AI for education and schools: from AI tutors to student progress analytics
Vocational, higher professional and university institutions are handling more and more students with the same teaching staff. At the same time, the pressure is growing to deliver personalised learning, spot dropouts early and assess open questions faster. Artificial intelligence gives educational organisations practical tools — from AI tutoring and automatic grading to student chatbots and learning analytics — that support teachers without taking over their pedagogical role. We build AI solutions that connect to your LMS, your SURFconext identity and the strict frameworks of the GDPR, the AI Act and the Dutch Inspectorate of Education.
Discuss your education AI case View use casesWhy AI in education is a logical step
The average teacher in vocational or higher professional education assesses hundreds of open questions, essays and internship reports every year, while classes grow and the demand for feedback per student gets louder. At the same time, students expect answers around the clock to their administrative questions — about enrolment, study progress, the binding study advice (BSA) threshold or internship regulations. AI is no miracle cure, but it is a set of tools that automates routine tasks and gives teachers room for what only they can do: teach, coach and act pedagogically.
Educational institutions work with large volumes of data that often go unused: grade history in Magister, SOMtoday or Eduarte, learning activity in Itslearning, Moodle or Canvas, progress data from Numo or formative assessments, and demographic data in BRON and DUO integrations. Combining these streams in a GDPR-compliant way produces models that flag students at risk of dropping out before they actually do, tailor practice material to each student's level, and give teachers a well-founded view of where catch-up or revision lessons deliver the most.
What sets AI in education apart from other sectors is how sensitive the target group is. A large share of students are minors, the content is publicly funded and the assessment process directly affects diplomas with civil effect. We build AI solutions that take this reality as their starting point: not as a compliance layer added afterwards, but as an architectural principle from the very first sketch.
Core applications for vocational, applied sciences and university education
The applications differ per level of education, but the common thread is the same everywhere: more personalisation per student, faster feedback and visible patterns where there used to be only gut feeling.
AI tutor for personalised learning
An AI tutor adapts practice material to each student's level, pace and error patterns. Not a generic sequence of questions, but a path that adjusts based on previous answers. We build tutors that run in your existing LMS via LTI 1.3 and write xAPI events back to your learning record store for analysis.
Automatic grading of open questions
Open questions, short essays and programming assignments can be pre-graded by language models based on a rubric set by the teacher. The teacher sees a draft grade with justification per assessment criterion and adjusts where needed. Awarding the official grade remains with the teacher: AI provides the first read, not the final verdict.
Plagiarism and AI text detection
Students use ChatGPT and similar tools, and that is a given. Our detection pipelines combine classic plagiarism checking with indicators for AI-generated text and stylistic deviations from the same student's earlier work. Not a black-and-white verdict, but a conversation starter for the teacher.
Content generation for teachers
Practice questions, summaries, example cases and draft exam questions can be generated by LLMs based on the syllabus and previous tests. The teacher selects, shortens and validates. Once again, responsibility for the content stays with a human, but the first version is ready in minutes.
Student chatbot for admin questions
"How do I unenrol from a course?", "When is the resit?", "Which documents do I need for my work placement?". A chatbot trained on your course guide, education and examination regulations (OER) and internal FAQ answers these questions 24/7 and escalates only truly complex cases to the student counsellor or the student services office.
Early detection of dropouts
By combining login frequency in Itslearning, Magister or Canvas with grade history, attendance and progress on assignments, you get an early picture of students at risk of dropping out. The study adviser doesn't get a black-box score, but a list of underlying signals to start the conversation.
Dutch as a second language (NT2) and language learning support
For non-native-speaking students or NT2 programmes, adaptive language models offer real-time speech and text correction, pronunciation analysis and level-appropriate practice material. Suitable for vocational colleges (ROCs) with language bridging programmes, international programmes and civic integration courses.
Study progress analytics
Dashboards that show teachers and programme managers which learning objectives the group has mastered, where there is wide variation and which topics need revision. Aggregation at class, year and programme level, with drill-down to the individual student where the teacher considers it necessary.
SURFconext and ECK integration
Our AI components log in via SURFconext, so single sign-on and federated identity stay in place. For digital learning materials, we connect to the Educatieve Contentketen (ECK), so that licences, deliveries and usage are recorded consistently.
