Superintelligence in Language Schools: What to Prepare

Superintelligence in language schools, in practice, is not a machine that teaches on its own. It is an AI layer that brings together student history, levels, materials and school rules to support teachers and coordinators, while assessment, rapport and teaching decisions stay with people. The result depends less on the tool and more on what the school prepares beforehand: organized data, written criteria and clear limits.

Key takeaways

• AI works as support that consults history, level and materials. It does not replace the teacher.
• Preparation matters more than the tool: standardized progress records, levels with written criteria and versioned materials.
• Speaking assessment, sensitive feedback, rapport with the student and decisions to change level remain human.
• Retention improves when the risk signal reaches the teacher or coordinator with context and a human contact has been agreed.
• Student data, including data on minors, needs a clear purpose, restricted access and human review under Brazil's data protection law (LGPD).

What does superintelligence mean in a language school?

The term tends to frighten or excite people more than it should. For a school, the useful version is far more modest: a layer that knows where student information lives, what the criteria for each level are, which materials exist and which actions the AI may or may not take. It answers to the teacher, the front office and the coordination team, always within the school's rules.

If you want to separate what the technology is from what is promised, read Is ChatGPT Superintelligence? What It Is and What It Is Not. The conclusion that matters here is operational: what your school runs today are assistants and agents with bounded tasks, and they are only as good as the context they receive.

This article does not repeat the admissions and enrollment workflow, which already has its own guide in AI for Schools: Admissions and Parent Communication Guide. The focus here is the teaching and retention side: what to prepare and what not to delegate.

What student history should you prepare before any agent?

An agent that answers well about a student needs a record worth consulting. In many schools, the history is scattered across the coordinator's spreadsheet, the teacher's notebook, WhatsApp conversations and the memory of whoever answers the phone. Before thinking about AI, consolidate the useful minimum:

• The student's goal (travel, career, certification, helping their children) and entry level.
• Class, current teacher and history of previous classes.
• Attendance and submitted assignments, with dates.
• Strengths and difficulties by skill: speaking, listening, reading and writing.
• Short teacher notes, in standardized fields and not only in free text.
• Pauses, suspensions and the reason given, when there is one.
• Relevant front office contacts, kept separate from sensitive financial details.

The common mistake is recording everything in free text, with every teacher doing it differently. AI can summarize free text, but it also summarizes the inconsistencies. Combine short fields with a free note, and make clear what is fact (was absent, submitted the work) and what is the teacher's impression (seems unmotivated).

Think about data minimization too. Teaching data serves the teacher and the coordinator. Financial data serves whoever handles billing. A teaching support agent does not need to see overdue payments. When the student is a minor, the processing calls for extra care and a well-defined legal basis. The ANPD, Brazil's data protection authority, publishes official guidance on data protection, and the legal basis decision should be validated with whoever handles compliance at your school.

How do you organize levels and materials for AI to consult?

Levels are where schools most often disagree with themselves. If two teachers understand an intermediate student differently, the AI will inherit that confusion. Preparation means writing down, for each level, what the student can do in each skill, with examples of expected output. If the school uses the Common European Framework of Reference (CEFR), record the mapping between your internal levels and that scale instead of leaving it implicit.

For materials, three precautions:

• Inventory: list books, workbooks, platforms, audio files and in-house activities, with level, skill, objective and duration.
• Rights: publisher content is licensed. Check what the contract allows before putting excerpts into any AI tool.
• Version: mark which material is current. An agent that cites an outdated workbook creates more rework than help.

For the AI to consult these documents, the design needs to stay simple. Anthropic's own guidance in Building effective agents recommends starting with the simplest solution and adding complexity only when there is proven need. At a school, that usually means an assistant that consults an organized set of documents and returns answers with the source indicated, not an autonomous system that rewrites the curriculum.

What stays human in teaching?

AI can help prepare the lesson, but it does not teach the lesson. In language schools, what sustains learning is interaction, and parts of it are hard to delegate. To organize the split, use this decision guide:

• Can support with review: lesson plan drafts, extra exercises at the right level, a summary of the student's history before class, conversation topics tied to the student's goal.
• Can support, with the teacher deciding: feedback on writing drafts, a list of recurring errors, suggestions for content review.
• Stays human: nuanced assessment of speaking and pronunciation, the decision to change level, feedback on performance, conversations with a discouraged or anxious student, adaptation for learning difficulties, mediating conflict in a class.
• Do not delegate: any decision with consequences for the student (failing, dismissal, class change) without human review.

