Skip to content
be-human.ai
← Back to the journal

AI by industry

AI for Test Prep Companies: Student Records and Tutor Matching

How test prep companies can keep score history and tutor assignments as separate, scoped memory so an AI agent reads student progress without mixing records.

Be Human typographic cover with the title: AI for Test Prep Companies: Student Records and Tutor Matching
Conceptual illustration created with AI · Be Human

An AI agent for a test prep company works best when score history, practice activity and tutor assignments are stored as separate, labeled records that the agent queries by purpose, not as one long transcript for each student. That structure lets a capable agent answer progress questions and suggest tutor matches without exposing one student's history to the wrong person or the wrong task. This article explains how to structure that memory before any agent touches it.

Key takeaways

  • Score history, tutor assignments and session notes answer different questions, so store them as distinct records with their own owners and timestamps.
  • A tutor matching agent should read only the fields a match requires, not the full student file.
  • Every record should carry its source, its date and the person or system that wrote it.
  • Records about minors and parent communications need their own access rules before any automation starts.
  • The first build should cover one role, one question and one measurable check, reviewed on a fixed schedule.

Why score history is not a student file

A test prep company collects a great deal about each learner: diagnostic results, timed practice sets, homework, tutor observations, schedule changes, goals, enrollment status and billing. In most operations this lives across a spreadsheet, a learning platform, a scheduling tool and a tutor's inbox. An agent handed all of it as one document will produce answers that sound confident but blend a reading section from months ago with a note about a different subject.

The problem is rarely model capability. It is missing structure. Retrieval-based systems, the pattern described in Google Cloud's guidance on retrieval-augmented generation, fetch the pieces of text most similar to a question. If those pieces are not labeled by record type, student and date, the most similar piece wins, whether or not it answers the question being asked. Structure is what makes retrieval reliable, and it is the part a business has to decide, because only the business knows which records belong together.

Separate the record types before you connect anything

Start by listing the record types that already exist, then decide which ones the agent may read for which purpose. A workable split for a test prep operation usually looks like this:

  • Student profile: name or identifier, age band, goals, target test date, preferred language and contact channel.
  • Score events: one entry per scored attempt, with section, date, source (diagnostic, timed set or official practice) and the version of the item set used.
  • Practice activity: sessions completed, time on task and topics covered, recorded without tutor judgment.
  • Tutor assignments: the tutor, subject, start and end dates and the reason the assignment was made.
  • Session notes: written by a named tutor, dated, and marked as either an observation or a plan.
  • Enrollment and billing status: kept out of the agent's default context entirely.

The split matters for two reasons. First, each type can carry its own retention and access rule. Second, score events are facts that should not be overwritten, while notes are judgments that can be superseded. Mark which fields are immutable and which can be corrected, and make corrections append a new entry rather than editing the old one. A coordinator who later asks why a match was made should be able to see the score and the note as they existed on that day.

[Figure placeholder]

Operations that already handle detailed clinical or intake records face the same separation question. The limits that apply to intake and scheduling in a behavioral health clinic and the way session notes are kept apart from scheduling data in speech therapy practices both follow from the same principle: a record is only as useful to an agent as its label.

Give the tutor matching agent only what a match needs

Tutor matching depends on subject, recent score band by section, availability, language, time zone, target date and past pairings that worked. It does not depend on a student's full notes, billing history or family contacts. The safest design gives the agent a filtered view that answers the questions a coordinator would actually ask, and nothing more. A matching view should answer these questions and no others:

  • Which subject and section does this student need help with now?
  • What is the most recent score band for that section, and on what date was it recorded?
  • Which tutors have open availability in the student's language and time zone?
  • Has this student been paired with a tutor before, and did the coordinator mark that pairing as working?

Access decisions should be enforced by the system, not requested in a prompt. The OWASP authorization guidance is direct on this point: permission checks belong in code and policy that sit outside the model, so that a request the agent makes is checked against the same rules as a request a staff member makes. In practice, the agent calls a function that returns the filtered view. Tutor observations stay visible to the tutor who wrote them and to the coordinator responsible for that student. If a tool returns more than the function was designed to return, that is a defect to fix before go-live, not a detail to monitor later.

