Pragmatic Revival

Walk the exhibit hall at any higher education conference this year and you will notice a peculiar uniformity. Booth after booth, banner after banner, the same promise rendered in the same sans-serif optimism: AI-powered engagement. The demo is always the same, too. A rep in a quarter-zip pulls up a university homepage — usually a fictional school with a name like Ridgemont State, its stock-photo students laughing at salads on a quad — and there, in the lower right corner, pulses a small purple orb. Click the orb. A chat window blooms. Type “what is the application deadline” and the orb, after a theatrical pause meant to simulate thinking, tells you the application deadline.

This is the product. This is, in an enormous number of cases, the entire product: a large language model, built by someone else at a cost of hundreds of millions of dollars, wearing the institution’s brand colors like a rented tuxedo, answering questions that were already answered on the FAQ page it was trained on. The industry has a term for this — the thin wrapper — and higher ed has adopted the thin wrapper with the enthusiasm of a sector that has been burned by every technology cycle since the CD-ROM viewbook and has apparently decided the solution is to get burned faster.

The diagnosis is not complicated. A chat widget bolted onto a legacy admissions site is a superficial application of a genuinely transformative technology. Worse, it is superficial in precisely the wrong direction. It attempts to automate the one interaction — human connection, guidance, the sense that somebody at this institution actually sees you — that prospective students are starving for. A seventeen-year-old navigating the most expensive decision of her life does not want a purple orb. She wants a person. The orb is a way of telling her she will not be getting one.

Here is the thesis, and it will occupy the rest of this essay: the true power of artificial intelligence in higher education does not live on the homepage. It lives in the back office, in the sediment. Universities need to stop deploying AI as a greeter and start deploying it as a structural engine — for predictive analytics, for continuous catalog auditing, for the grinding administrative work that consumes the institution from the inside. The goal is to mechanize the bureaucracy entirely, so the humans can finally do the thing only humans can do.

Section 1: Predictive Analytics and the End of Historical Yield

Consider the enrollment manager’s yield model, an artifact of institutional life as venerable and as trusted as the mace carried at commencement. It works like this: you take the last five or ten years of admissions data, you observe what percentage of admitted students in each segment historically enrolled, you apply those percentages to this year’s admit pool, and you present the resulting number to the president with the confidence of an actuary. For decades this worked, more or less, in the way that predicting a river’s flood stage from a century of gauge readings works — right up until somebody builds a dam upstream, or the climate shifts, and the historical record becomes not merely unhelpful but actively misleading.

Somebody has built the dam. Several dams, actually. The FAFSA rollout fiasco scrambled financial aid timelines in ways no historical model contemplated. The demographic cliff is not coming; it is here, walking around the admissions office, and it does not behave like the classes of 2015 through 2019. Test-optional policies rewired the composition of applicant pools. The historical yield model, in a market this volatile, is a nineteenth-century tide table handed to a ship’s captain in a hurricane. It is not that the math is wrong. It is that the ocean has changed.

The alternative is not a better spreadsheet. It is a fundamentally different instrument: an enterprise AI model trained not on what students did in 2017 but on what this student is doing right now. The raw material is unstructured and abundant. Website dwell time — not just that a student visited the financial aid page, but that she visited it four times in a week, lingering on the section about payment plans. Email cadence — the admitted student who opened every message in March and has opened nothing since April 12. The sentiment buried in a financial aid inquiry, the difference between “can you clarify my award letter” and “we were not expecting this number and don’t know what to do.” A properly built system ingests all of it and produces something the historical model never could: an individual melt probability, per student, refreshed weekly.

The operational shift is the point. Under the old regime, an enrollment team learns its class has shrunk when the fall census confirms it, at which moment the class is unshrinkable and the provost is unhappy. Under the new regime, leadership receives a daily risk assessment — a weather report for the incoming class — and can act while action still matters. A targeted aid adjustment in June. A phone call, from a human, to the student whose engagement flatlined. Proactive instead of forensic. The autopsy becomes a checkup.

