Agentic AI in schools: A problem for another day?

There is a lot of discussion about agentic AI at the moment. The promise is easy to understand. Rather than using AI simply to respond to a prompt, an agent can be given a task, use tools, search information, follow a sequence of steps and produce an outcome with less direct human input. In theory, this is where AI becomes more than a drafting tool or a chatbot. It becomes a digital worker, able to support planning, administration, analysis, communication and decision-making across the organisation.

At an individual level, this can be very useful. A teacher may create an agent to support lesson planning or resource creation. A school business manager may use one to help review supplier documentation. A technology leader may use one to summarise cyber security priorities. A data protection lead may use one to review Records of Processing, identify gaps or prioritise DPIAs. A Head may use one to test the logic of a paper before it goes to governors. For specific, repeatable or exploratory tasks, the value is obvious.

But individual usefulness is not the same as organisational readiness.

This is where I think schools need to be careful. The current excitement around agents can make it appear as though the next step is simply to build them and let them operate. In practice, agentic AI only really works when the rules of engagement are clearly defined, the objectives are understood, and the source material is accurate, current and organised. Even when that is true, the agent still needs to be maintained. It still needs to be checked. It still needs human judgement around it. It still needs someone to decide whether it has used the right information, understood the right context and produced something that can actually be relied upon. This is not a reason to dismiss agentic AI. It is a reason to treat it as an organisational change issue rather than a technology feature.

Embedding emerging technology into schools is rarely just about the technology itself. More often, success or failure comes down to the structure people are working within. Where authority is unclear, where the parameters are not defined, where reporting lines do not function, or where different teams are working from different assumptions, technology projects become harder than they need to be. AI makes this more acute because it does not sit neatly in one part of the school. It touches teaching and learning, safeguarding, privacy, cyber security, procurement, staff training, parent communication, inspection, accreditation and leadership accountability. That is why the organisational use of agentic AI is much harder than the personal use of agentic AI.

An individual can hold a lot of context in their own head. They can remember why they created an agent, what it was meant to do, what source material they gave it and when its output should be ignored. They can adapt prompts, correct mistakes and make informal judgements because the work belongs to them. An organisation cannot safely operate on that basis. If an agent is going to support a school-level process, the school needs to know what it does, what information it relies on, who owns it, who maintains it, what it is allowed to do, what it is not allowed to do, where human review is required and how decisions are recorded.

The first practical issue is source material. An agent can only work well from the information it can access and understand. This is one of the more uncomfortable truths about AI. It does not remove the need for good organisational information. It exposes whether that information exists.

Schools are full of contextual knowledge that is important but not always well documented. There may be tools that are widely used but not formally approved. There may be systems that are essential during exams but less important at other times of year. There may be policies that are technically current but no longer reflect practice. There may be contracts due for renewal, vendors changing functionality, departments piloting tools informally, or leadership decisions that have been made but not clearly recorded. A person working inside the school may understand these things. An agent will not, unless the school has created the structure for that knowledge to be recorded, governed and maintained.

Schools are also seasonal organisations. They do not operate in a consistent way across the year. The start of school is different from the middle of the year. Exam preparation is different from exams. Admissions, reporting, transition, safeguarding reviews, inspection readiness, staff training and budget planning all create different pressures. There are periods when normal teaching does not happen. Staff are away. Students leave. New students arrive. Priorities change. What mattered in June may not matter in September, and what was appropriate before exams may not be appropriate during them.

An organisational agent would need to understand this organisational context if it was going to support real delivery. It would need to know not only the task, but the context in which the task is being performed. It would need to understand the academic year, the school’s priorities, the relevant policies, the current risks, the people responsible and the boundaries of its own role. Without that, the agent becomes another tool that has to be supervised, corrected and maintained. Another person. Schools already have enough systems and people that require that kind of attention.

The third issue is the human in the loop. It is easy to say that AI outputs will be checked by a person. It is harder to make that meaningful in practice. Human review only works if the person reviewing has the time, expertise, authority and context to make a proper judgement. A teacher reviewing an AI-generated lesson resource may be able to do that quickly because they understand the class, the curriculum and the intended learning. But reviewing the output of an agent that has summarised safeguarding information, prioritised privacy risks, suggested vendor actions, drafted parent communications or interpreted organisational records is a different level of responsibility. The review burden grows as the consequence of the output grows.

