Round-the-clock enrolment answers grounded in an online university's own content

The enquiry an online university can’t afford to get wrong

An online university competes on its enquiry pipeline before it competes on anything else. A prospective student is weighing two or three providers at once, and the one that answers first, clearly, often wins the enrolment. The questions repeat across every intake. What are the entry requirements. What does the course cost. When is the census date. Does my diploma earn credit. How do I actually enrol. The answers all exist in the handbook and the policy library. The problem is getting them to the student at the moment they ask.

The volume is uneven and the timing is awkward. Enquiries spike in the weeks before each intake and again as census dates approach, and a large share arrive after hours, often from international students enrolling from another time zone. Many of those students are not asking in English. The enrolment team cannot staff every hour in every language, so enquiries sit in a queue, and a student comparing providers does not wait. A slow answer is a lost enrolment, and a wrong answer is worse.

The stakes are higher than a typical support line. Student-facing information has to be accurate under the Higher Education Standards Framework, which the Tertiary Education Quality and Standards Agency (TEQSA) enforces. International-student information sits under the ESOS Act and the National Code. Get a fee, an entry requirement or a census date wrong and the consequence is not a bad review, it is a student who enrols on the wrong basis or misses a deadline that costs them money. Most providers have tried an off-the-shelf chatbot for this and quietly retired it, because it could not be trusted with the answers that matter.

Why retrieval-augmented generation, not a trained chatbot

The build that works is a student enquiry agent on the web-chat channel, grounded in retrieval-augmented generation (RAG) so every answer comes from the university’s own published content rather than a model’s memory. The student asks in their own words. The agent retrieves the matching clause from the live handbook and policy index, phrases a plain-language answer, and cites the source it used. Anything outside its scope goes to a person.

RAG headlines the build because it is the only architecture that stays correct as policies change. A chatbot trained on a snapshot of the handbook goes stale the moment fees, entry requirements or census dates update at the next intake, and then it gives wrong, high-stakes answers with full confidence. A RAG agent retrieves from a live index instead, so when the source is updated the answer updates with it. The supporting pieces sit around that core. An OpenAI GPT model turns the retrieved passage into a clear answer within tight instructions on scope and tone, and the service runs on Microsoft Azure in an Australian region so student data stays onshore. The model writes the sentence. Retrieval supplies the fact.

The scope is set deliberately. The agent answers information requests, course, fees, entry, key dates, census dates, credit transfer and the enrolment steps, and triages everything else. Special consideration, financial hardship, a complaint or anything requiring a human decision is routed to staff, with the conversation summarised so nobody starts again.

Building it, and where it got hard

The model was rarely the hard part. The friction lived in the content and the languages, and two examples stand in for the rest.

The first was staleness. Published policies change every intake. Fees move, census dates shift, entry requirements are revised, and a snapshot-trained bot keeps answering with last term’s numbers. That is exactly the kind of error a future student acts on, and exactly what the Higher Education Standards Framework expects a provider to get right. The fix was to ground the agent on a live policy and handbook index rather than bake answers into the model. When the source document changes the retrieved answer changes with it, the agent cites the clause it used so a staff member can verify it, and where a policy is missing or ambiguous it refuses to guess and offers the right team instead. Refusing to answer is sometimes the correct answer.

The second was language. International students often enquire in their first language, and a naive translation of a policy is a quiet way to be wrong. A loosely translated entry requirement or a visa-linked condition can mislead a student in a way that has real consequences under the ESOS Act and the National Code. The fix was to keep the English policy as the source of truth and translate the explanation around it, not the rule itself, so a multilingual answer still maps back to the exact cited clause. We tested the agent against real enquiry transcripts across languages before launch, and we review escalations and low-confidence conversations each week.

Two constraints shaped the rest. Student enquiry data is handled under the Privacy Act 1988, so personal details are minimised before any conversation is logged and enquiry data is not used to train the model. And because a wrong enrolment answer carries real weight, the agent is tuned to escalate early rather than over-reach, with a clear handover to staff the moment a question needs a decision.

What changed

In a representative deployment the agent resolved around 58 per cent of inbound web-chat enquiries without a staff reply, covering the course, fees, entry-requirement and key-date questions that make up the bulk of the queue. Close to 40 per cent of conversations started outside business hours, many from international students in time zones where the enrolment team was offline. Every policy answer linked back to the handbook clause it came from, and on the questions it could not ground the agent handed over to staff with the conversation summarised rather than guessing.

These figures are illustrative. They describe the pattern we see rather than a published result for a named university. The shape is the point. The routine, after-hours and multilingual enquiry load comes off the enrolment team, prospective students get a fast, grounded answer in their own language at the moment they are comparing providers, and the answers that carry real consequences stay tied to the university’s own current policy.

Where this fits

A student enquiry agent is one application of our AI Agents service, built on a retrieval-augmented core, for an Australian online university. It is a contained, high-volume problem that a grounded agent is genuinely suited to, and a sensible first step before anything more ambitious in the student lifecycle. If your enrolment team is wearing the enquiry load, the place to start is to map your highest-volume questions, point the agent at your live handbook and policies, and decide where a person must stay in the loop.