That last point has a legal basis. The LGPD provides a right to review decisions made solely through automated processing. The article LGPD Automated Decision-Making: The Right to Review (Art. 20) explains how this shows up in a company's daily operations. At a school, the practical rule is simple: the AI suggests, a person decides and records.

There is also a teaching reason. People learning a language make mistakes in front of others, and how the teacher reacts determines whether the student comes back next week. That kind of reading of the room is not in any record.

How can AI help with retention without pressuring the student?

Retention at a language school is rarely lost on the day of cancellation. It is lost earlier, in small signals nobody put together: consecutive absences, stalled assignments, lower participation, slow replies to the front office, a question about pausing. AI is good at connecting those signals and taking them to someone who can act. It should not be the voice that tries to convince the student to stay.

A healthy design has four parts:

• Signals defined in writing by the coordination team, not invented by the model.
• An alert to the teacher or coordinator with a short summary of the student's context.
• An agreed human action: a conversation after class, a phone call, an adjustment of class or schedule.
• A record of what was done and the outcome, so the school learns.

When there is outbound contact on WhatsApp, follow the rules on permission and templates. The article WhatsApp Outbound Messages: Templates, Opt-In and Best Practices covers this in detail. A useful notice (class rescheduled, new material) is different from a sequence of payment reminders disguised as care.

Illustrative scenario: an adult student who loses momentum mid-semester

This is an illustrative scenario, unrelated to any real school. An adult English student, studying for meetings with overseas suppliers, starts missing every second class and stops sending the writing assignments.

Without support, the coordination team only notices when the student asks to cancel. With the layer prepared, the agent cross-checks attendance and assignments, sees the pattern and generates a note for the teacher: the student's goal, the skills where they felt most confident, the latest recorded difficulties and the recent change in pace. The agent does not write to the student. The teacher reads the note, talks to the student before the next class and learns that their travel routine has changed. The two agree on short listening activities and an alternative time slot, and the teacher records the agreement.

What the AI did was organize the information and bring the conversation forward. The listening, the proposal and the rapport were human. If the student decides to pause, that decision is theirs, and the process follows the front office's normal flow.

How to apply the method: map, limit, pilot, transfer

The method we use with companies fits a school without major adaptation, and each step has its own question.

• Map: how does a student enter, progress and leave today? Which records exist, where are they and who maintains them? What is the starting retention situation, measured the way the school already measures it?
• Limit: what data can the agent see, what actions can it take and what requires human approval? Who accesses what, why, and how is that logged?
• Pilot: pick one real workflow, such as the risk alert to the teacher, with every decision recorded and a success metric agreed before starting.
• Transfer: at the end of the pilot, the school decides to expand, adjust or stop, with documentation and named owners.

During the pilot, test the agent with anonymized real cases and expected results defined by the teachers. Anthropic's guide Demystifying evals for AI agents describes how to structure those evaluations. For thinking about risk in an orderly way, the NIST AI Risk Management Framework offers a public reference, useful for the coordination team to discuss what could go wrong and who is accountable.

Preparation checklist for the coordination team

Before buying or configuring anything, check:

• Do the levels have written descriptors mapped to a known scale?
• Does the student history use standardized fields, and does someone own keeping it up to date?
• Are materials inventoried, versioned and license-checked?
• Is it defined what the AI sees, what it does and what requires human approval?
• Is there a specific rule for minors and for communication with parents or guardians?
• Were the dropout risk signals defined by the coordination team?
• Is there a log of the agent's decisions and a path for human review?
• Does the pilot have a success metric and a date to decide whether to expand, adjust or stop?

If most answers are no, the best investment right now is organization, not tooling.

Frequently asked questions

Can AI replace the language teacher? Not in what weighs most for learning: live interaction, speaking assessment, reading the student's state and rapport. It can reduce preparation and record-keeping work, giving the teacher time back for what only a teacher can do.

Do I need a new system to start? Usually not. The first step is organizing the history, levels and materials the school already has. Once that is done, a small pilot on one workflow, such as the risk alert to the teacher, shows what is worth expanding.

How should we handle data on students who are minors? With a clear purpose, access limited to what is necessary, and special attention to the legal basis and to communication with parents or guardians. Consult the ANPD guidance and validate the decision with whoever is responsible for compliance at your school, since the exact path depends on the case.

Should AI contact a student who might cancel? Only within clear permission rules and with an agreed human action. In general, it is safer for the AI to alert the teacher or coordinator and leave the retention conversation to a person, who can hear the real reason.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework
• Anthropic: Building effective agents
• Anthropic: Demystifying evals for AI agents