"AI saves six weeks a year": why we stopped quoting that figure

Since summer 2025, one figure has appeared in every sales presentation in the sector: AI reportedly saves teachers 5.9 hours per week, or six weeks per school year. It appears at the opening of webinars, in sales pitches and in budget justifications.

The figure is real. It comes from a serious survey conducted by Gallup for the Walton Family Foundation. We cited it ourselves.

We no longer cite it. Not because it is false, but because it does not demonstrate what it is made to say, and because better evidence is now available.

Here is what this figure actually measures, why we removed it from our materials, and which data we rely on instead.

What the study actually measured

The report is entitled Teaching for Tomorrow: Unlocking Six Weeks a Year With AI. The survey was conducted online in April 2025 among more than 2,000 American primary- and secondary-school teachers, and was published in June 2025.

Three findings structure the report.

Six teachers in ten report having used an AI tool during the school year. Around three in ten use one at least once a week. And it is these weekly users, not all respondents, who report saving 5.9 hours per week.

The distinction matters. The gain does not apply to "teachers" but to the share of them who have incorporated the tool into a regular practice. Across all surveyed teachers, the average gain is much lower.

Gallup is transparent about the method: respondents estimate time saved themselves, task by task, in half-hour increments. It is therefore a self-reported measure, not field observation or an activity log. Teachers state how much time they believe they have saved.

This is not a disqualifying weakness; it is standard for this type of large-scale survey. But it calls for caution: it measures a perception of time saved, and perceptions of time are notoriously imprecise.

Where the time savings come from

The report details the uses, and the distribution is telling. AI is used primarily to prepare, adapt and produce materials, not to teach.

Reported quality gains concentrate in three areas:

Area Teachers reporting higher quality
Modifying materials for students 64 %
Using learning data 61 %
Grading and feedback for students 57 %

In other words, the gains occur in content-production and adaptation tasks. Not in learning design, not in relationships with learners, not in facilitation.

This clarification changes how the argument should be presented. Saying "AI saves six weeks" suggests a transformation of the job. Saying "AI speeds up the production of materials and the customisation of documents" describes what actually happens, and remains a considerable benefit.

The finding no one quotes

Here is the passage in the report that deserves more attention than the figure on its cover.

Teachers working in an institution with a formal policy on AI use report time savings 26% greater than those of other teachers. And only 19% of surveyed teachers work in such an institution.

This result says two things.

The main determinant of benefit is not the tool but the framework in which it is used. A clear policy on what is allowed, the recommended uses and what remains the teacher's responsibility has a measurable effect. Without this framework, everyone improvises and the return falls.

And the vast majority of institutions have no such framework. Four teachers in five either use AI or refrain from using it without institutional guidance.

This observation is consistent with a 2026 briefing from Stanford's National Student Support Accelerator, entitled AI Tutoring is Not a Monolith: What We Actually Know. In two controlled trials conducted across entire districts, 40 to 47% of students never used the AI platform made available to them. In another study of 181,000 students, 41% never logged in, and only 5% reached the recommended duration of use. Factors associated with the teacher, institution and district account for 57% of the variance in usage.

Deployment is not adoption. And adoption cannot be decreed through procurement.

What we use instead

A self-reported survey is not the best evidence available, and since 2026 it is no longer the only evidence.

Stanford's SCALE centre has published a review of causal studies on AI in education. Its overall assessment is severe: among more than 800 papers identified, very few rely on a design capable of establishing causality. But those that do are worth more than any survey.

One of them followed teachers who had access to ChatGPT and a user guide. The result: around 30% less time spent preparing lessons and resources, or around twenty minutes per week. More importantly, the quality of the lessons produced was assessed blindly by experts, without knowing which had benefited from AI. No detectable difference was found.

The difference in method is what matters. Teachers are not asked to estimate their gain: it is measured, and the quality of the result is assessed by third parties unaware of the experimental condition. A 30% gain established this way carries more weight than a self-reported saving of 5.9 hours.

The same review reports two findings that we find more useful still.

The first concerns learning to use the tool. Over the weeks, these teachers used AI less and less, in 39% of their classes initially and then 29%, while their time saving increased, from 27% to 31%. They identified where the tool was useful and stopped using it elsewhere. Competence is not about delegating more, but about delegating better.

The second is more uncomfortable, which is why it deserves to be cited. An automated system for grading papers did not reduce teachers' total working time. However, student outcomes improved and teachers discussed more papers individually with them. Time saved on routine tasks was reinvested in support, rather than turned into saved hours.

It is a less marketable promise than six weeks a year. Its advantage is that it corresponds to what people find once the tool is deployed.

What this means for French vocational training

Three differences must be acknowledged before any transposition.

The first concerns the audience: these are American teachers in the school system, whose work organisation, hours and administrative constraints differ substantially from those of an instructor in a French training provider or a corporate learning designer.

The second concerns the nature of the tasks. The gains concentrate on producing materials and differentiation. An instructor who spends most of their time facilitating in person does not have the same cost structure as a teacher preparing learning sequences for thirty pupils.

The third concerns the regulatory framework. A Qualiopi-certified provider must maintain records, document and justify its work. Time saved in content production can be partly reinvested in documentation requirements that do not exist in the American context.

That said, the underlying mechanism carries over. Rephrasing, adapting material for a different audience, producing variations of exercises and drafting instructions: these tasks exist everywhere, they are time-consuming, and they are precisely the tasks at which current models are effective.

The question that really determines the gain

An hour saved only has value in what is done with it.

If time freed up from producing materials is absorbed by other administrative tasks, the pedagogical benefit is nil. If it is reinvested in individual support, designing learning situations or following learners who are struggling, it becomes real.

This decision does not belong to the tool. It belongs to the organisation, and that is exactly what the 26% gap between institutions with a policy and the others measures.

For a training manager considering deployment, this suggests an unusual order of priorities: define the framework of use before choosing the tool. Which tasks are affected, which are not, what instructors remain responsible for validating, and what they are expected to do with the time recovered.

Without these answers, a high-performing tool will deliver mediocre results. With them, a competent tool will produce measurable results.

As for the six-week figure, we have stopped using it. A commercial argument that does not survive reading its own source always ends up turning against the person who uses it.

Discuss the right deployment framework for your teams →

References