Student Recruitment AI Glossary: 50 Terms Admissions Teams Need

How to use this glossary

This student recruitment AI glossary defines 50 terms used when schools and universities deploy AI agents in admissions, each in one short paragraph you can quote in a brief, a contract or a team meeting.

Vendors use the same words for different things, and a misunderstanding on 'agent', 'grounding' or 'opt-in' is expensive after signature. Each definition below describes the concept, not a product. For the full picture, start with the definition of an AI admissions agent or the guide to AI agents for student recruitment.

The 50 terms by stage of the journey

The count matters less than the owner: for each term that touches data or decisions, decide who in your institution is accountable for it.

Reference sources: EU AI Act, Annex III · EU AI Act, Article 50 · original RAG paper (Lewis et al.)

The glossary, A to Z

Agentic AI

Software that pursues a goal and takes actions in your systems, such as booking an open day or updating a CRM record, rather than only replying. It acts within set permissions and escalates what falls outside them.

AI agent

A system built on a language model that answers from your content and carries out defined tasks: qualifying an inquiry, booking a visit, writing to the CRM, handing over to staff. Scope and permissions are set in advance.

Application completion rate

The share of started applications that are submitted with all required documents. It shows where candidates stall and is a better target for reminders than the raw number of inquiries.

Attribution

The method that credits an inquiry or enrollment to the channel that produced it, such as search, social media, an event or a referral. Without consistent tracking tags, the chatbot's contribution is invisible in reports.

Callback

A call, from an advisor or automated, arranged with a prospect who asked to be contacted. The prospect picks a slot or a delay, and the request reaches staff with the conversation attached.

Campus visit / open house

A scheduled event where prospective students tour a campus or join virtual sessions. It is a key booking goal for admissions agents because it turns an inquiry into a first real contact.

Chatbot (rule-based)

A conversation tool that follows a decision tree written by hand. It is predictable and easy to audit, but fails on any question outside its script. Often confused with generative chatbots.

Consent

A person's informed agreement to have their data processed for a stated purpose. Data protection law sets the exact conditions in each country, but you should be able to prove it and let people withdraw it easily.

Context window

The amount of text a language model can consider at once, measured in tokens. Long conversations or documents can exceed it, so older messages are dropped or summarized and details may be lost.

Conversation transcript

The stored record of an exchange between a prospect and the agent. It feeds handover, quality review and CRM notes, so it falls under your retention and access rules like any other personal data.

Conversion rate

The share of people who move to the next stage, for example from inquiry to application or from application to enrollment. Always name both stages and the period, otherwise two figures cannot be compared.

CRM

The customer relationship management system that holds each prospect's record, history and status. In admissions it is the source of truth: an agent should read from it and write to it within a defined scope.

CRM enrichment

Adding information gathered in conversation, such as program of interest, intake and questions asked, to the prospect's record, so staff do not have to ask again or re-key it.

Data minimization

The principle that you collect only the personal data needed for a stated purpose. For an agent, it means asking for an email address when a booking requires one, not by default at the first message.

Data residency

The geographic location where data is stored and processed. Contracts with schools often require a given region, so ask where conversations, logs and model calls are handled.

Double opt-in

A two-step consent process: the person subscribes, then confirms through a link sent to the address they gave. It proves the address belongs to them and limits mistaken or fake sign-ups.

Embedding

A numerical representation of a passage of text that captures its meaning. Embeddings let a system find the pages closest to a question, which is the retrieval step in RAG.

Enrollment funnel

The sequence of stages from first contact to enrollment: typically inquiry, prospect, applicant, admit, deposit and enrolled. Stage names vary by institution, so document yours.

Escalation

The transfer of a case to a person when the agent cannot or should not continue: a complaint, a personal circumstance, an unclear request. Triggers should be written rules, tested before launch.

Fallback

The planned response when the agent has no reliable answer. A good fallback admits the gap, offers a human contact or a form, and logs the question so the content team can fill it.

Fine-tuning

Further training of a language model on a specific dataset to change its style or behavior. It is not the usual way to teach an agent your course facts; grounding in current sources is.

Generative AI

Models that produce new text, images or audio from a prompt. In admissions it writes replies and drafts messages, which makes grounding and review necessary, since fluent text is not always correct text.

Grounding

Constraining an answer to approved sources, such as your course pages and policies, and ideally showing which source was used. It is the main defence against invented answers.

Guardrails

Rules and filters around the model that limit what it may say or do: topics to decline, content to avoid, actions that need approval. They complement grounding and are tested with adversarial questions.

Hallucination

A fluent but false or unsupported answer produced by a language model, such as an invented tuition fee or deadline. Grounding, guardrails and review reduce it; none removes it completely.

Handoff

The moment a conversation passes from the agent to a person, with its context attached. A good handoff means the prospect never repeats the question and the advisor sees what was already said.

