I've been teaching high school English for eleven years. Last spring, our district rolled out an AI pilot — not mandatory, not well-explained, just "available." By fall, I was using it to draft lesson outlines and generate discussion questions. By winter, I was using it to write feedback on student essays before I edited it. By spring, I realized I hadn't written a parent email from scratch in months.
The work didn't disappear. It changed shape.
So I started paying attention. I asked colleagues in other districts, other subjects, other grade levels: What exactly is AI taking over first? And what is it not touching — at least not well?
Here's what I've found, both from my own experience and from the data that's emerging. I'd like other teachers to weigh in, especially if you're seeing something different in your building.
What AI Is Changing First
The pattern is consistent across the surveys I've read and the conversations I've had. AI is entering teaching through the preparation and paperwork channels, not through the relational ones.
Lesson Planning and Materials
This is the most common entry point. A Gallup/Walton Family Foundation survey of 2,069 U.S. K-12 teachers found that 58% receive encouragement to use AI for modifying student materials, 54% for making worksheets or assignments, and 53% for preparing to teach. The guidance is uneven — 69% of teachers receive no guidance at all on one-on-one instruction, and 58% receive none on grading — but where there is encouragement, it clusters around planning.
In practice, that means teachers are using AI to generate first drafts of lesson plans, differentiate reading levels, create practice problems, and draft rubrics. Anthropic and OpenAI have both launched teacher-specific products, with Anthropic's Claude for Teachers incorporating academic standards from all 50 states. The tools are getting better at alignment. The question is what teachers do with the time they save.
Grading and Feedback
This one is more contested. Only 8% of teachers in a recent YouGov survey said they use AI to mark or grade pupils' work. But the same survey found that 39% use it for student reports and 13% for SEND/EHCP contributions — the writing-intensive, compliance-heavy parts of the job that don't involve judgment about student learning.
LLM-assisted grading is being studied in higher education, with one computer science study finding a 40% reduction in grading time for handwritten exams. But K-12 teachers I've talked to are more cautious. The Gallup data shows that grading is the task where teachers are most likely to receive discouraging guidance — 10% are explicitly told not to use AI for it, the highest discouragement rate of any task measured.
My own experience: I use AI to draft feedback, but I never send it without rewriting. The AI is good at identifying patterns — "your thesis is unclear," "this paragraph lacks evidence" — but it doesn't know which students have IEPs, which ones are dealing with trauma at home, or which ones just need to hear that they're improving.
Parent Communication and Administrative Writing
This is the quiet workhorse of AI adoption in teaching. 39% of teachers use AI for letters and emails to parents, and 15% for reference letters. It's not glamorous, but it's a real time saver. Parent emails, behavior documentation, field trip permission letters, recommendation drafts — these are formulaic enough that AI handles them well, and the stakes of a slightly stiff tone are low.
The risk here isn't quality. It's over-reliance. If teachers stop practicing the skill of writing difficult parent communications — the ones where tone, timing, and word choice matter enormously — they may lose the ability to handle the 5% of conversations where AI output would be actively harmful.
What Still Depends Heavily on Human Judgment
Here's where the data and the experience diverge from the "AI will transform everything" narrative. There are parts of teaching where AI is either not being used, not trusted, or actively discouraged — and those parts are the core of the job.
One-on-One Instruction and Tutoring
69% of teachers receive no guidance at all on using AI for one-on-one instruction or tutoring. That's the highest "no guidance" rate of any task Gallup measured. And when guidance exists, only 35% of teachers say it encourages AI use for tutoring — the lowest encouragement rate of any task.
There's a reason for that. Tutoring is diagnostic. It requires reading a student's confusion in real time, adjusting the explanation on the fly, and knowing when to push and when to back off. AI can generate practice problems and explain concepts, but it can't see the face of a student who's about to give up, or hear the hesitation in a voice that says "I don't get it" when the words say "I'm fine."
The ETS "human-in-the-loop" framework puts it plainly: educators remain responsible for interpreting AI-generated information, weighing context, and making decisions that are clear, defensible, and fair. Tutoring is where that responsibility is most acute.
Classroom Management and Behavior
This is the domain teachers are most protective of. A study of vocational teachers found that decisions involving student behavior, urgent incidents, and tailored learning interventions "should be inherently human," with AI providing only limited or supporting input. Teachers in the study drew explicit "red lines" around these areas, preserving their professional agency.
