Tyler Walsh, Director of the Center for Higher Education Innovation at the University of Central Florida, joined the Use Case Lab alongside Jason Fife of Mainstay to open the box on Knightbot — UCF’s student-support AI. What I expected was a story about technology. What I got was a story about discipline, and one sentence that is the single cleanest statement of the 24 percent principle I have heard from any practitioner.

Walsh’s governing rule: before UCF sends any message, the team asks whether they can track the action they are asking the student to take. If they cannot track it, they question whether they should be sending it at all — because otherwise they will never know whether it worked.

That is the 24 percent, stated as an operating constraint rather than a theory. And it is why UCF can produce numbers most institutions cannot.

What Being in the 24% Actually Looks Like

Most institutions deploying student-support AI can tell you it exists. Walsh can tell you what it does. From Knightbot’s launch in January 2024, UCF tracks:

• Unique users — over 340,000, across both text and web

• Incoming messages — over 585,000 — roughly 650 a day

• Resolved without any human intervention — 85–88%, depending on the year, and improving as the system matures

• Staff time recovered — the equivalent of about six full-time staff a year

• Opt-out rate — 1.6% — roughly 5,000 students out of every user since launch

• Speed — basic questions answered in about seven seconds

• Campaign engagement and lift — in one registration campaign to just over 2,000 students, engaged students with registration holds registered at rates 32% higher than students with holds who didn’t engage — and roughly a week faster (about 6.5 days versus about 15)

• Across repeated campaigns — engaged students have registered 14% to 40% higher than non-engaged students

• Course-level outcomes — DFW rates, grade distribution, and performance in the next course, now three years into a four-year study.

That is not a dashboard. That is an institution that decided, before launch, what it was trying to change and how it would know. Meanwhile, three-quarters of the sector cannot answer the first question.

The Origin: A Gap, Not a Mandate

Knightbot did not begin as an AI initiative. UCF had been using Mainstay since 2019 for a narrow job — financial-aid outreach and drop-for-nonpayment notices. Then, around 2022, student government and student-support staff arrived at the same conclusion independently: there was nowhere for students to get basic questions answered. No one-stop shop. That was the gap.

UCF considered the full range of solutions, and this is the part worth pausing on. A brick-and-mortar one-stop costs staffing and a building, and it is only open when it is open. A website requires maintenance. A chat layer on a tool the institution already had could reach all 70,000-plus students, at any hour, at a fraction of the cost. AI was not the answer they went looking for. It was the option that best fit a problem they had already defined. Design and development ran from mid-2023 to January 2024, when Knightbot launched in earnest.

Today it operates on two layers. The first is the one-stop shop — web and SMS, answering the questions that used to consume three to five minutes of a staff member’s time each. The second, and the one Walsh calls the most powerful, is proactive: identifying students approaching a barrier — a registration hold, a nonpayment drop — and texting them before it becomes a crisis, with a way to raise their hand for help. A third layer, course-level bots, is now running in 13 faculty members’ courses, aimed at gateway math and English, where UCF is studying whether better organization translates into better grades and stronger performance in the next course.

The full Use Case Lab recording — Walsh and Fife, unedited — is below, along with the six lessons. Paid subscribers get the complete session and the full lab archive.

Below, for paid subscribers: the six disciplines that put UCF in the 24 percent — the ones that are transferable to your campus regardless of your budget or your vendor — plus the full recording of the session and the governance templates that operationalize them.

Read more

These are not three separate problems. They are one problem seen from three angles — a condition worth naming, because naming it is the first step to governing it. Call it adoption without authority: the state in which students are deeply embedded in AI but operate without institutional guidance, assessment alignment, or faculty preparation. Here are the three paradoxes that define it.

Paradox One: The Ban Signal

In the DEC data, 43% of U.S. and Canadian students say they would not be disappointed by an institution-wide AI ban — the highest ban tolerance of any region surveyed. In Latin America, just 15% feel the same. That is not students rejecting AI; it is students expressing exhaustion with using it in a vacuum. Because at the same time, 76% of students globally have never participated in any AI literacy training, and nearly half do not even know it exists at their institution.

The ban signal is what adoption without authority sounds like from the student side: widespread use, no scaffolding, and enough frustration that a clean prohibition starts to look preferable to the current ambiguity. The institutions reading that as “students want less AI” are misreading it. Students want less confusion.

Paradox Two: The Faculty Retreat

Globally, 73% of faculty worry that students are using AI at the expense of developing their own skills. The concern is legitimate — a separate AAC&U/Elon survey of 1,057 faculty put that figure at 90%. But in the U.S. and Canada, that worry is producing disengagement rather than redesign. Faculty intent to use AI in future teaching dropped nine percentage points in a single year, from 76% to 67% — the sharpest regional decline the DEC has measured.

This is the paradox that should most alarm a cabinet, because it is self-reinforcing. Faculty pull back, students receive less guidance, and the very skills erosion faculty worry about becomes more likely. The Gallup data closes the loop: students at institutions that discourage or prohibit AI are the most likely to feel undertrained — and the most likely to use it anyway. One in four students at schools that ban AI outright still report weekly use. Retreat does not reduce AI use. It removes the institution from the room while it happens.

Paradox Three: The Confidence Gap

The third paradox runs in the opposite direction from what you would expect. In the DEC data, 58% of U.S. and Canadian faculty say they are not worried that what they teach will be outdated by the time students graduate — the highest faculty confidence of any region. Yet only 19% of U.S. and Canadian students believe their program feels current on AI and future skills — the lowest student confidence of any region.

Faculty and students on the same campuses are reaching opposite conclusions about whether the curriculum is keeping pace. And students are not waiting for the disagreement to resolve: 47% of U.S. students have given serious thought to changing their major because of AI’s impact on the job market, and 16% already have. When students are restructuring their own education around a future faculty feel confident they are already preparing them for, the gap is not academic. It is enrollment behavior.

What the Three Paradoxes Add Up To

Put the three together and the diagnosis is precise. The DEC survey quantifies it: 72% of students globally say their assessments do not consistently reflect the work, skills, and judgment they will need in an AI-enabled workplace. In the U.S. and Canada, 48% say none or only a few do. Only 29% of students globally believe their instructors are well equipped to guide them on AI; in the U.S. and Canada, that drops to 17%.

This connects directly to an argument I have made in this newsletter and in Forbes: AI literacy is the floor, and fluency is the differentiator. The DEC literacy data now shows exactly where to build. The weakest dimension of student AI competency is domain expertise — applying AI within a specific discipline. Only 35% of students can use AI tools for discipline-specific tasks and identify which tools work best and why. The strongest dimension is human-centricity: knowing when human oversight is needed. Students already understand AI requires judgment. They have simply never been taught how to exercise it in context. That gap is the opportunity — and the institutional obligation.

This week’s companion piece. A version of this analysis runs in my Forbes column, written for a general leadership audience. This issue goes where Forbes doesn’t: the operational response — what a cabinet actually does in the next 90 days to convert adoption without authority into structured engagement. That’s below, for paid subscribers.

I’m sitting down with Brett Pollak of UC San Diego on Thursday, July 30 at noon Eastern to open the box on TritonGPT — the AI platform UCSD built on its own supercomputer rather than buying off the shelf, and now licenses to other universities. We’ll get into the decision that started it (a student at the IT help desk, of all places), why they kept their data in-house, and how any institution can think about build the control layer, buy the horsepower.

Below, for paid subscribers: the 90-day operating response to adoption without authority — the specific moves that replace the vacuum with structure, drawn from the institutions already doing it — plus the governance templates and the AI Governance Toolkit built to make it executable on your campus.

Read more

Two universities show the pattern most clearly. UC San Diego built TritonGPT on its own supercomputer — and now licenses it to UC Berkeley and other partner institutions. The University of Delaware built PATHS Engine to keep faculty intellectual property out of vendor hands. Neither rejected commercial AI. Both refused to let a vendor own the part that matters — the data, and the decisions about it.

UC San Diego: The Build That Started at the Service Desk

TritonGPT’s origin is the detail every leader should sit with, because it is the opposite of how these decisions are usually imagined. There was no master plan. In 2023, the chancellor asked IT Services whether the campus could use generative AI to make administration more efficient — deliberately starting with administration, because the academic senate held teaching and learning. UC San Diego had no contract at the time with a major AI provider. What it had was the San Diego Supercomputer Center, GPUs already running instructional and research workloads, and smart people.

The breakthrough came from a student. A staffer at the IT service desk — which UC San Diego staffs with students — downloaded an open-source orchestration framework called Onyx and used it to crawl the university’s Confluence knowledge base so he could answer support calls faster. He brought it to an internal AI community of practice, and the team realized the thing solving one student’s call-center problem could work enterprise-wide. Onyx, it turned out, was founded by two UC San Diego alumni. As Brett Pollak, the IT Services executive director who leads the AI program, put it to me: their solution let the model “draw information from our knowledge base” to answer questions. UC San Diego became the company’s first enterprise customer.

That is the build half. UC San Diego bought hardware, ran open-source models on its own infrastructure at SDSC, and wired in connectors to its enterprise systems and data warehouse through a retrieval-augmented architecture. Development began June 2023; a roughly 400-person administrative pilot ran by that October; the platform launched to all staff and faculty in spring 2024, with student access following in 2025. It now reaches 38,000 faculty and staff, with over 20 different agents in the ecosystem with a Socratic student tutor that integrates with their Canvas LMS and other repositories faculty use to teach courses.

The Part That Dissolves the Binary

Here is where build-versus-buy stops being a binary. UC San Diego did not refuse commercial models. It offers them via API— OpenAI through Azure, Anthropic’s models through AWS— but only inside contracted, single-tenant enclaves where, in Pollak’s framing, none of the data leaves. Open-source models run on-premises at SDSC for complete data control, while commercial models are accessed through agreements that keep UC San Diego data contained. Pollak’s reasoning—and the thesis of this piece—is simple: enterprise data should stay under UC San Diego’s control, without sacrificing users’ ability to choose the model that fits the task.

The other thing UC San Diego refused was a sealed product. It chose Onyx over competitors like Glean specifically because it wanted to build around the platform rather than inherit a black box. That single preference — flexibility over convenience — is what later let UC San Diego layer custom agents on top, and it is the difference between owning your AI strategy and renting it.

