NUUBI
Back to Blog

A Practical AI Governance Framework for Universities

· Nuubi Team

A university can publish an AI policy and still leave its most consequential decisions unresolved. A student services office wants a chatbot that can see enrollment records. An instructor wants students to get AI feedback on drafts before they submit. An admissions team wants help sorting applications. Each proposal raises different questions, yet all three may land in the same inbox with no clear owner, deadline or shared standard for approval.

AI governance for universities is the system that assigns those decisions, limits what gets approved and checks what happens afterward. A policy states expectations, and governance turns them into routes, records and accountable people. When it works, worthwhile uses move forward without every experiment being treated as equally risky or consequential choices being left to a vendor's default settings.

The recommendations below, written for presidents, provosts, deans and CIOs, draw on widely cited frameworks and published university examples. They are a starting point for institutional design and are not legal advice, so each institution will need to adapt them to its own governance structure, applicable law and the advice of counsel.

Reference Frameworks Offer a Common Vocabulary but Leave the Decisions to the Institution

Several shared frameworks give universities a common vocabulary. The NIST AI Risk Management Framework organizes review into four functions: Govern, Map, Measure and Manage. The OECD AI Principles, updated in 2024, emphasize human agency, actionable transparency and accountability. UNESCO's guidance on generative AI in education and research asks whether a use protects human agency and privacy, fits the pedagogy and matches staff capacity. ISO/IEC 42001 adds a management-system view, and the EDUCAUSE 2024 Action Plan on AI policies and guidelines connects governance to operations and pedagogy.

These frameworks help a university ask better questions, but the institution still has to decide what it permits, what evidence it requires and who answers for the outcome.

Decision Rights Should Be Spread Across the Institution

A common first instinct is to route everything through a new AI committee. A more durable approach gives each type of decision a clear home within existing shared governance. An executive sponsor owns resources and risk appetite, a cross-functional group sets common standards and resolves exceptions, specialists in IT, privacy, security, procurement, legal and accessibility review the controls in their remit, and faculty governance handles teaching, learning and assessment. Every live use also has an operational owner for support, monitoring and escalation. Institutions can name a chief AI officer as that sponsor where the role exists, or otherwise designate a provost or CIO.

Students and frontline staff belong in policy development and periodic review, since those who meet failures first often see risks a purchasing team cannot. Publish two routes: how a proposal reaches a decision, and how someone can challenge a decision or output. A committee without decision rights, staff time or an escalation path adds meetings without adding control.

Small universities can run a leaner version of the same system. Where one IT director covers privacy, security and procurement, those three reviews can be combined into a single checklist, as long as that director does not approve a use they also requested or run. The smallest workable version has three parts: an executive sponsor who sets risk appetite and signs off on high-stakes uses, a standing group of three or four people, often the IT director, a faculty governance representative and someone from academic affairs or student services, who review proposals on a fixed schedule, and a named operational owner for each live use. Outside counsel can be brought in only for high-stakes reviews. Consortia and shared services can carry part of the load through group purchasing agreements with negotiated data terms, shared vendor security reviews and common intake and contract templates. Accountability for each approval still stays with the institution.

Review Should Be Proportionate to the Use

The same AI tool can draft an event announcement or influence a student's academic standing, so approval works best when it attaches to a specific use, defined by purpose, data category, affected population and conditions, instead of a product name.

The three tiers below show how that can work in practice. They are typical examples and starting points that each institution can adjust, and most new ideas will land in the lower two and should move quickly. Whatever the tier, the approval should record the use, its owner and a review date. Some uses may stay excluded until the institution can show adequate controls, and the tiers are illustrative and do not reflect any statutory risk classification.

Low risk: Public information only, such as drafting event announcements. A supervisor signs off under the acceptable-use policy, and the vendor confirms university data will not train its models.

Medium risk: Student work with no change to records, such as AI feedback on course drafts. The academic unit approves with IT review and vendor data terms in place, students are told the feedback can be wrong and how to flag it, instructors spot-check, and feedback that affects a grade is reviewed before release.