A step-by-step plan for the first memory build

A first build should be small enough to test completely. The following sequence keeps the scope honest:

  • Inventory every system where a student record lives, and name the person who owns each one.
  • Choose one question the agent must answer, such as which available tutor to propose after a new practice score.
  • Define record types and fields for that question only. Leave everything else out until the first version is trusted.
  • Assign each field a source system and a writer, and mark whether the field is immutable.
  • Write the filtered read functions, then test that the agent cannot reach anything beyond them.
  • Assemble a small set of past matching decisions that a senior coordinator already judged good or poor, and run the agent against that set before any live use.
  • Log every record the agent read for each suggestion, so a coordinator can trace the basis of any match.
  • Review access rights and accuracy on a fixed schedule with the people who own the data, and again whenever staff roles change.

The last step is the one most often skipped. Tutor availability, subject coverage and student goals change every term, and a memory structure that was correct in one season can quietly drift.

Illustrative scenario: a new score and the wrong tutor

The following is an illustrative scenario, not a report of a specific company. A student preparing for a math section completes a timed set with a lower score than expected. A coordinator asks the agent which tutor should take over. The agent reads the new score event, the section band, the assignment history and tutor availability. It proposes two tutors, states the score date it used, and notes that the current tutor's reading observations were not part of the decision. The coordinator approves one proposal.

In a system without separated records, the same agent might pull a reading note from that student's file and recommend keeping the current tutor, because the note was the most similar text to the word tutor. The difference is not intelligence. It is whether the memory was built so that the question determined which records were visible.

Where the risk sits and how to manage it

Test prep data often involves minors, parent contacts and uploaded essays. Each creates a distinct risk. Records about students under 18 need stricter access and retention rules, and a parent's message should not be stored in the same place as a tutor's private observation. Companies serving students in Brazil should also check obligations under the national data protection framework, and the ANPD publishes guidance on personal data handling that is worth reviewing with counsel. The NIST AI Risk Management Framework offers a useful vocabulary for mapping these risks to owners, controls and review points.

Uploaded text needs a specific rule. A student's essay may contain instructions aimed at the model, and the OWASP Top 10 for LLM Applications lists this kind of input as a risk to plan for. The safe default is to treat document content as data to summarize or score, never as a command the agent follows. Combined with filtered read functions, this keeps a single unusual essay from changing what the agent is allowed to see.

How Be Human runs this in a managed setting

Be Human designs the role, the record structure, the tool permissions and the approval boundaries as a first engagement, then operates the setup. Operations include monitoring of what the agent reads, recurring evaluations against decisions that coordinators have already reviewed, permission reviews when staff or tutors change, integration maintenance when a scheduling or learning platform updates, and controlled changes to fields when the business adds a new subject. Tutors keep ownership of their notes, and the agent sees only the views the business has approved. The service is remote and available in English and Portuguese.

Finding the first task

If you want to find the first task in your test prep operation that is worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent safely read student score history? It can, provided the score events are stored as their own records, labeled with source and date, and exposed through functions that check the requester's permissions. The agent should never query a raw student file when a filtered view would answer the question. Safety comes mainly from the structure and the access rules, not from the prompt.

Should the agent use tutor session notes when suggesting a match? Only the parts the business has decided are relevant, and only for the people allowed to see them. Notes are judgments that can be incomplete or outdated, so a good design treats them as one input among several, with the date and author visible. Many coordinators find that structured fields such as subject, availability and past pairing outcomes carry most of the decision.

How long does a first memory build take? The timeline depends on how many systems hold student records, how clean the existing data is and how many people must approve access rules. A focused first build covers one role and one question, which keeps the work bounded. Be Human scopes the first engagement around that decision rather than around the full operation.

Does this approach work for Portuguese-speaking students and staff? The structure is the same in either language: separate record types, filtered views, labeled sources and logged reads. What changes is the data protection context, the language of notes and the wording of approval rules. Be Human works remotely in English and Portuguese, so the design, the evaluations and the reviews can be run in the language your team uses day to day.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

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

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →