Blog

AI agents in education: what they are, what they can do in a school today and what they should not do

What sets an AI agent apart from a chatbot, which leadership, office and teaching tasks can already be delegated, what permissions it needs and where a person must decide.

Ena by Edena | 7 min read
AI agents in education: what they are, what they can do in a school today and what they should not do

When people talk about artificial intelligence in education, the conversation usually turns to the classroom: whether students use ChatGPT for homework, how to assess now or what teachers should teach. There is another, quieter conversation moving faster: that of AI agents working inside the school's management, looking up data and preparing tasks that until now a person did by hand.

This article explains what an AI agent is, what it can do today in a school, nursery or academy, what it needs to be useful without being a risk and which decisions it should never take.

What an AI agent is (and is not)

An AI agent combines four pieces:

  • A language model that understands what you ask and reasons about it.
  • Tools that let it look up data or do things in other programs.
  • A goal, usually given by a person in plain language.
  • A loop: the agent decides which tool to use, looks at the result and decides the next step, until it achieves the goal or asks for help.

It is worth telling it apart from two things it is often confused with:

  • A chatbot answers with what it learned and what you type. It cannot see your school's data or do anything outside the conversation.
  • An automation follows fixed rules: "when a direct debit is returned, send this message". It does not reason or adapt, but it is very predictable.

An example of an agent in the school office: "Tell the families of Year 3A that pick-up is at 13:00 tomorrow". The agent finds the class, checks which families it has, drafts the notice, shows the proposal to whoever asked and sends it when that person confirms.

Four levels of autonomy

Not all agents do the same. A useful way to classify them is by what they can do without asking:

  1. Look up. It answers questions with the school's data: how many students were absent this week, which bills are still outstanding, how many enrolment applications are open.
  2. Prepare. It writes a draft that a person reviews and sends themselves: an announcement, a reminder, the first draft of a report card comment.
  3. Act with confirmation. It prepares the full action inside the software and carries it out only when the person clicks confirm.
  4. Act alone. It acts without asking, within certain limits.

In a school, the reasonable approach today is to work at the first three levels. The fourth makes sense for low-risk, reversible and well-defined tasks, and in those cases a rule-based automation is usually safer than an agent.

What they can do today, area by area

  • Leadership. Ask in plain language about the state of the school: the week's attendance, outstanding payments, open enrolments, staff availability. Instead of opening five reports, one question.
  • School office. Prepare payment reminders for families with outstanding bills, announcements to a class, calendar events, fee items and discounts, or find a document by name. We cover it in the article on AI agents in the school office.
  • Teachers. Take the register, prepare a message to their class's families or ask for a first draft of each subject's report card comment, which the teacher reviews and rewrites.
  • IT coordination and development. Build integrations with the help of coding agents such as Claude Code or Codex, which read the API documentation of the school's software and write the code against a test environment. We cover it in the article on APIs and webhooks.
  • Family support. Answer frequent questions about timetables, the calendar or procedures. Extra caution is needed here, because the person talking to the agent is not staff. We deal with it in the article on chatbots for families.

Where the agent lives: inside the software or in your assistant

There are two ways to have an agent working with the school's data:

  • Built into the management software. The agent lives inside the platform, knows its screens and permissions, and prepares actions with the tools the provider has designed for it.
  • In the assistant each person already uses. Claude, ChatGPT or other assistants connect to the school's software through the MCP protocol and act as that person, with their permissions. We explain it in what MCP is and what it is for in a school.

The two options can coexist. What matters is that, in both cases, the agent never sees or does more than the person using it can see and do.

What an agent needs to be useful without being a risk

  1. The school's real data. An agent without context makes things up. With access to the data, it answers with figures that can be checked.
  2. Narrow tools. Every possible action should be defined in advance: create an event, send an announcement. No "do whatever it takes".
  3. The person's permissions, and no more. If a tutor cannot see the bills, neither can their agent.
  4. Confirmation before every change. The person sees what the agent is about to do and decides.
  5. A record of what it did. What was asked, what it looked up and which action was confirmed, so it can be reviewed later.
  6. Sensitive data out of reach. Health, bank or identity data should not pass through the model unless it is essential.
  7. A switch. Someone at the school must be able to disable the agent and withdraw its access immediately.