High risk: Restricted records or consequential decisions, such as sorting admissions applications or a chatbot answering from enrollment records. This tier needs full review by IT, counsel and the data owner, sign-off from a named executive sponsor, bias and accuracy testing, and a tested kill switch.

The AI Coordination Committee in the University of Minnesota's College of Science and Engineering offers a concrete college-level model. It separates self-service innovation using public data from centrally managed tools involving sensitive data, high-impact changes or use across departments, which follow a more formal route. It is one college's approach, but matching review effort to consequence carries over well.

Northeastern's public Policy 125 sets a specific intake trigger for university operations. An AI system and its use case must go to both the AI Review Committee and Information Security when it processes confidential information, personal information or restricted research data, or when its actions or outputs may affect legal rights or physical safety. A separate public data-classification matrix classifies information by the potential harm from unauthorized disclosure, use, alteration or loss. Together, the two documents suggest treating data sensitivity and consequence as separate questions, since a harmless-looking task can still involve sensitive input data.

Stanford University IT's agentic AI guidance weighs tool access, decision influence, external exposure and autonomy alongside data sensitivity. Its high-risk tier includes hard-to-reverse actions and calls for human oversight, annual review and a tested kill switch. These classifications are Stanford's own and are not universal legal tiers. Other institutions can still adopt the underlying design question of whether a system only retrieves information or can also change a record or release a consequential decision.

Pepperdine University IT's tiered-trust AI evaluation process shows a working version. Every proposed tool is classified by the data it handles. A tool touching only public data needs supervisory approval, disclosure and a vendor guarantee that university data will not train a model. One handling limited personal or operational data adds review by IT or the Office of General Counsel, a SOC 2 Type 2 report and a data processing agreement. One involving restricted data, such as FERPA or health records, or core systems also brings in insurance and risk staff, the affected data owners and the AI Oversight Committee, plus a data protection impact assessment and a signed no-training addendum. The fast lane matters most, since routine uses move quickly and full review is saved for proposals that need it.

Two situations often fall through the cracks. The first is AI features added to software the university already licenses, such as a learning management system or productivity suite, where the original contract may not cover how those features handle data. A material feature change is a reasonable trigger for review. The second is free consumer tools adopted by individuals. An outright ban risks pushing that use out of view, where no one can see what data is being shared. A common alternative is to license an approved default with contractual data protections and make it easy to reach. Universities can pair that with a clear statement of which data must never enter unapproved tools, so the approved option becomes the simplest one.

Every Proposed Use Needs a Clear Review Record

A short intake record should identify the purpose, accountable owner, affected users, data, vendor and model dependencies, intended outputs and the decisions those outputs could influence. Reviewers then need predictable answers: which data the tool actually needs, where it goes and who can access it, what is retained and whether it trains a model, and what deletion or return means across the vendor and its subprocessors. Nuubi's article on questions to ask about student data in AI tools covers this in more detail.

The record should also address accessibility, total cost, copyright and ownership of outputs, procurement terms, support, performance testing, foreseeable misuse and exit arrangements. A demonstration with synthetic data can illustrate a workflow but does not show readiness for student records.

The approval should state its scope and conditions, the accountable owner, a review date and the events that would require earlier reassessment. That record proves its value when someone later asks why a use was approved and under what limits.

Institutional Safeguards Should Leave Room for Faculty Judgment

Central approval should set the floor for privacy, security, accessibility and acceptable risk while leaving teaching methods to the disciplines. Faculty are best placed to judge whether AI supports a learning outcome, and they need to communicate the rules at the course and assignment level.

The University of Rochester's guidelines for generative AI in education show one way to draw that line. Rochester preserves instructor choice over whether, when and how AI is used, expects course and assignment expectations to be stated, supports equitable access and calls for human oversight in grading and feedback.

In practice, a syllabus or assignment brief can explain permitted and prohibited uses, how students disclose AI use, what they must verify, what instructors can see in any course tool and where to get help. When a tool is required, students need a workable access plan, including accommodations and alternatives. Approving a tool for one purpose does not extend that approval to every category of institutional data or every assessment.

