If you've sat through a vendor pitch about AI in education in the last two years, you've heard about adaptive learning paths, AI tutors, predictive analytics, personalised pedagogy, and probably a "digital twin of the student." Most of these are either not in production anywhere, or in production somewhere and quietly failing.
Stripping away the optimism, only two AI applications have actually earned daily-use status in real classrooms. This post is a frank look at what those two are, why they work, and why the others haven't yet.
The two that work — and why
1. AI Grading
The single biggest win is grading. It works because the job has clean inputs (the paper, the answer key), a clean output (marks per question with reasoning), and — critically — a human in the loop who signs off. The teacher's override is the safety valve that lets a school adopt AI without betting the entire grade book on it.
It works because there's a structured artefact to verify: every grade comes with the matched key points, the missed key points, the OCR confidence. A teacher can audit any decision in 90 seconds. The thing being automated is the repetitive part of marking — pattern-matching the answer to the key — not the judgement part.
2. AI Content Generation, grounded in the school's books
The second one is content. Quizzes, worksheets, lesson plans, slide decks, comprehension passages — all the recurring material a teacher rewrites every term. But only when the generation is grounded in the school's own textbooks, not in a generic model's training data.
Generic content generation fails the second-week test: the teacher spends as long editing the output as writing it. Grounded content generation passes — because the page references, the difficulty calibration, and the vocabulary are already aligned to the book on the student's desk.
Why these two, and not the others (yet)
Both of the use-cases above share three properties that the failed ones don't:
| Property | Works in classrooms | Doesn't work yet |
|---|---|---|
| Verifiable in seconds | ✓ Teacher checks a grade or worksheet immediately | ✕ "Personalised learning path" cannot be spot-checked |
| Human in the loop by default | ✓ No mark or worksheet ships without sign-off | ✕ AI tutor talks to students unsupervised |
| Cost of being wrong is bounded | ✓ Override fixes a mark; edit fixes a question | ✕ Bad advice compounds for a month unnoticed |
The AI use-cases that work in schools are the ones where the AI is producing a draft and the teacher is shipping the final version. The ones that don't work are the ones where the AI is making decisions about students that nobody is auditing.
What about "AI tutors"?
This is the most hyped category and the most underperforming in production. The reasons aren't mysterious: tutors operate one-on-one with a student, often unsupervised, often outside the teacher's view. The same hallucination problem that makes generic chatbots dangerous in any classroom is amplified — the wrong answer goes into the student's understanding with no teacher to spot it.
The closest thing that does work is Textbook Chat — a Q&A interface grounded in the school's books, with every answer carrying a chapter and page citation. It looks similar to a tutor from outside. It's structurally different: it can refuse to answer, every answer is verifiable, and the teacher can hand it to a class without losing sleep.
What about "predictive analytics"?
The honest version of this — "students who score below X on inference questions in Term 1 tend to need remediation by Term 2" — is real and useful. But it isn't AI in any interesting sense; it's a regression on a clean dataset. The interesting version, "we predict this student will need intervention before they fall behind," over-promises and under-delivers because the underlying data is messier than the marketing makes out.
What actually moves the needle is descriptive, not predictive: show the teacher what's happening now, clearly. Most schools haven't fully consumed that level of analytics yet. There's no need to leap ahead.
The unglamorous shape of useful AI in schools
Strip away the demos and the conference talks, and the AI that's actually working in classrooms today looks deeply unglamorous: a teacher reviewing AI grades, a teacher editing an AI worksheet, a student looking at a chatbot answer with a page reference under it.
It's not adaptive. It's not magical. It's the AI doing the repetitive draft, and the teacher doing the judgement. That partnership is what survives the second-week test. Most of the rest still doesn't.
Send us a term-end paper. We'll grade it free.
No commitment. We'll upload it to Skora AI, run it end-to-end, and email you back a graded sheet with per-question audit trail, weak-area analysis, and our price for a pilot. Most schools get the grade back the same day.