Part of: Sovereign AI for Education Systems: A Practical Guide for Ministries, a guide to the decisions, infrastructure options and roadmap for ministries.
In October 2025, I was invited by UNICEF to design and deliver an AI in Education workshop for their Learning Pioneers Programme, bringing together education ministry leaders from 14 countries at the Innovation Camp in Helsinki. The session tackled a critical challenge: how can resource-constrained education systems harness AI’s potential without compromising cultural sovereignty, breaking budgets, or replacing educators?
The workshop addressed critical challenges facing education ministries deploying AI at scale – from infrastructure constraints to data sovereignty concerns. Through five live demonstrations and three hands-on activities, participants learned to build AI-assisted lesson planners, develop educator digital twins using context engineering techniques and explore infrastructure-appropriate solutions like WhatsApp chatbots for teacher support. A key focus was on sovereign pedagogy: ensuring AI tools respect and amplify local languages, cultural contexts, and curriculum frameworks rather than imposing Western educational models. I introduced UNICEF’s Learning Innovation team to open-source AI infrastructure frameworks that can reduce per-user costs from $16-20/month to $1-2/month while keeping sensitive education data on national servers.
This work reflects my deep interest in democratising AI in education: making cutting-edge tools accessible, affordable, and culturally appropriate for the Global South.
Workshop presentation here:
How the workshop was built
The session ran on 22 October 2025 and followed the same three-part structure I use in most workshops: instruction, demonstration, then practice. The aim was that every participant would leave having built something, not only having watched a presentation. The objectives ranged from AI fundamentals and global trends to prompting with free tools, context engineering, formative assessment, WhatsApp support for teachers and shareable lesson planners.
Part 1: what ministries need to decide
The first part covered how large language models work and why they hallucinate, the CRAFT approach to prompting (Context, Role, Action, Format, Tone), and a quick tour of how different countries are approaching AI in education. Four themes came up across every system we looked at:
- Teacher professional development comes before student tools.
- Student data protection is non-negotiable.
- Solutions have to be culturally adapted; one size does not fit all.
- Infrastructure matters, so solutions must match connectivity reality.
We then worked through the questions ministries actually have to answer: when and how students can use AI, what training teachers need, where data is stored and who can access it, how tools are vetted, how equal access is ensured, and how assessment changes. We also looked at the EU AI Act risk tiers, which place AI used for student assessment in the high-risk category, and at the UNESCO AI Competency Framework as a structured way to build capacity.
Why sovereignty matters
Sovereign pedagogy was the thread through the whole session. If every system uses the same general-purpose tools with the same defaults, there is a real risk of pedagogic flattening: one way of teaching, one cultural frame and one dominant language, applied everywhere. Context engineering is the practical answer. A ministry can write down its curriculum, languages, constraints and values in plain text files and make AI tools work inside that context, rather than adapting its classrooms to the tool.
Language support is uneven. In research on African language support that I shared in the session, the models we tested scored very differently, which is why I encourage every system to test tools in its own languages before adopting them. Open-source and on-device models, and infrastructure that keeps sensitive education data on national servers, give systems more control over cost and data.
Part 2: five demonstrations
- A text-file AI tutor that carries student context, for low-resource settings.
- A WhatsApp chatbot for teachers, built on curriculum documents, the annual teaching plan and textbooks, running on an open-source n8n workflow.
- An educator context file that tells any chatbot who the teacher is and what their classroom looks like.
- A lesson planner for a local curriculum, built from a template and context.
- A formative assessment assistant.
Part 3: hands-on practice
Participants built two things. First, an education system context file describing their own role, learners, languages, infrastructure and curriculum, which they then pasted into a chatbot to generate practical, constraint-aware plans. Second, a lesson planner Gem in Gemini that already knew that context, aligned to their curriculum and could be shared with colleagues. In the share-outs, participants extended the idea to tools for school leaders, district officials and policymakers.
What I took away
The main lesson is that the barrier is rarely access to AI; free tiers already exist. The harder work is deciding where humans add value, writing down local context so tools respect it, and choosing infrastructure that fits the connectivity and budgets systems actually have. That is also the starting point for the readiness work I describe in What 28 AI readiness assessments reveal about schools in Ghana.