AI Saved Planning Time—but Did It Improve the Lesson? How Schools Can Separate Efficiency, Resource Quality and Learning
A teacher uses an AI assistant to create a lesson outline, retrieval questions, examples and an exit ticket. Preparation takes forty minutes instead of ninety. The saving is valuable, especially in a profession where workload competes with attention for students, feedback and professional learning.
But the clock cannot tell us whether the lesson improved.
Schools need three separate judgements. Efficiency asks whether preparation required less time or effort. Resource quality asks whether the plan and materials were accurate, coherent, accessible and aligned with the curriculum. Learning quality asks what happened when those resources met these students in this classroom: whether explanations worked, misconceptions surfaced, participation widened and knowledge was retained or transferred.
AI may improve one outcome without improving the others. It can save time while producing materials of comparable quality. It can generate attractive resources that fail to address the class’s actual misunderstanding. It can free time that a teacher reinvests in adaptation and feedback, or create an additional layer of checking that shifts rather than reduces workload.
The right evaluation therefore follows a chain:
AI-supported preparation → teacher review and adaptation → enacted lesson → student experience → evidence of learning
What the Aila trial establishes
On 6 October 2026, the Education Endowment Foundation published findings from a randomised trial of Aila, Oak National Academy’s AI lesson-planning assistant. The study involved 464 Key Stage 2 teachers in 108 schools during the autumn 2025 term. Teachers encouraged to use Aila spent an average of two and a half hours per week on lesson planning and resource preparation, compared with three hours and nineteen minutes in the comparison group. That is a saving of approximately 49 minutes, or 24%.
The evaluation also reported that sampled lesson plans and resources were not of lower quality. Teachers earlier in their careers or less confident in their subject knowledge were more likely to report workload benefits. At the same time, use declined across the ten-week trial, experiences were mixed, and fewer than one in five teachers responding after the trial felt that the tool had meaningfully influenced their teaching and planning.
These are useful findings because they go beyond testimonials. They support a bounded claim: in this study, access to Aila reduced reported preparation time without detectable deterioration in the assessed quality of sampled materials.
They do not establish that every teacher will save the same amount, that all generative tools perform similarly, that the materials were better, that classroom instruction improved or that pupils learned more. The comparison group could also use other AI tools, and Aila’s main function—creating lessons from scratch—may be more relevant to some teachers and tasks than others.
The most responsible conclusion is therefore not “AI improves teaching.” It is “a particular tool reduced one part of workload in a defined context, while the effect on enacted teaching and student learning remains a separate empirical question.”
Why a good resource is not yet a good lesson
A resource can be accurate and well structured while remaining poorly matched to the class. Teachers interpret the room continuously. They notice that an example activated the wrong prior knowledge, that a pupil’s confident answer hides a misconception, that vocabulary has become a barrier or that an activity is easier than anticipated. They change pace, representation, questioning and support.
This adaptive work is not fully contained in the plan. It depends on subject knowledge, knowledge of learners and the evidence produced during teaching. AI can propose a sequence, but it does not automatically know which explanation this group has already heard, which student is masking confusion or which task will provoke productive rather than overwhelming difficulty.
The OECD’s work on high-quality teaching describes teaching through dimensions that include cognitive engagement, subject content, formative assessment, social-emotional support and classroom interaction. A polished slide deck is therefore only one component of a much larger educational event.
For the same reason, a better resource does not guarantee better learning. The material has to be enacted, interpreted and adjusted. Students must perform the intended cognitive work. The teacher must gather evidence and decide what to do next.
The three-ledger evaluation
A school can avoid inflated claims by keeping three connected ledgers. Each ledger answers a different question and requires different evidence.
LedgerCentral questionUseful evidenceMisleading shortcutTime and workloadDid AI reduce preparation effort, and where was the saving created?Time diaries, task-level estimates, checking and revision time, adoption over several weeks, teacher experience level.Asking whether the tool “felt fast” after one demonstration.Resource qualityWere the plan and materials accurate, coherent, accessible and usable?Independent review, curriculum alignment, subject accuracy, task demand, examples, accessibility and necessary teacher corrections.Treating fluent language or attractive formatting as quality.Teaching and learningDid the enacted lesson help these students learn?Observation, questioning, student work, misconceptions, participation, delayed recall, transfer and teacher adaptation.Assuming that time saved or a high-quality plan caused learning gains.
