The Real AI Bottleneck in Organisations Is Memory

Part of Expertise Encoding. This essay sets out the problem that method is designed to solve: organisations that store plenty of information but cannot remember their own work in a form AI can use.

A team leaves a meeting with a clear decision. Three weeks later that decision has gone soft. One person remembers the direction it was heading, someone else remembers the exception that got carved out, the document has moved folders twice, and the colleague who understood the trade-off has rolled onto a different project. When the question comes back round, the organisation does the only thing it knows how to do. It asks a few people to rebuild the context from memory.

That is usually the point where AI gets invited in. Someone opens a chat window and asks for help. But the assistant can’t do anything useful until a person has explained the project, dug out the document, recapped the decision, named the people involved, flagged the deadline and spelled out why any of it matters. By the time you have finished that briefing, you could very nearly have answered the question yourself.

None of this is an intelligence problem. The model would handle the question fine. What it can’t do is reach into an organisation whose memory was never kept in a usable state. In my experience that, far more than model choice or prompt skill, is why so much workplace AI stays impressive in demos and oddly weak inside real work.

AI sits beside the work, not inside it

Most of the teams I work with are experimenting with AI. People are trying ChatGPT, Copilot, meeting summarisers, document assistants and a handful of internal pilots. Some of it genuinely helps. A lot of it sits to one side of the actual work: useful in the moment and forgotten by the next meeting.

I used to think the gap was about better tools or better prompts. I have stopped believing that. The tools are already good, and they keep getting better. The problem is that they do not know enough about the work to be trusted with it.

What people need, most days, is not a sharper answer to a prompt. They need to know what was agreed in the last meeting, which version of the document is current, what the client is actually expecting, which deadlines are live this week, which decision is still open, and what needs their attention today. That is not a prompting problem. It is a context problem.

This is also why so many AI pilots stall after the demo. The demo impresses because it shows a model doing something clever in isolation. The pilot disappoints because the cleverness never connects to the surrounding work. The model answered the prompt, but it did not know the project, the history, the people or the stakes.

Storage is not memory

Here is a distinction a lot of organisations have quietly stepped over. Storage is where information ends up. Memory is the part you can pull back out, make sense of and act on at the moment it counts.

Almost everyone already has plenty of storage: SharePoint, Google Drive, Teams, Slack, email, Notion, the CRM, shared project folders, and now a fast-growing heap of meeting transcripts. None of these is short of files. What they are short of is memory.

Storage will happily hold ten versions of a document. It won’t tell you which one is current, what moved between version six and version nine, or why anyone bothered. It keeps the recording of the meeting and quietly loses the decision that came out of it. The information is technically there and practically out of reach.

So when the questions that actually move work forward come up, storage has nothing to offer:

  • What matters now?
  • Which decision is the current one, and why did we go that way?
  • What has changed since the last version?
  • Who owns the next step?
  • What should be back on someone’s desk today?

People answer these by remembering, by messaging a colleague, or by scrolling back through a channel until they give up and guess. That informal human memory carries far more of the load than most leaders realise. It is also brittle. It walks out of the door when someone resigns, goes dark when they are on leave, and frays a little every week even when nobody leaves.

Why memory is the bottleneck

AI becomes genuinely useful the moment it can reach context that is structured, current and trusted. Take that away and you are back with the workarounds we already lean on: the long prompt, the manual context dump at the start of every session, one person’s recollection standing in for the whole organisation’s.

That is why so much prompt engineering is really compensation. A long prompt is usually a symptom of a weak layer underneath it. We paste the background in because the system doesn’t already know the work. Every time someone rebuilds a project history from scratch to get one useful answer, the missing memory becomes visible. The effort hasn’t gone anywhere. It has been shifted back onto a person, every single time. I made a related argument about moving from prompt engineering to context engineering; this is the organisational version of the same point.

Prompting won’t fix it. You can write the most elegant instruction imaginable, but if the system has no idea which version of the proposal is live, who signed off the change of scope, or why the timeline slipped in March, you will get a confident answer resting on nothing. A confident answer resting on nothing is worse than no answer, because somebody downstream will act on it.

The wider evidence points the same way, with sensible caveats. Ofsted’s June 2025 study of early AI adopters in English schools and further education colleges, The biggest risk is doing nothing, interviewed leaders from 21 schools and colleges. The ones getting value had moved beyond one-off experiments into practice with clear leadership, approved tools, time for staff to learn and human oversight. That is an education finding, not a universal business rule, but the pattern travels. AI pays off where the surrounding work is organised enough to trust it. Where context is scattered, the cleverness has nowhere to land.

There is a parallel on the engineering side. A May 2026 survey by Xuying Ning and colleagues, Code as Agent Harness, frames durable agent value as coming from executable, verifiable and stateful systems, with planning, memory and tool use treated as core infrastructure rather than add-ons. In plain terms: systems that hold state, remember and can act reliably over time. Same instinct, approached from the other direction. The chat box is the interface. The layer underneath is where the value compounds.

Organisational working memory

The thing that is missing deserves a name. I call it organisational working memory: the live layer of context that helps people, and AI, keep hold of what is current, what has been decided, what needs attention and what has just changed.

In practice it draws on parts of the work that already exist but rarely talk to each other: live projects, the handful of documents that genuinely matter, the trail of decisions, meeting outcomes, open questions, deadlines, owners, dependencies and the obligations that come round every quarter. The aim is not to hoard all of it. It is to keep the parts that matter within reach at the moment they are needed.

This is where a knowledge base and working memory part ways. A knowledge base is somewhere you go and search, on the days you remember it exists. Working memory surfaces what needs attention today, before you have thought to ask. Think of it less as a bigger archive and more as something that puts the right context in front of the right person.

The missing operating layer

If you accept that framing, the priority shifts. The useful question is no longer “which AI tool should we buy?” It becomes “what does the AI need to know to be useful here?”