The proof that the build paid off: UC San Diego now sells it. The university licenses TritonGPT to peer institutions—Berkeley runs it as BearGPT, with UC Merced, UC Agriculture & Natural Resources, and Fairleigh Dickinson also on board—as a revenue-generating service that funds its own development team. The institution that built to control its data ended up with an asset it could sell to peers who couldn’t build it themselves. That’s what taking the power back looks like when it works: not just independence from vendors, but becoming the thing other institutions would otherwise have to buy.

Delaware: Buying Would Have Meant Surrendering Faculty IP

The University of Delaware reached the same sovereignty conclusion from the teaching side rather than the administrative one. PATHS Engine takes a faculty member’s own course materials — slides, PDFs, lecture transcripts — and turns them into structured learning objects the faculty member approves before any student sees them. The design principle, as the UD team described it in our May Use Case Lab, was a walled garden: faculty feel safe using AI and students feel permitted to use it, because both are working from the faculty member’s own data rather than a frontier model’s training set.

The reason this had to be built rather than bought is the same reason UC San Diego kept data at SDSC. To give students AI tools mapped to their actual course, UD would otherwise have had to send faculty intellectual property to a vendor — scraped, ingested, surrendered to a black box. UD’s answer was no scraping, an audit trail tracing every generated object back to its source, and faculty authority over what the system becomes. The newest phase, a federated learning pilot across Delaware institutions, extends the principle: institutions share how their models improved, never the underlying data. Same conviction as UC San Diego, different campus, different starting point — the institution keeps the data, and the vendor relationship is bounded to the parts that do not compromise it.

Data sovereignty is surrendered at a single moment: the day a cabinet leader signs a vendor contract without the language that specifies where the institution’s data goes and who may touch it — and by the time anyone notices, it cannot be recovered. What follows is not more of this argument; it is the diagnostic and operational apparatus that keeps that signature from being a mistake. It includes the contract language that forces a vendor to state exactly where data resides and who can access it; the data classification that determines which tier of institutional data may touch which model; the standing-committee structure that makes sovereignty a required question before any purchase clears; and the AI-Ready Institution Scorecard rubrics for the governance and operational-redesign pillars where these decisions are won or lost. Alongside them sits the Use Case Lab archive — the documented record of how real institutions actually made these calls, the precedent you will want in the room when it is your turn to decide.

Read more

I will share those, because they are genuinely inventive. But the entry that matters most is not any single project. It is a number I produced by analyzing the catalog as a whole — and it is the most clarifying thing I have found all year about why so much AI spending in higher education cannot be defended.

The Number

I analyzed 139 of the use cases in the tracker for one property: does the institution clearly define a metric for success? Not whether the project is good. Not whether it works. Simply whether anyone stated, in advance, what “working” would look like.

Twenty-four percent did. Three out of four AI deployments in higher education cannot tell you what success was supposed to be. No baseline. No target. No defined outcome. Which means that a year after launch, no one can say whether the tool earned its budget line or simply consumed it — because there was never a standard to measure it against.

This is not an AI problem. The tools largely work. It is an evidence problem, and it is the quiet line dividing two kinds of institutions: the ones building durable capacity and the ones accumulating expensive theater. The difference between them is rarely budget or sophistication. It is whether someone asked, before launch, a single question. What is the metric?

The Inventive Ones — and What Separates the Best of Them

First, the use cases worth admiring. These show how wide the aperture of serious AI work in higher education has become.

A “Grammar Laboratory” for deaf students at RIT. Developed by lecturer Erin Finton at the Rochester Institute of Technology’s National Technical Institute for the Deaf, with support from Google.org, the tool takes a bilingual ASL-and-English approach to a real pedagogical problem: ASL has its own grammar, and the gap between it and written English creates genuine learning barriers. The design choice I find most instructive is small and telling — the AI-generated questions are framed as an assignment, not an assessment, and every question includes an “I’m not sure” option that triggers further explanation. Someone thought about what the tool was for, which is exactly the discipline most deployments lack.

An institutional data dashboard at Florida State that unifies 47 databases. FSU’s Snowflake implementation pulls from 47 separate source systems — and growing — so a leader can ask a question like “how many first-year students are at high risk of dropping out based on attendance, mid-term grades, and payment status?” and get an answer in minutes instead of a three-week data request. Note what makes this one a model: FSU reports responding 95 percent faster to stakeholder data needs. That is a defined, measured outcome. It belongs to the 24 percent.

An extension chatbot for farmers and agricultural workers. Trained on 400,000 articles and fact sheets from 30 state extension networks and the U.S. Department of Agriculture, it puts a century of land-grant extension expertise within reach of farmers and gardeners at any hour. It is a reminder that the higher-education mission AI can serve extends well past the classroom — into the public-service obligations that land-grant institutions were built to carry.

Creative, specific, mission-grounded. But notice: the FSU example came with a number, and the other two largely did not surface one in public reporting. That is the pattern across the whole catalog in miniature. Inventiveness is everywhere. Defined success is rare. And the projects that pair the two are the only ones an institution can actually defend at renewal.

Why the Evidence Gap Persists

If defining a success metric were easy and obviously necessary, the number would not be 24 percent. So it is worth being honest about why three out of four projects skip it.

Defining success is harder than launching. A demo is exciting; a baseline is tedious. Writing down what you expect to happen creates the possibility of being wrong, and institutions are not in the habit of inviting that. It is safer, bureaucratically, to deploy something impressive and never specify what it was supposed to accomplish — because then it can never be said to have failed. The absence of a metric is not an oversight. It is, often, a quiet form of risk avoidance. And it is exactly backward, because the project with no defined outcome is the one most exposed when the budget tightens.

This is the Evidence and Accountability pillar of the AI-Ready Institution Framework, and it is consistently the lowest-scoring pillar I see. Institutions can show you activity — pilots, licenses, committees. Far fewer can show you outcomes. The 24 percent figure is what that pillar looks like measured across a large, real sample. It is not a flattering picture, and it is the most useful one a cabinet can start from, because it locates the problem precisely: not too little AI, but too little evidence.

The next Use Case Lab is Tuesday, June 30, noon ET (rescheduled from 6/25) with Tyler Walsh of UCF on Knightbot — a student-support system with exactly the kind of operating evidence most deployments lack. Paid subscribers get the live room and the recording. We open the box on one real deployment together, every month.

Below, for paid subscribers: the three-question test that moves a use case from the 76 percent to the 24 percent, the way to apply it before you sign anything, and the governance templates archive built to make the test enforceable on your campus.

Read more

On Thursday, June 25, Tyler Walsh — Director of the Center for Higher Education Innovation in UCF’s Division of Student Success and Well-Being — joins the Use Case Lab to take the system apart with us. Before that conversation, it is worth laying out why Knightbot is a case study that is worth studying, and the uncomfortable design question it raises for everyone currently writing a check to a generative vendor.

The Numbers That Make It a Reference Case

Per UCF’s own reporting, Knightbot resolves 85 percent of student queries without any human intervention, drawing on a vast pre-approved knowledge base spanning most colleges, many departments, and nearly all support units across campus. Since January 2023, that efficiency has freed roughly 12,000 hours of staff time — the equivalent of six full-time employees, in Walsh’s framing. The system answers questions about financial aid, advising, registration, housing, and campus resources, any hour of the day, for a student body north of 70,000.

Those are not projection numbers or pilot numbers. They are operating numbers, from a system that has been in production long enough to have a track record. That alone separates Knightbot from most of what crosses a provost’s desk. But the number that should stop senior leaders is not the 85 percent or the 12,000 hours. It is the architecture choice underneath them.

The Contrarian Design Decision

Knightbot is a partnership between UCF’s Center for Higher Education Innovation and Mainstay, and it runs on a curated knowledge base — a structured set of more than 1,200 pre-approved answers — not on an open generative model improvising responses. In an era where every vendor is racing to bolt a large language model onto student support, UCF’s most-cited success deliberately did not.

That is the design tension worth sitting with. A generative chatbot is more impressive in a demo. It speaks fluently, handles phrasing it has never seen, and feels like the future. A curated knowledge base is less dazzling and more accountable: every answer traces to something a human approved, the failure modes are visible, and the institution controls exactly what the system will and will not say to a student about financial aid or academic standing. The dazzle and the accountability pull in opposite directions, and UCF chose accountability. The outcome numbers suggest that for the specific job of resolving high-volume, high-stakes student questions, the less impressive architecture is the one that actually scales.

This is the Operational Redesign and the Evidence and Accountability pillars of the AI-Ready Institution Framework operating in the same system. Operational Redesign, because Knightbot is not a chatbot bolted onto a website — it is woven into how financial aid, advising, and student success offices actually route and resolve student need, with text-message reach directly to students about deadlines and enrollment. Evidence and Accountability, because UCF can tell you the resolution rate, the hours saved, the FTE equivalent, and the campaign-level outcomes — the questions most institutions deploying AI still cannot answer about their own tools.

It connects directly to the fluency argument from this newsletter’s last issue. The institutions that look like leaders in three years will not be the ones with the most pilots. They will be the ones that treated readiness as something built deliberately and measured as it goes. Knightbot is six years of exactly that discipline, in one system.

Below, for paid subscribers: the four questions to bring to your own student-support AI decision — the operational lens that separates a tool that scales from a tool that merely launches — plus the governance templates archive built to help you ask them before you sign anything.

Read more

The number that did not travel: 91 percent did not.

Most coverage ran the story as a cheating story. Read the study itself and it is something closer to the opposite: a faculty-capacity story wearing an academic-integrity headline. Roughly two-thirds of students reported using generative AI during the study period. Thirty-seven percent use it regularly — monthly or more. Regular use runs highest in computer science (62 percent), mathematics (53 percent), and business (51 percent). The tools are already inside the coursework of nearly every discipline that leads to employment. The institutional question is not how to catch the 9 percent. It is what the other 91 percent need from their faculty in the next twelve months — and whether faculty are getting the time, training, and craft to deliver it.

Take the 9 Percent Seriously — Including the Objection

The first objection to the 9 percent figure is that it is self-reported and therefore deflated: students under-admit cheating, and research on AI-assisted work shows students routinely misjudge how much of the output was theirs. The objection deserves a straight answer, and the study’s design partly supplies one. The researchers did not simply ask students whether they cheated. They used an indirect-questioning method built specifically to estimate sensitive behaviors that direct survey questions undercount. The 9 percent is a more defensible number than the usual self-report.

