AI in Hong Kong Schools: 10 Requirements a School-Based AI Platform Must Meet in 2026

Short answer: a school-based AI platform fit for Hong Kong primary and secondary schools must do four things at once — stay anchored to the school curriculum, protect students, keep teachers in control, and leave the data with the school. The 10 requirements below turn those four principles into items you can write into a tender.

Since the Education Bureau launched the AI Education Grant (『智』啟學教撥款計劃) in late 2025, many schools have had to do something they have never done before: choose one AI learning platform for the whole school within a few months. Offerings range from thin "ChatGPT wrappers" to whole-school systems, prices vary wildly, and feature names all sound alike. This article is not a product comparison. It is a requirements list you can drop into your specification and ask every vendor to answer line by line.

1. AI output must be anchored to the school curriculum

A general model's idea of "P4 fraction addition" may not match your textbook, pacing or learning objectives. A qualified platform has a school curriculum database — subjects, grades, units and objectives imported from CSV / Excel — and every AI tool (planning, question generation, tutoring) generates from that data.

Ask: "Pick a subject, grade and unit and show us how the lesson plan and worksheet reference that unit's objectives."

2. Students must not be spoon-fed: a scaffolded tutor is the floor

If students can get homework answers straight from the AI, the platform harms learning. A scaffolded hint ladder is essential: clarify the question first, give a directional hint, work through part of an example, and only reveal steps at the level the teacher allows. Teachers should set a "hint ceiling" per task.

3. Teachers need live monitoring, pause and flag

When primary pupils use AI, teachers must be able to see the conversations. The platform should provide a monitoring panel: view student–tutor dialogues live, pause individual students, flag messages for follow-up, and summarise common misconceptions automatically.

4. Multi-layer safety with human review

A single "content filter" is not enough. Mature designs have four layers: an input guard (inappropriate requests, personal data, jailbreaks), a system prompt (school policy, age band, Socratic rules), output review (suspicious content masked and flagged) and audit & quotas. Flagged content must be reviewed by an administrator, not silently deleted.

5. Roles and permissions: principal, panel head, class teacher, subject teacher, student

A school is not an organisation where "everyone is a user". You need at least six roles, and panel heads see only their subject; class teachers only their class. A permissions mistake is not an inconvenience — it is a privacy incident.

6. Encrypted personal data and no-training terms

Student phone numbers, student IDs, guardian contacts and SEN notes are sensitive personal data and should be encrypted at the database level. Calls to AI providers must use API terms under which data is never used for training, and the system prompt should forbid the AI from requesting or outputting student personal data. Ask for a mapping to the six PDPO Data Protection Principles.

7. The data always belongs to the school — on-premises or cloud

This is the requirement most often overlooked and the one that most affects a school's long-term autonomy. Whether you choose on-premises deployment (school server or private cloud) or cloud hosting, everything — student records, knowledge base, lesson plans, marking, conversation logs, results — must belong to the school and be downloadable in full, free of charge, at any time (JSON / CSV / original files) with no lock-in period. A confident vendor does not keep customers by holding their data hostage.

A direct question to ask: "If we switch vendors next year, how do we take all our data today? Show us."

8. Multiple models and a switchable AI provider

AI models change every few months. The platform should have a provider abstraction layer so OpenAI, Azure OpenAI, Anthropic, Poe or local models can be switched in the admin panel without touching teaching settings. The subscription should also carry no extra token charges, otherwise budgeting is impossible.

9. Dashboards and report export

The grant requires schools to account for outcomes. The platform must output AI usage (by subject / grade / role), objective mastery, tutor interactions and teacher training records — ideally through a PIE (Plan–Implement–Evaluate) module that produces interim and final report data directly.

10. Training, on-site support and sustainability

Software is only half of it. A qualified package includes teacher training, demo lessons, sharing sessions and consultant co-planning visits, with those records saved automatically as report annexes. The subscription should include all updates and model upgrades.

Checklist: turning the 10 requirements into tender questions

Takeaway

Choosing an AI platform is not about picking the "smartest model"; it is about choosing a system that lets the school feel safe, in control and sustainable. Ask the ten questions above one by one — only a vendor that can answer and demonstrate every one deserves your grant and two years of teaching time.