The honest answer is mostly the boring, structural material that organisations already half track in scattered places: which projects are live, who is responsible, what has been decided and why, which documents matter, what the deadlines are, and what needs attention next. Put together, that is an operating layer.

Here is the idea in one line. An AI operating layer is the persistent context and workflow infrastructure that sits between AI tools and the real work, so that the AI understands what is happening, what matters and what to do next.

It is less glamorous than a new model release and far more useful. It means your assistant can answer “what did we decide about the delivery timeline?” because the decision was captured when it was made. It means “what is overdue this week?” returns a real list, because tasks and deadlines live somewhere the system can read. It means a document doesn’t vanish into a folder, because the system knows what it is, which project it belongs to and when it was last reviewed.

This is not automation in the old sense of replacing a task. It is closer to giving the organisation a working memory and then letting AI operate against that memory, under human judgement. The tools answer questions. The operating layer makes sure they are answering them about the right things. On the Expertise Encoding page I describe this as four layers: the work as it exists, a knowledge layer that structures it, an operating layer of skills, retrieval and quality gates, and only then reusable AI capability.

What I am testing in NiallOS

I test this on my own work first, through NiallOS, which I treat as a working laboratory rather than a product. There is nothing to buy, and I am not claiming everyone needs my setup. It holds projects, tasks, decisions and contacts in a structured database, takes documents in and turns them into extracts that can be checked against the source, and surfaces what needs attention each day. AI agents work as bounded helpers in the background rather than as a single chat box I have to remember to open.

One lesson has held up as the system has changed. The structural work, retrieval, scoring and checking, increasingly runs as plain deterministic code, and a model is called only where a task genuinely needs judgement or language. That keeps the memory dependable and makes the model the replaceable part, which is the argument of Own the Memory, Rent the Intelligence. The origin story is in Why I Built an AI Twin to Run My Life.

Plenty of it is rough, and some parts I have rebuilt more than once. But the narrow claim holds: better organisational memory makes AI both more useful and more trustworthy, and its absence explains a good number of the pilots that quietly stall.

Building it has also been a lesson in fluency. Rick Dakan and Joseph Feller’s 4D Framework for AI Fluency (Delegation, Description, Discernment and Diligence) puts judgement and responsible practice ahead of the tool list. A working memory layer is where those habits become routine: what to hand to AI, what context to give it, how to check the result, and who stays accountable.

Questions for leaders

If you want to find your own memory problem before you sign off on another tool, these questions tend to expose it quickly:

  • Where does the organisation actually remember decisions, as opposed to storing files?
  • What context do people keep having to explain from scratch?
  • Which documents, meetings and decisions should be wired into live work, and currently aren’t?
  • Could an AI answer a question about a current project without someone rebuilding the background first?
  • If a key person left tomorrow, which decisions would leave with them?

The honest answers are usually uncomfortable, which is rather the point. They tell you what an AI would need to know before you could trust it to help.

Where this goes next

The practical takeaway is not “go and build a personal operating system”. It is to find out where your institutional memory is kept and whether an AI could reach it if you asked. For most organisations the answer is that memory is scattered across inboxes, drives, chat threads and individual heads. That is the real constraint, and it is fixable.

Connecting AI to project context, decision trails, meeting notes, documents and actions is unglamorous work. It is also where durable value sits, well ahead of another tool licence. If the first wave of workplace AI was about access to intelligence, the next will be about access to context. The organisations that benefit most won’t be the ones with the most tools. They will be the ones that make their memory usable, so that people and AI are finally working from the same version of the work.

If you are working through this in an education organisation or a professional team, Expertise Encoding sets out the method I use, and you can work with me on putting it in place.

Most AI Training Teaches Teachers to Prompt. We Taught Them to Build

Participants working on laptops during a KFAS Future Educator workshop

Part of How to Design AI Professional Development That Changes Practice. This post is a worked example of the Create level of UNESCO’s AI competency framework for teachers, drawn from the KFAS Future Educator programme in Kuwait.

Everyone is teaching teachers to use AI. Almost nobody is teaching them to build with it.

That distinction matters. Using AI usually means typing a prompt into a chat tool and getting a response. Building with AI means designing a system where the model is one step in a larger process: pull data from a spreadsheet, generate content against a rubric, run a quality check, and deliver the output to a classroom channel, automatically, on a schedule.

UNESCO’s AI competency framework for teachers defines three progression levels: Acquire, Deepen and Create. Most teacher AI training stops at the first two, learning to use tools and integrating them thoughtfully into practice. The Create level asks for creative configuration of AI systems, including combining or modifying open-source and customisable toolkits for local educational challenges. In practice, I have seen almost no programmes that take educators there.

In Kuwait, on the KFAS Future Educator programme, we tried. This post explains what we built, the teaching sequence that made it work, what participants produced in two hours, and what I would change next time. The case study for the wider programme is on Future Educator: AI for STEM teachers. The design principles sit in the AI professional development guide.

The ask

In November 2025, participants asked for something specific: more technical sessions. They wanted to move beyond prompting and explore what happens when AI becomes part of a larger system. We responded with a workshop focused entirely on building working AI automation workflows.

We used n8n, an open-source automation platform with a visual interface. You drag nodes onto a canvas and connect them. Each node does something specific: read data from a spreadsheet, call a model to generate content, send an email, save to a document. The power is in the connections. It is not what any single node does, but what happens when you chain them together.

We ran separate two-hour sessions for primary, secondary and higher education groups. Each worked with workflows designed for their teaching context:

  • Primary educators explored a differentiated reading comprehension generator.
  • Secondary educators worked with an exit ticket generator that produces both student and teacher versions at the same time.
  • Higher education instructors tackled batch essay feedback: a set of student submissions processed against structured rubric criteria, with feedback written back to a spreadsheet and a class-wide patterns report generated automatically.

