Technology
Responsible AI

AI agents in admissions, done responsibly

Jonathan Sabarre · , updated

A prospective student in another time zone has a question at two in the morning. An admissions officer has forty unread emails, most of them asking where to find the same page. AI agents can help both of them, but only if they are built with care.

The opportunity is real, and so is the risk

Admissions teams spend a large part of their week answering questions that already have good answers: entry requirements, application steps, accommodation, visa basics, open day dates. Applicants want answers quickly and in a form that suits them, whether that is a chat window, a phone call or a conversation in their own language.

Voice and chat agents can help close that gap. They can also do real harm if they are careless. An agent that invents a fee, misstates a deadline or pretends to be a person damages trust in the institution, and the applicant pays the price. Universities owe a duty of care to the people who apply to them, and that duty continues when the conversation is automated.

The principle I keep returning to is simple: automate the work, not the judgement.

Say it is an AI, at the start

The first rule is disclosure. Each conversation should open by making clear that the applicant is speaking to an AI, in plain words rather than small print. People make different choices when they know what they are talking to, and they are entitled to make those choices. Transparency obligations under the EU AI Act point the same way for systems that interact with people, which matters for UK institutions recruiting in Europe.

Offer a person

Disclosure is only meaningful if there is an alternative. A good agent offers a route to a person, recognises when a question is outside what it should answer and hands over with the conversation so far, so the applicant does not have to repeat themselves. Some topics should go to a person by design: personal circumstances, wellbeing, complaints, disability support and anything about an individual application decision.

Grounded answers, or no answer

The most important design choice is what the agent is allowed to say. A responsible admissions agent answers from the university's own approved content: the prospectus, programme pages, policies and published dates. Where it can, it shows the page it drew from so the applicant can check.

Just as important is what it is designed to decline. If the answer is not in the approved content, the agent should say so plainly and offer a person, rather than produce something plausible. Fees, deadlines, scholarship criteria and entry requirements are exactly the facts that should not be guessed. Test this deliberately with questions the content does not cover, and record how often the agent declines correctly. No model gets this right in every case, which is why the route to a person matters.

Consent and the rules on contact

Agents that collect contact details or make outbound calls and messages bring data protection and electronic marketing rules into play. Under UK GDPR, applicants should know what is being recorded and why; under PECR, marketing calls, texts and emails need the right basis, usually consent for individuals. In practice: clear information at the start of a conversation, explicit opt-ins before any follow-up and an easy way to say no. If a call is recorded, say so before it starts.

Keep records you can check

If an applicant later says "your agent told me the deadline was Friday", the institution needs to be able to check. Each response should be traceable: which version of the instructions was in use, which model produced it, what sources it drew on and what it said. Keep those records for a defined period, tell people they exist and give them a way to ask for deletion. Records are there to investigate, correct and learn, and to show your working to a regulator or governing body.

Fairer access, not just faster access

There is a strong equity case for well-built agents. Applicants who live far from the UK, who work during office hours or who are more comfortable in another language have historically found it harder to get answers. An agent that is available across time zones and can converse in an applicant's own language can narrow that gap.

Equity has to be designed in. Test agents with the accents and phrasings real applicants use, offer both voice and text, and make sure the route to a person works for everyone, not only for those who ask during UK office hours.

Staff time kept for judgement

The best argument for admissions agents is not cost. It is what happens to staff time. When routine questions are handled well, admissions officers have more time for work that needs a person: the applicant with an unusual qualification, the student whose circumstances have changed, the family that needs reassurance, the borderline case that deserves a careful look.

Decisions about admission should stay with people. An agent can explain a process, point to information and help someone reach the right team. It should not decide, or recommend, whether someone is offered a place.

A short checklist

  • The agent says it is an AI at the start of each conversation.
  • A person is on offer, with the context passed on.
  • Answers come from approved content, with sources shown where possible.
  • Questions outside the approved content are declined and tested for regularly.
  • Consent is clear, recorded and easy to withdraw.
  • Responses are logged, retained for a stated period and deletable on request.
  • Admissions decisions stay with people.

How our voice demo applies this

The Altiora voice demo at altiora.university/talk is a website demonstration, not a deployable admissions service. Demo, currently paused The assistant says it is an AI at the start of the call and points people to our contact form for a person. Call audio recording is switched off; the voice provider keeps a call transcript, which we delete on request. It is not connected to any university's records. Our AI transparency statement lists each provider involved and where data is processed.

Planning an applicant assistant? If you want to test disclosure, grounding and handoff with your own approved content, talk to us about Altiora Voice. We will be clear about what is available now and what is still in development.

Sources and further reading