Human-in-the-Loop Oversight Works Best as a Defined Workflow

Human oversight works only when a specific person reviews outputs at a defined point, with enough information and the authority to act. Leaders can test a proposed workflow with a few questions. Who checks the output, and when? Who can correct or withdraw a harmful recommendation? Which decisions require review before they affect a student? How does a student challenge an output and reach a person? A reviewer clicking approve under time pressure and without context provides little protection.

University of Iowa guidance on AI-generated feedback for student assignments takes a stricter line, saying a person should check accuracy before feedback reaches students and that student content should not be uploaded without consent. Some campuses may choose that standard for certain courses. For most draft feedback, disclosure to students and regular instructor spot checks are a proportionate starting point, with review before release reserved for feedback that affects a grade. Each campus can decide where it draws that line and staff the review accordingly.

Consider a workflow for draft feedback. With the student's consent, the student writes an initial attempt and the AI drafts feedback on it. The course tells students up front that the feedback can be wrong and gives them a simple way to flag a comment they think is mistaken, such as a button in the tool or a note to the instructor. The instructor or a teaching assistant spot-checks a sample of feedback on a regular schedule and follows up on anything flagged, and when the feedback counts toward a grade, a person reviews it before release. The student revises, noting which suggestions they accepted or rejected and why, while grades and integrity findings stay with people. The department sets what triggers escalation, and how well this works depends on course design and staff time.

Monitoring, Incident Response and Renewal Keep Approvals Honest

Approval is a decision made under conditions, and those conditions need checking. A few measures tied to the intended benefit and foreseeable harms usually serve better than a large dashboard: sampled answer quality, complaints, accessibility barriers, staff review time and whether people can reach human help. Usage counts show adoption but say little about learning or fairness. Failures can be examined across student groups without collecting more sensitive data than the analysis requires.

A monitoring record can separate vendor and institutional responsibilities. The vendor might investigate an output failure or access-control defect, while the institution assigns someone to respond to the student and decide whether an academic outcome needs correcting. Tracking concerns raised, concerns resolved and time to resolution helps, though reported errors do not measure overall accuracy, because unreported errors are never captured.

Material changes to the model, data sources, integrations, vendor terms or intended use should prompt reassessment, and contracts can specify the notice expected before such changes. Incident planning should settle who can suspend a workflow, how affected users are notified, how evidence is preserved with appropriate access limits and what must be fixed before a restart. Retention should follow a justified schedule tied to those purposes.

The cost of skipping that planning is no longer hypothetical. In February 2024, a British Columbia tribunal held Air Canada responsible after its website chatbot gave a customer wrong information about bereavement fares. The tribunal rejected the airline's argument that the chatbot was a separate legal entity and held the company liable for all information on its website. If a university chatbot misstates a deadline, a policy or a student's options, the institution owns the answer. Every approved use needs a fast way to correct or suspend a tool, with a named person who can act.

Renewal is the moment to record whether to continue, change, pause or end a use, and why. The exit plan should cover data return or deletion, records the institution must keep, a replacement process and continuity for students partway through a course or program.

The First Six Months Should Build and Test the System

A reasonable sequence spreads the work across six months so shared governance bodies have time to weigh in. In the first two months, leaders can inventory current AI uses, including features embedded in licensed software, and the people deciding about them. Months three and four can go to drafting common guardrails and a proportionate intake process and taking them through faculty senate and other governance bodies. In months five and six, the institution can test the process on a few contrasting cases, such as a public-data writing aid, a course feedback workflow and an administrative use involving student records, before expanding it.

By the end of the six months, a maintained directory of approved uses, conditions, owners and help routes should give faculty, staff and students one place to check, with sensitive security details kept restricted. Training and review time need a budget line, because oversight added invisibly to existing roles tends to erode quickly.

The test of the resulting system is practical. Someone proposing an AI use should be able to find the right route, a reviewer should reach a documented decision in reasonable time, and anyone affected by an output should be able to reach a person with authority to correct it. When those conditions hold, the structure can absorb changes in the tools themselves.

This is why at Nuubi we treat governance as a key design principle in our workflow design. Our course software gives students AI answers grounded in the course materials faculty control, and keeps faculty in the loop on every answer, so instructors can see what students ask, what the AI says back and where the class is getting stuck. Oversight becomes part of teaching, not another task on someone's list.