Within two hours, educators who had never seen a workflow tool were designing and running their own AI-powered systems.

Explore first, then build

The approach that made this work was not the tool. It was the teaching sequence.

Each group received two versions of their workflow: a full version that was production-ready, with quality loops and multiple outputs, and a starter version simplified to four or five core nodes. We showed the full workflow first. Participants then explored it through a structured scavenger hunt: where does the data come from? How does it flow through AI processing? Where are the quality checks? What are the outputs?

Only after understanding the complex version did they switch to the starter workflow and begin modifying it themselves. Primary educators added a difficulty dropdown. Secondary educators added a quality check loop. Higher education instructors added new rubric criteria.

This sequence, study the expert version, then modify the simple version, is deliberate. It draws on the worked example effect from cognitive load theory: for novices, studying a complete worked example before attempting their own work produces better learning and transfer than jumping straight into problem-solving. The cognitive load of understanding a new tool, a new paradigm and a new task all at once is too high. Separating “understand the pattern” from “apply the pattern” made each step manageable.

Exploring the full workflow built understanding. Modifying the starter workflow built confidence. Participants left with both: a professional tool they could use the next day, and the skills to customise it.

What participants built

The hands-on session produced workflows that went well beyond our templates.

A primary teacher created a workflow triggered weekly to generate a curriculum-aligned, locally contextualised story for her class. It pulls from her curriculum standards, generates age-appropriate content set in familiar Kuwaiti contexts, and automatically posts to her Microsoft Teams classroom channel every Sunday evening.

A secondary teacher built a daily research briefing: the workflow searches for recent news on topics she teaches, filters for relevance, generates a personalised summary, and emails it to her each morning before school.

Both workflows took shape within the 30-minute hands-on period. Neither teacher had used n8n before that morning.

One group discussed automating parent communications in both Arabic and English. Another explored generating differentiated homework based on that day’s lesson performance. The ideas went further than what we had prepared. Participants were not asking whether they could do things, but how.

Why this matters beyond the workshop

There is a widening gap between what the frameworks call for and what teacher training actually delivers.

UNESCO’s Create level is explicit about customising and modifying AI toolkits for local challenges. The OECD and European Commission’s AILit Framework includes a Manage AI domain: decide whether AI is appropriate for a task, choose an approach, decompose a problem so that some steps are automated and others stay with the person, and monitor the result. That is workflow design described in policy language.

The problem is not that educators lack the ability to work with AI systems. It is that we are not teaching them how. We have collectively decided that “AI training for teachers” means prompt engineering and stopped there. Prompting is a useful skill. But it keeps educators in a reactive loop: type a request, get a response, evaluate, repeat. Workflow design is different. It asks: what process do I want to automate? What inputs does it need? What checks should it include? Where should the output go? That is systems thinking, not just tool use.

The speed at which participants in Kuwait moved from “I have never seen this tool” to “here is how I would automate my weekly lesson prep” suggests the barrier is not technical aptitude. It is exposure. Give educators a visual tool, a clear pedagogical sequence and a realistic use case, and they will build things that surprise you.

What I would do differently

The starter workflows were probably too simple. We had worried about technical overload, but participants consistently wanted to add more features: additional quality checks, extra outputs, integration with other school systems. Next time, I would pitch the starter version slightly higher and see what happens.

I would also add a collaborative design phase. The strongest ideas emerged from table conversations, not individual work. Building in structured time for pairs to design a workflow on paper before touching the tool would likely produce even better results.

And I would connect the workshop more explicitly to the rest of the programme. In Kuwait we already used a simple Human-AI Partnership pattern (Define, Draft, Refine, Validate, Deliver) so that review sat inside every classroom workflow. The automation session works best when that judgement loop is already familiar, so teachers treat quality checks as part of the design rather than an optional extra.

What this means for AI training design

This workshop reinforced something I keep coming back to in the professional development guide: educators do not need to be convinced that AI is useful. They need structured pathways to move from using it to building with it.

The two-workflow approach, a professional version to study and a simplified version to modify, provides that pathway. It works because it respects both the complexity of the tool and the intelligence of the learner. It also matches Design principle 3 in the guide: practise on real tasks teachers already own, rather than toy demos.

The question I am sitting with now is about scale. This worked with educators in a well-resourced workshop setting. What does it look like when a school or ministry tries to build this capacity internally, without an external facilitator? That is the design challenge I am working on next, and it is why the train-the-trainer path in the Kuwait case study matters as much as the workshop itself.

If you are designing AI professional development and want to push past prompting into Create-level practice, start with one real workflow your teachers already need, show them a finished version, then give them a simpler one to change. The tool can be n8n or something else. The sequence is the point.

Related reading: How to Design AI Professional Development That Changes Practice, Future Educator: AI for STEM teachers, AI in Education: Lessons from Kuwait, and Designing a 24-week AI training programme.

Own the Memory, Rent the Intelligence

Part of Sovereign AI for Education Systems: A Practical Guide for Ministries. This post applies the same control question to any organisation choosing an AI vendor.

Twice this year I have sat with the strategy team of a Big Four firm and heard the question their clients keep asking: which AI model should we standardise on? GPT, Claude or Gemini? A US model or a Chinese one? Closed or open-weight?

I hear the same question from ministries, assessment bodies and publishers across Africa and the Gulf. Model choice matters. Different models bring different capabilities, costs, risks and constraints. But it is the wrong first decision.

The assets that compound over time are the memory you accumulate and the methods you build around the model. Those should remain yours. The model should be replaceable. This post sets out why, what the three layers are, what it looks like in my own work, and a three-question test you can run on any AI deployment this week.

What you are really buying when you “adopt AI”

When an organisation, or an individual, adopts AI, it is usually acquiring three different things at once. Most products bundle them so tightly that the boundaries disappear. That is convenient at the start. It also makes it hard to see where the lasting value sits, and what you would lose if you left.

