Blog Industries

AI in education: admissions, course support and rules (2026).

What AI can safely do for applicants, students, teachers and administrators, where a person has to stay in charge, and the rules that tighten around admissions, assessment and children’s data.

Fig. 0Course material in, a cited answer out

For schools, universities and training providers, the AI worth building first answers applicants’ and students’ questions at any hour, from your own published course and policy information, and hands over to staff when a question turns personal. It also takes routine admin off teachers and staff. Anything that decides admission, grades work or involves children’s data needs far more care.

This guide covers the uses that work, a starter AI-use policy, the rules by market (as of October 2026), and how to build and measure an admissions assistant.

Where does AI help in education?

Admissions and applicant questions

Entry requirements, deadlines, how to submit transcripts and English-language test results, accommodation, fee policies, intake dates: applicants ask the same questions in the same weeks, from many time zones. An assistant that answers from your published programme pages and policy documents, in the applicant’s language, is the most winnable first project for a university or college. It is especially useful for international recruitment, where an applicant’s evening is your office’s night. It hands over to an admissions officer, with the conversation attached, for anything personal, such as a particular application, a borderline qualification or a scholarship.

Course support

An assistant grounded in one course’s own materials answers questions about the content and points to the lecture, reading or page it came from. It helps students learn rather than write their assignments, and staff control what it can and can’t do. It usually sits inside the learning platform you already run (Moodle, Canvas, Blackboard, Microsoft Teams and similar) rather than replacing it.

Teacher and lecturer workload

Drafting lesson materials and quizzes, producing differentiated versions of a resource, summarising feedback and preparing reports, all for the teacher to review and adapt. England’s Department for Education says it sees “more immediate benefits and fewer risks from teacher-facing use of generative AI” than from pupil-facing use, which is a useful order in which to plan.

Administration

Reading applications and supporting documents, checking them against a checklist, answering staff questions from policies, and drafting routine correspondence. The output is written into your student information system (Banner, Workday Student, a CRM such as Slate, or whatever you run), with a person approving anything that changes a record.

Corporate training and academies

For training providers and internal academies: assistants that answer from course content, let learners practise, and draft materials from subject experts’ notes.

Which kind of AI fits which job?

Most education problems need a document or retrieval system, not an open-ended chatbot. The more a task affects a person’s future, the more a person should review it.

TaskRight kind of AIData it needsWho stays in charge
Applicant and student FAQsRetrieval-based assistant (RAG) with citationsPublished programme pages, policies, calendarsStaff review conversations; handover for personal cases
Application status and document checksAssistant with read-only record lookup, plus document AIApplication record, uploaded documentsAdmissions officer approves anything that changes a record
Course Q&ARetrieval over one course’s materialsSlides, readings, syllabusLecturer sets what it may and may not do
Lesson and quiz draftingLanguage model with the teacher’s own inputsCurriculum, teacher notes (no pupil personal data)Teacher reviews and adapts every output
Deciding admission or grading workNot delegated to AINot applicablePeople decide; high-risk under the EU AI Act

What should an AI-use policy cover?

Before anything goes live, write down how staff and students may use AI. The TeachAI toolkit suggests addressing AI inside your existing policies, such as academic integrity, and updating responsible-use policies, rather than starting only a separate document. It also lists AI literacy in professional development and the curriculum as part of the same effort. A starter policy covers:

  • Academic integrity. What counts as acceptable help on each type of assessment, and how students disclose AI use. In UK qualifications, the JCQ guidance on AI use in assessments expects students to acknowledge the AI source and date, and expects teachers to be sure the work is the student’s own.
  • AI detectors as evidence. JCQ says detection tools may be used as a check or to verify concerns, and that results should be weighed with all other information. Treat a detector score as a prompt for a conversation, never as the sole basis for a finding.
  • Personal data and tools. The DfE recommends not putting personal data into generative AI tools, and warns that many free tools use what is typed in to train their models. It also says students’ original work must not be used to train a model without permission, unless a copyright exception applies.
  • Who may approve a new tool, and what that review checks: where data is stored, whether inputs are used for training, accessibility and age suitability.
  • Training for staff. What teachers and administrators need to know before they use or supervise an AI system.

Which rules apply?

The rules differ by market and by who the system serves. Check the current text with a qualified adviser before launch; the points below describe what the official sources say.

