A large vocational and certificate college teaches thousands of students a year, and its course coordinators spent their days answering the same handful of them: when is the exam, how do I appeal a grade, am I eligible for a grant. The answers were all written down somewhere in the college's own regulations, and a person was still typing them out one enquiry at a time.
Every answer a student needed already existed. Exam dates, appeal procedures, grant eligibility, attendance rules, all of it sat in the college's regulations and its student system. What did not exist was a way for a student to get to it, so they did the obvious thing and asked a person.
Hundreds of enquiries a day landed on the back office and the course coordinators, and most of them were repeats. The coordinators were hired to look after students, and they spent their time looking things up and pasting them into replies. Anything that needed real judgement, a student in trouble, a case that did not fit the rules, waited in the same queue behind "when is the exam".
Students felt it as slow answers to simple questions, which is the fastest way to make someone who is already working and studying feel like an administrative burden. The college felt it as expensive hours burned on copy-and-paste, hours that were not going into keeping the students it had or recruiting the next intake.
A service agent grounded only in the college's own sources, allowed to answer what it can prove and required to escalate what it cannot.
The agent retrieves from the college's official regulations, procedures and course material before it writes anything, so a student gets the actual rule that governs their case rather than a plausible-sounding paraphrase of it.
Hard boundaries stop the agent inventing. Where the retrieved material does not cover the question, it says so and moves the student to a human instead of filling the gap, which is what makes an official answer safe to publish at this volume.
The agent is wired into the college's business logic and student records, so eligibility, deadlines and status come back for the person asking. "Am I eligible for a grant" gets an answer about them, which is the difference between a chatbot and a service desk.
For anything with a real decision in it, a request to withdraw from a course, an appeal, a special case, the agent stops and assembles a summary of the student's file, then hands it to a coordinator who decides and signs. The AI drafts; a person owns the outcome.
Most of these students work. Their questions come up at night and at the weekend, when the office is shut and the enquiry used to wait until Sunday. The agent answers immediately, at any hour, from the same approved sources the office would have used.
Questions the agent could not ground are logged as gaps in the college's own documentation. The result is a running list of what the regulations do not actually cover, so the source material gets better instead of the model being asked to guess harder.
The repeat questions resolve on their own; the coordinators keep the cases that need a person.
Time spent handling support and registration enquiries fell by 60%. The repeat questions that used to fill a coordinator's day are answered from the regulations in seconds, and what reaches a person now arrives with the student's file already summarised.
Student satisfaction rose 40%, driven by accurate answers available immediately rather than after a queue. The hours the coordinators got back went where they earn: keeping current students enrolled, and getting the next intake registered.
A student services desk cannot afford a confident wrong answer about a deadline or an eligibility rule, which is why the grounding and the guardrail matter more here than the fluency does. Every reply traces back to the college's own regulations and the student's own record, anything outside that goes to a coordinator with the file already summarised, and the questions the agent cannot ground come back as gaps in the documentation. It is the same grounded-answer discipline behind our tax advisory copilot, aimed at thousands of students asking about their own status rather than partners researching a legal position.
A veteran Israeli university was losing hundreds of first-year students to silent dropout. Early-warning student profiles flagged risk weeks ahead and cut first-year dropout by 25%.
Read the case → Smart facilities · Higher edA large university burned energy on empty halls and lost research to lab failures. Smart scheduling and predictive maintenance cut energy and upkeep ~18%, with zero lab downtime.
Read the case →
Reef TRH