But grant the objection entirely. Suppose the true figure is double, or triple. The institutional conclusion does not move. The study found misuse rises to 26 percent among daily users — the more embedded the tool, the more the line blurs — and that use and misuse vary so much by discipline that the authors’ own prescription is discipline-specific assessment reform, not blanket bans and not universal detection regimes. “Assessment reform is necessary and urgent,” Kizilcec said when the study released. Whatever the real cheating number is, the answer runs through the people who design assessments. That is faculty. And almost no institution is resourcing them to do it.

The Support Gap, Quantified

The Digital Education Council’s Global AI Faculty Survey — 1,681 faculty across 52 institutions in 28 countries — found that just 6 percent of faculty fully agree their institution has provided sufficient resources to build their AI literacy. Only 17 percent rate themselves at an advanced level with the technology they are now expected to design assessment around; 40 percent describe themselves as just beginning. The Tyton Partners Time for Class study, drawing on more than 3,000 instructors and administrators, adds the structural version of the same finding: most institutions still have no generative AI policy at all, and instructors report that monitoring student AI use is adding work to their load, not removing it.

Put the two datasets side by side and the diagnosis writes itself. Student use: roughly two-thirds and climbing. Faculty support: single digits, by faculty’s own account. The assessment-validity problem the Science study surfaced is not first a student problem. It is a faculty-support problem. The 9 percent who cheat are the symptom. The architecture that makes cheating the easier path than learning is the cause — and faculty are the only people who can rebuild that architecture.

Same Year, Same Use Case, Opposite Design: SUNY and CSU

Two public systems made faculty-AI decisions this spring. The design contrast is the most instructive thing happening in higher ed right now.

SUNY set a systemwide AI framework, reported by Inside Higher Ed on May 4, that runs the integration through faculty by design. A cohort of twenty AI for the Public Good Fellows — faculty and staff across disciplines, in compensated fellowship roles — consults with colleagues, instructional designers, and librarians to build AI into coursework across the system’s 64 campuses. AI literacy is embedded in the Information Literacy core competency of SUNY’s general education framework, so the integration work has a curricular destination rather than floating as professional development. Every campus must adopt or update its own AI guidelines by December 31. Faculty are the integration layer, the fellows are the support structure, and the deadline forces the governance work.

The California State University system renewed its OpenAI contract in May at $13 million a year for three years — $39 million to put ChatGPT Edu in front of more than 470,000 students and 63,000 faculty and staff — while the system faces $144 million in budget cuts and while its own survey of more than 94,000 students, faculty, and staff found 59 percent of faculty skeptical that AI is benefiting education at all. Faculty delivered a petition opposing renewal; NPR reported that internal planning documents had described the partnership as a branding opportunity. Whatever else the renewal is, it is a sequencing decision: tool access first, faculty enablement second.

Same use case. Same year. Opposite design choice. One system bought capacity in its people. The other bought seats. The Science study is effectively a referendum on which sequence holds up — because assessment reform cannot be procured. It has to be built, by faculty, on paid time.

Below, for paid subscribers: the three moves the institutions getting this right are actually making — with the operational detail to bring to your own provost’s council — plus the governance templates archive: the procurement-gate language, AI review-body charters, and FERPA-aligned data-tier definitions your campus can adapt before its own December deadline arrives.

Read more

Three surveys, one finding. The full diagnosis is in my recent Forbes piece: AI literacy is the floor. Fluency — judgment applied through the tool — is the differentiator. And the institutions most at risk are not the ones that know they are behind.

The Missing Word Is Applied

The Pearson data names the institutional version of the problem with painful precision. Seventy-eight percent of higher-education leaders believe they are meeting employer expectations on AI readiness. Only 28 percent of employers agree. That fifty-point perception gap is the most useful number a provost will see this fiscal year. It tells you the institutional self-assessment is wrong, and it tells you in which direction.

Strada finds AI literacy ranks last in employer importance — not because employers do not need it, but because they assume graduates can already open an LLM and type a prompt. The bar moved without most institutions registering it. What employers cannot find is the elevation above that floor: graduates who can apply AI to domain expertise, pressure-test an analysis, catch a flawed recommendation, and defend why they overrode the tool’s first answer. That is fluency, and only 14 percent of graduates demonstrate it.

The volume of entry-level hiring is holding. Strada finds AI is increasing entry-level hiring at nearly three times the rate it is decreasing it. IBM announced in February it would triple US entry-level hiring in 2026, rewriting junior roles around judgment, customer engagement, and AI oversight. What has changed is the selection inside the volume. Stanford economists documented a 13 percent relative decline in employment for 22-to-25-year-olds in AI-exposed occupations — graduates who can elevate work using AI are getting hired; graduates whose preparation suits the old routine work are not.

That is the diagnosis. A graduating class of students who can use AI but cannot apply it is structurally underprepared for the entry-level labor market that already exists. And no curricular bolt-on solves it, because producing AI-fluent graduates is not a curriculum problem. It is a readiness problem.

Why This Is an Institutional Readiness Problem

If the fluency gap were solvable by adding an AI literacy module to first-year seminar, institutions would have closed it by now. They have not, because the work required to produce AI-fluent graduates touches every layer of the institution: who decides what counts as AI-appropriate assessment, whether faculty have the fluency to teach it, whether programs are aligned to the labor market it serves, whether operations support it at scale, and whether anyone is measuring whether it is working. That is five different organizational capabilities. Most institutions have one or two.

This is the gap I built the AI-Ready Institution Framework to make visible. Five pillars: Governance, Human Capacity, Workforce Alignment, Operational Redesign, Evidence and Accountability. The full framework is at edgenerative.com; the diagnostic version is the AI-Ready Institution Scorecard™ at edgenerative.com/scorecard. What follows is what each pillar looks like when an institution takes it seriously, with the two pillars most directly engaged by the fluency gap developed first.

The part nobody selling an “AI transformation” wants to say out loud

I have yet to see an institution that would score a clean 30 on the scorecard. Most land in Emerging or Developing. That is the starting line, not failure — and treating it as failure is how leaders end up chasing tools instead of building capacity.

Workforce Alignment: The Pillar the Fluency Gap Most Directly Engages

Workforce Alignment asks whether programs, credentials, advising, and employer partnerships are preparing learners for AI-shaped work — and how the institution would know. The diagnostic question is simple and most institutions cannot answer it: when did you last review your top programs by enrollment against current AI-impacted labor market data, and what changed as a result?

Georgetown’s Center for Security and Emerging Technology built an answer worth borrowing. Its PATHWISE project, launched in fall 2023 with the NobleReach Foundation, maps the supply and demand of the AI and cybersecurity workforce by region and names the institutions producing graduates into those fields. CSET’s own analysis put the AI-ready U.S. workforce at roughly 17 million in 2022, up from 14 million in 2018. The lesson for a provost is not to replicate a national research tool. It is that program-to-labor-market review can be evidence-driven rather than anecdotal, and the data to do it now exists.

The harder Workforce Alignment work, surfaced by Strada, is the math-intensive STEM shortage. Twenty-nine percent of Strada employers rank math-intensive STEM graduates as the hardest field to hire from. This does not respond to a fix inside existing STEM programs — it responds to gateway-course architecture. Most of the attrition happens between calculus, organic chemistry, and the second-year programming sequence. Wright State University’s EGR 1010, an NSF-funded first-year engineering math course taught by engineering faculty with embedded lab and recitation, has been the most-studied example: four-year graduation rates for students who took the course averaged 23 percent across the first three cohorts, compared with under 8 percent for those who did not. That is roughly thirty additional sophomores retained per year. Workforce Alignment lives in gateway courses as much as in program review.

Human Capacity: Fluency You Can Document, Not Assume

Human Capacity is the second pillar the fluency gap most directly engages, and it is the one institutions most often confuse with training. A webinar is not human capacity. A signed attendance roster is not human capacity. Human Capacity is whether faculty, staff, students, and leaders actually have the fluency and support to apply AI well — and whether the institution can show it.

The University of Manchester treated that as a workforce-development problem rather than an awareness campaign. In December 2024 it launched a Data and AI Academy with Multiverse, a 13-month, apprenticeship-levy-funded program enrolling an initial cohort of roughly 70 professional-services staff. Participants reported saving about four hours a week. The number matters less than the structure: role-specific training, funded through an existing mechanism, with outcomes someone is tracking.

For students, the Human Capacity question runs through course design. Strada found that 33 percent of employers report AI has reduced foundational, skill-building tasks for entry-level employees — a transfer of pedagogical responsibility back to the academy. New hires used to learn to write, analyze, and judge by doing those things badly under correction. If AI is doing that work in the workplace, students need to do it somewhere else. Minerva University built its entire undergraduate model around exactly this architecture: small seminars on its Forum platform delivering, in the institution’s own framing, frequent feedback and low-stakes evaluations through continuous session-level data capture. Other institutions now license the methodology, suggesting the design scales beyond a single niche school. The Human Capacity pillar asks whether your institution has rebuilt the formation environment AI removed from the workplace.

Governance, Operational Redesign, and Evidence: The Three Pillars That Make the Other Two Possible

Workforce Alignment and Human Capacity are where the fluency gap shows up. The other three pillars are what allow an institution to address it at scale rather than in one department.

Governance is the pillar that ends the Reactive pattern of vendor-led, crisis-driven AI adoption. Northwestern established a Provost’s advisory committee on generative AI in 2023; it began meeting that February. Purdue went further: its AI Use policy, published November 1, 2025, requires that any use case touching Purdue data or making decisions about people be reviewed and approved by a Data Ethics Committee, with the depth of review scaled to the risk level. Two different institutions, same underlying move — ownership, decision rights, and risk review in a standing body, not a single administrator’s inbox.

Operational Redesign asks whether you have rebuilt a core process around AI, integrated into the systems people already work in, rather than letting disconnected pilots accumulate. UC San Diego’s TritonGPT is the cleanest example I have seen at scale: a suite of in-house assistants now in the hands of roughly 38,000 faculty and staff, hosted on the university’s own infrastructure at the San Diego Supercomputer Center rather than a commercial cloud, doing concrete administrative work and beginning to extend access to neighboring systems. That is redesign, not a subscription.

