape of a good rollout.