A Business Academy Gave Every Course Its Own AI Assistant, and Its Own Rules

A European business academy teaching business, finance and IT programs runs 209 assistants in production, each one scoped to a single course or department and grounded only in that scope’s own documents.

Across the life of the account the estate has taken 95,028 questions and carried at least one question on 1,076 of 1,084 days. Teaching staff have made 628 revisions to assistant instructions. In the latest quarter, of the questions the academy’s own material covers, 97.1% of the answers landed well.

The academy shared all of this anonymously.

A rule that changes between semesters cannot live in one assistant

A first-semester finance student wants to know whether a late hand-in means a lower grade or a re-sit, and the answer depends on which semester’s regulations she is under. A lecturer in business law answers that same question forty times a term, then answers a different version of it for the second-semester cohort whose rules changed.

A study adviser in the counseling office fields everything nobody else caught, for programs she does not teach.

The academy already had the answers written down. They sat in course plans, semester descriptions, assignment briefs and study regulations, one set per program per intake, revised whenever a program was revised.

A single assistant trained across all of it would have been worse than none, because it would have blended two semesters’ rules into one confident answer. An answer drawn from the wrong semester leaves a student relying on a rule that was never hers.

What got built: one assistant per course, and per intake

Each assistant is grounded in its own course’s material and nothing else, so the scope of the answer matches the scope of the rule. The same subject runs one assistant for first semester and another for second, because the regulations differ between them.

Department assistants run alongside the course ones. Student counseling and continuing education each have their own, scoped to a department’s guidance rather than to a single course.

And all 209 live assistants came from academic staff rather than from a central integration project. No learning platform was replaced, nothing was migrated, and source documents went in as the teaching staff had already written them.

Five things the academy would not compromise on

An academy carries an obligation a general knowledge base does not. What it tells a student about a deadline, a re-sit or a grading rule is what that student will rely on, and the rule that applies is the one written for their program and their intake.

So scope had to follow the regulation rather than the subject. Business law in first semester and business law in second semester are one subject and two rule sets, and they needed two assistants.

Teaching staff had to build their own, because a central team building two hundred assistants would have become a queue. Source material had to go in unchanged, so the academy would not end up maintaining a second version of every course plan.

An assistant had to stop at the edge of its material rather than reaching past it. And nothing else in the stack could move.

The capability that decided it: scoping as a first-class thing

Each course assistant is grounded in that course’s own documents and no others, and that boundary is the entire product. Platforms built around a single institutional knowledge base could not express it, and the alternative was asking staff to trust a blend of two semesters’ rules.

Building hundreds of narrow assistants is a normal shape here rather than an edge case, which is what let the estate grow a course at a time. The first assistants covered research methodology and the high-traffic shared modules every program sends students through.

When a program runs again under revised regulations, the new cohort gets its own assistant rather than an edited old one. Departments followed once the course assistants were established.

The results across three years of teaching

Read the 1,076 first. Across 1,084 days the estate went quiet on eight of them, through vacations, between semesters and across three intakes. A launch spike would have decayed long before the third intake arrived. The estate has taken 95,028 questions in all, with the busiest single assistant carrying 15,897 of them on its own.

Separately, in the latest quarter, 97.1% of the answers to questions the academy’s own material covers landed well, measured across 1,543 in-context questions. Questions that arrive on subjects no course’s documents cover are excluded from that figure rather than counted against it.

They are recorded instead as a list of what students want in writing and the academy has not written down yet, which is a more useful thing to hand a program director than a lower number would be.

What 405 minus 209 means

Just under half the assistants ever created are no longer live. For an estate scoped per intake, that is the design working rather than attrition. A semester ends, its assistant retires with it, and the next cohort gets one built against the regulations that now apply.

The 628 revisions to assistant instructions tell the same story from the other side. Each one was made by the person who owns the course rather than by a central team, which is the only way a number that size accumulates at all.

What this means for you

Count your rule sets, not your documents. The right number of assistants is the number of places where the same question has a different right answer, and if your answers change by cohort, by region or by contract, one assistant cannot hold them without flattening the difference.

Put the build where the knowledge is. Two hundred assistants exist here because two hundred were never anyone’s central project.

The people who know what a course actually requires are the people who teach it, and a platform that needs a specialist in the middle will cap your estate at whatever that specialist can get through.

The finance student asks her own course and gets her own semester’s rule. The business law lecturer answers the question once, into the assistant, rather than forty times to forty students.

The study adviser spends her time on the cases that were never in a document at all, which is the part of her job nobody could have automated.

Measure	To date
Assistants in production	209 of 405 created
Questions asked, raw count across the estate	95,028
Days with at least one question	1,076 of 1,084
Revisions to assistant instructions	628
Busiest single assistant, questions	15,897
Satisfied answer rate, latest quarter	97.1%
In-context questions behind that rate, latest quarter	1,543

Figures from CustomGPT.ai product usage records. Shared anonymously at the customer’s request.