Section 2: The Forensic Catalog Audit

Every academic catalog is a geological formation. Courses are deposited, layer upon layer, by the slow tectonics of faculty interest, accreditation requirements, long-departed deans’ pet initiatives, and the occasional grant that funded a course nobody remembered to sunset when the grant expired. Nothing erodes. A university founded in 1887 is carrying strata from every decade since, and if you core-sample the catalog you will find them all: the Cold War area-studies layer, the 1990s “computers and society” layer, the mid-2000s stratum of courses with “globalization” in the title. There are courses in most catalogs that have not enrolled a student in years and courses that exist in triplicate across three colleges, taught by three departments that have never spoken, each convinced theirs is the real one.

The reason this persists is not stupidity. It is scale. A provost overseeing two thousand courses does not have the time — nobody has the time — to read two thousand syllabi and cross-reference their stated learning outcomes against what the labor market is actually paying for. So the catalog review, when it happens at all, happens the way most institutional self-examination happens: a committee is convened, a semester passes, a report is produced, and the report goes on a shelf, where it joins the reports of previous committees in a stratum of its own. The catalog, meanwhile, keeps accreting. Ask a registrar how many active courses the institution actually offers and watch the pause before the answer; the honest ones will tell you the number depends on how you define “active,” which is another way of saying nobody knows.

This is precisely the kind of problem — vast, textual, tedious, pattern-rich — that machine intelligence was built for. Picture a forensic audit tool, internal, unglamorous, no purple orb anywhere in sight. It ingests the entire catalog: every syllabus, every course description, every stated outcome. It cross-references them against live workforce data — job postings, skill taxonomies, wage trends — and it produces something no committee ever has: a definitive, data-backed map of the academic portfolio as it actually exists, laid against the economy as it actually is.

What the map reveals tends to surprise people in both directions. Yes, it flags the redundancies — the three intro statistics courses, the writing-intensive seminars that are the same seminar wearing different prefixes — and yes, it exposes administrative bloat with an indifference no internal committee could sustain, because the algorithm does not have to eat lunch with the department chair. But the more interesting findings run the other way. The audit routinely uncovers hidden labor-market value in the humanities: the philosophy course whose outcomes map almost perfectly onto what employers describe when they post for compliance analysts, the rhetoric seminar that is, functionally, a UX-writing bootcamp that predates the term by forty years. The humanities do not need charity in this analysis. They need translation, and the machine turns out to be a decent translator.

The operational shift: faculty lines and instructional budgets get reallocated on the basis of empirical alignment rather than departmental seniority and the volume of whoever complains loudest at the budget hearing. That sentence will make some readers flinch, and it should. It is also how survival works now.

Section 3: Clearing the Administrative Sludge

Here is a small story that repeats itself several hundred thousand times a year in this country. A twenty-six-year-old — call her the transfer student, though she is also a shift manager, a parent, a person with exactly one free evening a week — decides to finish her degree. She requests her community college transcript. She sends it to a four-year institution. And then she waits, because somewhere in the registrar’s office, a human being must sit down with her transcript and a course-equivalency database and a set of articulation agreements last updated during a previous presidential administration, and manually determine that her Intro to Sociology is your Intro to Sociology. This takes three weeks. Sometimes six.

The transfer student does not know it takes three weeks because of staffing shortages and a byzantine equivalency process. She knows only that she has heard nothing. And in that silence, the anxiety compounds. Maybe her credits won’t count. Maybe she’ll have to retake everything. Maybe this whole idea was a mistake. Somewhere around week four, a competing institution — often a large online operator that has spent lavishly on exactly this problem — tells her within a day that her credits transfer and she can start in eight weeks. The four-year school never hears from her again and never knows why. Bureaucratic friction does not appear on any yield report. It just quietly eats the class.

The machine-learning solution here is almost embarrassingly tractable. Transcript evaluation is document processing: read an incoming PDF, extract the courses, map descriptions against the institutional database, apply the articulation rules, and generate a preliminary degree audit. A trained system does in seconds what takes a human registrar weeks — not because the registrar is slow, but because the registrar is doing four other jobs and the transcript sits in a queue. The human still reviews the edge cases. The queue simply ceases to exist.

