Every conversation about AI in education eventually arrives at the same place: what will happen to teachers? The framing is usually either optimistic displacement — AI will free teachers from routine tasks so they can focus on higher-order work — or anxious displacement — AI will reduce the need for teachers altogether. Both framings misunderstand the actual structure of the teaching role, and both miss the more interesting and more tractable question: what are the specific things teachers spend time on that AI could meaningfully reduce without touching the work that only teachers can do?
The teaching role, at its most granular, is three things: preparation and planning, direct instruction and facilitation, and assessment and feedback. Of these, direct instruction and facilitation is the part that requires human presence, real-time responsiveness, and the kind of relational attunement to a group of specific young people in a specific moment that is genuinely irreducible to an algorithm. A teacher who reads the room and modifies her explanation because she can see that three students in the back are confused is doing something that no current AI can replicate in a real classroom setting. But preparation and planning is a different matter — and assessment and feedback, depending on the type and level, spans a wide range from highly automatable to irreducibly human.
The typical secondary teacher in New Zealand spends somewhere between one and two hours preparing for every hour of teaching. That preparation time includes reviewing the curriculum requirements for the lesson, selecting or creating activities and resources, differentiating for students with identified learning needs, and reviewing student work from previous sessions to inform what to focus on. This is where AI has the most obvious and immediate leverage: tools that can surface relevant curriculum-aligned resources, generate differentiated task variants, and summarise patterns in recent student responses can compress preparation time without touching the direct instruction work. Kami's annotation tools reduce the time teachers spend setting up digital documents for student work. Pedagogue's curriculum design tools compress the time from curriculum requirement to lesson outline. The leverage is in the preparation layer, and it's real.
The assessment and feedback dimension is more nuanced. Marking multiple-choice and short constructed-response items can be fully automated with high accuracy, and the time recovered from that automation is genuinely useful — a teacher who would otherwise spend three hours marking a set of diagnostic comprehension tasks can have that time back with high-confidence auto-marking. But the extended writing assessments, the project-based learning evaluations, the portfolio reviews that constitute a significant portion of NCEA-style assessment work require something different: not just scoring, but qualitative commentary that is specific to a student's work, that identifies the precise aspects that meet the standard and those that fall short, and that frames the feedback in a way that this particular student is likely to act on. AI-generated feedback that is generic, formulaic, or produced without genuine engagement with the student's actual work is worse than useless — it erodes student trust in the feedback process and gives the teacher a false sense that the feedback loop is closed when it isn't.
What teachers actually need from AI, based on what we hear from the educators and school leaders we talk to regularly, is tools that handle the preparation and tracking overhead precisely and reliably, without requiring the teacher to spend significant time managing the tool itself. Teachers are not looking for a powerful AI partner that requires substantial professional development to use well. They are looking for tools that take a task they currently do manually — generate a differentiated worksheet for the students who have already mastered the standard content, summarise which students submitted incomplete work this week, create a reading quiz from the text they assigned — and do it in under sixty seconds without errors. The AI capability required for that is modest by current standards. The product design required to make it genuinely frictionless within a teacher's actual workflow is not modest at all.