Most Irish schools I talk to are stuck in the same place: a senior management team drowning in policy updates from the Department, a SET department buried in IEPs and PPCRs, and a principal who can't get a straight answer about which students are sliding before the Christmas tests. The technology in the building hasn't really moved on from VSware, Office 365, and a shared drive that nobody fully trusts. An "AI strategy" usually means somebody pasting marking schemes into ChatGPT and hoping the Data Protection Commissioner doesn't notice. That's the gap I've been working on.
Why generic AI doesn't fit an Irish school
The problem with using public LLMs in a school setting isn't that they're bad at the work — they're often quite good. The problem is the data path. When a teacher pastes a Junior Cycle CBA into a consumer chatbot, that prompt leaves the school, leaves the jurisdiction, and in most cases leaves the EEA. Under GDPR, and under the Department of Education's own guidance on processing student personal data, that's a controller-level decision being made by an individual teacher at 9pm on a Sunday. Schools have no audit trail, no DPIA, and no way to revoke it.
The second problem is context. A general-purpose model doesn't know your school's Code of Behaviour, your Critical Incident Plan, your DEIS plan, or the difference between a Stage 2 and Stage 3 SEN intervention under the Continuum of Support. It will confidently invent answers based on UK or US schooling assumptions — GCSEs, IEPs in the American sense, "grade levels" — none of which map cleanly onto Irish primary or secondary practice.
The third problem is procurement. A school principal cannot reasonably be expected to evaluate a SaaS vendor's sub-processor list, EU Standard Contractual Clauses, and Schrems II transfer impact assessment between supervising yard duty and chasing a leaking radiator in the prefab. So the practical state of school AI Ireland-wide is: nothing official, plenty of unofficial, and nobody quite sure what's allowed.
What an intelligence brain actually is
I use the term intelligence brain deliberately, because it's not a chatbot and it's not a "copilot". It's an on-premise or sovereign-cloud layer that sits inside the school's own boundary and ingests the documents, policies, and structured data the school already owns: VSware exports, Aladdin reports for primary, Department circulars, Croke Park hours documentation, board of management minutes, child protection records (with appropriate access controls), SEN files, and the school's own policy library.
The brain indexes that material into a private vector store, attaches a permission model that mirrors the school's actual roles (principal, deputy, year head, SET coordinator, class teacher, secretary), and exposes it through a single query interface. When a year head asks "show me the attendance pattern for 2B over the last six weeks against their CAT4 scores and flag anyone whose pattern has changed", the brain answers from the school's own data, with citations back to the source rows. Nothing leaves the building unless the school explicitly chooses to use a hosted model — and even then, only the minimum necessary tokens, with PII redaction in front.
That architecture is what makes it defensible under GDPR. The data controller (the board of management) keeps full control of the personal data. The processor relationship, if there is one at all, is bounded and contractual. You can produce a DPIA for it. You can show the DPC what's happening.
The technical stack, briefly
For schools that have asked, here's roughly what sits underneath. A small server — genuinely small, this runs on a single mid-range box in the comms room, not a data centre — hosts a quantised open-weights model in the 8B to 30B parameter range, depending on the school's hardware. For most secondary schools, a model in that range, properly fine-tuned on Irish curriculum content and policy language, outperforms a generic frontier model on the tasks that actually matter, because it isn't guessing about Irish context.
Around the model sits a retrieval layer (a vector database — usually pgvector or Qdrant, depending on what the school's IT provider is comfortable maintaining), a document ingestion pipeline that handles the messy reality of school documents (scanned PDFs, Word files from 2009, photos of handwritten notes), and a permissions layer that ties every query to a logged-in identity through the school's existing Microsoft 365 or Google Workspace tenant. Every query, every answer, every source citation is logged. If a parent submits a Subject Access Request, the school can produce a complete record of what the system did with that child's data.
Connectors to VSware and Aladdin are read-only. The brain doesn't write back into MIS. That's a deliberate choice — it keeps the system of record clean and means the brain can be removed tomorrow without leaving fingerprints in the school's official data.
Where it earns its keep day to day
The use cases I keep seeing land hardest in Irish secondary schools are unglamorous and high-volume.
- SEN documentation. Drafting Student Support Files, Stage 1/2/3 plans, and reviewing them against the school's existing IEP language for consistency. The SET coordinator stops being the single point of failure for every document.
- Subject department reviews. Pulling assessment data, attendance, and behaviour incidents into a single brief for a department meeting, instead of a teacher spending a Sunday in spreadsheets.
- Policy queries. "What does our anti-bullying policy say about online incidents that happen at the weekend?" answered in seconds, with the policy paragraph cited, instead of a deputy principal hunting through the shared drive.
- Department circular triage. When a circular drops, the brain summarises what's changed against the previous version and flags which school policies need updating. This alone saves a principal hours per term.
- Parent communication drafting. Letters that sound like the school sounds, not like a chatbot. The brain learns the principal's tone from the existing letter archive.
Primary schools use a thinner version of the same thing. The volume is lower, the document estate is smaller, but the principal-as-everything problem is worse. A primary school AI deployment that handles Aladdin queries, Department circular summarisation, and policy drafting is, in my experience, a bigger relative win than the equivalent in a 1,000-pupil post-primary.
What schools should be careful about
I'll be blunt. There are things this technology should not be doing in a school in 2025, and any vendor telling you otherwise is selling you a problem.
It should not be making decisions about students. Not subject level recommendations, not SEN diagnoses, not behavioural classifications. The model can summarise, draft, and surface — a human decides. This isn't a philosophical point, it's an Article 22 GDPR point about automated decision-making affecting individuals, and it's a child welfare point.
It should not be marking state exams or CBAs autonomously. It can support a teacher's marking, give a second opinion, check consistency across a class — fine. Pressing the button is the teacher's job, with the teacher's professional judgement, under the SEC's framework.
It should not be ingesting Tusla referrals, child protection concerns, or Designated Liaison Person records into a general-access index. Those documents need their own access control, their own audit trail, and in most cases shouldn't be in the brain at all. Keep them in the locked filing cabinet they belong in, physical or digital.
And it should not be deployed without the board of management understanding what it is. The board is the controller. They need to sign off on the DPIA, not the IT coordinator.
What this looks like for an Irish school AI rollout
A realistic deployment for a secondary school takes weeks, not months, when the school already has reasonable digital hygiene. The sequence I generally recommend: scope the document estate first, agree the permission model with the principal and deputy, run a DPIA, deploy the brain in read-only mode against policies and circulars only, prove it on low-risk queries for a half-term, then expand to MIS connectors and SEN workflows. Most of the value shows up in the first phase. The later phases compound it.
For schools that want to see the shape of this in more detail, I've written up the education-specific architecture and use cases at the intelligence brain for education, and the broader platform sits at the intelligence brain overview.
Where to start this week
If you're a principal or deputy reading this, do one thing this week: ask your staff, anonymously if you have to, what AI tools they're already using for school work. You'll be surprised. That list is your real starting point — not a strategy document, not a policy, but an honest map of where school data is currently going. Once you know that, the conversation about doing it properly, on your own terms, inside your own boundary, becomes a much easier one to have with your board.