1. The memory layer

This holds the records, documents, decisions, correspondence and history of how you or your organisation works. It also includes the structure around those records: relationships, provenance, metadata and version history.

Everyone agrees they should control their data. Many assume they already do because a product offers a CSV export. But a pile of rows is not a working memory. If the relationships, history and structure disappear on export, much of the useful context disappears with them.

2. The method layer

This is the layer that gets the least attention, and it is where your working methods become explicit and executable. It includes the schemas that structure information, the workflows that move it, the prompts and rules that guide the model, the permissions and review gates that constrain it, and the evaluations that tell you whether it is working. It also includes the routing logic that decides which task goes to a model and which needs a person.

When a team spends six months getting an AI assistant to handle intake properly, the value is not only in the model or the records. It sits in the method the team has worked out: what to collect, how to classify it, which exceptions matter, when to escalate and how to check the result. That method is often the least portable part of an AI deployment. It is also the part I care most about, because it is where professional judgement lives. I write about capturing it in Expertise Encoding.

3. The intelligence layer

This is the model itself, the part everyone argues about: which lab, which benchmark, which release. It is also the layer that changes fastest. A model that leads this quarter may be overtaken next quarter, and prices move just as quickly.

What ownership means in practice

For me, ownership is practical, not legal. Can I inspect it, change it, version it, export it intact and move it without asking a vendor for permission? You can own the method while renting the infrastructure it runs on. Control matters at the points where your judgement and institutional memory build up.

Build inside a vendor’s walls and your workflows may be written in its configuration language. Your automations may live in its runtime. Your prompts, review steps and accumulated judgement may exist only inside its interface. You may be able to export the records, but not take the method with you in a form another system can run.

This is no longer only an architecture preference. The European Union’s Data Act, which has applied since 12 September 2025, includes rules to help customers switch between providers of data processing services and to improve interoperability. Portability has become a policy question.

Renting more of the stack can still be the right call. Managed platforms provide support, security, integrations and reliability. Building your own method layer brings maintenance, versioning and key-person risks, and managed dependency can be worth paying for. The risk is letting that dependency deepen without noticing it, testing it or pricing the exit.

What this looks like on my desk

I have spent the past year running a small version of this architecture across my own work, in a system I call NiallOS. My working memory sits in a structured database on infrastructure I control. It holds projects, documents, extracted evidence, decisions, commitments, contacts and journal entries. Around it sits a set of scripts, plain-text instructions and scheduled routines. A morning brief is ready before my working day starts. Meeting context is prepared before I ask. New research in my field is scanned each week. At the end of a working session, the relevant updates are written back into the system.

No single part is the breakthrough. The value comes from every routine sharing the same memory and the same methods.

Since I built it, I have changed the underlying models several times. I have used US frontier models, Chinese open-weight models routed through an aggregator, and different models for different roles depending on cost and capability. Changing the provider connection took minutes. No records had to move. The core workflows did not have to be rebuilt. The morning brief arrived the next day as usual, sometimes slightly smarter and sometimes slightly cheaper.

That does not mean a new model is automatically safe for every task. Models behave differently, and prompts, structured outputs, tool use and safety behaviour all need testing. The US National Institute of Standards and Technology’s AI Risk Management Framework treats evaluation as contextual and ongoing, which is the right expectation. A replaceable model still has to earn its place in a live workflow. What the architecture changes is the cost of that evaluation: I can test a different model against the same records and methods without rebuilding the system around it first.

Some models reach my system through an aggregator, a service that routes requests between providers. That is another dependency, but the records and methods still live elsewhere, and the route can be replaced.

This is the operational fact underneath the argument. The model provider can change. The memory and the method keep compounding. Compare a team whose prompts, automations and review steps live only inside one vendor’s assistant. For them, changing provider means more than choosing a new model. They have to reconstruct how the work was being done.

When control has to extend into the model

For many individuals and organisations the pattern is simple. Hold the memory and method somewhere you control, then call the model through an interface you can replace.

Some institutions need to move the control boundary further. I work with ministries, regulators and assessment bodies in the Gulf and across Africa. In some of these settings, law, data classification policy, contracts or institutional risk appetite require certain learner or citizen data to stay in-country or on premises. The intelligence has to come to the data, rather than the data going out to the intelligence.

That can mean running open-weight models on infrastructure the institution controls. This gives more operational control over the model layer, but it is not complete sovereignty. Licences, hardware, operators, security and update paths still matter. The sovereign AI guide for ministries covers those infrastructure choices in more detail.

The capability picture has also shifted. Stanford’s 2026 AI Index reported that in March 2026 the leading closed model was 3.3% ahead of the leading open-weight model on the Arena leaderboard. That does not make every open model suitable for every task, and running frontier-scale models locally still needs serious infrastructure and skills. It does mean open-weight models are now credible options for many workloads where local control matters.

The layers can be separated, but only by design. You have to draw the boundaries, write down the dependencies that remain and test whether switching actually works.

What this means for education systems

Education has a particular reason to care. A ministry’s curriculum, assessment frameworks, marking guidance and years of learner records are exactly the kind of memory that compounds. The way its officials moderate marks, review content or support teachers is exactly the kind of method that should not live only inside a vendor’s product. Procurement cycles are long, contracts outlast model generations, and the people who negotiated a deal have often moved on by the time it needs renewing.

So when an education organisation buys an AI service, I would want four things written into the agreement or the design:

  1. The right to export records together with their structure, relationships and history, in a documented format, at any time and at the end of the contract.
  2. Prompts, workflow definitions, rubrics and evaluation sets held in files or systems the organisation governs, not only in the vendor’s interface.
  3. A model interface that can be pointed at a different provider, or a locally hosted model, without rebuilding the service.
  4. A tested exit plan: what would be moved, retested and approved again, by whom, and how long it would take.

None of this stops a ministry from using commercial tools. It means the ministry chooses the dependency rather than discovering it later.