Over the summer, I have built agents in OpenAI, Anthropic and Grok. Some have been useful. For brainstorming, prototyping and scalable tasks, they can be genuinely helpful. But anything with substance still requires review, correction and approval. Not a quick glance, but proper checking. After a while, that becomes tiring. There is a limit to how much a person wants to build, maintain, prompt, test, correct and approve before speaking to another human would have been quicker, safer or more useful.

The same is true of software development. We have built software using Claude Code and Codex. Both have been impressive in showing what is possible. They are useful for prototyping, testing ideas and accelerating parts of the development process. But the quality of the output still needs careful review. From a distance it looks impressive. Underneath, there will be bugs, inconsistent code, weak architecture and cyber security vulnerabilities that need to be identified, tested and fixed.

That may be acceptable for one person trying to complete a small, low-risk task, in the same way that an old run-around car may still get you to the shops. It is very different when the requirement is reliability, safety, scale and enterprise-level assurance. At that point, the question is not whether AI can produce something that looks useful. It is whether the organisation has the people, processes and controls to check whether it is genuinely safe, reliable and fit for purpose.

The more agentic AI becomes, the more governance it requires. A chatbot that drafts a paragraph is one thing. An agent that plans, connects systems, draws on organisational records, recommends action or produces outputs that affect pupils, staff, parents or leadership decisions is something else. It needs a defined purpose, accurate information, permissions, data protection review, security controls, safeguarding consideration, human oversight, ownership, maintenance, escalation routes and evidence of decisions. Without those things, the agent may look productive, but the organisation may not be in control.

I am not convinced that standalone agents will become the default route for most schools, at least not in the short term. The more likely path is that existing EdTech and productivity platforms will use AI to make their current products more useful. The focus will be on better search, better summarisation, better drafting, better reporting, better insights and more role-specific intelligence. This is less dramatic than the idea of a fully autonomous agent, but it is more likely to be adopted because it sits inside systems people already use.

Microsoft and Google have an obvious advantage here. Their platforms are already in schools. They sit across email, documents, meetings, calendars, storage, collaboration and administration. As their AI capabilities improve, they may become the natural interconnector between users and the wider EdTech stack. Not necessarily because their models are always better than OpenAI, Anthropic, Grok or others for every task, but because adoption is not only about model capability. It is also about access, familiarity, security, procurement, cost, governance and whether staff can use the tool without having to build and maintain something separate.

For most users in a school, the AI features provided through existing platforms may be enough. A teacher does not necessarily want to build an agent. A member of the operations team may not want to maintain one. A safeguarding lead may not want another system to supervise. They want the tools they already use to become easier, quicker and more helpful, within boundaries the school understands.

This may be where we reach peak agentic AI experimentation in schools. There will always be a small group of people who enjoy building agents, testing models, refining workflows and seeing what is possible. Schools need those people. They help the organisation learn. But they are not the majority. For most staff, the question is not whether they can build an agent. The question is whether AI can help them do their job better without creating more work, more risk or more complexity.

The sensible position for schools is not to reject agentic AI, but to put it in the right place. For many schools, it may be a problem for another day. The priority now is to understand what AI is already in use, assess AI-enabled vendors, define approved tools, train staff and students, agree where human review is required, connect AI use to safeguarding, privacy, cyber and academic oversight, and report progress and risk to leadership. Once those foundations exist, a school can make a much better decision about where agentic AI may have a role.

Agentic AI may become valuable in schools, but it will not become valuable simply because the technology improves. It will become valuable where schools have clear strategy, practical policy, accurate source material, defined ownership, appropriate controls, human review and a governance structure capable of learning and adapting.

The work is not just building the agent. The work is maintaining the conditions in which the agent can be trusted.

That is why, for many schools, the next step should be less about chasing agentic AI and more about building the foundations for governed AI. That may sound less exciting than the future currently being sold, but it is more realistic. And in schools, realism matters.

Technology only becomes valuable when it is safe, understood, governed and genuinely useful to the educational mission of the school.