AI Agent for Language Schools: Placement and Enrollment Records

An AI agent for a language school can answer students about their current level, retake options and enrollment steps, but only when it reads a verified placement record and a clear enrollment file. The agent should suggest, prepare and route requests, while a teacher or academic coordinator confirms every level change. The records that make this safe are mostly ordinary ones, and most schools can start keeping them without new software.

Key takeaways

• Store each placement result as structured fields (test version, skill scores, assessor, date, scale version), not as a single level word.
• Mark every level as confirmed, provisional or under review, and let the agent read only confirmed values.
• Give the agent read access to enrollment status and a write path to a request queue, never to the level field itself.
• Write a one or two sentence reason for each placement so students and staff hear the same explanation.
• Hand over one narrow task first, such as placement retake questions, and check its answers against staff-approved cases before widening the scope.

Why the placement record comes before the agent

Most language schools hold placement information in several places at once: a test sheet in a teacher's folder, a level code in the enrollment system, and a message thread noting that a student moved up after a speaking review. An agent asked about a student's level will answer from whatever it can reach, and it will repeat the first answer it finds with full confidence. Before any agent is connected, the school needs one record per student that states what was decided, on what evidence, by whom, and whether the decision is final.

The technical pattern here is retrieval. The agent looks up the relevant record at the moment of a question instead of relying on what it remembers from earlier conversations. Google Cloud's reference architecture for retrieval-augmented generation describes this separation between the model and the grounded data, and the same logic applies to a school's own files. Retrieval helps only when the records are clean. If three documents disagree, the agent will present the disagreement as fact. Operations teams that already keep disciplined per-client memory will recognize this problem from managed service provider memory. Preparing for more capable agents mostly means cleaning up the records they will be asked to trust.

The placement fields to keep now

Add these fields to every new placement from today. Backfill older students only when a teacher can verify the entry.

• A student identifier that matches the enrollment system, so one person is not recorded under two spellings.
• The placement test name and version, because a retired test cannot be compared directly with a current one.
• Scores by skill where the school assesses reading, listening, speaking and writing separately.
• The assessor's role, name and the date of the assessment.
• The level assigned, written on the school's own scale, with the scale version noted.
• One or two sentences giving the reason for the level.
• A status field: provisional, confirmed or under review.
• Any open request that affects scheduling, such as a need for evening classes or a pending speaking review.

Keep the scale itself in a document the agent can read. If the school uses A1 to C2 labels plus an internal sub-level for exam preparation, record both and state which one students see. Agents handle explicit labels well and struggle with labels that only staff understand from habit.

Enrollment records: what the agent reads and what it never writes

Enrollment questions are where students take action. They ask whether a seat is held, whether a deposit arrived, whether an identity document was received, and when the next cohort starts. Those answers should come from the enrollment system, not from the agent's memory of an earlier chat.

• Course, cohort, start date and capacity status, as recorded by staff.
• Seat hold status and its expiry date, as shown in the enrollment system.
• Payment status as a reported state (received, pending, overdue), with no card or bank details stored in the agent's memory.
• Documents received, such as identity papers, prior certificates and signed consent forms, with the date each arrived.
• The version of the refund and cancellation policy the student accepted.

Records about students are personal data. In Brazil, the national data protection authority, ANPD, publishes guidance on obligations under the LGPD, and the school should check its own duties with its advisers before deciding which fields an agent may see. Write that decision down, because the list of readable fields is part of the agent's design.

Who may change a level: a decision guide

The safest default is that the agent cannot change a level at all. Permissions in this area should follow deny-by-default logic, the principle in OWASP's authorization guidance that an action is allowed only when a rule explicitly grants it. The guide below turns that principle into school decisions.

• The agent may state the confirmed level, the date it was confirmed and the scale it uses.
• The agent may explain what the placement test measures and what a review involves.
• The agent may open a review request with the placement record and the student's message attached.
• The agent may propose a level only as a labeled suggestion, with the evidence it used.
• The agent may not write to the level field, waive a retest, or tell a student that a class has a seat.
• A teacher or academic coordinator makes every change, and the system records who approved it and when.