A three-question portability test

When the “which model?” conversation starts, in a boardroom, a team meeting or your own head, start with three questions instead.

  1. Where does our memory live? Could we take the records, relationships, history and structure with us tomorrow?
  2. Where is our method written down? Is it held in files and systems we govern, or only in a vendor’s configuration screens?
  3. If we changed model provider next month, what would we have to rebuild, retest and approve again?

If the answers are “inside their product”, “inside their product” and “almost everything”, you have made a deep platform commitment. That may still be the right commercial choice. But it should be explicit, bounded and revisited as the market changes.

You do not need to self-host everything to apply the principle. A one-person system can use a structured database, plain-text workflows and a thin routing layer. An institutional version will also need identity and access controls, audit logs, monitoring, evaluation, backups, disaster recovery and clear procurement terms. The scale changes. The ownership test does not.

Try the test on one live workflow. The parts that break will show you where the dependency sits.

Own the memory, own the method, rent the intelligence

Model capability will keep moving, and that is good news if you are set up to benefit from it. The method you develop and the memory you accumulate should still be there when it does.

If you are working through these choices for a ministry, an assessment body or an education organisation, see Work with me or get in touch. For the wider policy picture, start with Sovereign AI for Education Systems, and for how I capture working methods so AI can use them, see Expertise Encoding.

Start With Governance, Not Technology: Where Schools Should Begin With AI

Part of AI Readiness for Schools: A Practical Framework for African and Global South Contexts. This post covers the first dimension, leadership and governance.

When school leaders ask me where to start with AI, they almost always ask about technology. Which tools should we use? Should we pay for ChatGPT or Gemini? Which platform is best for our learners?

These are fair questions, but they are the wrong first questions. The schools making the best decisions about AI did not start by choosing tools. They started by deciding who is responsible, what the rules are and how decisions get made. They started with governance. This post sets out what that means in practice, why it comes first, and three things any school can put in place within a week, at no cost.

The tool trap

The tool-first approach usually looks like this. A school buys a licence for an AI writing assistant. A few teachers try it. Some love it, some ignore it, some are sceptical. Learners are already using ChatGPT on their phones anyway. A parent complains about AI-written homework, a head of department asks what the school’s position is, and nobody has an answer, because nobody was asked to develop one.

This is happening in schools everywhere. In a 2023 UNESCO survey of more than 450 schools and universities, fewer than 10% had formal guidance on generative AI. Meanwhile College Board research found that 84% of high school students were using generative AI for schoolwork by mid-2025. In England, a 2025 Sutton Trust survey found that only 9% of state school leaders had an official AI strategy.

That gap is not a technology gap. It is a governance gap.

What the Ghana results showed

I have since seen the same pattern in data of my own. At Cambridge International Schools Day in Accra in February 2026, I asked a room of about 200 school leaders how many had a written AI policy. A handful of hands went up. In the workshops that followed, 24 schools completed the AI readiness self-assessment, giving 28 responses.

  • Infrastructure and Resources was the strongest dimension, at 58.8%.
  • Leadership and Governance averaged 41.2%, and Policy and Ethics 35.4%, the two weakest.
  • 89% of schools had one of those two as their weakest dimension.
  • Even the highest-scoring school, at 84% overall, scored lowest on Policy and Ethics.

Schools had connectivity, devices and teachers experimenting. What they lacked was someone in charge and a set of rules. The full analysis is in What 28 AI readiness assessments reveal about schools in Ghana.

What governance means here

I am not talking about bureaucracy. You do not need a committee that meets once a term and produces a 40-page policy. School leaders are too busy for that, and the technology moves too fast for it. Governance, at school level, means three things.

1. Someone is responsible

Who in your school is accountable for AI decisions? Not “interested in AI” or “good with technology”. Accountable. If a parent emails tomorrow asking what your position is on learners using AI for coursework, who replies?

In most schools the honest answer is that nobody has been asked. That is the first gap to close. Name an AI lead. It can be a part-time responsibility for a deputy principal, a curriculum coordinator or a confident head of department. The title matters less than the clarity: this person owns the question.

2. AI is on the agenda

Make AI a standing item at leadership meetings, not a one-off discussion. Five minutes is enough. What tools are staff using? What are parents asking? What are learners doing? What needs deciding?

This sounds small. It is not. A recurring agenda item signals that the issue matters, and it gives the conversation a place to happen before a crisis forces it.

3. There is a process for decisions

When someone wants to use a new AI tool, what happens? Who looks at it, against what criteria, and who says yes? In most schools individual teachers make these choices one tool at a time with no coordination. That is how learner data ends up in six different AI platforms, none of which has been checked for privacy.

A one-page process fixes this. A teacher proposes a tool, the AI lead reviews it, and the answer is approved, approved with conditions, or not approved. The review can be five questions:

  1. What problem does it solve for teachers or learners?
  2. What data does it collect, where is that data stored, and does that meet your country’s data protection law (POPIA in South Africa, the Data Protection Act in Ghana or Kenya, for example)?
  3. Is there an age limit in the tool’s own terms, and do your learners meet it?
  4. Does it work on the devices and connectivity you actually have, including phones and patchy data?
  5. What does it cost now, and what will it cost when the free tier ends?

Why this comes before technology

These three moves cost nothing. No licences, no hardware, no consultants. A school can do all three within a week, and they improve every AI decision that follows. You stop reacting and start leading.

I have seen schools with limited infrastructure make excellent AI decisions, and well-resourced schools with a device for every learner struggle because nobody had clarified who was responsible. Infrastructure matters, but it is rarely the main constraint. Clarity usually is. This matters especially in African schools, where budgets are tight and a tool bought without a plan is money that cannot be spent twice.

