“AI saves teachers 49 minutes a week.”
I expect to see that line on a good number of staff briefing slides this half term. It comes from the Education Endowment Foundation’s new evaluation of Aila, Oak National Academy’s AI lesson assistant, and it deserves the attention. The trial was well designed, the finding is secure, and the quality of teachers’ resources held up.
Underneath the headline sits a second finding that interests me more. The teachers who saved the most time were those newer to the profession, or less confident in their subject knowledge and pedagogy. Most readers will take that as good news: AI helps the people who need help most.
It may well be good news. I think it’s also a reason to look more closely. The trial counted minutes. What it couldn’t count is slower to show: whether the teacher using it is getting better.
A result worth taking seriously
Start with what the trial found. Some 460 Key Stage 2 teachers in 108 schools were randomly allocated and followed for ten weeks in the autumn term of 2025. Teachers using Aila spent about 2.5 hours a week preparing lessons, against 3.3 hours for colleagues who didn’t. That’s 49 minutes a week, roughly a quarter of their planning time. Independent reviewers judged the resources no worse in quality. The EEF gives the finding its highest security rating.
For a primary teacher, 49 minutes saved a week is not to be sniffed at.
Most schools will read this as confirmation of what they already believe: AI is an efficiency tool, and if it saves time, it’s working. That’s the assumption I want to question. Saving time and adding value are two different things, and the teachers who save the most time may be the ones with the most to lose.
The pressure that shapes every choice
The case against my concern is strong, and it starts with the school day. A teacher plans, marks, assesses, writes reports, organises the Year 5 trip, does break duty, meets a parent and fills in the data spreadsheet before Friday. Under that pressure, a tool that saves time will be used to save time. Shortcuts will get found. Pretending otherwise would be a little naive.
My worry sits inside that reality. Good planning involves a lot of trial and error: the activity that flopped, the explanation that needed a third attempt, the misconception you only spotted because you built the lesson yourself. That struggle is how judgement forms. If a new teacher receives finished lessons before they’ve had the chance to struggle with them, what happens to their judgement?
I should be honest here: that’s a hypothesis. Nobody has tested it, including me. But it leads to the question the trial raises without answering. What is AI for? Is it there to save us time, or to make us better at the job?
The order matters
Most people will answer “both”, and they’d be right, as long as the order is right.
My rule of thumb is teacher first, AI second. The teacher decides what the lesson needs to do, what pupils already know and where they’re likely to trip. Then AI comes in to test that thinking, fill the gaps, produce the materials and suggest the alternative the teacher hadn’t considered. The design stays with the teacher, and the machine takes on the production.
Done in that order, the 49 minutes becomes a professional choice. If you didn’t have time last week to sit with Lucy, who struggles with maths, now you do. That’s a gain. If you’d rather spend ten of those minutes on the coffee break you haven’t managed since September, you’ve earned that too. Either way the teacher decides, because the teacher is still the one doing the thinking.
The tool that asked teachers to think
Here’s what interests me most about the trial. Aila was designed along these lines. It deliberately pauses at key points to draw on the teacher’s expertise, and it takes you through the planning process step by step.
Teachers found that hard going. Some described it as “slow”, “inflexible” and “cumbersome to use”. In an average week it was used for 7% of lessons, and use fell as the trial went on: 53% of teachers included Aila in their planning in the first week, and 33% in the tenth. Many preferred general tools such as ChatGPT, which simply hand you an answer. Fewer than one in five felt Aila had meaningfully influenced their teaching or planning.
I’ve worked with education technology for most of my career, and I recognise the pattern. A tool is built to solve one problem. It gains features, then steps, then settings, until people have to manage it. Eventually they go back to something simpler, which often means pen, paper and a photocopier.
AI planning tools face the same pull, with a twist. The effort Aila asked of teachers is the point. Wrestling with a plan is how teachers get better at the job. It’s also the first thing to go when you’re pressed for time. In the heat of the moment, people want the answer. I feel this myself. I’m building a tool called Whyzo, and the temptation to make it solve more and more problems never goes away. I’m trying hard to make it solve one problem well.
The effort that makes teachers better is the very effort AI is so eager to take off their hands.
The researchers flag the risk themselves. The bigger time savings for newer teachers, they write, do “not necessarily imply that the quality of the lesson delivered is as good as if it had been planned without Aila”. That’s the part the headline leaves out.
Maybe they should just be told
There’s a serious counterargument here, and it comes from an unexpected direction.
A decade ago, one strand of the profession had a slogan: just tell them. Explicit instruction, they argued, beats leaving pupils to discover things for themselves. I was never wholly in that camp. I’ve always believed the best learning sits on a middle path, where pupils work things out with guidance and the teacher makes sure they get there. But the research behind “just tell them” is strong, and it deserves a hearing.
Two findings matter. The worked-example effect shows that novices learn more from studying a fully worked solution than from solving problems unaided. The expertise reversal effect shows that this flips as expertise grows: the guidance that helps a beginner gets in an expert’s way.
Apply that to teachers and the EEF pattern makes sense. Perhaps an AI-generated lesson plan is a worked example for a new teacher. No wonder novices benefit most. Perhaps they should just be told.
Except that the same research also explains how worked examples work. They teach only when the learner studies them: asking why this step comes before that one, comparing the example with their own attempt, noticing what they would have missed. And the support is meant to fade. Good teachers withdraw the scaffolding as the pupil becomes capable.
A plan downloaded at 10pm and taught at nine the next morning gets used without being studied, and nothing fades. If we’re not careful, it becomes the crutch we can’t teach without.
So the research the “just tell them” camp relied on ends up supporting teacher first, AI second. A worked example helps the novice who thinks first and uses it second.
Caution-informed, all of us
When there’s no evidence either way, a decision can’t be evidence-informed. At best it’s caution-informed.
That applies to me too. My concern about new teachers is caution-informed. So is the confidence that time saved costs nothing over a career. The EEF trial was never designed to settle this. It ran for ten weeks and measured hours. Whether teachers grow or stall over five years is a different question, and it takes longer to answer.
What we haven’t measured, we can still measure. A school could try it over a term: show some staff what AI does well and what it does badly, ask them to plan teacher first and AI second, and see whether their planning, their confidence and their pupils’ work change. That’s the evidence I’d like to see, and I’d be glad to be proved wrong by it.
Until then, the rule of thumb holds. Teacher first, AI second. The aim is a teacher who is better at the job because of how they spent those 49 minutes.
Comments
Loading comments…