  • EU AI Act. Education appears in Annex III as a high-risk area: systems used to determine access or admission, to evaluate learning outcomes, to assess the appropriate level of education a person will receive, or to monitor and detect prohibited behaviour of students during tests. The application date for Annex III systems is 2 December 2027 after the Digital Omnibus amendment. Under Article 50, providers must ensure that people talking to an AI system are told they are doing so, so a student-facing chatbot should say it is one. And Article 4 requires providers and deployers to take measures to support the AI literacy of their staff and others who operate AI systems on their behalf. See our EU AI Act guide.
  • United Kingdom. The ICO’s Children’s code sets fifteen standards for online services likely to be accessed by children, and the ICO publishes guidance specific to education technology. For decisions made solely by automated means with a significant effect on a person, the Data (Use and Access) Act 2025 changed the UK GDPR rules, and the ICO’s draft guidance on automated decision-making (updated March 2026) covers the safeguards, including human intervention and the right to contest. Our UK AI rules guide covers the wider picture.
  • United States. The Department of Education’s July 2025 Dear Colleague letter says AI use should be educator-led, accessible, transparent and data-protective, and that systems must comply with federal privacy law including FERPA, which protects student education records. COPPA applies to online services collecting data from children under 13, and the FTC’s COPPA FAQ says a school’s ability to consent for parents is limited to the educational context and to collection for the school’s use and benefit, with no other commercial purpose.
  • Italy. Under Article 4 of Law 132/2025 (Disposizioni e deleghe al Governo in materia di intelligenza artificiale, “provisions and delegations to the Government on artificial intelligence”), in force since 10 October 2025, access to AI technologies by children under 14, and the related processing of their personal data, requires the consent of whoever holds parental responsibility. A young person aged 14 or over may consent to the data processing themselves, provided the information given is easy to access and understand.
  • Global guidance. UNESCO’s guidance on generative AI in education and research calls for data-privacy protection and an age limit for independent conversations with generative AI platforms, without naming a number.
  • Gulf institutions. Data protection and residency rules shape where a system can run and where student data can sit. Our guides to building AI in the UAE and Saudi Arabia cover them.

Where AI is the wrong tool, and where a person stays in charge

  • Admission decisions. An assistant can check that a file is complete and summarise it. The decision to admit or reject stays with an admissions officer, partly because it is a high-risk use in the EU and partly because borderline cases are judgement calls.
  • Grading and feedback that counts. A model can help a teacher draft comments. It shouldn’t set a mark that goes on a record without a teacher’s review.
  • Misconduct findings. A detector score is not proof. A person looks at the evidence and speaks to the student.
  • Exam monitoring. Systems that watch students during tests are named in Annex III, and raise privacy and fairness questions. Think hard before building one.
  • Wellbeing and safeguarding. If a student tells an assistant they are at risk, the system must recognise it and route it to a trained person straight away. It should never try to handle it alone.
  • Pupil-facing tools for young children. The DfE says pupils should only use generative AI in education settings with appropriate safeguards, and that safety should be the top priority. Start with teacher-facing uses.

How is an admissions assistant built to be safe?

This is the project we would start with, and the design choices are mostly about limits.

  • Answers only from approved material. Each answer cites the programme page or policy it came from and is scoped to the right intake year and programme. When the material doesn’t cover a question, the assistant says so and offers a person (see retrieval that knows when to say “I don’t know”).
  • Limits in code, not in the prompt. It never predicts whether an applicant will be admitted, never promises a scholarship or a place, and never quotes a deadline that isn’t in the source. Those rules are checked after the model writes, not hoped for.
  • Read-only access to personal data. If applicants can ask “where is my application?”, the assistant looks up only that person’s record after they have verified who they are, and it can’t change anything.
  • Uploads are untrusted. Transcripts, statements and emails can contain text that tries to steer the model. Treat them as input to validate (see our note on prompt injection).
  • Handover that works. If the channel is WhatsApp, Meta’s business policy allows automation during the 24-hour customer service window but expects a prompt, clear escalation path (such as a transfer to a person), and outside that window only approved message templates can be sent. Design the handover and the follow-up around that.
  • Tested on real questions. Collect a few hundred questions from past enquiries, in each language you support, have admissions staff mark the right answers, and re-run the set before every change (see evals before features).
  • Accessible and age-appropriate. It works with screen readers and on a low-end phone, and anything minors use follows the children’s data rules above.

How to start, and what to measure

  1. Pick one audience and one season. For example, international applicants for the next intake, or new-student questions in the first month.
  2. Collect the sources and the questions. Gather the pages and policies the assistant may use, and the real questions staff answer most.
  3. Agree the numbers before building. Baseline them today.
  4. Run it alongside staff. Staff see every conversation at first and correct the answers.
  5. Widen carefully. Add languages, then record lookups, only when the numbers hold.

The numbers worth tracking:

  • Median time to a first useful answer, including out of office hours.
  • Share of questions resolved without staff, and how often applicants were handed over when they should have been.
  • Answer accuracy on your test set, per language.
  • Enquiry-to-application conversion for applicants who used the assistant, compared with a baseline.
  • Staff hours spent on repeat enquiries during the peak weeks.

Complex integrations, such as an older student information system with limited APIs, are usually the biggest piece of work, which is why the first version starts with published content and adds record lookups later. Our services page covers how we build and run these systems.

Questions educators ask

Will AI encourage cheating?

A general chatbot might. A course assistant you build controls what it does: it can explain concepts and point to materials without producing assessed work. Clear policy on disclosure and human review of suspected misconduct matter as much as the tool.

Can an AI assistant decide who gets admitted?

It shouldn’t. Under the EU AI Act, systems used to determine access or admission are high-risk, and UK and EU data rules limit decisions made solely by automated means. Use AI to answer questions and prepare files, and keep the decision with admissions staff.

Can it work in several languages?

Yes, which matters for international applicants and multilingual student bodies. Each language gets its own test questions and is checked before launch, so you know the accuracy rather than assume it.

Where should student data be kept?

In regions and with providers your data protection obligations allow, never used to train public models, and accessible only to the staff who need it. The DfE warns that many free tools train on what users enter, so avoid putting student data into them.

Working in education or training?

Modulus Labs AI builds applicant, student and course assistants grounded in your own material, with staff in control. Tell us what you’d like to support, and we’ll tell you what we’d build first.