Making AI Work in Higher Education

AI spend is rising faster than returns

Higher education is investing in AI at a pace that would have been hard to imagine two years ago. Ellucian's third annual AI survey, fielded in fall 2025 with about 779 higher education administrators across more than 300 U.S. and Canadian institutions, found institution-wide AI adoption rising from 49% to 66% in one year. Roughly two-thirds of executive leaders said their institution now sets aside budget specifically for AI.¹

EDUCAUSE found the same strategic shift from another angle. In its 2025 AI Landscape Study, 57% of roughly 800 respondents said AI is now a strategic priority, up from 49% the year before.²

Wasabi's 2026 education cut of the Global Cloud Storage Index, based on 241 education IT decision-makers, found that 98% of education organizations expect AI infrastructure budgets to hold steady or grow, with 46% planning increases.³

The returns are not keeping pace. Wasabi found that only 37% of education AI projects already in place were delivering positive return.³

Outside education, MIT NANDA's 2025 GenAI Divide report reviewed more than 300 enterprise AI initiatives and found that roughly 95% of generative AI pilots produced no measurable profit-and-loss impact, despite an estimated $30-40 billion in spending.⁴

Gartner reached a similar conclusion from the agentic AI side, predicting that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of cost, unclear business value, and inadequate risk controls.⁵

The pattern is not that the models are too weak. The recurring failure is what MIT called the learning gap: AI tools and organizations fail to integrate the technology into real workflows, data, and operating rhythms. Higher education is now arriving at that same inflection point, with more money going in than ever and a data foundation that remains fragmented on many campuses.⁴

Why making AI work is not optional for colleges

The demographic math is no longer hypothetical. Economist Nathan Grawe's widely cited projections anticipated a roughly 15% decline in traditional college-age demand between 2025 and 2029.⁶

WICHE projects that the total number of U.S. high school graduates will decline 13% from its 2025 peak through 2041.⁷

Fitch has kept a deteriorating outlook on the higher education sector, while Federal Reserve research has sharpened the sector's view of closure and financial-distress risk.⁸⁹

When institutions cannot recruit their way to stability, retention becomes the lever. National Student Clearinghouse data show that first-spring persistence for the fall 2023 cohort reached 86.4%, while first-spring retention at the starting institution reached 83.7%. That is an improvement, but it still means roughly one in seven first-year students did not return anywhere in higher education the next spring, and roughly one in six did not return to the institution where they started.¹⁰

Catching those students early, before they slip away, is exactly the kind of problem AI is supposed to help solve. The pressure to deploy is real. The question is why so much of that deployment stalls.

The real reason campus AI stalls

It usually is not the technology. The models available to a regional public university are broadly the same models available to a global bank. What differs is the ground they land on, and on most campuses that ground is fragmented.

The institutions themselves point straight at it. When education respondents in Wasabi's survey were asked what most challenged their AI projects, the single most-cited answer was data: the cost of storing it and the difficulty of accessing it. Two-thirds of AI infrastructure spending went to data, storage, and compute rather than to the models on top. Wasabi is a cloud-storage vendor, so the survey naturally foregrounds storage, but the direction of the finding matches what MIT and EDUCAUSE reach from outside the data-infrastructure business: the bottleneck is the foundation, not the algorithm.³

Part of that foundation problem is tool sprawl. K-12 gives the clearest public quantification: LearnPlatform by Instructure found that districts accessed an average of 2,982 distinct edtech tools annually in 2024-25. Higher education has its own version of the same dynamic: software adopted department by department, with overlapping tools, duplicate contracts, and a map of privacy obligations no one fully owns.¹¹

That sprawl produces something worse than wasted budget: a fractured view of the student. Admissions data sits in one system, advising notes in another, financial aid in a third, and LMS activity in a fourth. The early-warning signals exist; they are scattered across systems that do not talk to one another.

The sector's strategy data confirms the pattern. EDUCAUSE found that about 55% of institutions described AI strategy as happening in pockets around the institution, while fewer than 40% had an acceptable-use policy in place.²

This is the institutional version of MIT's learning gap: AI bolted onto a foundation that was never unified to begin with. Drop a sophisticated model on top of siloed, inconsistent data and one of two things happens. It stalls, or it produces answers no one trusts.⁴

What the exceptions actually do

Across the research, organizations that get AI to work converge on a recognizable set of behaviors. The evidence is not perfectly higher-ed-specific. Scale AI and Reuters Insights, for example, surveyed nearly 500 senior enterprise AI decision-makers across industries. But the pieces we can check against higher education data, especially the centrality of the data foundation and executive sponsorship, line up closely enough to treat the broader pattern as a serious operating model.¹²

