It’s a little after one in the morning and a sessional tutor is on essay 140 of 300, sitting at a kitchen table with a cold cup of tea and a marking rubric open in one window and the submission in the other. She gives this one four minutes, because four minutes is what the maths allows, and she reads fast, and she ticks the boxes — argument, evidence, structure — and she types a sentence she has now typed ninety times tonight: good effort, tighten your thesis. Then she opens the next one.

She can see, even at this speed, that the student had a real idea in here and buried it in the third paragraph. In another life she’d write them half a page about it. Tonight she writes tighten your thesis and moves on, and she knows the student will look at the mark, not the comment, and she knows why.

This is the quiet grief of the job. She didn’t get into teaching to become a marking machine. Almost nobody does. You become an educator because somewhere along the line a good teacher wrote something in the margin of your work that changed how you thought — and then you arrive on the other side of the desk and discover that the one thing that made you want to teach is the one thing the workload will not let you do. Full-time instructors report something like fifteen hours a week just marking, on Crowdmark’s read of faculty surveys. The feedback that actually teaches (the response, not the tick) is the first casualty of scale. Not because anyone stopped caring. Because there are only so many hours, and the essays keep coming.

The argument everyone’s having is the wrong one

Meanwhile, upstairs, the conversation about AI on campus has curdled into a single question: is this cheating? Students are certainly using the tools. In the UK, the HEPI/Kortext survey found 88% of undergraduates used generative AI for assessments in 2025, up from 53% a year earlier, with 92% using it in some form. That’s a British sample, not an Australian one, but the picture rhymes, and every dean I talk to knows it rhymes.

So the reflex is to buy a detector and go hunting. The problem: detection doesn’t work well enough to base a misconduct case on. Turnitin’s own chief product officer acknowledged a false-positive rate of around 4% at the sentence level, and a Stanford study by Liang and colleagues found AI-writing detectors flagged 61.22% of essays by non-native English speakers as machine-written. Read that twice. The tool you bought to protect academic integrity systematically accuses your international students of cheating. On a campus running under an international-student cap, that’s not a rounding error, it’s a governance incident waiting to happen. This is also why TEQSA’s posture is deliberate governance of AI, not a ban you can’t enforce anyway.

The integrity panic isn’t wrong to exist. It’s just aimed at the wrong target. It asks whether to allow AI, which is a question the students have already answered for you. The question that actually governs what you get back is narrower and far more practical: which of your systems can an agent already read and write?

What “reachable” actually means

This is where the hype meets the plumbing, and the plumbing is more honest than the hype. We spent a phase last year mapping the higher-ed software stack the way an integrator has to: not “does it have AI on the brochure” but “can a tool I build reach in and do work.” What we found sorts cleanly:

The agent-reachable APIs cluster in the LMS and CRM tiers; the SMS/SIS tier is real-but-gated, and the two government rails are consume-only.

In plain English: your learning platform (Canvas, Moodle, Brightspace, Blackboard) and your admissions CRM are open and reachable today: real read-and-write APIs you can build on. Your student-management system, where the records actually live, is real but gated: the VET and CRICOS platforms expose proper provider APIs, the big university platform hands you an API only when the vendor provisions it, and one or two are closed altogether. And the two government collections (PRISMS for CoE and attendance, TCSI for HELP reporting) are consume-only by law. Your student system reports to them through the government’s own gateway; nobody re-exposes that data. That’s not a limitation to route around, it’s the correct boundary, and it’s why the honest picture is a staged climb rather than a big-bang platform.

Myth two: it comes for the academics

The second thing people say is that if AI can mark, it’ll be used to mark instead of academics, and casualise an already casualised workforce even further. It’s a fair fear and the design answer is specific: the tool drafts the first pass; the academic shapes it, signs it, and owns the final grade. No grade is auto-assigned. Ever.

And the machine is genuinely good at the first pass. On a benchmark of 1,600 graded answer pairs across 16 courses, a fine-tuned model’s median grading error came out 44% smaller than human re-graders’ — more consistent than the people, in other words. That’s a preprint (Gobrecht and colleagues, 2024) from authors with a commercial interest, so hold it lightly. But “more consistent first draft” is exactly the right job for it: it reads every submission against the rubric, drafts criterion-by-criterion comments, and hands the academic a marked-up starting point instead of a blank box at 1am. The tutor stops being a tick-generator and goes back to being an editor of feedback, accepting, correcting, adding the half-page about the buried idea in paragraph three. The student finally gets a response. That’s not headcount coming out. That’s the craft coming back.

