Keeping an Admissions AI Agent's Knowledge Base Up to Date

An admissions agent is only as accurate as its knowledge base

An admissions chatbot's knowledge base goes stale exactly where mistakes hurt: fees, entry requirements, application routes and dates. Keeping it current is an organisational job more than a technical one: one source of truth, an owner for each type of content, a review rhythm tied to the recruitment cycle, and a short route for urgent changes.

A note on method: no paid keyword-volume tool was available, so this article was scoped from public search results. Every external source cited was opened. Nothing here comes from Skolbot or its customers, and we state no figures beyond sourced ones.

Why the knowledge base ages faster than the website

In a retrieval-augmented set-up, the agent retrieves passages from an index of your documents and writes its answer from them. The original retrieval-augmented generation paper presents this as a way to reach external knowledge and keep it updated. The benefit is practical: change a document and the answers can change, with no retraining.

The risk is the same fact seen from the other side. If last year's prospectus stays indexed beside this year's, the agent can quote either with equal confidence. Our guide to chatbot hallucination guardrails covers invented answers; this article is about answers that were right last cycle and are wrong now.

Typical causes:

• Duplicates: an old PDF, a blog post, an exported FAQ. Nothing says which one counts.
• Orphan copies: a fees table pasted into a working document nobody maintains.
• Changes outside the process: an entry requirement altered in a committee, announced by email, never updated in the index.

Set the source of truth and the document hierarchy

For each kind of information, decide which document is authoritative. The agent should read that document, not a copy.

Then write a conflict rule. If two passages disagree, the agent does not pick one. It flags the clash to your team and, meanwhile, offers the applicant a human adviser.

A review rhythm tied to the UCAS cycle

Content does not age at a constant rate. Tie reviews to the stages of the cycle, not to an arbitrary quarter. For UK providers, UCAS publishes the key dates for each application cycle; use that page as your reference, and update your own pages before the agent talks about them.

1. Before applications open: a full review of fees, course pages, entry requirements, supporting documents and process pages. Archive anything tied to the previous entry year.
2. During the application period: short, frequent checks on dates and routes.
3. Offers and replies: check messages on reply deadlines, deposits, conditions and waiting lists. These are the questions applicants worry about most.
4. Results season and Clearing: vacancies and entry routes change quickly, so ownership and turnaround matter most here. If you cannot keep vacancy information current, have the agent point to the live page and a person.
5. Before arrival and after: arrival information, induction dates, accommodation; then retire what has expired and review the questions handled badly.

Owners, versions and traceability

Content without an owner goes stale. Three roles are enough for most providers:

• Content owner (fees, courses, student services) fixes the source and declares the change.
• Knowledge-base manager reindexes, removes the old document and checks the result.
• Approver, usually the head of admissions, signs off anything that commits the institution: fees, entry requirements, dates.

For each document keep an upload date, an expiry or next-review date, the entry year it covers and the owner's name. A change log answers "why did the agent say that yesterday?". Check that your platform supports this metadata and rollback; do not assume it.

Expired documents should be removed from the index, not merely down-weighted. If you must keep the history, archive it outside what the agent can search.

When a date or requirement changes overnight

A moved deadline, a cancelled interview day, a corrected fee: these arrive mid-campaign. Prepare a short route:

1. The owner corrects the source page.
2. The manager reindexes and removes the old version the same day.
3. A test question on the changed point confirms the new answer.
4. Meanwhile, a temporary instruction ("for this topic, hand over to an adviser") can suspend the automatic reply. Confirm your tool allows it before you need it.
5. The log records the time of the fix.

Set your own target for how long a wrong fact may circulate; the aim is that it never survives a working day.

Test freshness instead of hoping

Build a small question set whose correct answer changes over time: "What is the application deadline for next year's entry?", "What are the fees for the BA?", "Is there an interview?". After every significant update, rerun it and mark each answer correct, cautious or out of date. Only the first two pass.

Add trap questions that ask about last year, or about a figure that no longer exists. A good answer names the entry year and points to the source page. For the initial build of the corpus, see how to train an AI chatbot on your school's data, and for the wider picture the guide to AI chatbots in student recruitment.

Personal data and consumer law

Keep applicant files, grades and emails out of the knowledge base. It should hold publishable information. UK providers also work under consumer law: the government's advice for higher education providers states that "students have consumer rights", so treat the agent's answers as part of the information you give applicants, held to the standard of your website. Agree retention and data handling with your data protection officer.

The EU AI Act lists systems intended to determine access or admission to educational institutions as high-risk. That is EU law; UK providers recruiting in the EU should take advice. An information agent answering from published pages is not automatically in scope, but it moves closer if it ranks or screens applicants. So no screening rules belong in the knowledge base.