AI agents in the school office: from assistant that suggests to system that executes

AI agents in the school office: from assistant that suggests to system that executes

Deloitte ranks agentic AI as the top technology trend of 2026. Gartner estimates 74% of companies will deploy autonomous agents within the next two years, and McKinsey projects that by 2027 between 60% and 70% of administrative tasks in services will be automated. The market reached 45 billion dollars in February of this year.

These are impressive figures and almost none of them land on what a school office actually does on a Tuesday morning. This article tries to land it.

The distinction that matters

An assistant answers. An agent completes the work. The difference is not model power, it is design:

• Assistant: takes a question and returns an answer or a draft. The work still belongs to the person, who now types faster.
• Agent: takes a goal, breaks it into steps, executes each one by consulting the system, and only stops when the goal is met or when it hits something it is not entitled to decide.

A school office example

Asking an assistant to draft a reminder for an unpaid invoice saves five minutes. Asking an agent to chase that invoice until it is paid saves the entire process:

• Check whether the failed payment is real or a bank data error.
• Send the right reminder given the family's history.
• Wait and retry the charge on the correct date.
• Escalate to the responsible person if it remains unresolved.

Against the automations you already have

There lies the difference from classic automations, which many schools already have. An automation executes a fixed rule: if A happens, do B. It works perfectly until the first exception, and in a school exceptions are half the job.

Which processes suit an agent

Not everything is suitable. Processes that work well with an agent share three traits:

• They are repetitive: the chain runs dozens of times a month with small variations.
• They have a verifiable outcome: the invoice was paid or not, the document arrived or not, the statement balances or not.
• Their exceptions are frequent but bounded: they show up often, and it is clear in advance which ones a person must resolve.

Three office workflows you can already delegate

With those three traits, a mid-sized school can already hand three workflows to an agent:

• Arrears recovery: a returned payment today triggers a manual chain — spotting it on the statement, checking bank details versus insufficient funds, calling, noting, retrying, updating status. An agent walks that whole chain and only raises its hand when the family asks to spread payments or when the failure repeats a third time.
• Incomplete enrolment: every September there are files missing a document. An agent knows what is missing from each, chases it through the channel that family actually uses, checks what arrives and escalates only the cases still open past the deadline.
• Reconciliation: matching the statement against issued invoices is mechanical apart from a handful of odd cases. An agent reconciles everything obvious and leaves on the desk only what genuinely requires judgement, usually under 5% of transactions.

Where a person still has to decide

There are three boundaries worth not crossing, and one of them is no longer merely a recommendation:

• A student's access or assessment: the EU AI Act classifies those systems as high-risk under Annex III, with obligations deferred to December 2027. Designing workflows today so an agent gathers information and a person resolves avoids having to redo it.
• Exceptions with financial or social impact: spreading a debt, waiving an amount, activating a safeguarding protocol. An agent can prepare the complete file, but the decision has a name attached.
• The first communication of bad news: letting an automated message inform a family of a serious problem before someone from the school has spoken to them is a judgement error no efficiency gain compensates for.

The three conditions that keep it from getting expensive

An agent working unsupervised turns time savings into incidents. Three non-negotiable requirements:

• A full record: what the agent did, when, with which data and with what outcome. Without traceability you cannot audit or correct, and facing a complaint you have to reconstruct the exact sequence.
• Explicit limits: what it may never do without human approval. Sending to more than X families at once, changing amounts, deregistering anyone. It is easier to widen permissions than to repair a mistaken mass send.
• A real stopping point: a named person who reviews exceptions. If nobody looks at that queue, the agent ends up being a system that fails silently.

Case study (Spain)

A language academy with four hundred students spent around two working days a month on the arrears cycle. One person reviewed the returned payments, cross-checked against the student list, wrote to families and recorded on a sheet what each had replied.

After configuring an agentic workflow over the same process, the load dropped to about three hours a month. What was interesting was not the saving but its distribution: the agent resolved 78% of cases without intervention, mostly because they were bank data errors or one-day delays. The remaining 22% reached the person with the context already assembled: family history, amount, previous attempts and preferred channel.

The effect nobody was looking for

The side effect was the most valuable part. Because reminders always went out at the right moment rather than when somebody had time, arrears at forty-five days fell steadily without anyone having tightened anything.

Conclusion

Agentic AI is neither a distant promise nor a threat to the school office job. It is a shift in which part of the work a person does: less executing chains of predictable steps and more deciding on the exceptions, which is where human judgement is worth something. The schools that benefit will not be those buying the most powerful tool, but those whose processes are clear enough to be described.

Edena lets you build these workflows on data already in the platform — invoices, records, communications — with a full log of every action and stopping points where a person decides. Book a demo and we will build your school's arrears workflow with you.