Myth three: “our LMS already has AI”

The third myth is the most expensive, because it feels like a reason to do nothing. Your LMS advertises AI. Your student system advertises AI. Aren’t you covered?

Not for this. Almost every vendor AI in the stack (Canvas’s, D2L’s Lumi, Blackboard’s assistant, the CRM copilots) is in-product only. It helps a user inside one screen, and it was never built to be called from outside its own walls. It cannot reason across your LMS and your student system and your Microsoft tenancy at once, because reaching across systems is precisely the thing it can’t do. The leverage isn’t the AI bolted onto each product. It’s the data APIs underneath them (Microsoft Graph, Canvas, the Australian student systems) where one agent, under your direction, works across all of it.

The climb, one rung at a time

Follow the reachability and the sequence writes itself.

Rung one, read-only. A handbook-grounded assistant answers routine student questions with citations to your own policy (“when’s the census date,” “how do I apply for special consideration”), writing nothing back. Georgia State’s chatbot handled around 200,000 messages in a single summer with under 1% needing staff, and their randomised trial found it cut summer melt about 21% and lifted enrolment 3.3 percentage points. Vendor-reported on the volume, peer-reviewed on the enrolment lift. La Trobe did a local version of the read-only idea with a Copilot Studio assistant over its own cleaned data, giving staff step-by-step procedures inside Teams. Category proof, not our client’s numbers. We’ll say that every time.

Rung two, draft-and-sign. The marking assistant above, reading LMS submission and rubric data, drafting feedback, writing nothing back to the gradebook. The academic owns the grade.

Rung three, surface-and-decide. Signals from the LMS, attendance and assessment converge, and a student drifting toward failure gets flagged around week two instead of at the exam. The tool surfaces the at-risk signal and quotes the evidence. It does not email the student. A human decides who to contact, and how. This rung is the one gated by your student-system reachability — which is exactly why we map it before we promise it.

Those magnitudes (30–50% less first-pass marking time, 80-plus percent of routine queries self-served, at-risk students visible by week two) are illustrative of what we build toward, not results we’ve banked for you.

What it’s worth, honestly

Take a 60-staff provider with 30 people who mark, carrying roughly 450 marking hours each a year. Reclaim the conservative end, 30%, and that’s around 4,050 hours back, near $243,000 of teaching capacity at a $60 internal hourly cost. Be honest about that figure, though: most of it is salaried time reinvested into teaching and retention, not cash you can bank. The cashable slice is narrower and real — the casual, paid-by-the-hour marking load you stop buying.

Then there’s the lever nobody prices properly. Under Ministerial Direction 111, international commencements are held to a National Planning Level (270,000 in 2025, rising to 295,000 in 2026) allocated provider by provider. Every CRICOS student’s progress and attendance is monitored and reported through PRISMS, and a reporting failure can put your registration at risk. In that world, a student you catch at week two and keep is a place you don’t lose — and under the cap, a place you lose is fee revenue you cannot back-fill. I’m not going to invent a dollar figure for that; it depends on your fees and your melt, and that’s precisely the number the diagnostic produces on your own data. A focused build here is a $10k–$60k job that ships in three to six weeks, not a two-year platform.

Where to start

You don’t need a transformation programme to find out if this is real. You need an honest look at two things: where your marking hours actually go, and which of your systems an agent can actually reach.

• Read the map first. We’ve written a plain-English brief on the higher-ed software landscape: which of your LMS, CRM and student systems are reachable, by what mechanism, and where the gated and consume-only boundaries really sit. → https://realmindsai.com.au/guides/higher-ed/
• Book a free 30-minute discovery call. We walk one real path and name your institution’s own baseline (your marking load, your query volume, your at-risk and melt rates) before anyone builds anything; any build is quoted fixed-scope after that. Proof on your data, not ours. → https://outlook.office.com/book/[email protected]/?ismsaljsauthenabled

There’s a tutor marking essay 140 of 300 tonight who knows exactly which students deserved a real response. The only question worth asking is whether anything in your systems is ever going to let her write it.