How Appfront builds AI for educational institutions
We don't build generic education SaaS. Our approach starts with your existing infrastructure (Magister, SOMtoday, Eduarte, Itslearning, Moodle or Canvas) and connects AI components through standards such as LTI 1.3, xAPI, OneRoster and the SURFconext federation. No parallel system that teachers have to learn separately, but an enhancement of what they already use.
An ICT coordinator at a vocational college (mbo) with nine locations has different constraints from a digitalisation manager at a university of applied sciences with its own IT department, or a university (wo) faculty that already takes part in a SURF pilot. We tailor scope, governance and technology choices to your situation: which data is available via DUO and BRON, how mature is the data in your LMS, which data protection officers are involved, and which decisions still have to be made by the executive board (CvB)?
Our way of working is iterative and pedagogically informed. We start with one concrete use case, for example a chatbot for study advice or an AI tutor for mathematics, build it first as a proof of concept, validate it with a small group of teachers and students, and only expand once the pedagogical and technical results justify it. No mega-tender for a platform that may not fit your educational vision.
From first conversation to working education AI
Our approach to AI projects in education follows four phases. Each phase delivers a concrete, testable result, with no year-long analysis phase during which teachers and students notice nothing.
1. Pedagogical and technical assessment
We map out your educational vision, data landscape and legal framework: which data sits in Magister, SOMtoday, Itslearning or Canvas, which data processing agreements are in place, and which arrangements exist with the data protection officer and the executive board.
2. Proof of concept
Within a few weeks, we build a working prototype of the chosen use case on a ring-fenced dataset. An AI tutor for one subject block, a chatbot covering part of the study guide, or an early-warning model for one programme.
3. Integration and pilot
The validated model is connected to the LMS, SURFconext and, where relevant, ECK and BRON via LTI 1.3 or an API integration. We run a pilot with a defined group of teachers and students and measure impact and acceptance.
4. Scaling up and monitoring
After a successful pilot, we scale up to a broader rollout, with ongoing monitoring of model performance, bias and data drift. We retrain when new school years or curriculum changes call for it.
Technology we use for education AI
The choice of technology depends on the application. For adaptive tutors, we use knowledge tracing models and transformer-based dialogue systems. For automatic grading, we combine rubric-based evaluation with large language models, and we always keep the teacher in the loop. For learning analytics, we work with classical statistical models and gradient boosting. A well-founded regression model often outperforms a deep neural network in predictive value, and it is also easier to explain.
We connect AI components through open education standards rather than proprietary integrations: LTI 1.3 for embedding in the LMS, xAPI for learning activity events, OneRoster for course and user data, and SURFconext for identity and single sign-on. This keeps your institution independent: the AI component can be replaced in the future without teachers or students noticing any change in the user experience.
Test your idea first: a working prototype in 1 day
With OneDayBuild, we make your idea tangible in a single day for €1,150, so you know whether further development is worth the investment. Decide to go ahead with the full build? Then we deduct the cost in full.
View OneDayBuild →Education data, GDPR and the AI Act: what you need to know
Educational institutions process special category personal data, often of minors, with a direct impact on students' educational careers. Our AI architectures take the GDPR, the AI Act, the SURF Normenkader (SURF's security and privacy framework) and the guidelines of the Dutch Inspectorate of Education as their starting point, not as a box-ticking exercise after the fact.
GDPR for minors
For students under sixteen, many forms of processing require consent from a parent or guardian. We build consent flows that respect this age limit, with clear information provision and the option for parents to object or request access. Data subject rights are implemented technically, not just on paper.
AI Act: high-risk classification
Under the European AI Act, AI in education and vocational training falls partly into the high-risk category, particularly for admissions, assessment and proctoring. Article 5 also prohibits specific forms of emotion recognition in educational settings. We assess every use case against the AI Act framework and document the risk assessment, data quality, human oversight and transparency as required.
SURF Normenkader and the Inspectorate of Education
The SURF Normenkader for information security and privacy and the guidelines of the Dutch Inspectorate of Education are leading for secondary vocational education (mbo), universities of applied sciences (hbo) and universities (wo). Our solutions align with these frameworks: pseudonymisation choices, logging, audit trails and retention periods are made explicit and are auditable.