Evidence and Accountability is the pillar institutions skip and the one boards and accreditors ask about first. The University of Hawaiʻi has started putting numbers on the board. In a report covering the first quarter of 2026, its student-support chatbots surpassed 100,000 student messages, with 51 percent of students engaging across the system’s ten campuses; UH reported the tools saved staff roughly 165 hours over the period and flagged 1,924 students for proactive follow-up. Engagement, time saved, early intervention — measured, reviewed, and reported. You can argue about whether those are the right metrics. You cannot argue that measuring beats not measuring.

The Pattern: One Pillar Caps the Whole Institution

Put the five pillars together and a pattern shows up almost every time. The constraint on producing AI-fluent graduates is not tooling. It is that one pillar — most often Governance or Human Capacity — is lagging far enough behind the others to cap the institution’s overall readiness. That is the single most useful thing a diagnostic surfaces. You do not close the fluency gap by buying more AI. You close it by strengthening the pillar currently holding the other four back.

Which is why a low score on the scorecard is not an indictment. It means AI activity is moving faster than institutional infrastructure — a common, solvable pattern, and the most honest place to start. The institutions that will look like leaders in three years are not the ones with the most pilots today. They are the ones treating readiness as something built deliberately, pillar by pillar, measured as it goes.

Take the Scorecard

If you want an honest read on where your institution stands across all five pillars, take the AI-Ready Institution Scorecard™. Ten questions, scored 0–30, organized around Governance, Human Capacity, Workforce Alignment, Operational Redesign, and Evidence and Accountability. Four tiers: Reactive, Emerging, Developing, Mature. Built for a leadership team to do together. Available at edgenerative.com/scorecard.

With gratitude,

Dr. Aviva Legatt

Founder, EdGenerative • Affiliated Faculty, University of Pennsylvania • Forbes Contributor

About Dr. Aviva Legatt

Dr. Aviva Legatt is the founder of EdGenerative, where she advises university boards, presidents, and systems on AI governance, adoption strategy, and microcredentials. She holds a doctorate from the University of Pennsylvania, where she serves as affiliated faculty teaching organizational dynamics. A Forbes Contributor (top 2% designation), she has covered education leadership and interviewed figures from Simone Biles to Adam Grant. She serves on the Montgomery County Advisory Council on Artificial Intelligence for Public Good and is the author of Get Real and Get In (St. Martin’s Press). Her AI Use Cases in Higher Education: A Community Handbook is the open-access resource behind this newsletter.

What follows is a partial recap of that conversation. The full recap, with quotes, the federated learning pilot mechanics, the three governance lessons, and the operational through-line for institutions wanting to build something similar, continues below for paid subscribers.

Mark Your Calendar: Tyler Walsh, UCF, Thursday June 25

The next Use Case Lab is Thursday, June 25 at noon ET, featuring Tyler Walsh, Director of the Center for Higher Education Innovation in UCF’s Division of Student Success and Well-Being. Tyler leads the team behind Knightbot, the AI-powered student-support system serving UCF’s 70,000+ students — and the work he is doing is one of the cleanest examples I have seen of operational redesign and measurable outcomes happening at scale in the same institution.

UCF’s reported numbers are the kind of evidence the sector needs more of. Per UCF News, Knightbot has fielded more than 459,000 incoming messages from 277,000 unique users since launching in January 2023, resolving 85 percent of student queries without human intervention and freeing approximately 12,000 hours of staff time — equivalent to six full-time employees. Students engaged in a recent enrollment campaign were 50 percent more likely to enroll than those who did not respond. A drop-prevention campaign reached over 1,000 students and reduced the drop rate by 15 percent among those who engaged.

And This Week: AiM Higher 2026 at the University of Delaware

Before June 25, there is another room worth being in this week. The University of Delaware — the same team behind PATHS Engine — is hosting AiM Higher 2026 on June 9–11. It is not a watch-the-slides conference. It is a problem-based learning event built around a Five Scenario Framework: institutions come in with a real problem (AI policy, AI literacy across the curriculum, scaling support services, professional development, governance), get paired with Pathfinder Fellows who have lived through similar work, and leave with a proof of concept built in the AI Makerspace. Nova and Lauren are at the center of it. Sponsors include Zoom and EBSCO.

I attended last year; the actionability of what people walked away with was the difference between this and most AI-in-higher-ed events. Conference site: sites.udel.edu/aim-higher • Five Scenario Framework: sites.udel.edu/aim-higher/the-five-scenario-framework-2026

Now, the May 28 Lab: PATHS Engine and the Federated Learning Pilot

Back to Delaware. PATHS Engine takes unstructured academic content — PowerPoints, PDFs, Canvas courses, lecture-capture transcripts — and turns it into structured learning objects the faculty member approves before students see anything. The AI works from the professor’s own material, not a frontier model’s training set. Students get tools sourced from the course. Faculty keep authority over what their content becomes.

Nova described the original move plainly: “We decided that, what if we built a kind of walled garden where we can have the faculty member feel safe using AI and the students feel like they have permission to use AI — but both will be sourced from the faculty data.”

That sounds modest. It is not. It is the answer to a question most institutions have not yet figured out how to ask, much less solve: how do you give students access to AI tools that actually map to their course — without sending faculty intellectual property to a vendor, without scraping, without surrendering institutional data to a black box?

The first faculty partners were not random. Nova started with what every institution has: the early adopters who had already trusted the office’s work on lecture capture and classroom technology. “Every institution has their superstars or their super users,” Nova said. The faculty member she called first had years of lecture-capture recordings already in the system — a deep pool of source material to test on, and a faculty member who already trusted the team. That sequencing matters: the institutions that try to launch AI tools by pushing them out to all faculty at once fail. The institutions that start with the faculty who already trust the technology office build something durable.

PATHS Engine is the Operational Redesign pillar of the AI-Ready Institution Framework, made concrete: a core institutional process (teaching with AI) rebuilt with privacy, consent, and faculty authority designed in from the first decision, not bolted onto a commercial chatbot after the fact. It is also a Governance story, because of how UD built faculty trust along the way, and an Evidence story, because of how UD is starting to measure whether the tools actually help students study.

Paid subscribers continue below: the three engineered choices that built faculty trust, how UD measures whether the tools actually help students, the full mechanics of the federated learning pilot across Delaware, three governance lessons, and the operational through-line for institutions that want to build something similar — plus access to the June 25 Tyler Walsh lab live.

Read more

Read more

On May 11, Instructure confirmed it had paid a ransom to ShinyHunters, the extortion group that twice breached Canvas in nine days, exposed data from approximately 8,809 institutions and 275 million users, and brought the most widely deployed learning management system in North American higher education offline during finals week.

The company said it received “digital confirmation of data destruction (shred logs)” and assurance that no Instructure customers would be extorted. The monetary value was not disclosed. According to Wikipedia’s ongoing accounting of the incident, unconfirmed rumors place the payment at $10 million.

The ransom is the story drawing cybersecurity attention. It is not the story higher ed leaders should be paying attention to.

The Attack Vector Tells the Story

The most consequential detail of the May 2026 Canvas breach is not the ransom payment. It is how ShinyHunters got in.

Sit with what that attack vector means. ShinyHunters did not exploit some exotic zero-day. They entered through a free-tier signup pathway that had been operating, with weakened verification, for years. The architecture of multi-tenant SaaS — logical isolation between tenants sharing the same underlying infrastructure — is industry standard. It is also industry-fragile. When the verification gap was exploited, the boundary between unverified teacher accounts and paying institutional customers failed.

This is the architecture of every major SaaS platform higher education depends on. Canvas. Workday. Banner. Salesforce. The Microsoft and Google productivity stacks. Every institution that has migrated systems to a cloud vendor in the last decade has accepted some version of this exposure as the price of operational efficiency. Most have not assessed the exposure deliberately. The Canvas breach is what deliberate assessment looks like — applied in retrospect, under duress, with student data already in criminal hands.

The operational consequence was immediate. When Canvas login pages were defaced on May 7, the University of Washington, UC Riverside, and several thousand other institutions lost simultaneous access during finals. Per CalMatters, California was hit especially hard: all 22 California State University campuses, all 116 California community colleges, six of ten UC campuses, and Stanford were on the affected list. The North Carolina Department of Public Instruction cut access to Canvas across the state’s K-12 system. Durham Public Schools and Wake County Public Schools pulled the plug independently while waiting for clarity that did not arrive for days.

This is what happens when a single vendor’s security posture determines whether 8,809 institutions can hold class.

Vendor Concentration Risk: The Category Higher Ed Boards Have Not Been Governing

Higher education has been quietly concentrating its operational dependencies into a small number of SaaS platforms whose failures cascade across thousands of institutions simultaneously. A single LMS vendor for academic operations. One or two HRIS vendors for personnel data. A handful of student information systems. A short list of identity providers.

The board-level institutional risk register that lists facilities, endowment volatility, enrollment decline, and accreditation status almost never lists “the company that runs our LMS gets breached during finals.” That risk is now demonstrable, current, and unresolved.

The diagnosis is straightforward: Vendor concentration risk is a category financial regulators have governed in banking and insurance for years. It belongs on the higher education risk register now. The Canvas breach is the event that forces the conversation.

Cliff Steinhauer, director of information security and engagement at the National Cybersecurity Alliance, told Inside Higher Ed that Instructure’s ransom payment can create “a dangerous feedback loop where attackers are effectively rewarded for successful breaches.” He is right about the ransom. But the larger feedback loop is structural. Each year higher ed concentrates more operational dependency in fewer vendors, and each year the institutional cost of any single vendor’s breach grows.

The ransom is the symptom. The concentration is the disease.

What Provosts, Presidents, and CIOs Need to Do This Quarter

Four specific moves in the next 90 days.

1. Run a vendor concentration assessment. Identify every SaaS platform on which institutional operations depend. Score each on three dimensions: how many other institutions depend on the same platform, what recovery looks like if the platform is offline for five days, and what data the vendor holds that a breach would expose. The assessment is not theoretical. It is a one-page document the cabinet should be able to read in a meeting.

2. Review the LMS contract specifically. Does the contract include adequate breach notification timelines? What counts as a material breach for purposes of termination or renegotiation? Does it require the vendor to maintain cyber insurance proportional to the exposed data volume? Does it indemnify the institution for downstream harms from a vendor breach? Most LMS contracts were negotiated when the LMS was a content delivery tool, not the operating system of academic life. The contract should reflect what the platform has become.