Governance builds teacher confidence

There is a second, less obvious reason to start here. Governance builds confidence. A 2025 Gallup and Walton Family Foundation survey found that 68% of US teachers had received no training on AI tools, and the gap is likely wider in many of the schools I work with. A compulsory training day does not close that gap on its own. It closes when teachers feel supported: when they know who to ask, and when clear boundaries make trying things feel safe rather than risky.

“Try this tool, within these guidelines, and talk to the AI lead if you are unsure” is a very different message from “AI exists and we have not decided what to do about it yet”. Confident teachers make better decisions about AI than compliant ones. For how to build that capacity once the ground rules exist, see How to Design AI Professional Development That Changes Practice.

From governance to policy

Governance is the first dimension of the readiness framework, and policy and ethics is the second. Once someone owns AI and there is a decision process, that person’s first job is to write the rules down: a short acceptable use statement for staff and learners, guidance on academic integrity, and a position on learner data. Assessment comes soon after, because the first real test of a policy is usually a piece of AI-assisted coursework. Assessment in the AI Era covers that step.

The same logic applies above the school. For a school group, district or ministry, governance means deciding who owns AI policy across the system, how tools are procured, and where learner data is hosted. Sovereign AI for Education Systems looks at those choices for ministries.

What to do this week

Before your next leadership meeting, answer three questions.

  1. Who is your AI lead? Put a name to it, and tell them.
  2. When will AI next be on the leadership agenda? Put a date on it.
  3. What happens when a teacher wants to use a new AI tool? Write down the process, even if it is three steps.

You do not need a full AI strategy, a chosen platform or a budget to begin. You need someone responsible, a regular conversation and a way to make decisions. The rest follows.

To see where your school stands across all five dimensions, take the five-minute AI readiness assessment or read the full framework. I run readiness workshops and leadership sessions for schools, school groups and education organisations; see Speaking or get in touch.

AI in education: practical lessons from implementation

Niall McNulty giving the PAIDEIA keynote in Buenos Aires, with a slide on readiness across the system

On 13 October 2026 I gave a keynote at the PAIDEIA International Conference in Buenos Aires, with UNICEF Argentina, to government representatives from all 24 of Argentina’s jurisdictions. The question behind the talk is one I hear most weeks: where should we start with AI in education?

The honest answer is that AI is already in education. Teachers, students and parents are using new tools in their own ways. Schools are building their own approaches, and governments are running pilots alongside system-wide initiatives. The question now is how to make that use useful, responsible and sustainable.

I am a strong advocate for AI in education, and I use it in my own work every week. Used well, it could help us understand each learner more deeply, give teachers targeted support and give leaders better information for decisions. The potential is greatest for learners who have not been served well by traditional methods. But none of that is guaranteed. In many cases, the evidence from classrooms simply is not there yet, although there are early signs of value in back-office work.

The PISA 2025 results point the same way. Students who used AI to draft their writing assignments did less well than those who did not. But frequent users did better when they used AI to help them learn and had also been taught to judge the quality of what it gives them. The teaching and judgement around the tool matter as much as the tool.

Here are four lessons from my work with leaders and educators at Cambridge, and what they mean for a federal system.

1. Start with the educational need

The first thing we do in the HP Cambridge Partnership for Education EdTech Fellowship is step back from the technology. Leaders arrive with limited resources and a huge, growing choice of tools. We put the brochures and demo videos to one side and ask what needs to improve in their system.

Learners needing more timely feedback, a ministry needing clearer data on student progress and teachers leaving because their days are overwhelming are different problems. Each needs its own diagnosis, including its root cause, informed by data and by the views of teachers and students.

Only then do we ask how AI could help, and whether a simpler, cheaper or more established approach would do better. Better teaching materials, a change in workflow, more professional development or stronger coordination may be the right answer. AI should earn its place. We also ask what must stay human-led: decisions about a learner’s progression, wellbeing, assessment or access to support need meaningful human responsibility.

Finally, define success before you start. Agree what evidence would lead you to continue, adapt, pause or stop, and put in place the systems to collect and review it. That gives leaders a far stronger basis for procurement than a list of features. The AI readiness framework for schools uses the same logic at school level.

2. Make pilots part of a learning system

A good pilot is small enough to reveal problems and serious enough to produce meaningful evidence. A demonstration in ideal conditions may show that a product works technically. It tells us much less about whether schools across a system can use it well. I want to see approaches tested in schools that are not yet getting strong results, that struggle with teacher retention or that serve more vulnerable communities. Otherwise we widen the gaps we are trying to close.

Slide: pilots as part of a learning system, from diagnosing the need to deciding whether to proceed, adapt, pause or stop
From the PAIDEIA keynote, Buenos Aires, 13 October 2026

In one pilot we are working on, the evaluation asks whether teachers plan deliberately for the roles of the teacher, the learner and AI, and whether students develop stronger critical thinking, curiosity and metacognition. It looks at how workload changes and whether wellbeing is protected, and compares participating classes with classes continuing their current approach. Login counts and time in the system cannot answer those questions. Did learning improve? Was the use accessible? Were there differences between learner groups? What did it really cost to implement well?

Every pilot should end in a genuine choice: proceed, adapt, pause or stop. Scaling up is one possible path, not the default. Evidence should change the next decision, rather than justify the last one.

3. Treat human judgement as infrastructure

Cambridge’s Prompt the future: Humans at the heart of education argues that people bring what AI cannot: knowledge of their context, ethical responsibility, relationships, reflection and the ability to judge when the right answer is not obvious. That judgement decides where AI adds value, how its output is read and when it should not be used at all.

So the question is not only whether a tool works. It is whether the people using it have the knowledge, authority and confidence to question it, adapt its use and stay accountable for the outcome.

I saw this in the Future Educator programme with the Kuwait Foundation for the Advancement of Sciences, where Cambridge and Educate Ventures Research worked with about 50 STEM educators. Teachers moved from exploratory or efficiency-driven use of AI towards more deliberate design. One described the change as moving from curiosity to clarity. A mathematics teacher needed students to tell two concepts apart, used one AI tool to create questions and another to turn them into a game. The game was not the point. The teacher had found a precise need, chosen tools for different jobs and stayed responsible for checking the result. There is more on designing this kind of training in AI professional development that changes practice.

