AI for education
In an educational institution the expensive error is not a slow answer: it is an invented date. An applicant who misses an enrolment deadline because the assistant got it wrong does not come back. We build on the real calendar and regulations, with explicit refusal.
Key takeaways
The expensive failure mode here is the invented date or requirement. An applicant who misses an enrolment window over a wrong answer does not complain: they leave.
So the rule is that every date and requirement comes from the in-force official document, cited, and when it is not there the system says so and escalates.
Where it pays most is admissions: high volume, repeated questions, documented answers and a narrow window in which the applicant decides.
We do not build it to grade or assess students. The academic decision stays with the teacher, and that boundary is defined in writing before building.
The five gaps in an educational institution
Four are in admissions and one is in teaching operations. All five have an answer written down in some institutional document.
- Deciding where to study has a short window, and whoever answers first enters the conversation.
- Requirements, costs, dates, credit transfers, scholarships. Expensive people doing repeated work.
- Error-prone manual processes at exactly the highest-volume moment of the year.
- Much of the evening and online student body asks questions when the office is closed.
- Students find out late what applies to them, and that feeds dropout.
A date is not an opinion
Almost everything an educational institution publishes is dated and versioned: the calendar changes every term, admission requirements change by cohort, regulations get amended. A system that indexes it all together answers with whichever version sat closest in vector space, and in this domain that answer costs an applicant.
The fix is the same one we apply in insurance and has the same shape: effective dates as mandatory metadata at ingestion, filtering by period before ranking, and the document’s date visible in the answer. With two calendars loaded, the system does not pick the most similar one: it picks the one in force, or says there is ambiguity and escalates.
The second rule is refusal. When the question is about an exception — an unusual credit transfer, a disciplinary case, a particular situation — the correct answer is to route it, not to approximate from the general regulation. That gets measured: the evaluation set includes cases whose only correct answer is an escalation.
The numbers before opening it to applicants
| What is measured | What it means | Why it matters here |
|---|---|---|
| Date and requirement accuracy | Against a set validated by the admissions office: whether every date and requirement matches the in-force document. | It is the headline metric. A wrong date costs an applicant a whole term. |
| Currency of the cited document | That the answer comes from the calendar or regulation of the right period, with its date visible. | With two versions loaded, a system without a currency filter answers with last year’s and sounds just as confident. |
| Escalation on exceptions | Questions about transfers, disciplinary matters or particular cases whose only correct answer is routing. | General regulations do not settle an exception, and answering from them creates an expectation the institution has to walk back. |
| Groundedness | Every statement holds up against the cited document, adding nothing. | It is what lets the admissions office defend an answer when a parent pushes back. |
We do not build systems that grade work, assess students or make admission decisions. An academic decision has a named person behind it, and a system that automates it moves that responsibility to a vendor.
We also do not recommend starting if the documentation is out of date or scattered across offices with no clear owner. A retrieval system over an old calendar hands out old dates with far more authority than a posted PDF, and that is worse than not having it.
Related services
AI chatbots and assistants
With the evaluation set run before it ever reaches production.
View serviceRAG systems
Chunking, reranking and hybrid search, evaluated with recall@k and NDCG.
View serviceIntegrations and APIs
MCP servers against the systems already in operation.
View serviceAI automation
Applied where there is volume and stable rules, not where there is expectation.
View serviceMore from the blog
Twelve questions to ask an AI vendor
Every proposal reads the same.
Keep readingThe 15 best AI development companies in Latin America (2026)
This list came out of reading the public pages of more than eighty providers across Colombia and the region.
Keep readingRAG vs fine-tuning: which one you need
The question usually arrives framed wrong, as if these were two routes to the same place.
Keep readingFrequently asked questions
How do you stop it inventing an enrolment date?
Three things, and none of them is the prompt. Effective dates go in as mandatory metadata at ingestion, so the system filters by period before ranking instead of taking the closest-looking document. The answer cites the document and shows its date, so the reader can verify. And it gets measured: against a set validated by admissions, what share of dates and requirements matches the in-force document. That number is handed over before it opens to applicants.
Can it grade assignments or assess students?
We do not build it for that, and it is a decision rather than a technical limit. A grade is an academic act with a responsible person behind it, and automating it shifts that responsibility onto a system that cannot carry it. What we do build is the administrative layer around it: requirements, dates, procedures, request status and first-line support, which is where the volume is.
Does it integrate with our LMS or student information system?
Yes, over API or an MCP server against Moodle, Canvas, Blackboard or an in-house system. With integration it can say "your application has been under review since the 12th"; without it, it can only explain how the process works. We recommend starting with the document layer — which already covers most admissions volume — and adding integration afterwards, measuring what it adds.
Is this for a school or only for a university?
It depends on query volume and the state of the documentation, not on size. A school with a concentrated admissions season and several hundred families asking the same things has a clear case. One where queries run in the dozens and the front office copes does not, and we say so on the first call.
What about student data?
Data on minors is sensitive and the design assumes it. The index holds institutional documentation — calendar, regulations, requirements — not student information. When an answer needs a specific person’s data it is fetched over API at that moment and not persisted. We use enterprise plans where data sent is not used for training, and deploy inside the institution’s infrastructure when the case requires it.
How long does it take, and when in the year should we do it?
Four to six weeks to production for an admissions scope, with something running from week one. Timing matters more here than in other sectors: it should be live before the enrolment season opens, not during it, because the peak is exactly when you cannot be adjusting things. It is quoted with fixed scope, price and date in a proposal 48 hours after the first call.
Book 15 minutes
Tell us what you are building, or what stopped working. You leave the call with a concrete answer: it can be fixed, it can be built, or it isn't worth it.