The ledgers should remain separate in reporting. If preparation time falls by 20%, say that. If independent reviewers judge that resource quality was maintained, say that. Do not convert those findings into a claim about attainment unless student evidence supports it.
The ledgers can still inform one another. If saved time is consistently reinvested in analysing prior assessment, adapting examples or giving feedback, a plausible mechanism for improvement exists. If resources require extensive correction, the apparent saving may disappear. If a tool produces reasonable materials but teachers stop using it, implementation has failed even if the prototype was promising.
Evaluate the plan as a professional object
A lesson plan produced with AI should be reviewed against educational criteria rather than against the sophistication of the prompt.
DomainReview questionLearningIs the objective precise, worthwhile and connected to prior and future knowledge?ExplanationAre concepts, examples and representations accurate and likely to make the content clearer?Cognitive workWhat will students retrieve, reason about, practise, compare or create?AssessmentHow will the teacher discover what students understand during and after the lesson?InclusionAre language, accessibility, prior knowledge and participation barriers anticipated without reducing the intellectual goal?EnactmentWhere is teacher judgement required, and what evidence might justify changing the plan?
A plan that passes this review is ready to be tested, not declared effective. Classroom evidence may still show that the explanation failed, the practice sequence was too easy or the assessment did not reveal the misconception that mattered.
Five productive roles for AI in planning
AI is often most useful when it expands the teacher’s field of options without replacing the final educational decision. It can generate contrasting examples, alternative representations or possible misconceptions that a subject expert then evaluates. This is especially useful when the teacher is planning unfamiliar content, a pattern also suggested by the Aila trial.
It can expose assumptions by challenging a draft plan: Which prior knowledge is being taken for granted? Which pupils may struggle to enter the task? What alternative explanation would reveal a different feature of the concept? The output should be treated as a prompt for professional thought, not an authoritative audit.
It can create controlled variation for practice. A teacher may ask for examples that vary one feature while preserving another, then check them for accuracy and difficulty. This can reduce formatting time while preserving teacher control over the learning progression.
It can also support critique. Comparing two generated explanations and identifying why one is clearer may be more educationally useful for planning than requesting a single “perfect” answer. Finally, AI can remove low-value production work such as reformatting teacher-approved content, provided the saved time is not consumed by correction or by generating excessive material.
How AI can weaken the plan
The first risk is generic alignment. A generated lesson may repeat the vocabulary of a curriculum while missing the conceptual sequence, local specification or particular knowledge students need. It sounds aligned without being instructionally precise.
A second risk is fluent error. A plausible explanation can contain a subtle misconception, fabricated reference or misleading simplification. Teachers with weaker subject confidence may benefit most from planning support while also finding these errors harder to detect. That tension makes review design essential.
A third risk is excess. Generative systems can produce more activities, questions and explanations than a lesson can use. Abundance may increase choice, but it can also increase checking time and crowd out deliberate selection.
A fourth risk is cognitive outsourcing. If the tool decides the objective, progression, examples, questions and assessment, the teacher may lose contact with the reasoning that holds the lesson together. This matters when adaptation becomes necessary: it is difficult to depart intelligently from a plan one did not understand.
The final risk is deprofessionalisation. A planning system that treats teachers as approvers of machine output can reduce judgement to editing. A better design uses AI to support comparison and reflection while keeping purpose, diagnosis and responsibility with the professional.
The lesson-improvement test
Before and after using AI, the teacher can complete a short written comparison.
Before planningAfter the lessonWhat do students need to learn?What evidence shows what they learned?What prior knowledge and misconceptions matter?Which anticipated or unexpected misconceptions appeared?What cognitive work must students perform?Did they perform that work, or did the resource/tool do it for them?Where might adaptation be needed?What did the teacher change, and why?What would count as a better lesson?Which part improved, remained unchanged or worsened?
The comparison prevents retrospective storytelling. If “better” was never defined, any successful moment can be attributed to the tool. A credible test states the expected mechanism in advance.