Closer to Argentina, Uruguay’s Plan Ceibal is a public digital ecosystem, and Cambridge is at the start of a teacher-development programme there, again with Educate Ventures Research. It is too early for results, but the design choice is worth borrowing: build teacher judgement inside a system that already exists.

Design matters just as much when students use AI. In a science task, students first draw a process in their own words. A tool may then suggest a missing step, and each student must accept or reject it and say why. If the tool writes the explanation, you have assessed the tool. If the student judges the suggestion, you have assessed the student. I write more about this in assessment in the AI era.

Cambridge’s Shaping AI Policy and Practice in English Language Teaching recommends involving teachers and learners in the design, procurement and evaluation of AI tools. We have found the same across subjects and stages. Teachers also need permission to adapt or decline a use when their judgement requires it.

4. Leaders must support, and review with teeth

If we expect people to exercise judgement, leaders must create the conditions: approved tools and uses, clear expectations, time to learn, technical support and safe routes for raising concerns. That applies to back-office staff as much as to teachers.

Slide: leader support and institutional review, with govern and enable, build capability, teach learn and protect, and assure and improve
From the PAIDEIA keynote, Buenos Aires, 13 October 2026

Review is not the last administrative step after adoption. It informs the next decision. If a supplier changes how a tool handles data, adds a new AI feature, or evidence shows it is unsafe or ineffective, institutions need the contractual and practical ability to restrict, suspend or withdraw it. Adoption should never be irreversible.

Coherence without uniformity

For a federal system like Argentina, the challenge is how jurisdictions can learn together while responding to different priorities, infrastructure and communities.

A common national spine can set out educational purpose, rights, minimum safeguards, shared competency expectations and common evidence questions. Jurisdictions then build their own pathways within it. The spine does not need to prescribe one product. A shared commitment to safe, human-led use might mean younger learners meet AI only through an adult, while older learners use approved tools under defined conditions, and higher education and vocational settings set expectations suited to their disciplines.

I have seen this work in the Digital Education Leadership Programme, co-developed by Cambridge, UNICEF and the Global Partnership for Education. It brought together more than 40 senior ministry officials and UNICEF staff from 17 countries. Delegates applied shared frameworks to their own contexts through workshops, peer exchange and a practical capstone, and kept responsibility for local adaptation. Some used total-cost-of-ownership thinking to weigh maintenance, training and infrastructure alongside the purchase price. Others strengthened monitoring and evaluation, teacher development or national roadmaps. Diversity was treated as a source of evidence, not a barrier. (More on my work on the programme.)

The goal is not identical activity. It is dependable quality, local ownership and a way for learning to travel across the system.

Three pages to start on Monday

If I were working with one jurisdiction next Monday morning, I would not start by choosing a platform. I would start with three pages:

  1. The educational need you will actually support this year. One need, not a catalogue.
  2. The boundary that stays human: assessment decisions, safeguarding and anything that changes a student’s path.
  3. The evidence you will look at in six months, and who has the authority to pause the work if that evidence is poor.

Those pages give jurisdictions something comparable to exchange without forcing the same implementation. Our task is not to change education in the face of AI tools. It is to shape AI around the future we want education to create.

Thank you to PAIDEIA and UNICEF Argentina for the invitation to take part in this national conversation.

AI literacy in African schools: from policy ambition to classroom learning

Ghana’s proposed primary curriculum offers a useful case for considering what children should learn about AI, how schools should assess it, and what teachers need to make it work.

A curriculum commitment worth examining

Artificial intelligence literacy deserves a place in school education. Its value depends on whether it helps children understand technology, make informed decisions and take responsibility for their work. Curriculum policy should make those capabilities visible in what pupils learn and how their learning is assessed.

Ghana offers a concrete starting point. In July 2026, its curriculum review committee presented proposed revisions for kindergarten to Primary Six that explicitly included AI literacy, computational thinking, coding and digital citizenship. The direction is reinforced by Ghana’s National AI Strategy 2025-2035, which calls for primary-level data science and coding, practical use of existing AI, ethics education and teacher training.

There is an important qualification. An Education Ministry statement reported on 10 August said the revised curriculum was still undergoing review before formal adoption. The syllabus documents examined here are drafts. They show an intended direction, not evidence of completed implementation.

Read more

ChatGPT won’t fix education systems. But it has become shorthand for AI.

ChatGPT is a useful starting point for a conversation about AI in education. It gives people something concrete to try: ask a question, draft an explanation, adapt a resource. The difficulty comes when that experience becomes the basis for an institutional strategy.

For a ministry, education publisher or NGO, the question is how those capabilities fit into the work of delivering education. Who approves a generated resource? What information can staff share with a provider? How will a translated explanation be checked? What happens when a funded pilot ends?

My view is that ChatGPT-class tools can make useful contributions across that work. But the programme around them needs as much attention as the technology. Without it, institutions risk expanding production faster than their capacity to check what they produce.

Read more

What AI means for Mzansi’s classrooms

Niall McNulty speaking on the Lefa AI webinar, What AI Means for Mzansi's Classrooms, September 2026

On 10 September 2026 I joined the first of three national webinars in Human Studios’ Lefa AI series, What AI Means for Mzansi’s Classrooms. The session was Navigating AI, Learning and Safety in South African Basic Education. Lefa is a ten-month national initiative convened by Human Studios, with UNICEF South Africa and the DG Murray Trust. It is a sense-making process, not an adoption campaign: research, dialogue and practical conversation while South African schools are already using AI, and while national guidelines are still being written.

