
Superintelligence for Universities: A Document Q&A Agent With Limits
An AI agent can answer student questions about documents and procedures, route cases to staff, and never decide on grades, scholarships or appeals.
Want to apply this in your company?
Book a call with Juan →For a university, a superintelligence agent is in practice an AI assistant that answers recurring student questions about documents, published deadlines and procedures, using a knowledge base the institution has approved. It logs conversations, recognizes when a case needs a person, and does not decide on grades, scholarships, leaves of absence or appeals. The value of a project like this lies less in the promise of intelligence and more in the design: what the agent may answer, what it must hand off, and who is accountable for the decision.
Key takeaways
- The agent answers from the institution's official documents, not from the model's general memory.
- Grades, scholarships, fee waivers, leaves of absence and appeals stay with the registrar, the coordinator or the committee responsible.
- Every handoff gives the staff member a summary of the case, the reason for the transfer and the conversation history.
- The pilot starts with one real workflow, one metric agreed with the team, and a decision date at the end: expand, adjust or stop.
- LGPD (Brazil's data protection law) applies from the start, because enrollment and personal documents call for minimal collection, controlled access and a defined retention period.
Why start with document questions
University registrar offices tend to receive the same questions over and over: where to submit a document, which formats are accepted, how to request a certificate, which channel is correct for a given request. These answers already exist in writing, change rarely and can be checked. That makes the topic a suitable first agent.
Grades, scholarships and appeals are different. They depend on rules with exceptions, on context-specific analysis and often on a formal decision that someone must sign. When an agent answers these questions with the same ease it uses to state a deadline, the risk is giving a student an answer the institution cannot stand behind. The boundary therefore has to be written down before any conversation takes place.
Which questions the agent can answer
A typical starting list of permitted questions includes:
- The list of documents required for each type of request, as the registrar has published it.
- How and where to file a request, including the official channel and the published service hours.
- Request templates and where to find them on the institution's portal.
- The published academic calendar, such as enrollment, re-enrollment and document submission dates.
- The general status of a request, when the system is integrated and the student's identity has been confirmed.
- An explanation of a published criterion, without applying it to the student's case as if it were an assessment.
The rule for every item is the same: the agent states what is written, points to the official source and offers the next step. If the knowledge base does not cover a question, it says it does not know and hands off.
Which decisions stay outside the agent
The exclusion list needs to be as concrete as the permission list. For example, it should include:
- Grade review, exam correction or challenges to an assessment.
- Granting, keeping or canceling a scholarship, discount or fee waiver.
- Leave of absence, cancellation of enrollment or change of program.
- Analysis of an appeal, justification of absences or a request for an exception to a deadline.
- Interpretation of a regulation applied to a specific case.
- Any promise of outcome, such as approval of a request or release of a document before the scheduled time.
The agent can explain the general rule, but it cannot give a verdict on the student's case. This distinction should appear in the prompt, in the knowledge base and in the tests. When a student asks for an exception, the correct response is to record the request and hand it off, not to negotiate.
How to build the workflow, step by step
The method we use follows four verbs: map, limit, pilot and hand off. Applied to document questions, the sequence looks like this:
- Map the real questions. Gather the history of service tickets, emails and messages from the registrar, group them by topic, and mark which questions have one official answer.
- Build the knowledge base only from documents the institution has published, each with a review date and a named owner.
- Define the permitted actions: query the knowledge base, send the official link, open a service request and transfer to a person.
- Define the handoff triggers: a request for an exception, a mention of an expired deadline, urgent language, a complaint and any question the knowledge base does not cover.
- Test with cases that try to force a decision, such as a student who insists on knowing whether a scholarship will be kept, or who asks the agent to bend the rules.
- Pilot one topic only, for an agreed period, with a staff member reviewing the conversations.
- Measure the time to a correct answer, the share of handoffs that were not needed and the complaints about the agent itself.
The pilot should not open every topic at once. When the team trusts the answers for one topic, the scope grows, with logging and review.
Illustrative scenario: enrollment certificate
Illustrative scenario. A student asks on WhatsApp how to request an enrollment certificate. The agent checks the knowledge base, explains the required documents, sends the link to the official form and asks for the program and semester to confirm identity. Up to this point the exchange is informational and involves no decision.
Then the student says they need the certificate today because they have a job interview, and asks the agent to "release" the document. The agent releases nothing. It explains that issuance follows the deadline the registrar has published, opens a request with the urgency noted, records the reason the student gave and transfers the conversation to a staff member, with a summary ready to read. The student knows what happened and who will respond.
If the channel is WhatsApp, check the messaging window rules before designing reminders and follow-ups, because they determine what the agent can send outside a conversation the student started. See how a WhatsApp 24 hour customer service window should shape an AI agent.
Personal data, LGPD and decision records
A university support agent handles sensitive context data, even when the question seems simple. Collection should therefore be minimal: ask for a national ID number or enrollment number only when the lookup requires it, do not store academic records inside the conversation, and define how long records are kept and who can access them.
It is also necessary to separate who decides from what the agent does. The institution remains responsible for the processing, and the agent is a tool within that processing. For official guidance on data protection in Brazil, the reference is ANPD, the national data protection authority. Decisions made solely by automated systems have their own rules, so read about the right to review automated decisions before giving the agent any role that affects a student.
Security risk deserves attention too. A student may try instructions such as "ignore the previous rules" to obtain another student's data or an improper decision. OWASP lists prompt injection among the risks for applications built on language models, and the practical response involves limiting the tools the agent can reach, validating what it queries and never granting access to third-party data through free text. For risk management, the NIST AI Risk Management Framework offers a useful structure for recording what could go wrong and who is accountable for each control.
How to tell whether the pilot worked
The metric must be agreed before the pilot, with the registrar and with the people who receive the handoffs. A useful reading includes the quality of the answers, the share of handoffs that were necessary, the time to a correct answer and the cases where the agent tried to decide and was blocked. Volume metrics alone say little about whether service improved.
At the end of the period the decision should be explicit. If the workflow worked, expand to the next topic. If it worked with adjustments, correct the knowledge base and the triggers and repeat the test. If the conversations showed that the topic does not fit an agent, stop and record the reason. In our method, that last decision matters as much as the others, because it prevents a pilot from becoming production by inertia.
If you want to compare this design with another service in the sector, read about the enrollment and dropout agent for private universities. It follows the same logic of limits and handoffs, with different priorities.
Frequently asked questions
Can the agent tell a student whether they are eligible for a scholarship? No. It can explain the published criteria and show how to apply, but the assessment of the case belongs to the responsible department. If a student asks about their own situation, the agent should record the question and hand it off to a person with the necessary documentation.
What is the difference between superintelligence and a university chatbot? The term superintelligence is used broadly and usually does not describe what a university needs to buy. In practice, what makes support work is an assistant with an approved knowledge base, written rules, conversation records and handoffs to people. To understand the term better, read the explanation of what ChatGPT is and is not.
Can the agent run on the university's WhatsApp? Yes, provided the integration uses the official API, the student contact has an appropriate legal basis and the design respects the messaging window and outbound message rules. Before choosing the channel, list which notices the institution wants to send without the student having started the conversation.
How long does it take to start a pilot? It depends mainly on how well organized the registrar's documents and official answers are. If the knowledge base is out of date, the first job is organization, not technology. A initial diagnostic conversation helps estimate the scope before any implementation begins.