Education pseudonym legislation and BRON data
Exchanges with DUO and BRON are subject to specific rules on education numbers and pseudonyms. AI models trained on these datasets must meet strict data minimisation requirements. We keep training data separate from production data, use pseudonymous identifiers and remove directly identifiable data from model input wherever possible.
No proctoring with emotion recognition
AI-based online exam proctoring is a sensitive category. Under the AI Act, emotion recognition on students is prohibited, and proctoring itself is classed as high-risk AI. We generally advise against automated proctoring models and, where needed, look for solutions that support human oversight rather than replace it.
Hosting and data sovereignty
Education data stays within the EU, preferably on Dutch infrastructure. We work with SURF-approved cloud providers, our own on-premises deployments and privacy-friendly model hosting. No uncontrolled exchange with US-based LLM APIs for student or teacher data.
Practical scenarios for mbo, hbo and wo
AI in education is no longer a thing of the future. These are realistic projects that fit in with existing systems and frameworks, and that can be up and running within one or two academic years.
Early dropout warning at a regional training centre (ROC)
A secondary vocational education institution with 18,000 students sees structural dropout in the first year. By combining grade history from Eduarte, login frequency in Itslearning and attendance data, we create a dashboard for study career counsellors. Not a score that "writes students off", but a list of signals, for example a student who has not logged in for three weeks and is slipping in one subject, that the counsellor can use to start a conversation.
AI maths tutor in higher professional education (hbo)
A university of applied sciences with highly varied maths intake builds an AI tutor that runs in Canvas via LTI 1.3. The model adapts practice material to each student's level, offers explanatory videos for specific error patterns and reports aggregated progress to the lecturer. In the dashboard, the lecturer sees which topics the group has not yet mastered and adjusts the next lecture accordingly.
Student chatbot for a university faculty
A university faculty receives thousands of emails every year about teaching and examination regulations (OER), the binding study advice (BSA), resits and internship procedures. A chatbot trained on the study guide, the OER and the internal FAQ, with SURFconext login so the bot knows which programme and cohort the student belongs to, answers standard questions 24/7 and escalates genuinely complex cases to the study adviser.
Automatic grading of programming assignments
A university of applied sciences computer science programme has hundreds of programming assignments to assess every week. An AI pipeline runs the code against test cases, assesses code quality against a fixed rubric and produces a draft grade with justification. The lecturer reviews the output and adjusts it. Marking time per assignment drops considerably, while the lecturer retains responsibility for the substance.
Dutch as a second language (NT2) support for civic integration
A regional training centre (ROC) with language bridging programmes uses an adaptive speaking and writing coach for NT2 students. Students practise pronunciation and phrasing at their own pace, receive instant feedback and follow a personal learning path. The teacher uses aggregated data to focus classroom teaching on the points the whole group struggles with.
Content generation for question banks
A programme team wants a larger practice and question bank without individual lecturers spending hundreds of hours writing questions. An LLM generates draft questions based on the syllabus and previous exams. The subject group reviews, shortens and validates them, so the question bank grows faster while quality assurance of the content is maintained.
Why Appfront for AI in education
Architecture built for education
We understand the difference between a student information system (SIS), a learning management system (LMS) and a learning record store. Between DUO integrations, BRON data exchange and SURFconext federation. Between formative and summative assessment. We translate that domain knowledge directly into AI models and integrations that fit your landscape.
GDPR and AI Act compliance from day one
Compliance is not an afterthought but a design principle. Risk assessment, data minimisation, human checks and explainability are built into the architecture, not added as documentation afterwards. Your data protection officer and executive board get technical answers to their questions, not just legal ones.
Open standards and no lock-in
LTI 1.3, xAPI, OneRoster and SURFconext: we integrate through standards your institution already invests in. That means AI components are replaceable and your data landscape does not get stuck in the proprietary formats of a single vendor.
Frequently asked questions about AI in education
Using AI in your educational organisation?
Discuss your case with us. Together with your ICT coordinator or digitalisation manager, we map out where AI delivers the most value, both pedagogically and technically, and which compliance requirements take priority. No obligation and no strings attached.
Schedule a call