Wednesday morning, a senior lecturer in the Faculty of Commerce at the University of Cape Town opens her inbox. Forty-seven new messages since Monday. Thirty-one of them are from students. Twelve concern the same three things: a clarification on the Week 8 assignment brief, an extension request because of a death in the family, and the same question about readings on the LMS that she answered on Friday. Six are administrative — a curriculum committee chasing learning-outcome mapping, a journal asking about her revision deadline, an HR query she has now ignored for four working days. The teaching for the week starts at 14:00. The actual lecture preparation has not begun.
This is the unglamorous shape of an SA academic's week, especially in second semester. The teaching is the part everyone sees. The admin around it, for a lecturer carrying a 150-student first-year course and a 40-student honours seminar, is between fifteen and twenty hours of pure non-teaching, non-research work. Course coordination meetings, marking moderation, student support emails, ethics-clearance paperwork, departmental committee work, line-management of two tutors, the term-end mark spreadsheet, and the postgraduate students who genuinely need supervision.
I have had a fair number of conversations with academics across UCT, Stellenbosch, Wits, NWU, UJ, UWC and a handful of TVET colleges about whether AI can take some of this load. The answer is yes, in a narrower band than the hype suggests, and with a non-negotiable line around the things a university exists to do — teach, assess, examine, and certify. The line matters. Cross it and the institution starts to mean less.
Where AI actually earns its keep in faculty admin
The unglamorous, repetitive work of running a course module is exactly the kind of structured, predictable, high-volume task that automation handles well. Three patterns I have seen pay back inside a single semester.
The first is the student query triage layer. A bounded handler in front of the lecturer's course-related email, or a dedicated module address, reads incoming messages, classifies them into a small set of categories — assignment clarification, extension request, marks query, LMS access problem, general FAQ — and either answers the genuine FAQs directly from the course handbook, or routes the others to the lecturer with a one-line summary. The handler does not pretend to be the lecturer. The auto-reply is explicit that it is a course assistant providing first-line information, and that the lecturer will respond personally to anything that needs judgement. The handler never resolves an extension request itself. That decision stays with the academic.
The second is the moderation and marks-admin layer. The actual marking stays with the lecturer and the tutor team. What AI is genuinely useful for is the surrounding paperwork: pulling the moderator's spreadsheet into the right format, drafting the moderation report from the marker's spreadsheet and the model answer, generating the per-student feedback paragraph that the marker writes the bullet points for. None of this changes a single mark. It compresses the four hours of clerical work that surrounds a 150-paper mark cycle into something closer to forty minutes.
The third is committee paperwork. Meeting agendas, minutes, action lists, the curriculum committee's learning-outcome mapping exercise, the HEQC self-evaluation that comes around every five years. An AI working from the previous cycle's documents and the current term's emails drafts version one. The convenor reads, corrects, and circulates. The draft was never the bottleneck for an academic — committee work was eating Sunday afternoons because nobody had time to do the typing. Take the typing away.
Student support without pretending to teach
This is where most South African universities are at risk of doing the wrong thing. The temptation to deploy a "campus AI assistant" that answers any question a student asks is strong, especially with every vendor pitching one. I have not seen this work in practice. The university that lets a chatbot answer "How do I prepare for the test?" is one bad answer away from the front page of GroundUp.
A narrower scope works. The places I would put AI student support are:
- Administrative wayfinding — where to submit a deferred-exam application, what the late-registration deadline is, which forms a student in financial difficulty needs from the Student Funding office, how to book a study room in the library.
- LMS technical troubleshooting — the password reset, the "I can't see the Week 6 folder" problem that turns out to be a release-date setting, the Turnitin upload that silently failed.
- After-hours acknowledgement — a student emailing at 22:40 the night before submission deserves a confirmation that the message arrived and a sensible next step, not silence until 09:00.
- Pointer to the right human — for anything involving mental health, financial distress, harassment, or academic appeal, the handler's job is to get the student to the correct office today, not to attempt support itself.
Anything academic — the actual content of a course, the right way to approach an essay, how to read a journal article — sits with the lecturer or with peer tutors. Not with a model. The student is paying R75,000 a year, more at some institutions, for human teaching. Replacing the human at the point of confusion is precisely the moment they need the human most.
