The ROI Question on AI in Higher Education
Halfway into 2026, only 29% of campus technology leaders say their institution's AI investments have met or exceeded ROI expectations. Half say the return is unclear or falling short. That comes from the Inside Higher Ed/Hanover Research survey of 130 campus technology leaders, published in May.
Those numbers are a symptom. Many of these deployments were announcements. Few were placements, with a defined job in a specific course, and that makes their return hard to pin down.
I saw a version of this problem before, in enterprise settings.
Enterprise Teams Learned to Start With a Specific Problem
In my consulting work with Fortune 100 clients, the AI initiatives that stuck shared a profile: one team, one use case, and a problem someone was already struggling with. On the manufacturing and robotics side, the questions were exact. What will this system do on the floor? What safeguards keep it safe? What measurable improvement should we expect?
Teams answered a narrow question in detail before anything launched. That specificity is what made the result measurable later. A baseline needs something concrete to attach to, like one team's problem. "We're deploying AI" gives it nothing.
Higher Education Gets Its Scoreboard Years Later
Enterprise had a forcing function. A factory floor either runs more efficiently or it doesn't. Vague AI initiatives became untenable quickly because the scoreboard arrived on a short timeline.
Higher education's returns show up in enrollment, reputation, and student outcomes, on a much longer lag. When half of campus technology leaders say they can't tell whether AI delivered a return, they are describing a feedback signal that arrives years after the decision. A tool can look strong at launch and stay in place a long time before the evidence to evaluate it exists.
Without a baseline, the gap gets filled with an explanation about people. Faculty weren't ready. Students didn't take to it. That conclusion requires no data, and it is frequently wrong, since the same faculty are often using general-purpose AI tools every day on their own.
Copying a Peer Does Not Copy the Problem They Solved
Organizational researchers have a name for why deployments go out without a use case: mimetic isomorphism. When goals are ambiguous and a technology is poorly understood, institutions copy visible peers. Given how much enrollment depends on perception, that is a rational response. Being visibly behind on AI carries a cost of its own.
In Inside Higher Ed's reporting, strategist and former university administrator Brian Fleming describes higher education's long habit of innovating through imitation. The institution down the street may be solving a different problem, or no named problem at all.
Copying a peer imports the license and the announcement. The use case and implementation plan that would make the tool get used stay behind.
Institutional goals like improving retention are real, and hard to turn into a deployment decision. A use case is easier to name and check. Support students in a high-DFW course with an AI tutor trained on that course's curriculum, and a faculty member can tell within a term whether it is being used.
Course Integration Made a Visible Difference in My Early Pilots
I run Nuubi, an AI-integrated learning tool for higher education, and its early pilots overlapped with my consulting work. In one course, the tool was introduced a few weeks into the term as an optional resource. It saw almost no usage.
In a later course, it was integrated from the start around a specific use case. Two-thirds of enrolled students were still active in the final month of the term, ungraded and entirely voluntary.
The product also improved between those terms, and this is one course with my own data, so I hold it loosely. It matched what I saw in enterprise: how a tool fits into the existing workflow shaped adoption as much as what it could do.
Teaching Method Moved Outcomes More Than Model Choice
A 2026 three-level meta-analysis by Fan and colleagues pooled 36 studies comparing students who used generative AI against control groups, covering 7,229 students. AI use produced a medium positive effect on learning outcomes.
Of seven moderators tested, teaching method was the only significant one. Collaborative and blended course designs showed the strongest results. Which tool or model students used made no significant difference.
One Task Can Hide Several Different Jobs
Model choice still matters in a different sense: fit to the task. With a PhD in applied math, most of my career has been spent matching models to specific jobs. Most failures I've seen trace back to a model being asked to do something it was never suited for.
Grading looks like one task and isn't. Applying a rubric consistently, drafting feedback, and flagging submissions for human review are different problems with different tolerances for error.
When a Cambridge-led team tested three frontier models on 761 undergraduate psychology essays, the models matched human-assigned degree classifications only 35% to 65% of the time. They favored writing style over substance and clustered marks in the middle, so they were least accurate at the boundaries where classification decisions get made.
A general license expected to handle layered tasks like this off the shelf will run into that mismatch. It is one reason broad campus licenses struggle to show ROI, even as institution-wide student licenses jumped from 27% to 61% in a year. Fewer than a quarter of technology leaders put teaching and learning among their top three areas of AI value.
Measure the Outcome You Bought the Tool For
For a course-level pilot, I'd want the use case and baseline agreed before launch, a plan for how students encounter the tool, and a review point at the end of the term. If the goal is better learning, the evaluation needs evidence of learning. If the goal is retention, activity in the tool is not enough to claim success.
AI belongs in the budget. The harder work is setting up implementation so that months later people are still using it and student outcomes have measurably improved.
What would you need to see by the end of a term to know your AI investment was helping students?
Sources Are Listed Below
• Kat Wifvat, "The ROI Question on AI Initiatives in Higher Education", LinkedIn, September 10, 2026.
• Kathryn Palmer, "Half of Campus Tech Leaders Question AI's ROI", Inside Higher Ed, May 12, 2026. Reports the Inside Higher Ed/Hanover Research survey of 130 campus technology leaders.
• Fan, C., Ke, L., Chen, Z., and Lv, P. (2026). "Exploring the effect of GenAI on learning outcomes in higher education: a three-level meta-analysis". Frontiers in Psychology, 17, 1758670.
• University of Cambridge, "AI not yet good enough to mark university essays, rewarding 'style over substance'", May 22, 2026.
• DiMaggio, P. J., and Powell, W. W. (1983). "The Iron Cage Revisited: Institutional Isomorphism and Collective Rationality in Organizational Fields". American Sociological Review, 48(2), 147 to 160.