The operational shift is a promise the institution can print on the homepage — a real one this time, not an orb: every transfer credit evaluation returned within 24 hours, guaranteed. That single sentence repositions the institution in its market more decisively than any brand refresh. You win transfer enrollments not with billboards but by deleting the wait, because the wait was never neutral. The wait was the competitor.

Section 4: Elevating Human Capital

Now the objection, which arrives on schedule in every faculty senate and every staff meeting where this subject comes up: you are describing job replacement. It is a reasonable fear with a century of industrial precedent behind it, and it happens to be aimed at the wrong crisis.

The actual crisis in higher education’s front lines is not that the jobs are being automated. It is that the jobs have become unbearable and the people are leaving on their own. Ask any dean about advisor turnover. The admissions counselor position — once a launching pad, a job people held for five years while they figured out they loved this work — now turns over in eighteen months. Exit interviews tell the same story with numbing consistency: nobody quits because they got tired of talking to students. They quit because they never got to. The job they signed up for — the relational work, the mentoring, the phone call that changes a kid’s trajectory — was buried under data entry, CRM hygiene, compliance reporting, scheduling logistics, and the endless duplicative paperwork that constitutes the connective tissue of the modern university. They took a people job and discovered it was a spreadsheet job wearing a lanyard.

Viewed through this sociological lens, enterprise AI is not a replacement mechanism. It is an absorption mechanism. The machine exists to soak up the sludge — the data entry, the routine scheduling, the report that must be filed in triplicate for an accreditor who will never read it — because the machine does not burn out, does not resent the work, and does not quit in eighteen months to go sell software.

The arithmetic is startling. Absorb the bureaucratic load and the typical academic advisor reclaims something on the order of twenty hours a week. Half a job, returned. And here is where institutional discipline matters, because those hours will evaporate into new administrative sediment unless leadership mandates otherwise. The mandate should be explicit: the reclaimed hours go entirely into relational outreach. Deep, unhurried conversations with the struggling sophomore. The proactive call to the student whose melt probability just spiked — flagged, of course, by the predictive system from Section 1, because these pieces are meant to interlock. The machine finds the student in trouble; the human sits with her.

There is a phrase for what this produces, and it sounds like marketing but is actually mechanics: automation scales the empathy of the institution. The chatbot vendors had it exactly inverted. You do not automate the human connection. You automate everything else, so the connection can finally happen.

Conclusion: The Operational Moat

Return, briefly, to the exhibit hall and the purple orb. Ask the fatal question: what happens when the college down the street buys one too?

They will, of course. They may have already. The thin wrapper’s defining feature — the reason it can be demoed in ninety seconds and deployed in a week — is that it requires nothing of the institution. No process change, no data discipline, no hard decisions about the catalog or the org chart. And anything that requires nothing of you is, by definition, available to everyone. A public-facing chatbot confers a competitive advantage with the approximate half-life of a press release. It is the higher-ed equivalent of every airline installing the same seatback screens and calling it differentiation.

Moats are not built from features. They are built from operations — from the accumulated, compounding, hard-to-replicate competence of an institution that actually works. The university that trains predictive models on its own behavioral data owns something no vendor can sell its rival. The university that has forensically audited its catalog knows things about its own academic portfolio that its competitors do not know about theirs. The university that guarantees a 24-hour transcript turnaround has re-engineered a process its peers have not even mapped. The university whose advisors spend twenty reclaimed hours a week in real conversation with real students is building loyalty that no orb, however purple, can simulate.

None of this demos well. There is no booth for it. It is plumbing, and plumbing does not photograph. But the institutions that survive the next decade — and it will be survival, for a meaningful fraction of the tuition-dependent sector — will be the ones that understood the assignment: the AI revolution in higher education is not a widget in the corner of the homepage. It is the boring, bespoke, structural rebuild of everything behind it. Execution is the moat. Everything else is an orb.