A step-by-step plan for the first pilot

A narrow first version is easier to check than a broad one. The sequence below is one practical order of work.

• Export current placement and enrollment data and list every level label in use, including informal ones.
• Map each label to one school scale and have the academic coordinator sign off on the mapping.
• Add the new fields to incoming placements first, then backfill older records only where staff can verify them.
• Set the agent's access: read on confirmed records and enrollment status, write only to the request queue.
• Choose one question for the first version, such as whether a student can retake the placement test.
• Write fifteen to twenty test questions taken from real messages, and have staff write the correct answer for each. OpenAI's evals guide describes this kind of repeatable check against expected outputs.
• Review every answer that sent a student to the wrong place, record the cause, and fix the record or the instruction, not only the reply.
• Widen the scope only after the narrow task has stayed correct across several review cycles.

In a managed setup, Be Human designs this kind of agent as an ongoing service. For a language school, that means defining the one role the agent plays, the record fields it may read, the requests it may open, and the approval steps a coordinator must complete. The success criterion is written before launch, for example whether every level answer cites a confirmed record. After launch, managed operations cover monitoring, recurring evaluations of the test questions, permission reviews when staff roles change, and incident response when an incorrect answer reaches a student. The service is remote and runs in English and Portuguese.

Illustrative scenario: a student asks to move up

The following scenario is illustrative and does not describe a real school. A learner in an evening general English class writes to the school's assistant asking to move to the next class because a friend is already there. The agent finds a confirmed B1 record on the current test version, along with a provisional note that a speaking review is pending. It tells the student the confirmed level and the review status, offers the available review times, and opens a request for the coordinator with the placement record and the student's message attached. The coordinator listens to the speaking sample and makes the decision. The agent never changes the level, and it never says a seat is free in the upper class, because capacity is a staff fact it does not hold. If the student wrote in Portuguese, the reply would be in Portuguese, with the same record and the same labels.

Watching what the agent says about levels

Each answer should leave a trace: which record version was read, which tools were called, whether a request was created, and what the student was told. Observability concepts such as traces and logs, described in OpenTelemetry's documentation, are the practical basis for reviewing these traces later. Without them, a wrong answer can be noticed only when a student complains.

Students will sometimes try to override the rules by typing instructions such as telling the assistant to ignore its policy and set them to a higher level. This kind of prompt injection is covered in OWASP's prompt injection prevention guidance. The defense is structural: the level field is not writable from the conversation at all, so no wording of a message can change it. The NIST AI Risk Management Framework offers a shared vocabulary for mapping and measuring these risks, which a school can adapt without building a large program. If you are not sure which placement or enrollment task is worth handing to an agent first, take the free Be Human AI diagnostic to find it.

Frequently asked questions

Can an AI agent change a student's level in a language school? It can be built to do so, but this is not recommended. The level is a judgment supported by evidence, and it should stay with a teacher or academic coordinator. The agent can read the confirmed level, open a review request and explain the process to the student.

What is the minimum record a school needs before using an agent? One record per student is enough to start, containing a stable identifier, the test name and version, the assessor, the date, the level on the school's scale, a short reason and a status. Enrollment status should be readable from the enrollment system. Without the status field, the agent cannot tell a confirmed level from a provisional one.

Should the agent answer in English and Portuguese? Yes, the agent should reply in the language the student uses while drawing on the same record. Keep the level labels identical in both languages, or map them clearly in the scale document so translations do not create a second scale. Be Human's managed service runs in both languages.

How does a school know the agent is working correctly? Build a set of fifteen to twenty real questions with answers approved by staff, run them on a regular schedule, and review every wrong answer. Track whether each level answer cites a confirmed record and whether every request reached the right person. Compare the results with the success criterion that was set before launch.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.