3. Build an incident response procedure for sustained LMS outage. Existing IT incident response plans almost certainly do not include the scenario of the LMS being down for four to seven days during a high-stakes academic window. ShinyHunters demonstrated that this is real, not hypothetical. The procedure should include pre-authorized faculty communication templates, alternate assessment protocols, deadline extension authority, and a clear chain of decision-making for academic continuity.

4. Brief the board on SaaS dependency exposure. The institutional risk register needs a SaaS dependency line item. Boards have been asked to govern cybersecurity for years; most have done so at the level of “do we have a CISO and a phishing training program.” The Canvas breach forces the conversation up a level. A board that does not know what happens to the institution if its LMS, HRIS, or SIS goes offline for a week is not currently equipped to govern the operational risk it is responsible for.

None of these moves require new technology purchases. None require waiting for federal guidance. They require institutional time and institutional decision-making, which are the resources senior leaders actually control.

The EdGenerative AI Governance Toolkit

Vendor concentration risk is one piece of a larger governance gap. Through my work advising institutions, I developed a policy-neutral, audit-ready AI governance framework mapped to NIST AI RMF, DOJ Title II, FERPA, HECVAT, and ISO/IEC 42001. It is built around Ten Core Components that give institutions a clear path from pilot to practice — and a 90-Day Starter Plan sequenced across three phases (governance charter, risk assessments and procurement safeguards, monitoring and training infrastructure) that produces auditable artifacts boards and accreditors actually want to see

.What’s Ahead

A reminder: the next Use Case Lab is Thursday, May 28 at noon ET — a live working session for paid subscribers where we walk through one real higher-ed AI scenario together. This month’s session will focus on building a vendor concentration assessment for your own institution, using the four-move framework from this issue. If you have been reading as a free subscriber, this is the issue to upgrade on.

Issue #6 will return to the workforce alignment thread from Issue #4 — specifically, how microcredential governance interacts with the SaaS dependency question, and why the institutions building stackable credentials need to think about vendor concentration at the credentialing layer too.

We will also continue documenting how senior leaders are operationalizing AI governance under real pressure — the contract clauses that work, the board conversations that landed, and the governance moves that scaled.

And we will go deep on the next round of federal action: the Workforce Pell launch in July, the accreditor guidance taking shape this summer, and the emerging breach disclosure standards that will shape what “prepared” actually means at the institutional level.

The era of treating LMS choice as a procurement decision is over. The era of treating it as a board-level risk decision has begun.

With gratitude,

Dr. Aviva Legatt

Founder, EdGenerative • Affiliated Faculty, University of Pennsylvania • Forbes Contributor

About Dr. Aviva Legatt

Dr. Aviva Legatt is the founder of EdGenerative, where she advises university boards, presidents, and systems on AI governance, adoption strategy, and microcredentials. She holds a doctorate from the University of Pennsylvania, where she serves as affiliated faculty teaching organizational dynamics. A Forbes Senior Contributor (top 2% designation), she has covered education leadership and interviewed figures from Simone Biles to Adam Grant. She serves on the Montgomery County Advisory Council on Artificial Intelligence for Public Good and is the author of Get Real and Get In (St. Martin’s Press). Her AI Use Cases in Higher Education: A Community Handbook is the open-access resource behind this newsletter.

A handful of institutions are already three years in.

The University of Delaware is one of them.

So on Thursday, May 28 at 12pm ET, I’m bringing two of the people quietly building it into the Higher Ed AI Playbook Use Case Lab — live on Substack.

What You’ll Hear Inside the Room

Why I Want My Paid Subscribers in This Room

This is not a conference panel. Not a vendor pitch. A working lab where institutions show their actual work — and you bring your actual problem.

Paid Higher Ed AI Playbook subscribers get the live conversation, the recording, and the working notes. That alone is the upgrade math.

Another deeper reason to go paid right now: the policy starter documents.

Every paid subscriber gets the growing archive of governance templates Higher Ed AI Playbook publishes alongside each issue — procurement-gate language, AI tool review-body charters, faculty-student disclosure frameworks, FERPA-aligned data-tier definitions, AI evaluation criteria, and the AI-Ready Institution Scorecard™ scoring rubrics.

These are not consulting deliverables I sell separately. They are the starter drafts your governance committee can pick up, redline, and run with on Monday.

If you’re sitting down to write or defend an AI governance framework this month, this is the toolkit I’d want in front of me.

Upgrade to Paid

And If You Want to Go Deeper — Block June 9–11

UD is hosting AiM Higher 2026, the East Coast AI Conference, two weeks after our live.

It is not a watch-the-slides conference. It is a problem-based learning event built around UD’s Five Scenario Framework:

You pick the scenario that maps to your institution’s biggest gap. You work in teams with peers facing the same problem. You leave with a proof of concept built in the AI Makerspace, ready to pilot back on campus.

Nova and Lauren are at the center of it. The May 28 lab is your preview.

The Five Scenario Framework: sites.udel.edu/aim-higher/the-five-scenario-framework-2026

AiM Higher 2026 Registration

Why It’s Worth Being in the Room May 28 — Not Catching the Replay

The questions you ask Nova and Lauren shape the hour.

Nearly 4,000 higher ed leaders are already subscribed to the Higher Ed AI Playbook — the most aligned cohort I know of for this work. The live Q&A is where the real value compounds.

Subscribe. Show up. Bring the AI problem keeping you up at night this week.

Resources

PATHS Engine

The Delaware Federated Learning Pilot

With gratitude,

Dr. Aviva Legatt

Founder, EdGenerative · Affiliated Faculty, University of Pennsylvania · Forbes Contributor

About Dr. Aviva Legatt

Dr. Aviva Legatt is the founder of EdGenerative, where she advises university boards, presidents, and systems on AI governance, adoption strategy, and microcredentials. She holds a doctorate from the University of Pennsylvania, where she serves as affiliated faculty teaching organizational dynamics. A Forbes Senior Contributor (top 2% designation), she has covered education leadership and interviewed figures from Simone Biles to Adam Grant. She serves on the Montgomery County Advisory Council on Artificial Intelligence for Public Good and is the author of Get Real and Get In (St. Martin’s Press). Her AI Use Cases in Higher Education: A Community Handbook is the open-access resource behind this newsletter.

#HigherEd #AIinEducation #AILiteracy #HigherEdLeadership #AIGovernance #AiMHigher2026

Last issue we documented the alignment crisis: only 37% of employers say higher ed is preparing students with the skills they need, while 50% are now assessing candidates for AI fluency at the point of hire (WGU Workforce Decoded, 2026). The fix isn’t another general-education tweak. It’s a credential infrastructure question: who issues which badges, what do they signal, and how do they stack into something employers will actually pay attention to?

The case for microcredentials is strong. The risk is portfolio chaos - a thicket of unrecognized certificates students collect with no clear hierarchy, no learning-outcome evidence, and no employer endorsement. This issue is about how to build microcredentials with governance baked in: the kind that survive an accreditation visit and an HR screening at the same time.

These are the numbers from the same WGU survey of 3,147 employers. They aren’t a forecast - they’re the current employer screening posture, already in motion.

The Credential Recognition Gap

Two data points reframe the conversation. From the WGU Workforce Decoded survey: 86% of employers view certificates as valuable indicators of readiness, and 78% say work experience is equal to or more valuable than a degree. Lightcast’s analysis of more than 100 million worker profiles tells the parallel story from the other side: formal education is no longer the sole on-ramp into AI careers. Most workers reach AI roles through nonlinear transitions, self-directed learning, and stackable credentials picked up along the way.

The implication for institutions is uncomfortable. If employers are signaling that bundled four-year degrees aren’t the only - or even the primary - proof of readiness they care about, then the institutions that capture credibility in this market will be the ones that issue meaningful, granular credentials alongside the degree, not in opposition to it.

Federal policy is reinforcing the shift. Workforce Pell takes effect July 1, 2026 and requires 70% completion and 70% job placement for participating short-term programs (U.S. Department of Education). Most current short-term offerings won’t clear that bar, which is exactly the point. And on January 9, 2026, the AHEAD rulemaking concluded with a historic 12-0 consensus vote - what the Department called the most significant accountability shift in three decades. The federal posture is moving from credentialing-as-input to credentialing-as-outcome, and microcredentials sit squarely in the new line of sight.

Expense Letter Template

What ‘Stackable’ Actually Means

Three properties separate a coherent microcredential portfolio from a chaotic one.

Verifiable competency. Each badge is tied to specific learning outcomes that can be assessed and replicated. Purdue set the U.S. bar last fall when its trustees approved an AI working competency as a graduation requirement for every student - not a vague aspiration but a defined standard the IDA+A data science team is accountable to measure against. That’s what verifiability looks like in practice: a named competency, a documented assessment method, and an institutional team responsible for proving the bar got cleared. Without it, a badge is a participation trophy.

Architected stacking. Badges have explicit prerequisite relationships and explicit pathways into degrees, certificates, or job classifications. The Medical University of South Carolina’s competency-based AI literacy framework, grounded in the Digital Education Council AI Literacy Framework, defines a shared progression across awareness, competency, and application. An awareness-level badge in one school maps cleanly onto an application-level requirement in another - which is what ‘stackable’ has to mean if students are going to combine credentials across schools and over time.

Employer-recognized. A badge has portability only if employers in the relevant region say so. Georgetown CSET’s PATHWISE tool maps AI and cybersecurity job demand at the metro level, giving institutions a way to ground credential design in real-time labor-market signals rather than internal assumption. Carnegie Mellon’s Heinz College takes the validation a step further: in its employer-validated capstone program, employer partners sponsor semester-long projects, teams tackle real challenges, and both faculty and the employer evaluate. The credential records what the employer signed off on - not what the syllabus said it should. For institutions wanting an external proficiency scale to anchor against, the University of Manchester’s AI literacy partnership with Multiverse uses the European Digital Competence Framework’s six levels (Newcomer through Pioneer) - a useful comparative reference for U.S. institutions designing their own.

When all three are present, microcredentials stop competing with the degree and start completing it.

