1 in 3 teachers used AI at work (OECD).

Students moved first. Teachers are catching up fast.

Around one in three lower-secondary teachers used AI in their work across OECD systems in 2024, the TALIS survey found. The spread is wide: about three in four teachers in Singapore and the United Arab Emirates, 14% in France. Most use it to learn about or summarise a topic (68% of AI users) and to generate lesson plans or activities (64%). Only a quarter use it on student participation or performance data.

Germany moved within a year. The share of teachers who use AI for work several times a month or more rose from under a third to 52%, the German School Barometer reports. Students are further ahead still: 84% of German 12- to 19-year-olds have used ChatGPT and 74% use AI for homework. In the UK, 95% of undergraduates use AI and 94% use it for assessed work, according to HEPI. In the US, 54% of teens have used chatbots for schoolwork.

52% of German teachers use AI for work several times a month or more (Schulbarometer 2026)

The question for schools, universities and company academies is no longer whether AI is in the classroom. It arrived on students' phones. The question is whether it improves learning.

Design decides the effect

The evidence is now clear on one point. AI designed for teaching raises learning; unguided use can weaken it.

Sources: as listed; effect sizes as reported by the authors.

The guided-tutoring study in Nigeria and the Harvard physics trial used tools with pedagogical guardrails: they ask, prompt and check rather than hand out answers. The PNAS study from Turkey gave students plain chatbot access; they practised faster and learned less. The OECD reaches the same conclusion in its Digital Education Outlook 2026: tools built for teaching work better than general chatbots.

For a buyer this is the most useful filter in the market. Ask what the tool does when a learner asks for the answer.

A profession under pressure

The world needs 44 million more teachers by 2030, UNESCO estimates. A quarter of teachers in OECD systems work in schools where shortages limit teaching. In Germany the education ministers project a gap of about 49,000 teachers to 2035, concentrated in lower secondary and vocational schools.

Skills are the barrier, not interest. Three in four teachers who do not use AI cite a lack of knowledge, and 52% of German teachers say they need training. Teachers are also cautious: two thirds expect negative effects on critical thinking, while 59% see benefits for individual support. Training is where the time returned by AI should go first.

That makes the people side part of every pilot, not an afterthought. A workforce annex names the teachers, trainers and support staff a pilot touches, measures preparation, marking and administration time before anything changes, and sets a training plan with an owner. It brings in staff councils, school leadership, parents and learner representatives early. It names the teacher or trainer accountable for every grade and learning decision, and it says how learners and parents can challenge an AI-supported result. Finally it decides where returned time goes: individual support, feedback or professional learning. Saved time is not realised value until it demonstrably goes into learning.

What the market offers

The Good AI Prism maps 24 offerings in Education, each with a dated source from the provider's own pages. Seven serve corporate learning and course creation, among them 360Learning, Area9 Rhapsode, Docebo Harmony and Sana Learn from Workday. Six help teachers prepare lessons: Brisk Teaching, Eduaide, fobizz, MagicSchool, TeacherMatic and to-teach. Six tutor students directly, including AIS.chat, Khanmigo from Khan Academy, simpleclub and Skye from Third Space Learning. Three are course, language and study tools such as Duolingo Max and NotebookLM, and two support assessment: Gradescope and Kahoot!. 12 of the 24 offerings come from providers headquartered in Europe, five of them in Germany. Ownership matters too: Workday owns Sana, Turnitin owns Gradescope, and Jisc owns TeacherMatic, so a school buying a small tool may be contracting with a much larger group. The order is alphabetical, not a ranking. The full map sits in the Education sector brief.

Germany shows the shared route. AIS.chat, formerly telli, is one AI chatbot for schools run by FWU for all 16 Länder, hosted in the EU, with pupils' data not used for training, as the FWU announcement sets out. DigitalPakt 2.0 adds €5 billion from the federal government and the Länder from 1 September 2026, with AI focused on teacher training.

Grading: high-risk duties from Dec 2027

Europe's AI Act, as amended by the Digital Omnibus on AI, already bans emotion recognition in education institutions since 2 February 2025. Since 2 August 2026, learners must be told when they interact with an AI system. By 2 December 2026, generative systems already on the market must mark their output in a machine-readable way. From 2 December 2027, AI used for admission, evaluating learning outcomes, steering the learning process, deciding the level of education or proctoring exams carries high-risk duties under Annex III.

2 Dec 2027 EU high-risk duties for AI in admission, grading and exam proctoring apply

The AI Act classifies uses, not products. A lesson-planning assistant is minimal risk; the same vendor's feature that grades essays falls under the high-risk duties. Elsewhere the rules come from ministries and regulators. England's exam regulator asked exam boards in March 2026 to strengthen detection of AI misuse, and 65% of UK students say assessment has already changed because of AI.

In practice

Use case: one chatbot for all Länder. AIS.chat, formerly telli, is one AI chatbot for schools, run by FWU for all 16 Länder, hosted in the EU, with pupils' data not used for training. A single shared contract settles hosting, training and exit for every school that joins.

Scenario: an essay-feedback tool. A secondary school pilots a tool that comments on essays before they are handed in. As long as it gives hints and teachers set every grade, the school records it as learning support. Once its score enters the report card, it evaluates learning outcomes and carries high-risk duties from 2 December 2027. The school decides and documents which side of that line the pilot sits on before the first essay is uploaded.

What to do Monday

1. List every tool in use. Include those teachers and students bring themselves, and note where pupils' data is stored and whether it is used for training.
2. Separate support from grading. Lesson preparation and tutoring stay low-risk; admission, grading, steering learning and proctoring need high-risk controls by December 2027.
3. Prefer tools built for learning. Ask for evidence of learning gains for your age group and for the guardrails that stop the tool from handing out answers.
4. Train teachers first. Document AI literacy measures, funded through programmes such as DigitalPakt 2.0, and keep a teacher accountable for every grade.
5. Redesign assessment. Adapt assignments and exams to a world where AI is always available, and measure learning gains, teacher time and equity before scaling.

The Public Sector topic area applies the same AI Act calendar to administrations, and the global market study sets education against ten other sectors.