I was a panellist on the AI and learning conversation. Shirley Eadie facilitated. My contribution was practical examples, with a limit I wanted on the table early: AI does not improve learning because it provides answers. It helps when it is designed to preserve productive struggle, work within a learner’s zone of proximal development, and leave teachers in charge of judgement. One public example was Cambridge’s Future Educator programme with KFAS in Kuwait, where STEM teachers applied AI through their own subjects rather than sitting through a tour of tools. That programme has no published learning-outcome data, so I offered it as a design story, not as proof of attainment.

What Lefa is trying to do

Lefa AI describes itself as a national process to help South Africa make sense of what AI means for basic education, from Grade R to Grade 12, and to shape a response that is informed, collaborative and grounded in local context. Over ten months it brings together schools, communities, researchers, policymakers and organisations through rapid research, dialogue, webinars, roundtables and accessible resources. The premise on the Lefa AI site is simple: AI is moving faster than schools, systems and policy can adapt, and South Africa’s existing challenges of inequality, language diversity, teacher workload and uneven access to resources make delay expensive.

I like that framing because it does not start from a product. Most AI conversations in education begin with a tool and work backwards to a use. Lefa starts with what schools are already experiencing and asks what responsible use would look like here.

The argument I made on the panel

The point I wanted on the table early was that answers are not the same as learning. A tool that gives a learner the answer can remove exactly the effort that produces understanding. AI helps when it is designed to keep productive struggle in the task, to pitch support at the edge of what a learner can do with help, and to leave the teacher in charge of judgement about what counts as good work.

That has practical consequences for anyone designing or buying AI for schools:

  • Ask what the learner still has to do. If the tool does the thinking, it is a shortcut, not a tutor.
  • Ask where the teacher sees the work. Support that happens privately between a learner and a chatbot is hard to assess and easy to misuse.
  • Ask what evidence exists. A good design story is useful, but it is not proof that learning improved.

The Kuwait example I used is a case in point. Teachers on the Future Educator programme applied AI inside their own subjects, which is the right design choice, but there is no published learning-outcome data yet, so I described it as a design story and nothing more.

Why the questions here are local

Language is the clearest example. South Africa has eleven official languages, and most AI models are trained mainly on English text produced elsewhere. A tool that works well for an English-medium school in a well-resourced suburb may be far less useful for a learner whose home language is isiXhosa or Sepedi. The same is true of infrastructure: a system that assumes constant broadband and a device per learner describes only part of a bimodal school system.

Teacher capacity matters just as much. Teachers are already stretched, and AI training that adds workload without removing any will not last. The most useful uses are the ones that save teachers time on preparation and assessment while keeping their professional judgement visible.

What to watch next

This was the first of three webinars in the series, and Lefa also runs roundtables and publishes research and resources as it goes. If you are working on AI in South African schools, the series is a good place to test your assumptions against what teachers, researchers and officials are actually seeing. For a practical starting point at school level, the Ghana readiness findings show the same pattern I expect here: tools arriving faster than the policies and people needed to use them well.

Lefa is worth paying attention to because the South African questions are local: eleven official languages, a bimodal school system, teacher capacity that is already uneven, and tools trained somewhere else. If you work in a school, a ministry or an education organisation here, start with the Lefa AI site.

Related resources

What 28 AI Readiness Assessments Reveal About Schools in Ghana

On 21 February 2026, I ran AI readiness workshops with school leaders at Cambridge International Schools Day in Accra, Ghana. Twenty-four schools completed an AI Readiness Framework, a self-assessment tool that scores schools across five dimensions of AI preparedness. Twenty-eight responses were collected in total, as several schools had more than one participant complete the assessment.

The results tell a clear story. Schools are investing in hardware and tools. They are not investing in the governance and policy structures needed to use them well.

Niall McNulty running an AI readiness workshop for schools in Accra

Read more

The question is not which AI tools to buy

What 200 school leaders in Ghana taught me about AI readiness

“How many of you have a written AI policy for your school?”

I asked this to a room of 200 school principals and senior leaders at Cambridge International Schools Day in Accra, Ghana. A handful of hands went up. Maybe ten. In a room of 200.

That ratio is not unusual. In a 2023 UNESCO survey of over 450 schools and universities, fewer than 10% had developed institutional policies or formal guidance on generative AI. The numbers have likely improved since, but not by much. Meanwhile, RAND’s 2025 survey of US schools found that more than half of teachers are now using AI for planning, feedback, and content. College Board research puts the figure even higher for students: 84% of high school students reported using generative AI for schoolwork by mid-2025.

The gap between what we have planned for and what is already happening. That’s what this session was about.

Niall on-stage at the Cambridge International Schools Day in Accra, Ghana

Read more

Your School Isn’t Behind on AI. You Just Don’t Have a Map Yet.

A practical readiness framework for international schools in Sub-Saharan Africa

I spend a lot of time talking to school leaders about AI. Principals in Zimbabwe and South Africa, heads of school in Kuwait, curriculum directors in the UAE. The conversation almost always starts the same way: “We know we should be doing something. We just don’t know what.”

It’s not a lack of intelligence or ambition. These are people running complex institutions, managing multilingual communities, navigating international curricula, and doing it across infrastructure realities that range from fibre-connected campuses to schools where the Wi-Fi drops out twice a day. What they lack isn’t capability. It’s a map.

Niall McNulty

Read more

AI in Education: Lessons from Kuwait

What happens when educators move beyond the hype and get practical about AI in the classroom?

Last week, I was in Kuwait for four days of intensive AI training with Cambridge University Press & Assessment, Educate Ventures Research, and Kuwait Foundation for the Advancement of Sciences (KFAS). It was the final in-person sessions for a cohort who had already completed three months of hybrid learning – and the energy in the room showed just how far they’d come.

The participants were teachers, trainers and educators from across primary, secondary and higher education. What brought them together was a shared question: how can AI genuinely support better teaching and learning?

Carla Aerts leading a session with teachers.

Read more