Three Programs Worth Studying

University of Hawai’i - AI Foundations Badge. UH launched ‘Artificial Intelligence for Hawai’i’ in March 2026, a free 12-chapter interactive program for the UH ‘ohana and surrounding communities. New chapters release weekly; completion earns an AI Foundations Badge - a digital credential demonstrating foundational AI literacy, ethical practice, and responsible use across the UH system. Two design choices stand out. First, the badge is system-wide, not departmental, which gives it the same brand weight a UH degree carries. Second, it pairs an entry-level credential with a free public on-ramp - the badge becomes the institution’s invitation to non-traditional learners, not just to students already paying tuition.

Purdue University - Charlie, AI Writing Feedback Tool. Purdue’s data science team rebuilt a faculty research project into Charlie — an AI writing coach live in 40 classes serving ~1,400 students. It abandoned essay scoring entirely in favor of written coaching feedback only, running GPT-4 via Azure with RAG from course PDFs. Build cost: roughly one full-time data scientist on Purdue’s general budget.

Two choices stand out. First, they built a separate evaluation model, calibrated to a faculty expert’s judgment, that scores Charlie’s feedback quality every time the tool changes. That’s how they caught that RAG improved accuracy but made the tone harsher. They chose accuracy. Most teams wouldn’t have known there was a choice. Second, development is currently paused while the commercial market catches up. Purdue has pre-committed to stepping back if a vendor matches their quality. Knowing when to stop building is as rare as knowing how to start.

Vanderbilt - Multi-Tool Stack with Grow with Google. Vanderbilt has expanded community-wide access to a coordinated multi-tool AI stack: ChatGPT Edu, the Amplify 2.0 platform built in-house, and Grow with Google certificates spanning analytics, project management, UX, cybersecurity, digital marketing, and IT support. Reported adoption: more than 70% of the campus community uses generative AI tools regularly. The lesson is interoperability. Vanderbilt isn’t trying to issue every credential itself; it’s issuing the integration - wrapping vendor certificates, internal platforms, and OpenAI-grade access in a single recognizable institutional package. The university becomes the trust layer over a credential ecosystem its students will keep using long after graduation.

Each of these programs makes a different bet about where institutional value sits - in the badge itself, in the credit-employer co-design, or in the institutional integration of third-party credentials. All three are defensible. What none of them can skip is the governance scaffolding that makes the credential trustworthy.

Governance Moves That Make Microcredentials Stick

Microcredentials fail not because students can’t earn them but because institutions can’t vouch for them. From advising work with university systems, four governance moves separate credentials that travel from credentials that don’t.

1. Define the credential ladder before issuing the first badge. Decide upfront how each microcredential maps to (a) credit and degree progression, (b) external proficiency frameworks (DigComp, Digital Education Council, NIST AI RMF role definitions), and (c) job classifications in your local labor market. Without this map, badges accumulate without composing - which is exactly the chaos employers complain about and accreditors flag.

2. Treat assessment design as a governance artifact, not a course matter. Each badge needs a documented competency definition, an assessment method, an inter-rater reliability target, and a periodic review cadence. Borrow Purdue IDA+A’s discipline of triangulating evaluation across an LLM rater, a human subject-matter expert, and a calibration loop. Defining ‘good’ is hard; credentials live or die on whether your institution can defend that definition under scrutiny.

3. Bring procurement and vendor due diligence into the credential conversation. When the badge runs on a third-party platform - a Coursera Coach, a Multiverse cohort, a Grow with Google certificate - the vendor’s data practices, accessibility posture, and IP terms become institutional liabilities the moment your seal is on the credential. Credential governance needs the same rigor as enterprise procurement: HECVAT-plus-AI annex, FERPA-aligned data tiers, audit logging, and clear language on who owns the verification record if the vendor relationship ends.

4. Publish outcomes data. Workforce Pell will require 70% completion and 70% placement. Don’t wait for the rule to bite - institutions that publish completion-to-placement data on every microcredential build credibility with both employers and accreditors well before the threshold becomes a compliance question. The institutions doing this best treat outcome reporting as a credentialing feature, not a regulatory burden.

What’s Ahead

Issue #6 will turn to AI hiring committees - how institutions are setting up cross-functional review boards for AI-related faculty and staff appointments, what those committees actually evaluate, and how to staff one without creating a bottleneck.

We’ll also start a recurring ‘Use Case of the Month’ feature, anchored to a single validated implementation drawn from the open AI Use Cases handbook.

And we’ll continue tracking federal policy: the AHEAD rulemaking implementation timeline, Workforce Pell program approvals, and where state AI frameworks are landing - Digital Promise’s December 2025 review of 32 states found most still nascent and exploratory, which means the institutions moving now are the ones writing the templates everyone else will eventually adopt.

With gratitude,

Dr. Aviva Legatt

Founder, EdGenerative - Affiliated Faculty, University of Pennsylvania - Forbes Contributor

What AHEAD/STATS Actually Does

Four facts every cabinet should hold steady.

The earnings premium becomes the test, and it covers everything. STATS replaces the debt-to-earnings ratio with a single accountability metric: do completers earn more than working adults aged 25–34 without that level of education? The standard now applies to all programs at all Title IV institutions—public, nonprofit, and proprietary alike (Federal Register, April 20, 2026).

Failure has a defined consequence, and it is fast. Programs that fail in two of three consecutive years lose Direct Loan eligibility. Institutions where more than half of Title IV recipients or more than half of Title IV volume sit in failing programs go on provisional status and lose Pell access for those programs. The institutional consequence is not just per-program defunding; it is a Pell-side existential risk for institutions whose volume concentrates in failing programs.

Reporting is on a tight calendar. Institutions must submit program-level tuition, fees, and financial-aid data—by program—by October 1, 2026, covering the two prior award years, then annually each October 1.

Read more

Welcome to the first recap of The Use Case Lab — our monthly live conversation series where we open the hood on how universities are operationalizing AI.

For our inaugural episode, I was joined by Ian Pytlarz, Principal Data Scientist at Purdue University’s Institutional Data Analytics + Assessment (IDA+A) team. Ian’s team sits under the Office of the Provost and has an unusually wide mandate: from AI models that saved Purdue $1.27 million annually on campus utility bills, to student-facing learning tools. He’s also co-chair of Purdue’s Data Ethics Committee, which, as you’ll see, makes him exactly the kind of balanced voice this conversation needed.

We spent nearly an hour on one specific use case: Charlie — Purdue’s AI-powered writing feedback tool currently in use in 40 classes, serving approximately 1,400 students. The conversation went well beyond Charlie. It became a masterclass in how to evaluate, build, govern, and scale AI tools in higher ed without making expensive mistakes.

Here’s what you need to know.

What Is Charlie?

Charlie started as a faculty research project by Dr. Lindsey Hamm, who saw a simple but stubborn problem: there’s never enough time to give students feedback on writing drafts before their final grade. Students submit work. They wait. They get a grade. They don’t learn as much as they could.

Her original idea? Use AI to score essays by rubric. That approach mostly fell flat, not because the idea was wrong, but because the execution needed to be refined. Giving a student a score without explanation isn’t helpful. And TAs don’t grade consistently, which confused the underlying models.

When Ian’s team got involved around 2023, they tore the idea down and rebuilt it. The new version shifted entirely away from scoring and toward written feedback — analyzing the student’s draft against the rubric and generating guidance to help them improve, not just judge where they landed.

The next challenge: getting the AI to stop trying to do the assignment for the student.

Getting it to coach rather than complete took significant prompt engineering, iteration, and what Ian describes as a deliberate feedback loop.

The Key Insight: Building a Feedback Loop to Know If It Works

This is where Ian said something I want every higher ed leader to hear:

Most people deploying AI have no idea if their tool is actually working.

Ian’s team built a rigorous evaluation process.

Read more

Read more

Something quietly shifted in the employer landscape over the past year, and most higher ed leaders haven’t noticed yet.

Employers are adopting AI at a pace that makes innovation cycles look quaint. 44% of U.S. businesses now pay for AI tools—up from just 5% in 2023. But here’s the tension: only 37% of these same employers say higher education is actually preparing their future hires with the skills they need. That’s not a minor perception gap. That’s a crisis of institutional relevance.

And it’s directly tied to enrollment, reputation, and your institution’s ability to serve the workforce it claims to prepare.

The Numbers Behind the Alignment Crisis

The WGU Workforce Decoded Report: What Employers Actually Need

The most comprehensive look at this alignment problem comes from the WGU Workforce Decoded Report, which surveyed 3,147 employers in 2026. The findings are both clarifying and sobering:

This is the moment when you need to sit with the implications. Employers are clearly asking for AI fluency. Faculty report that graduates aren’t prepared. And institutions have yet to coordinate a systematic response.

The problem isn't that institutions don't care about workforce outcomes. The problem is that they're not designed to move at the speed employers are changing.

To grasp the full scale: up to 80% of the U.S. workforce could have 10% or more of their job tasks exposed to large language models (OECD Skills Outlook 2025). Meanwhile, 63% of faculty report that spring 2025 graduates are not prepared to use AI effectively in the workplace (AAC&U / Elon University survey). And despite all this urgency, 90% of institutions say they’re integrating AI into teaching and research—suggesting that integration and alignment are two entirely different things.

The disconnect is real, measurable, and growing. So what are the institutions that are getting this right actually doing?

Purdue in Practice: Inside the IDA+A Data Science Team

Graduation requirements set the expectation; the harder work is what sits behind them. Purdue’s Institutional Data Analytics + Assessment (IDA+A) team — led by Lead Data Scientist Ian Pytlarz — is where that work becomes operational. Two recent projects Pytlarz shared with me show how a workforce-aligned institution embeds AI into both the back-of-house and the classroom at the same time.

Energy & Utilities predictive modeling. Purdue’s campus E&U operation spends roughly $39 million a year on utilities, and its purchasing priority has historically been reliability over economic efficiency. Pytlarz and Senior Data Scientist Darshan Bhansali built models that forecast campus demand for electricity, steam, and chilled water 96 hours out — refreshed hourly — with consistent accuracy since November 2021 (2.83% MAPE on campus electric, 3.93% total; 7.56% on steam). Paired with Duke Energy’s real-time price data, the system surfaces the lowest-cost operational plan that still meets predicted demand. When the team compared 2022 model recommendations against actuals, the approach pointed to roughly $1.27 million in yearly savings — the kind of measurable AI outcome that moves a board conversation from ‘should we?’ to ‘how fast can we scale it?’

