Articles/Field notes

Field notes

AI Won't Transform Education Unless We Can Prove It Works

18 Nov 2026 ·Helsinki, Finland


EdTech / AI & Education

AI Won't Transform Education Unless We Can Prove It Works

Nectora Conversations · 10 November 2026 · OnlineDr. Wayne Harrison, co-founder and CEO of WhatWorked Education, argues that the next generation of educational technology needs more than adoption and engagement metrics. It needs evidence that students are actually learning.

AI can now generate exercises, personalise feedback and support students at a speed that would have been difficult to imagine a few years ago. But there is a question that too often gets lost in the excitement: does any of it improve learning?
For Dr. Wayne Harrison, co-founder and CEO of WhatWorked Education, this is not an abstract research problem. It is the question that led him from teaching secondary-school science to researching how schools can evaluate what works in their own classrooms.
In a conversation with Nectora, Harrison made the case for faster, teacher-led experiments, more rigorous product development and a future in which AI supports educators rather than bypasses them.

The question behind the research

Harrison spent a decade teaching before moving into university research at Durham, where he worked with Professor Steve Higgins, a co-founder of WhatWorked Education and a leading figure in evidence-informed education.
Large educational studies can tell policymakers what works on average. That is valuable, but Harrison argues that it often leaves teachers with a more immediate question:

"Does it actually work in my school?"
Traditional randomised controlled trials (RCTs) are designed to provide robust answers, but they can be expensive and take years to complete. That becomes especially problematic for AI-powered tools. By the time a long evaluation is published, the version being tested may no longer resemble the product students are using.
Harrison believes schools need a way to generate evidence that keeps pace with both classroom practice and technological change.

Turning classrooms into evidence generators

WhatWorked Education's approach centres on micro-randomised controlled trials, or micro-RCTs: small experiments that teachers can run as part of ordinary school practice.
Harrison offers a simple example. A teacher identifies students struggling with spelling, measures their starting level and enters anonymised scores into the platform. Students are randomly assigned either to a targeted spelling intervention or to continue with normal teaching. After six weeks, both groups are assessed again.
The teacher can then receive an impact report showing what happened in that classroom, without waiting for a lengthy academic publication. The platform provides accessible measures alongside the statistical analysis.
A single small trial cannot settle whether an intervention works everywhere. The more interesting idea is to repeat the same experiment across many schools, then combine the results. Each school contributes its own evidence, while researchers can examine the overall effect and the differences between contexts.
Harrison describes each micro-trial as a building block. Together, they create a cumulative evidence base that can evolve as products, teachers and classrooms change.
This also changes who produces educational research. Teachers are no longer just handed findings from elsewhere; they help create the evidence they will ultimately use.

Can research keep up with AI?

One example is WhatWorked Education's evaluation of Medly, an AI tutoring platform for secondary-school science.
The research involved 39 classroom trials across biology, chemistry and physics. Of 929 students completing an initial assessment, 644 completed the follow-up assessment. The study reported a positive short-term effect for students assigned to Medly compared with usual self-directed revision, and Harrison says the evaluation was completed within around 12 weeks.
That speed matters. Instead of freezing a fast-changing product for years, a company can repeat trials with new classrooms and track whether the evidence continues to hold as the technology develops.
But Harrison is careful about what the results mean. The Medly findings were preliminary: the study had substantial attrition and used curriculum-aligned rather than standardised assessments. Micro-RCTs can reveal an early signal, help refine implementation and inform a decision to invest in larger studies. They do not automatically replace longer-term, large-scale trials.
For Harrison, the goal is not to choose between speed and scientific rigour. It is to build a system in which evidence arrives early, is replicated and improves over time.

Engagement is not the same as learning

A central concern in Harrison's argument is that EdTech companies are often rewarded for the number of users they attract rather than the learning outcomes they improve.
A student can open an app, complete exercises and generate impressive usage statistics without necessarily making meaningful progress. The reverse problem is equally important: a tool may help highly motivated students while most of its intended users never engage enough to benefit.
Harrison points to an evaluation in which students needed to complete roughly 120 questions over a year for the product to show an improvement on a standardised assessment. That insight was useful, but it also raised a practical question: could schools realistically get students to use the product that consistently, especially when the work was assigned at home?
His point is that a successful learning product needs both effective use and measurable outcomes. A tool that students enjoy but that does not improve learning is not enough. Neither is a tool that works in ideal conditions but cannot be implemented in a real school.

"If students don't use it, you will not see the impact."
The consequences of getting this wrong may be less visible than a failure in medicine or aviation, Harrison argues, but they are real: schools can waste scarce resources and students can miss out on more effective support.

Why human teaching still matters

Harrison is optimistic about AI's potential to support learning, but sceptical of the assumption that it can replace the whole educational experience.
He points to peer tutoring as a useful counterexample. When students help one another with mathematics, the benefits do not stop at mastering a concept. The students doing the tutoring also practise communication, build confidence and reinforce their own understanding.
Those human interactions are part of what schools are for. AI may make some tasks faster and offer useful support, but Harrison argues that education also develops capabilities that cannot be reduced to answering questions on a screen.
There is a parallel with AI-assisted work outside the classroom: when a system makes a mistake, people still need enough expertise to recognise it, understand it and respond.

What EdTech founders should do before scaling

Harrison's advice to anyone building an educational product is to start evaluating it from the beginning, not after acquiring a large user base.
First, articulate a theory of change: why should the product improve learning at all? Then test its individual components. How frequently must it be used? For how long? What happens when it is introduced in different classrooms? What barriers do teachers encounter?
He sees independent feedback from teachers as especially valuable. A product may look promising in a controlled demonstration but create too much friction in actual school routines. That feedback should inform the product roadmap, followed by another round of evaluation to see whether the changes improve outcomes.
Harrison also has a proposal for investors: set aside 10% of EdTech funding for evidence-building, with increasingly rigorous evaluation expected at successive investment stages. It is his suggested model, rather than an established industry requirement, but it addresses what he sees as a basic incentive problem: if companies are rewarded only for growth, proving learning outcomes can remain optional.

The takeaway

The education sector has already lived through several waves of technology promising to transform classrooms. Harrison's concern is that AI could repeat the pattern: rapid investment, widespread adoption and too little understanding of what genuinely improves learning.
His alternative is practical rather than anti-technology. Give teachers a role in generating evidence. Test products while they are being built. Repeat experiments across different contexts. Scale promising interventions, but keep measuring what happens to students.
For EdTech founders, the challenge is not simply to build something students will use. It is to build something schools can implement and prove that students are better off because of it.

Conversation

Dr. Wayne Harrison — Co-founder & CEO, WhatWorked Education
In conversation with Nectora — 10 November 2026

Further reading

About the author

Julio Cordal
Julio Cordal

Founder | European Startup Media & Network | AI, Founders & Tech Events

Im the founder of Nectora, a European media company focused on business, technology, entrepreneurship and innovation. We believe the most interesting stories come from the people building the future. At Nectora, we sit down with founders, CEOs, investors and industry leaders to explore their journeys, the decisions behind their businesses, the challenges they face and the ideas shaping tomorrow. Through podcasts, in-depth interviews and short-form content, we bring these conversations to a wider international audience.

More by Julio →