There is also a quieter point about access. Students from under-resourced schooling backgrounds — and SA still has many — often need scaffolding the AI is the worst at giving. The model will produce a confident, fluent-sounding answer that misses what the student actually misunderstands. A first-year tutor who knows the cohort will not.
The academic-integrity line
Universities do one thing the rest of the economy does not. They certify that a graduate has demonstrated the knowledge and skills the qualification claims. The minute that certification means less, the institution means less. AI puts pressure on this in two directions, and the institution has to be deliberate about both.
The student-facing direction is well known. ChatGPT writes plausible-sounding undergraduate essays. The marker who has not adjusted assessment design is grading a model, not a learner. The honest response is not to try to detect ChatGPT — the detectors do not work reliably enough, and Turnitin themselves have walked back early claims. It is to redesign assessment. Oral examinations. In-class writing. Iterative assignments where the working-out is the assessable evidence. Projects with a live defence. The good departments at SA universities are already doing this. The ones holding onto the take-home essay as the main assessment will look foolish in three years.
The faculty-facing direction is less discussed. A lecturer who has the AI draft the marking rubric, then has the AI mark against that rubric, then has the AI write the feedback, has not assessed the student. They have certified a chain of model outputs. Senate is entitled to expect that the human grader engaged with the human work. The line we hold for our clients: AI assists with the paperwork around assessment. AI does not produce or finalise the assessment itself.
That is the line. There is no clever workaround.
POPIA, DHET reporting, and the student data question
Students at SA universities are data subjects. POPIA applies. Layered on top are the DHET reporting requirements (the HEMIS submissions), the CHE accreditation cycle, and at most institutions an internal information-governance policy that pre-dates POPIA and needs revision. Anything AI that touches student personal information has to fit inside all four.
Three concrete things that come up every time we scope a project with a university department:
- Student names, student numbers, marks, and any health or disability disclosures do not pass through a consumer AI tool whose terms permit training on the input. The legitimate path is a contracted processor with explicit data-processing terms, or an in-house deployment.
- Anywhere a student's personal data is fed into a model for any reason, the consent basis has to be clear. The University's existing Privacy Notice probably covers analytics. It almost certainly does not cover "we put your essay through a third-party AI service to grade it". If you want to do the second thing, get specific consent first.
- Audit trails. Who accessed which student's data, when, and through what tool. This is now the first thing the Information Regulator asks for if a complaint lands. It is also the first thing your institutional Information Officer will ask at the next compliance review.
A specific note for postgraduate supervisors. PhD candidates' research material — interview transcripts, drafts, unpublished findings — is sensitive both under POPIA and for academic-integrity reasons. Putting it through consumer AI tools to "help with editing" can compromise the candidate's claim to original work, future publication options, and any research-ethics approval. Don't.
What I would not bother with
Three things I would actively push back on if a university client raised them.
A general-purpose campus chatbot. The institution looks like it is doing something. The students get a bad experience inside a week, work around it, and the project produces a poor return on a substantial spend.
Auto-grading of anything other than narrow, well-validated technical exercises — multiple-choice, code that compiles and runs against a test suite, accounting calculations with a deterministic right answer. The reputational risk of a single misgrade outweighs the time saved on a hundred correct ones.
Predictive models that try to identify which students will fail or drop out, deployed to staff without serious governance. The models reproduce structural bias from the training data. The interventions triggered by them affect young people's lives. SA's higher-education sector cannot afford to be the place where this lesson gets learned the hard way.
A realistic first project for a department
If you are a head of department or a faculty associate dean reading this and asking where to start, the answer is the same one I would give to a small law firm or a vet practice. Pick one structured, high-volume process and build it properly.
For most SA university departments that is the student query triage layer for the largest first-year module. The course coordinator agrees a small list of FAQs the handler can answer directly, with the exact wording vetted. Everything else routes to a human, faster and better-organised than before. The metric is not "AI replied to X percent of queries". It is "course coordinator now finishes the week's email inbox by Friday lunchtime". For the modules we have helped with, that is a real, repeatable outcome inside one semester.
The next step, once the triage layer is steady, is the marks-admin compression. Then committee paperwork. Then, if it makes sense for the department, a narrow LMS-troubleshooting handler.
The work that AI lets you stop doing is the work nobody became an academic to do. The teaching, the supervision, the research — those stay yours. They should.