Charlie and the ‘codebook’ approach to LLM evaluation. Charlie is a student writing assistant: faculty upload a grading rubric, students submit drafts, and Charlie returns rubric-grounded feedback. The interesting problem isn’t the tool — it’s the evaluation layer behind it. Pytlarz’s team has built a disciplined evaluation method — an LLM rater, a human subject-matter expert, and a human rater triangulating on whether a prompt change actually made feedback better, with a target of roughly 0.5 Mean Absolute Error between human and LLM judge. Their biggest takeaway, stated plainly on their own slide, is one every academic leader should internalize: “Defining ‘good’ is difficult. Criteria drift — evaluating the output changes the evaluation criteria. Simple word swaps (‘bad,’ ‘good,’ ‘excellent’) are not enough.”

Why this matters for workforce alignment: Purdue’s board set the competency standard, but the IDA+A team provide infrastructure to make it happen. Students graduate having encountered AI as a rubric-grading tutor in their writing classes and as a $1.27M operational savings model inside the institution they attend. That is what ‘prepared’ looks like when it is real — not a syllabus update, but an entire data science function treating LLMs as research artifacts requiring the same calibration, measurement, and criteria-discipline that employers now expect of new hires. With thanks to Ian Pytlarz, Lead Data Scientist, Purdue IDA+A, for sharing the E&U Walkthrough and Charlie & Codebook presentations that informed this section. The work described is his team’s.

The EdGenerative AI Governance Toolkit

The workforce alignment mismatch isn’t just about curriculum — it’s about whether institutions have the governance infrastructure to move quickly enough. Employers are adopting AI faster than most universities can update a syllabus, and the institutions getting it right are the ones with frameworks that connect academic programs to workforce signals in real time. Through my work advising institutions, I developed a policy-neutral, audit-ready governance framework mapped to NIST AI RMF, DOJ Title II, FERPA, HECVAT, and ISO/IEC 42001. It’s built around Ten Core Components that give institutions a clear path from pilot to practice.

For institutions tackling the workforce alignment crisis profiled in this issue, three components are especially relevant: Component 9 (Training & Culture) provides the infrastructure for role-based AI literacy — not a one-size-fits-all workshop, but differentiated micro-modules that prepare faculty to teach with AI, staff to work alongside it, and students to enter a workforce that expects AI fluency. Component 1 (Clear Governance & Roles) ensures workforce alignment isn’t siloed in career services or the provost’s office, but coordinated across academic affairs, employer partnerships, and institutional research. And Component 6 (Procurement & Vendor Due Diligence) matters more than you might think — the AI tools you bring onto campus for workforce preparation need the same governance rigor as any enterprise system, especially when they touch student data and learning outcomes.

The toolkit includes a 90-Day Starter Plan sequenced across three phases: Q0 (Days 0–14) for publishing a governance charter and drafting your first operational AI assessment; Q1 (Days 15–45) for standing up risk assessments and procurement safeguards; and Q2 (Days 46–90) for launching monitoring, training, and management system infrastructure. It’s designed to produce auditable artifacts at every stage — the kind accreditors and boards actually want to see.

Try This Week

Longer-Term Strategy: Build an Employer Council (60-90 days)

Your 90-Day Implementation Plan

Days 1–30: Discovery and Landscape Scan

Days 31–60: Design Employer-Aligned Competency Framework

Days 61–90: Pilot Integration and Feedback Loop

What’s Ahead

Issue #5 will dive into microcredential governance—how to build stackable, badged credentials that employers recognize and that pair with degrees without creating portfolio chaos.

We’ll also continue documenting case studies from the handbook: how institutions have redesigned curriculum for AI, how they’ve built employer feedback loops, and the governance moves that made it stick.

And we’ll go deep on the federal policy shifts—Workforce Pell, accreditation guidance, and the emerging labor standards that will shape what ‘prepared’ actually means.

With gratitude,

Dr. Aviva Legatt

Founder, EdGenerative • Affiliated Faculty, University of Pennsylvania • Forbes Contributor

About Dr. Aviva Legatt

Dr. Aviva Legatt is the founder of EdGenerative, where she advises university boards, presidents, and systems on AI governance, adoption strategy, and microcredentials. She holds a doctorate from the University of Pennsylvania, where she serves as affiliated faculty teaching organizational dynamics. A Forbes Contributor, she has covered education leadership and interviewed figures from Simone Biles to Adam Grant. She serves on the Montgomery County Advisory Council on Artificial Intelligence for Public Good and is the author of Get Real and Get In (St. Martin’s Press). Her AI Use Cases in Higher Education: A Community Handbook is the open-access resource behind this newsletter.

When I started the Higher Ed AI Playbook, I had one goal: give senior higher ed leaders the analysis they actually need to govern AI well — not the hype cycle, not the fear cycle, just the work.

So I’m building something I’ve wanted to build for a while.

The Use Case Lab

The last Thursday of every month at 12 PM ET, live on Zoom (~60 minutes), we sit down together and work through one real institutional AI deployment in depth.

How each Lab works:

• One real use case, end to end. What the institution built, how they governed it, what worked, what broke, and what they’d do differently.

• Live discussion, not a lecture. You’ll be in the room with other higher ed leaders wrestling with the same decisions. Bring a question from your own campus; you’ll leave with something you can use Monday.

• Recorded and archived. Can’t make it live? The full session and a written summary go in the archive, searchable and yours to keep.

This is the piece I get asked for most. I’ve been doing it informally for client campuses for two years. Now it runs every month, for paying subscribers, at a fraction of what consulting costs.

What else comes with paid — $15/month or $150/year

• The Use Case Lab (above)

• Monthly governance briefings — what’s shifting at the federal, accreditor, and institutional level, and what it means for your roadmap

• The full archive and resource library, including every framework and template I’ve built

• Comment and discussion access on paid posts

Founding Member — $250/year

For readers who want to be inside the work:

• Everything in paid, plus:

• A private founding member discussion thread to troubleshoot your own institutional use cases between Labs and get input from peers doing the same work at other campuses

• Your use case, considered for a future Lab or newsletter feature. After you subscribe, you’ll get a short form to pitch what you’re building — credited or anonymized, your call

• Occasional surprise drops: interview clips, early reads, members-only conversations

What doesn’t change

Everything the Playbook has been stays free. Regular posts on governance, policy, and pedagogy. The Use Case Handbook.

My Forbes analysis translated for practitioners. If you never upgrade, you lose nothing. I’m not moving existing content behind a paywall.

Why I’m telling you this directly

Paid subscribers make this work sustainable. They let me run the Lab every month, which takes real prep. They let me keep writing in public instead of disappearing into client work.

But I won’t pretend every reader needs to pay. If the free tier serves you, stay free. If you’ve been waiting for the paid version because you wanted something worth paying for — the Lab is here.

Depending on your role, this subscription may be tax-deductible as a professional expense, or eligible to be run through your institution’s professional development budget. Worth a quick note to your finance team or admin.

— Aviva

When I sit with enrollment leaders and student affairs VPs, the same number comes up in almost every conversation: the advising ratio. Your institution has roughly the same number of advisors it had a decade ago. But the complexity your students face — navigating majors, financial aid, course scheduling, career pathways, mental health resources, and policies that change every semester — has tripled. Many students don’t even know where to start asking for help.

This is the paradox at the heart of modern higher education: rising student need meets constrained institutional budgets. In my work advising institutions on AI strategy, I’ve seen this gap widen at every type of institution — from flagship research universities to regional comprehensives. And it’s the reason some institutions are quietly running experiments that are reshaping what student support looks like.

The Larger Landscape: What Staff Need to Know

Before you launch any student support AI initiative, your staff need to understand the bigger picture. Here’s what the data shows about institutional readiness:

Translation: more than half your staff are already using AI, but they don’t know what your policies are. This is a finding I bring into every governance conversation I lead — because it reframes the stakes. Student support AI matters beyond just saving advisor time. It’s a chance to align staff practice with institutional values. When you deploy a governed, institutional AI tool, you’re not just solving a workflow problem. You’re setting a precedent for responsible AI use across campus. And institutions that establish that precedent early have a significant advantage when it comes time to scale.

Case Study 1: University of Hawaii System — Campuswide Chatbot Deployment

The University of Hawaii System serves 10 campuses and 33,000 students across the archipelago. In 2025, they launched an AI chatbot system designed to answer the most common student support questions across all campuses. The results offer a clear lesson in scaled student support.

But here’s what made the Hawaii deployment succeed; they didn’t just build a chatbot and turn it loose. They integrated it into existing staff workflows. The system was trained on actual institutional data — course catalogs, financial aid policies, campus calendars — so answers were contextual, accurate, and campus-specific. And critically, it was designed to escalate, not replace. When a student question required human judgment, the chatbot logged it and routed it to the appropriate department.

The 2,533 students flagged for follow-up weren’t a failure. They were precisely identified cases where the chatbot recognized a student needed human support — financial hardship, mental health concern, academic probation — and flagged them for proactive outreach. In other words, the chatbot didn’t just save staff time. It surfaced students who needed help in ways the institution would have missed otherwise.

This is a pattern I’ve seen across every successful student support AI deployment — and one I explicitly build into the governance frameworks I design for institutions: the goal isn’t to eliminate human judgment. It’s to preserve it for the decisions that matter most.

Case Study 2: Georgia State University — Predictive Analytics at Scale

Georgia State University serves 50,000 students and has become a national leader in using AI for retention and enrollment management. But their approach is notably different from Hawaii’s. Where Hawaii deployed reactive support (answering questions), Georgia State built a predictive system designed to intervene before problems emerge.

A 23% improvement in 6-year graduation rates means hundreds of additional students graduating who otherwise would have dropped out. And that improvement came not from a new program, but from better information reaching advisors at the moment it matters most.

Georgia State’s system works by feeding advisors timely, relevant risk signals. When an advisor sees that a student is suddenly missing class, when financial aid disbursement hasn’t posted, when a student’s grades drop in a prerequisite course, the system flags it. The advisor reaches out. The student doesn’t disappear. Graduation rates go up.

But — and this is crucial — the system doesn’t make the decision. The advisor does. The AI is doing what it’s actually good at: pattern recognition at scale. The human is doing what only humans can do: understanding context, building relationships, and exercising judgment.