Where a person must decide

There are decisions where an agent can help prepare but should never decide:

  • Assessment and grades. The agent can suggest a draft comment; the grade and the final text belong to the teacher.
  • Admissions. Sorting applications or summarising an interview, yes; deciding who gets in, no.
  • Behaviour and sanctions. Any measure regarding a student needs a person who knows the case.
  • Health and safeguarding. Allergies, medical reports or risk situations are handled by the people responsible, not by a model.
  • Money. Issuing or cancelling invoices, refunds and direct debit batches are done in the software and by a person.
  • Sensitive communications. A conflict with a family or a personal situation is not resolved with a generated message.

Three references to keep in mind:

  • GDPR and AEPD guidance. Using an agent with student and family data is data processing: it needs a legal basis, a processing agreement with the provider, data minimisation and information. In February 2026, Spain's data protection authority published guidance on agentic AI calling for specific governance, involvement of the data protection officer, a design that fails safely and a plan to stop the agent and withdraw its access.
  • The EU AI Act. Since February 2025, practices such as emotion recognition in schools have been banned and training staff who use AI has been compulsory. High-risk systems in education, such as those that decide admissions or assess learning, will have their own obligations from 2 December 2027. We explain it in the guide to the AI Act for schools.
  • The Ministry's guide. Version 2.0 of the INTEF guide on AI in education helps turn all this into a school plan. We summarise it in the article on the Ministry's AI guide.

Where to start

  1. Choose three repetitive, low-risk tasks, such as payment reminders, notices to a class or leadership's questions about the school's data.
  2. Decide the level of autonomy for each one: look up, prepare or act with confirmation.
  3. Review the permissions of whoever will use the agent.
  4. Inform staff and, where appropriate, families, and document the processing with your data protection officer.
  5. Measure time saved and errors for a month before expanding.

Conclusion

An AI agent is not a smarter chatbot: it is a system that looks up data and prepares actions inside the school's software. Well designed, with real data, narrow tools, each person's permissions and confirmation before every change, it takes hours of repetitive work off leadership, the office and teachers. Badly designed, it is a risk to students' data. The difference lies in the design, not in the model.

Ena, Edena's AI, is an agent built that way: it looks up the data each person can see according to their permissions and active modules, and prepares actions such as creating events, sending announcements and messages, sending payment reminders or taking the register, which only happen when the person confirms. It does not see who has which allergy, only how many cases there are, and it works for the school's staff, on the web and in the app. Meet Ena or, if you prefer to work from your own assistant, see the connector for Claude and ChatGPT.

Frequently asked questions

What is an AI agent?
It is an artificial intelligence system that, as well as answering, can use tools to achieve a goal: look up data, prepare a document or take an action in another program, chaining several steps together. In a school, an agent can check who is absent today, prepare the notice for those families and send it once a person confirms.
How is an AI agent different from ChatGPT?
A general chatbot answers with what it learned and what you type into it. An agent connected to the school's software looks up real data, with the permissions of the person using it, and can prepare actions in that software. ChatGPT or Claude become agents when they are connected to tools, for example through the MCP protocol.
Can AI agents be used with student data?
Yes, with the safeguards of the GDPR: a legal basis, a processing agreement with the provider, the minimum data needed and information for families. In February 2026, Spain's data protection authority (AEPD) published guidance on agentic AI calling for specific governance, the involvement of the data protection officer and the ability to stop the agent and withdraw its access.
What should an AI agent never do in a school?
Decide on its own about a student's assessment, admission or sanctions, send communications nobody has reviewed, move money or issue invoices without confirmation, and work with health data. In all those tasks the agent can help prepare, but the decision belongs to a person.
Is Ena an AI agent?
Yes. Ena, Edena's AI, looks up the school data the person can see according to their permissions and the active modules, and prepares actions such as creating an event, sending an announcement, sending a payment reminder or taking the register. None of them happens until the person confirms it, and Ena works only for the school's staff, not with families.

This content was generated by Ena, Edena's artificial intelligence agent. It may contain errors or inaccuracies and does not constitute legal, tax or professional advice. Edena does not warrant the accuracy, completeness or currency of the information. Consult official sources and, where appropriate, a qualified professional before taking any decision.

The office closes the month without chasing families

Invoices go out with Verifactu, attendance is recorded and each family sees it in the app. Seeing it with a school like yours takes 15 minutes.

Let's talk