AI tools can monitor engagement, flag behavioral patterns, and surface data about participation. But the decision to intervene — to call a student aside, to adjust seating, to escalate to a counselor, to let something slide because you know the kid is having a terrible week — is not an algorithmic decision. It's a judgment call made by someone who knows the student.
Emotional Support and Relationship Building
This is the part of teaching that doesn't show up in any productivity metric, and it's the part AI is least equipped to touch.
One analysis of AI in education put it directly: AI should take away "mechanical lesson prep, repetitive grading, process reminders, and inefficient communication" and give back to teachers "the ability to see students' states, the patience to understand students, and the guidance of students' development". That framing is useful because it names the trade-off: if AI handles the mechanical work, teachers can spend more time on the human work. But that only happens if the time saved actually goes to the human work, not to higher paperwork quotas.
A recent study found that 51% of teachers said AI eased their workload, but only 35% said it reduced their hours. The hours saved on grading and planning were swallowed by other demands. The human work — the relationships, the mentoring, the informal check-ins — didn't automatically expand to fill the gap.
Ethics, Privacy, and Student Data
This is the area where teachers are most exposed and least supported. 54% of teachers say they have not received clear district-level policies on AI use in classrooms, and 38% say their school or district has provided little to no support for AI integration.
That matters because the decisions teachers make about AI — which tools to use, what student data to enter, how to handle AI-generated work in assessment — carry real ethical weight. The American Federation of Teachers has proposed vendor requirements and limits on AI use in schools, including banning "student-facing" AI in elementary schools and social companion chatbots for students under 16. Microsoft and two major teachers unions announced a legally enforceable AI privacy standard for U.S. schools in September 2026. But district-level policy is still catching up.
In the absence of guidance, teachers are making it up as they go. That's not sustainable.
The Workload Paradox: Less Work, Same Hours
The most important finding from the data — and the one that matches my experience exactly — is that AI is reducing the cognitive load of certain tasks without reducing the total hours of work.
A YouGov survey of 1,033 teachers found that 80% now use AI for their work, with 52% using it weekly and 12% daily. Among users, 51% said AI eased their workload, but only 35% said they worked fewer hours. More than half (55%) said their hours were unchanged, and 4% said they worked more.
Teachers typically save one to two hours a week through AI use, with 7% saving more than five hours and 15% saving less than one hour. But the time saved on discrete tasks — the letter, the worksheet, the report — gets absorbed by other demands. As one report put it: "Teachers feel AI takes the weight off particular jobs but, for most, the hours saved are swallowed by other demands rather than handed back as a shorter week".
This is not a failure of AI. It's a failure of how AI is being integrated. If the efficiency gains go into higher output expectations rather than reduced hours or reinvested time, teachers experience the tool as a speed-up, not a relief.
One teacher development lead put it sharply: "If our ambition is simply to help teachers complete existing tasks faster, we have set the bar remarkably low. A prompt can be taught and a workflow modelled; neither tells a teacher whether using it is the right decision for these pupils, for this purpose, at this moment".
What I'm Doing (and What I'm Still Figuring Out)
I don't have this solved. But here's where I've landed:
I use AI for preparation, not for relationships. Lesson drafts, differentiation, parent email templates — yes. Tutoring decisions, behavior interventions, emotional support — no. I'm not anti-AI. I'm protective of the parts of the job that made me want to teach.
I keep a correction log. Every time AI-generated feedback is wrong or tone-deaf, I note it. It's helped me get better at prompting, and it's given me evidence when colleagues ask whether AI is "good enough."
I ask for policy, not permission. When our district opened the AI pilot, I asked three questions: What data can we enter? Who reviews AI output? What's the escalation path when it fails? I got partial answers. I'm still asking.
I track my time — not for productivity, but for evidence. If AI saves me two hours on planning and I spend those two hours on one-on-one check-ins, that's a win. If I spend them on more paperwork, that's a signal that the integration isn't working.
I'm skeptical of "efficiency." The most important question isn't "How much faster can I do this?" It's "What should I do with the time?" And that's a question teachers should answer, not administrators or vendors.
No feedback yet — submit the first.