What Others Are Learning: Peer Institutions and Emerging Patterns

Hawaii and Georgia State aren’t alone. A growing number of institutions are experimenting with AI student support. Here’s what the broader landscape looks like:

Notice the pattern: every institution that’s getting this right is deploying governed AI. Not just turned loose. Governed. In our governance pilots, we start by mapping exactly these elements: clear goals, stakeholder input, bias monitoring, escalation pathways, and regular audits of what’s working and what isn’t. Institutions that build this infrastructure before scaling find that adoption accelerates — because trust is already in place.

Who Needs to Be in the Room

In my work with advising directors and student affairs leadership, I’ve learned that the stakeholder map is what separates a chatbot pilot that stalls from one that scales. Before you deploy student support AI, you need to know which stakeholders will make or break the implementation. Here’s the map:

Each of these groups has different priorities, different data security concerns, and different definitions of success. In the governance engagements I lead, I’ve found that the single biggest predictor of pilot success is whether advising staff were involved in the design — not just consulted after the fact. If you bring them all in early, you’ll move faster than you’d expect.

The EdGenerative AI Governance Toolkit

Chatbots and student support systems touch some of the most sensitive data on campus — academic records, mental health disclosures, financial aid status, advising notes. Deploying them without governance infrastructure isn’t just risky; it’s a compliance liability. Through my work advising institutions, I developed a policy-neutral, audit-ready governance framework mapped to NIST AI RMF, DOJ Title II, FERPA, HECVAT, and ISO/IEC 42001. It’s built around Ten Core Components that give institutions a clear path from pilot to practice.

Your Action Plan

Your 90-Day Implementation Plan

Coming Next in Issue #4

With gratitude,

Dr. Aviva Legatt

Founder, EdGenerative • Affiliated Faculty, University of Pennsylvania • Forbes Contributor

About Dr. Aviva Legatt

Dr. Aviva Legatt is the founder of EdGenerative, where she advises university boards, presidents, and systems on AI governance, adoption strategy, and microcredentials. She holds a doctorate from the University of Pennsylvania, where she serves as affiliated faculty teaching organizational dynamics (Coursera). A Forbes Contributor, recognized among the top 2% of writers, she has covered education leadership and interviewed figures from Simone Biles to Adam Grant. She serves on the Montgomery County Advisory Council on Artificial Intelligence for Public Good and is the author of Get Real and Get In (St. Martin’s Press). Her AI Use Cases in Higher Education: A Community Handbook is the open-access resource behind this newsletter.

If you’re a provost, dean, or academic leader trying to figure out where your institution stands on AI—and where it should be heading—you’re not alone. The pace of change in generative AI has outstripped most institutions’ ability to develop thoughtful policy, let alone keep track of what peers are doing or what’s actually working in classrooms.

That’s why I built two open resources designed to give higher education leaders a running start. Today I’m sharing both of them with you as downloadable spreadsheets—living documents that I update regularly as the landscape evolves.

Resource 1: AI Usage Policies in Higher Education Handbook

What it is: A comprehensive, curated database of AI usage policies from colleges, universities, and higher education organizations around the world. Each entry captures the institution, policy type, key provisions, scope, and a direct link to the source document—all independently verified.

How to use it: Use this resource to benchmark your institution’s AI policy against peers. Filter by policy type to find models that match your governance structure. Share the statistics tab with your board or faculty senate to ground conversations in data rather than speculation. And pay close attention to the enterprise risk section if your institution is exploring AI agents or autonomous tools—this is where the next wave of policy challenges is emerging.

Resource 2: AI Use Cases in Higher Education Handbook

What it is: A curated catalog of validated, real-world AI implementations across higher education. Every entry documents a specific use case, the institution behind it, the AI tools involved, the current status of the initiative, and a source link for further reading.

How to use it: If your institution is debating whether to pilot AI in a particular area—say, student advising or course design—start here. Look at what peer institutions have already implemented, what tools they chose, and whether the initiative is still active. This can save your team months of exploratory work and help you build a stronger case for investment by pointing to validated precedents.

Why I Created These Resources

I created these handbooks because I kept having the same conversation with academic leaders: they knew AI was transforming higher education, but they didn’t have a reliable, centralized place to see what was actually happening across the sector. Policy documents were scattered across institutional websites. Use cases lived in conference presentations and press releases. And the gap between what vendors promised and what institutions had actually implemented was enormous.

These resources are my attempt to close that gap. They’re designed to be practical—not theoretical—and to evolve as the field does. I update them regularly, and both include a submission tab so that you can contribute your own institution’s policies or use cases to the growing database.

The most useful resource in a fast-moving landscape is one the community builds together.

Download and Share

Both resources are available as downloadable Excel files attached to this issue. I encourage you to share them with colleagues, reference them in committee discussions, and use them to accelerate your institution’s AI strategy. If you find them valuable, reply to this email and let me know—I’d love to hear how you’re using them.

And if your institution has an AI policy or use case that should be included, use the submission tabs in each handbook to contribute. The more complete these resources become, the more useful they are for everyone.

With gratitude,

Dr. Aviva Legatt

Founder, EdGenerative · Affiliated Faculty, University of Pennsylvania · Forbes Contributor

Here’s the tension that defines this moment in higher education: 90% of faculty say AI is weakening critical thinking. Ninety percent. That number keeps me up at night, because it reflects real concerns about how students learn and what education is for. And yet, in the same landscape, a peer-reviewed randomized control trial published in Nature Scientific Reports shows that AI tutors achieved double the learning gains compared to in-class active learning, with higher engagement and motivation.

Both things are true. How do we reconcile this? .

The Evidence: What the Research Actually Shows

The paradox starts to resolve when you look at the details. It’s not AI itself that drives learning gains or learning loss. It’s how AI is deployed.

Notice what’s different about these implementations: they’re not unstructured AI chatbots answering any question in any way. Harvard’s AI tutors were designed with pedagogical intent — deployed to supplement, not replace, active learning, with clear learning outcome alignment. Dartmouth’s NeuroBot uses retrieval-augmented generation (RAG) to ground responses in course materials that faculty have vetted. In both cases, the AI is structured by human expertise. And that structure is where the learning gains come from.

The studies the research community has published show a consistent pattern: well-designed, pedagogically informed AI tutoring produces measurable learning gains. But “well-designed” is the operative phrase. Generic AI assistance doesn’t cut it.

Three More Data Points from the Field

UniDistance Suisse (Switzerland): A semester-long study with 51 psychology students using AI-generated microlearning questions derived from course materials showed an average of 15 percentile point improvement. The AI wasn’t generating original content—it was generating learning activities from course materials. That distinction matters.

Coursera AI Coach: Their AI-powered personalized coaching system supports 1M+ learners globally and shows concrete performance gains: a 9.5% higher quiz pass rate on first attempt and 11.6% more lessons completed per hour. At scale, with 1 million learners, this is evidence of what AI tutoring can do.

Carnegie Mellon University: Researchers developed a methodology that enables teachers to create intelligent tutoring systems in approximately 30 minutes by simply demonstrating problem-solving approaches. This is crucial: CMU’s work shows that faculty don’t need computer science backgrounds to build effective AI tutoring systems. They need structured design thinking and a clear pedagogical goal.

The Numbers: Faculty Concerns vs. Research Evidence

Here’s where the paradox becomes visible:

Faculty are concerned, yet committed to expanding AI use. They sense both the risk and the opportunity.

Who Needs to Be in the Room: A Stakeholder Engagement Map

Deploying AI tutoring at scale isn’t a technology project. It’s a governance and change management project. These conversations need to happen:

The institutions that are getting this right don’t approach AI tutoring as a pilot. They approach it as an implementation that requires governance. That means representation across all seven stakeholder groups, with explicit decision rights and communication protocols.

The Governance Layer That Determines Success or Failure

Before you launch an AI tutoring pilot, your governance framework needs to address five critical areas:

These questions aren’t obstacles. They’re the structure that makes sustainable implementation possible. Institutions that skip them run into faculty resistance, union grievances, and sustainability problems.

The EdGenerative AI Governance Toolkit

Every AI deployment — whether it’s a tutoring system, a chatbot, or a grading platform — eventually surfaces the same question: who governs this? Through my work advising institutions, I developed a policy-neutral, audit-ready governance framework mapped to NIST AI RMF, DOJ Title II, FERPA, HECVAT, and ISO/IEC 42001. It’s built around Ten Core Components that give institutions a clear path from pilot to practice.

For institutions deploying AI tutoring systems like the ones profiled in this issue, three components are especially critical: Component 3 (Risk & Impact Assessment) ensures you evaluate the tutoring tool’s pedagogical claims, data handling, and bias risks before scaling. Component 7 (Human Oversight & Recourse) defines when students can override AI recommendations and how faculty maintain pedagogical authority. And Component 9 (Training & Culture) ensures faculty aren’t just handed a tool — they’re supported in integrating it into their teaching practice.

Your Week: Build the Business Case

Your Longer-Term Strategy: Design a Rigorous Pilot

Your 90-Day Implementation Plan

At the 90-day mark, you’re not looking for a final verdict. You’re looking for evidence of learning and a clear sense of the institutional change required to scale. That’s when the real work begins.

Why This Matters Right Now

The timeline matters. According to a 2026 study by the Brookings Global Task Force, involving over 500 stakeholders across 50 countries: AI risks to learning currently outweigh demonstrated benefits. That doesn’t mean AI in tutoring is bad. It means the unstructured, ungoverned use of AI in tutoring is risky. Structured, pedagogically informed AI tutoring—the kind Harvard and Dartmouth are running—is exactly the evidence institutions need to demonstrate responsible innovation.

But here’s the catch: you can’t build that evidence without running the pilot. And you can’t build sustainable pilots without the governance structure in place first. That’s why the next 90 days are critical. Institutions that start now will have semester-long outcome data by fall. That data becomes your institutional narrative—the difference between being a fast-follower and a leader in responsible AI adoption.

Coming Next: Issue #3

“The Chatbot That Saved 357 Staff Hours” — how the University of Hawaii and Georgia State built AI student support systems that freed human staff to do higher-value work. We’ll break down the economics, the implementation path, and what happens when you get the governance right.

With gratitude,

Dr. Aviva Legatt

Founder, EdGenerative · Affiliated Faculty, University of Pennsylvania · Forbes Contributor