• They consolidate instead of accumulating. The strongest programs avoid a drawer full of single-purpose apps and combine internal capability with integrated external platforms.¹²
• They treat the data layer as the first decision, not the last. High-quality proprietary data was the single most-cited factor in Scale AI's research on successful enterprise deployments, and the higher-ed numbers show the same pressure in budget form.¹²
• They buy a partner, not a product. Successful organizations want deep co-development or managed outcomes, not a vendor who hands over a login and calls it transformation.¹²
• They frontload the unglamorous work. Executive sponsorship, workflow redesign, change management, and outcome measurement have to be decided before go-live, not after the pilot starts to drift.⁵¹³

It works when the foundation is there

The most instructive higher education proof point is also one of the oldest. Georgia State University set out a decade ago to attack summer melt, the gap between admission and enrollment. By pairing proactive conversational texting with guided enrollment tasks, the university reduced summer melt from roughly 19% to 9%, according to Georgia State's National Institute for Student Success. In the first summer, the Pounce chatbot interacted with incoming first-year students 185,000 times.¹⁴

The same work has evidence beyond a case-study headline. Brookings summarized randomized controlled trial findings from the Georgia State effort, and later Georgia State research reported course-level chatbot work associated with higher grades and stronger retention outcomes.¹⁵¹⁶

A Mainstay case study reported that fewer than 1% of more than 50,000 incoming student messages required Georgia State staff attention.¹⁷

That last detail is the point. The technology did not replace the staff; it routed routine work away from them so their attention could go where it mattered. And it worked because the assistant was connected to actual enrollment tasks and embedded in a redesigned process. A bare chatbot dropped onto the same campus today, disconnected from those systems, would not reproduce the result, however capable the underlying model.

Where the value actually sits

It helps to stop treating AI for student success as one thing. In practice it operates on two layers, and they are worth different amounts.

The first is the deflection and automation layer: answering repetitive questions, sending timely reminders, and nudging students through required next steps. This is real and necessary, and it produces the fastest, most legible returns. Georgia State's summer-melt work sits here, but even that result did not come from the conversational surface alone. The surface worked because it was tied to a student workflow.¹⁴

The second is the intelligence and orchestration layer: the system that reads across admissions, advising, LMS, engagement, and financial-aid context to assemble a coherent picture of a student, decide which intervention matters, and route it to the right person or channel. This is where genuinely agentic value lives. The reasoning may run on models everyone can buy. The defensibility sits in what the reasoning runs on: proprietary, connected student data and institution-specific integrations into systems of record.¹²¹⁸

The mistake is to pit one layer against the other. The value is in the loop between them. Student interaction generates signal. The orchestration layer turns scattered signals into a decision. Action flows back out through a student-facing or staff-facing channel. Remove the surface and the orchestration layer goes blind and mute. Remove the orchestration layer and the surface becomes a faster help desk.

Campus Technology described the same gap from the ERP and analytics side: institutions can generate insights, but the work often stalls between insight, ownership, and action. AI does not solve that execution layer by itself. It has to be wired into the operating system of the institution.¹³

What this means for decision-makers

If you are a provost, CIO, enrollment leader, or vice president for student success weighing where to place an AI bet, the research points to a short and unsentimental checklist.

• Anchor a named executive sponsor before funding anything. Pilots run from a departmental line, without leadership attention beyond procurement, are the ones most likely to stall.¹²
• Treat the student-data layer as the first architectural decision. Resolve ownership, access, permissions, and integrations before choosing a model or locking the use case.³¹²
• Resist the point-tool reflex. Another single-purpose app is more sprawl, more silos, and another source of truth to reconcile.¹²
• Redesign the advising or support workflow before go-live. The integration failures that kill pilots are usually designed in at this stage.⁵¹³
• Measure persistence, retention, and conversion, not adoption. Logins and messages sent are activity metrics. They are not outcomes.
• Keep humans in the loop where judgment belongs. AI can classify, summarize, draft, nudge, and route. High-impact student decisions still need accountable human review.

The window is still open

None of these moves is technically exotic. All of them are organizationally hard, which is why so few institutions make them. The campuses that come through the next several years in the best shape will not be the ones chasing the newest model. They will be the ones that did the foundational work on data, integration, sponsorship, workflow, and measurement, then built AI into how the institution actually runs.

The fastest, most measurable returns sit on the surface, which makes student communication and support automation a practical place to prove value quickly. The durable value sits underneath, in the connected student record and the orchestration layer that can convert signal into action. Institutions that win will build both.

The window to do that work is open. Given the demographics bearing down on the sector, it will not stay open indefinitely.