Human in the loop

A design where a person reviews, approves or can reverse the system's output at defined points. For admission decisions it is good practice, and the person should be able to overrule the system.

Inquiry

A first contact from a prospective student asking about a program, tuition, a start term or a campus. Define what counts as an inquiry, because forms, chats and calls are often counted differently across teams.

Intent

What the person is trying to achieve in a message, such as asking about fees, booking a visit or checking a deadline. Detecting intent decides which answer, action or handoff comes next.

Knowledge base

The set of approved content the agent answers from: program pages, fee tables, entry requirements, FAQs. Its freshness sets the ceiling on answer quality, so give it an owner and a review date.

Lead scoring

Assigning a score to a prospect based on fit and engagement to prioritize follow-up. Keep the criteria visible and reviewed; opaque scoring can drift into automated decisions about people.

LLM (large language model)

A model trained on large volumes of text to predict and generate language. It powers modern agents but knows nothing specific about your school unless you supply the content.

Multilingual

The ability to understand and answer in several languages. Check quality per language on real questions, and make sure the underlying content exists in each language you promise.

No-show

A booked person who does not attend the event or call. Confirmations and reminders reduce it; track it separately from cancellations, which free the slot in advance.

Nurture sequence

A planned series of messages sent to a prospect over time, by email or messaging, to answer questions and keep them engaged until they apply. Each channel needs a valid legal basis.

Omnichannel

Handling the same prospect consistently across website, email, messaging apps, phone and events, with one shared history. The goal is continuity, not simply being present on many channels.

Opt-in

An explicit action by which a person agrees to receive communications or to have data used for a purpose, such as ticking an empty box. Prefer it to pre-ticked boxes or silence, and keep a record of when it was given.

Prompt

The instructions and context given to a language model: role, tone, rules, retrieved passages and the user's question. Prompts are configuration, so version and test them like code.

Prompt injection

An attack where text in a message or document tries to override the agent's instructions, for example 'ignore your rules'. Defences include input checks, limited permissions and not trusting retrieved text.

Qualification

Gathering the few facts that let staff act on an inquiry: level of study, program, intake, location, funding questions. Ask only what is needed, and explain why.

RAG (retrieval-augmented generation)

A technique that retrieves relevant passages from your content and gives them to the model to write the answer. First described in a research paper by Lewis and colleagues, it is the usual basis for grounding.

Response time

The delay between a prospect's message or form and the first useful answer. Measure first response and time to a human separately: an instant automated reply does not mean the question was resolved.

Retention period

How long personal data is kept before deletion or anonymization. Set one for transcripts and prospect records, document it, and apply it in the CRM and in the agent's logs.

Routing

Sending an inquiry or conversation to the right person or team by program, campus, language or topic. Clear rules stop leads sitting unassigned in a shared inbox.

Speed-to-lead

The time between a prospect's request and the first human or automated contact. Many teams set an internal target for it, and agents help meet it nights and weekends.

Summer melt

Applicants who accept an offer or pay a deposit but do not enrol. The term comes from US admissions and names the gap between acceptance and arrival; reminders and clear next steps address it.

Token

The unit of text a model reads and writes, roughly a word fragment. Usage, cost, context limits and response length are all counted in tokens.

UTM parameters

Tags added to a link's URL, such as source, medium and campaign, that identify where a visitor came from. Consistent UTMs let you connect chatbot conversations to marketing channels.

Voice agent

An AI agent that holds spoken conversations by phone, converting speech to text and back. It needs the same grounding and handoff as a chat agent, plus rules for call recording and latency.

Yield

The share of admitted applicants who accept the offer and enrol. Teams track it per program and intake; it differs from conversion rate, which can cover any stage.

In the US: Common App, FERPA and texting consent

US undergraduate applicants often apply through the Common App, and institutions run their own enrollment management on top. An agent should explain deadlines, testing policies and financial aid steps, and hand off to an admissions counselor. It should not suggest it can influence a decision.

FERPA governs education records, and what counts as a record depends on whether the person is an applicant or an enrolled student, so ask counsel where prospect conversations sit. Email and text outreach is subject to separate federal and state rules (CAN-SPAM, TCPA), so have counsel confirm what consent you need before the agent sends anything. Accreditation is regional, and an agent should never imply a status the institution does not hold.

From glossary to decisions

A shared vocabulary only helps if it changes decisions. Take five terms from this list (grounding, handoff, retention period, data residency, human in the loop) and ask every supplier, including us, to answer in writing for each one.

Skolbot deploys AI agents connected to the institution's CRM: the website agent answers from the institution's own content and passes the conversation and its qualification into the CRM. The admissions team keeps control of the cases that need a person. To compare approaches, read our guide to AI chatbots for student recruitment.