For example, a teacher may predict that AI-generated contrasting examples will reduce preparation time and help students distinguish two scientific concepts. The plan should then record both outcomes: time spent preparing and evidence from student explanations or transfer questions. If time falls but the distinction remains unclear, the efficiency claim may succeed while the learning claim fails.
Track where the saved time goes
Time is not automatically converted into educational value. A teacher who saves 49 minutes may use it for feedback, collaboration, adaptation, recovery from workload or unrelated duties. Each use may be legitimate, but schools should not claim a particular benefit without evidence.
A small study can ask teachers to assign saved time to broad categories: checking the generated plan, improving the same lesson, preparing other work, communicating with students or families, professional collaboration, administrative tasks, or reducing overtime. The categories should not become surveillance. Their purpose is to understand the mechanism.
If the intended theory is that AI creates more time for responsive teaching, the school should look for that response. Were misconceptions analysed more carefully? Were examples adapted for the class? Did feedback become more timely? Without this link, the organisation has evidence of workload reduction, not of improved teaching.
A six-week school pilot
In the first week, define the claim precisely. “AI will transform planning” is not testable. “For Year 5 science lessons planned from scratch, the tool will reduce preparation time while maintaining independent ratings of resource quality” is clearer. If the school also expects learning benefits, it must state the proposed mechanism and evidence separately.
The second week establishes a baseline using ordinary practice. Record preparation time, the type of lesson, teacher experience and a sample of resources. Identify a small set of learning indicators that fit the subject.
During weeks three and four, teachers use the tool and document the same evidence. They should record checking and revision time, not only generation time. Short reflective notes can capture when the tool helped, created extra work or changed a professional decision.
Week five examines enacted lessons. Review a small number of observations, student work samples and teacher adaptations. The purpose is not to rate teachers but to see whether the plan survived contact with the classroom.
In week six, separate conclusions by ledger. Continue, adapt or stop the use according to the evidence. A finding that the tool saves time for novice teachers planning unfamiliar content may justify a targeted use rather than a whole-school mandate.
Questions for leaders, vendors and project teams
Leaders should ask what part of planning the tool is intended to improve, what evidence supports that claim and which outcomes have not been studied. They need to know whose curriculum and content underpin the system, how errors and bias are monitored, what information enters the platform, how accessibility is addressed and whether teachers can inspect or override the output.
Vendors should distinguish measured outcomes from aspiration. “Teachers saved time in a ten-week trial” is different from “students learned more.” Evidence from one phase, subject or product version may not transfer to another. Schools should also examine sustained adoption because a tool that performs well under encouragement but is rarely used may have limited organisational value.
Teachers, meanwhile, should ask whether the plan reflects what they know about these students, whether they can explain the instructional sequence, which part required the most correction and what they will look for during the lesson. These questions keep the professional relationship to the plan intact.
The learning-coach perspective
Efficient planning becomes educationally valuable when it creates better opportunities to observe, question, adapt and transfer responsibility. A learning-coach perspective asks what the learner will notice, decide, practise, explain and carry into a new situation.
This emphasis complements the Learning Coach guide to student independence. The plan should not merely organise teacher activity. It should design conditions in which students perform meaningful cognitive work and receive support calibrated to evidence. From Brain Games to Real Learning provides a complementary test: can students apply the reasoning beyond the supported exercise? Academic Coaching vs. Tutoring for Teens similarly distinguishes assistance that completes work from support that develops autonomous methods.
Educators who want to strengthen diagnostic observation, scaffolding and evidence-informed reflection can explore the Become a Learning Coach training pathway. Further English-language resources on AI and teaching are available through the Learning Coach article hub.
The honest conclusion
The Aila trial provides credible evidence that one specialist AI planning tool reduced preparation time by approximately a quarter without lowering the assessed quality of sampled resources in a defined primary-school context. It also found mixed views, declining use and limited perceived influence on teaching. That combination is more useful than either celebration or dismissal.
A faster plan is not automatically a better lesson, and a better resource is not automatically better learning. Schools should measure efficiency, resource quality and classroom learning separately, examine the teacher’s adaptation and trace the evidence into student thinking.
