Dr Guy W. Bate is Thematic Lead for Artificial Intelligence, Director of the Master of Business Development (MBusDev) programme, and Professional Teaching Fellow in Strategy and Innovation at the University of Auckland Business School. He chairs the AI in Education Technology Stewardship Group for EdTechNZ, and has two decades of international industry experience in health technology, biotechnology and pharmaceuticals.
———————
Across the EdTech ecosystem, the debate is now less about whether AI belongs in education. It is already reshaping products, classrooms and policy. The harder and more consequential question is what “good” use looks like in practice.
I recently wrote about a related challenge in a business context: helping a pharmaceutical executive team move from scattered AI experimentation toward genuine organisational capability. The underlying logic of that work (distinguishing automation from augmentation) applies just as directly to education. But the second axis of that original matrix, built around operational efficiency versus strategic capability, needs to change. In education, the more consequential question is not how AI affects strategy, but how it affects learning itself.
The concern behind automation
Much of the anxiety about AI in education is, at heart, anxiety about automation. If AI can write the essay, solve the problem, generate the lesson plan or produce the policy analysis, what happens to the thinking that people previously had to do themselves?
This concern is often expressed through ideas such as cognitive offloading, deskilling and dependency. In education, these risks carry particular weight because completing a task is not always separate from learning. The process of forming an argument, working through an unfamiliar problem or constructing an explanation may be the very activity through which understanding and judgement develop. When AI performs that work, the learner may receive a better output without developing the capability that the task was intended to build.
But this does not make automation inherently undesirable. Education systems contain a great deal of work that supports learning without itself being the object of learning. Automating routine communications, organising evidence, formatting resources or producing a first-pass timetable can reduce administrative burden without diminishing anyone’s understanding. Even cognitive offloading is not automatically harmful. The important question is what is being offloaded, by whom, and for what purpose.
This is why education needs a more precise distinction than whether AI is being used or whether it saves time. We need to distinguish between automation that removes unnecessary work and automation that bypasses necessary learning. We also need to recognise a different kind of AI use: augmentation, where AI does not take over the thinking but helps a person extend, challenge or improve it.
A matrix for EdTech: automation versus augmentation, task versus capability
One axis therefore distinguishes automation from augmentation. Automation uses AI to perform or accelerate a task on someone’s behalf. Augmentation uses AI to extend human capability by helping a learner, teacher or policymaker consider alternatives, test their reasoning or notice a blind spot.
The second axis distinguishes operational efficiency from learning capability. It separates work that supports the system around learning from work through which understanding, skill and judgement are developed.
This produces four kinds of EdTech use case for AI:
Lighten the load combines automation and operational efficiency. This is AI handling the administrative weight around teaching and learning: drafting routine communications, generating first-pass timetables or rosters, summarising long policy documents, organising evidence for a curriculum review, or producing a first draft of a report. None of this touches learning directly. It clears space around it.
Sharpen the everyday combines augmentation and operational efficiency. Here AI is still working on operational tasks, but it is extending judgement rather than replacing it. A teacher might use AI to draft three different versions of a parent newsletter and choose the sharpest one. A product team might use it to stress-test a feature brief from three different user perspectives before committing engineering time. A policy analyst might use it to surface objections a consultation paper hasn’t yet addressed. The person still owns the decision; AI gives them more to work with.
The hollowing-out risk combines automation and learning capability, and it deserves the most caution of the four. This is where AI performs the cognitive work that was meant to build someone’s understanding or skill: writing the essay a student needed to write to learn to argue, solving the problem set a student needed to struggle with to learn to reason, generating the explanation a beginning teacher needed to construct themselves to develop their own pedagogical content knowledge. In a business context, automating a task usually frees capacity for higher-order work elsewhere. In education, the task itself is very often the point. This is the quadrant where convenience and capability development pull in opposite directions, and it’s where EdTech tools, classroom norms, and assessment design need the most careful thought, not blanket avoidance, but a clear answer to one question: is this task the thing the learner is supposed to be building?
Deepen the thinking combines augmentation and learning capability, and it is where the most valuable EdTech work happens. This is AI as a thinking partner rather than a thinking substitute: a Socratic tutor that asks the next good question instead of supplying the answer, a feedback tool that helps a student see the gap in their own reasoning, a system that helps a teacher rehearse how to explain a tricky concept three different ways, or a tool that helps a curriculum team pressure-test whether an assessment actually measures the capability it claims to measure. The learner (student, teacher, or system) still does the work. AI helps them do it with more range, more challenge, and sharper feedback than they would get alone.
Three audiences, one underlying question
Developers make this call at the point of design, often without realising it. A feature that defaults to producing a finished answer is making an automation choice. A feature that defaults to a question, a hint, or a partial structure is making an augmentation choice. Neither is automatically right. A finished answer may be exactly what an administrative task requires. But the way a tool is designed guides how people use it, and most users will follow the path it makes easiest. A product that routinely supplies completed answers may therefore encourage cognitive substitution even when its designers intended to support learning.
Educators are increasingly the ones deciding, lesson by lesson, which quadrant a given task sits in for their students. This is hard to do well from the outside; it requires educators to have used AI enough in their own work to recognise the difference between an AI-polished output and genuine understanding. Just as in the business context, this cannot be delegated downward. A teacher who has only ever seen AI used to generate worksheets is not well placed to judge how it should or shouldn’t be used by their students.
Policymakers are working at the level of the whole system: assessment design, procurement standards, professional learning, and guidance for schools. Their lever is less about any single tool and more about the incentives the system creates. An assessment regime that only rewards an AI-polished final product, with no visibility into the thinking behind it, will keep pushing schools and students toward the hollowing-out quadrant regardless of what guidance documents say. Procurement and curriculum decisions that ask vendors and developers, explicitly, which quadrant their tool is designed for would do more to shape good practice than most awareness campaigns.
Keep the thinking visible, not just the output
There is a caution that runs through all of this, and it’s the same one I raised in the business context: AI can make weak thinking look far more convincing than it is. A fluent, well-structured piece of AI-assisted writing can mask a shallow understanding just as easily as it can showcase a deep one. The output looks the same either way. This is precisely why the matrix matters more in education than almost anywhere else: the thing being assessed, in a classroom or in a curriculum review, is rarely the output by itself. It’s the thinking that produced it.
Start with a few cases, then ask the harder question
For developers, educators, and policymakers working through their own AI roadmap, the practical next step is the same one that I would suggest to any executive team: pick a small number of meaningful use cases and test them properly. But in education, add one extra question to the usual list of what’s faster, what’s clearer, and what’s riskier: what capability is this task meant to build, and does this use of AI build it, bypass it, or replace it?
The EdTech tools, schools, and systems that benefit most from AI won’t necessarily be the ones that adopt the most of it. They will be the ones whose developers, educators, and policymakers have a clear, shared answer to that question, and the discipline to keep asking it as the technology keeps changing.