When generative AI first took off a few years ago, teachers were understandably cautious. It seemed transformative, but it was uncharted territory. Schools and teachers were wary of early tools, with their uneven mix of quality and reliability, uncertain about what AI could do for them and whether they could trust it. It was the same at Oak. We suspected the technology would develop rapidly and have a significant impact, but we weren’t sure exactly how, and whether quality and safety could be built in. So in 2024 we launched Aila, our AI lesson assistant, deliberately placing it in a standalone “labs” website to reflect its experimental nature. Teachers embraced it and since then more than 60,000 have used it to help develop lessons. Three years on, the landscape looks very different. Around 93 per cent of teachers now use AI in their work. New capabilities are emerging all the time, alongside innovations from MATs and schools experimenting and building tools of their own. Teachers’ confidence and proficiency have grown enormously and, by trying a wide range of tools, they have discovered a much broader range of tasks where AI can help. That experimentation is beginning to tell us something important about how teachers actually want to use AI. Time savings When we looked closely at how teachers were using our own tools, a clear pattern emerged. Teachers didn’t necessarily want to visit a separate AI destination and ask it to produce an entire lesson plan with accompanying resources. Often the need was smaller and more immediate: some extra quiz questions, a resource to reinforce a particular point or an adaptation to make an existing lesson work better for their pupils. New research published today from the Education Endowment Foundation (EEF) adds significantly to that picture. Its randomised controlled trial, the gold standard of research, found that teachers using Aila saved an average of 49 minutes a week on lesson planning, with no difference in lesson quality. That’s around 32 hours over the school year, and almost double the time saving found among teachers using ChatGPT for lesson planning in an earlier EEF trial. But there was another important finding. Aila use declined during the trial and many teachers preferred adapting existing materials to creating new lessons from scratch. Taken together, that should give anyone developing AI for schools pause for thought. Teachers’ use depends on how well it fits with their work The question isn’t simply whether an AI tool “works”. AI can save teachers time. But whether teachers choose to use it, and keep using it, depends on something more fundamental: how well it fits into the work they’re already doing. We’re seeing the same thing in our own usage data. Our “create with AI” feature, which puts AI alongside the lessons and resources on Oak’s main site, has generated roughly five times as many interactions as the equivalent feature within standalone Aila over the same period. The underlying technology is the same. What is different is where teachers encounter it. That’s helped shape a principle for how we think about AI. Rather than asking teachers to change how they work to accommodate the technology, we should be asking how the technology can fit around them. For Oak, that means embedding AI directly into the resources teachers are already using. We will shortly introduce a feature providing additional scaffolding options within existing Oak lessons. That’s a practical example of AI supporting teachers at the point they need it, rather than requiring them to go somewhere else. It also means Aila’s role is changing. The safety, quality and sector expertise behind it remain, but its capabilities are increasingly becoming part of the tools and resources teachers already use rather than sitting apart as a standalone product. Wider lesson There’s a wider lesson here for all of us developing AI for education. The question is no longer simply what generative AI can do. It’s identifying where it genuinely helps teachers, how it fits into their working lives and whether it solves the problems they have. And we need better evidence, faster. There is no shortage of surveys and eye-catching claims about the hours AI can save teachers. Those can tell us something useful, but rigorous independent trials like the EEF’s tell us much more about what happens when teachers use AI. The challenge is that AI is developing so quickly that evidence based on how teachers were using it a year ago can already feel dated. But if schools are going to make good decisions about AI, we need more high-quality research like this, rigorously conducted, openly shared and produced at something closer to the pace at which the technology is evolving.