A practical framework for protecting institutional accountability without restricting academic choice.

Centralize assurance, not necessarily choice.

IT manages identity and infrastructure, procurement handles contracts, and data protection teams assess privacy. Faculties decide how technology enters teaching, while researchers need enough freedom to test emerging models and methods. Generative AI cuts across every one of those boundaries, often before the people involved have agreed who owns the decision.

That is why AI governance in higher education is proving difficult. A university can make a new tool available in weeks, yet the operating model around it (the approvals, evidence, safeguards, budgets and accountability - takes far longer to establish.

One platform will not solve that. Neither will one policy.

Universities need shared guardrails and better visibility, but they also need to protect the autonomy on which teaching and research depend. The most workable principle is simple: centralize what the institution must stand behind, while leaving room for informed academic choice.

Key takeaways:

• AI governance in higher education is an organizational challenge as much as a technology challenge.
• Universities need clarity about which AI decisions belong centrally and which should remain with faculties and researchers.
• The EU AI Act does not classify every educational AI application as high-risk; intended use and institutional role both matter.
• GDPR and AI governance must be considered together, although a DPIA is not automatically required for every AI tool.
• Centralized AI access can improve visibility over software, licensing, entitlements and spending without forcing everyone onto the same model.
• Shadow AI becomes a governance problem when institutions lose sight of the tools in use, the data entering them and the people paying for access.
• AI literacy in higher education has to develop alongside access.

AI governance gets complicated when responsibility is distributed.

Universities are decentralized by design. Faculties retain autonomy over teaching, researchers choose tools that suit their work, and central teams take responsibility for security, procurement, identity and data protection. That arrangement exists for good reason.

Generative AI does not fit neatly inside it.

Consider a researcher who wants to use a new AI service. Their immediate concern may be whether the tool improves the work. The university has a longer list: what information will be processed, where it will go, which terms apply, who pays, whether another department already licenses something similar, and whether the use case calls for additional review.

Very few of those questions sit comfortably with one team. AI decisions cross institutional boundaries faster than university decision-making structures do.

A workable AI governance framework for universities therefore needs more than a catalogue of approved products. It needs clear ownership: who can decide, which evidence they need, and when a local request becomes an institutional matter.

What should universities centralize in their AI governance model? 

Centralizing AI should not mean forcing an entire university onto one model. A computer science research group testing foundation models has little in common with a lecturer preparing course material or an administrative team supporting students; standardize those use cases too aggressively and people will simply look for tools outside the institutional environment.

The more useful question is which parts of AI use the university must govern centrally.

Identity and access

Universities should know who can access institutionally provided AI services and under which conditions. Connecting those services to institutional identity makes entitlements easier to grant, change and remove as people move through the university, while a structured higher-education onboarding process avoids yet another disconnected collection of accounts for IT to manage.

Vendor information

A faculty evaluating an AI service should not have to begin from zero if another part of the institution has already reviewed the same provider. Contracts, hosting, data processing, retention, subprocessors, security and relevant safeguards can be documented centrally and reused where appropriate; the individual use case may still require assessment, but the underlying vendor research often does not. This connects naturally with broader software distribution in higher education.

Cost visibility

AI is also changing the way universities buy software. Some products use familiar per-user licenses; others charge by tokens, credits, API calls, compute or consumption. Research groups may maintain their own accounts, faculties may purchase subscriptions separately, and employees may expense premium tools.

No single cost looks alarming. Taken together, however, they can become difficult to see and even harder to manage.

Central visibility helps IT and procurement understand what the university already buys, where services overlap and which tools people actually use. A centrally managed software deployment platform can support that visibility without prescribing one AI model for every academic.

Baseline safeguards

Some rules must apply across the institution, regardless of faculty or model. A university may decide that certain information cannot be entered into particular services, for example, or that specific uses require additional review and human oversight.

Academic choice still has room to breathe. The aim is governed choice, not enforced uniformity.

Faculty autonomy is not an AI governance failure 

It is easy to see decentralized AI use and label all of it shadow IT. That misses an important distinction: researchers are expected to experiment, and faculties are expected to make decisions that reflect the needs of their disciplines. 

A team testing foundation models will work very differently from a lecturer exploring generative AI for course preparation. Central IT cannot realistically anticipate every worthwhile use of AI in higher education, nor should it try. 

Its more valuable role is to establish safeguards that remain consistent even when the technology varies. Done well, central governance makes responsible experimentation easier; it does not turn IT into the approval desk for every academic decision. 

Shadow AI: when adoption moves faster than governance

There is still a point at which experimentation becomes an institutional problem. Students, researchers and staff can start using powerful generative AI services in minutes, often with little more than an email address or payment card, and the university may never know that the service has entered its working environment.

That is shadow AI.

The risk is not the mere existence of an unapproved tool. It is the loss of visibility that follows: the institution may not know what data is being shared, which contractual terms apply, whether departments are buying overlapping subscriptions, whether access disappears when someone leaves, or whether users understand the system’s limitations.

Here, AI adoption in education and research can move ahead of governance. Policies set expectations, but they cannot reveal activity taking place beyond the systems a university can see. That gap between policy and practice is where shadow AI becomes an accountability issue.

AI policies are only the starting point

Universities have already made considerable progress on guidance for generative AI. A UNESCO survey found that 19% of responding institutions had a formal AI policy and another 42% were developing guidance. Because the respondents came from institutions connected to UNESCO Chairs and UNITWIN Networks, the figures should not be read as representative of every university worldwide.

Even so, the direction is clear: higher education is moving beyond experimentation toward more formal approaches to responsible AI in education. Making those policies work is the harder step.

A policy can warn users not to enter sensitive information into an unapproved service, yet it cannot tell IT which services people actually use. An approved-tools list may guide behavior, but it does not manage access or licensing. Responsible-AI principles can establish an institutional position; they do not tell procurement what to do when a faculty requests a new model next week.

Policy defines expectations. Governance makes those expectations workable.

That requires ownership, repeatable processes, reliable evidence and, where appropriate, technical controls. Academic Software’s work on ethics and privacy in AI provides further context for institutions shaping those controls.

How the EU AI Act affects AI governance at universities

The EU AI Act makes it even more important to understand how an AI system is used rather than treating every AI tool alike. Not every application in higher education is automatically high-risk.

Specific educational uses can fall within the high-risk framework, including certain systems used for admission or access to education, evaluating learning outcomes, determining an individual’s educational level, and monitoring or detecting prohibited behavior during tests.

A generative AI assistant used to brainstorm teaching ideas is not automatically in the same regulatory category as a system involved in evaluating student learning outcomes. Classification begins with a plain question: what is this system actually being used to do?

Institutional role matters as well. The AI Act distinguishes between providers and deployers, whose responsibilities are not identical. Providers of high-risk systems have obligations covering areas such as risk management, technical documentation, logging capabilities and design for human oversight; deployers have responsibilities of their own, including following instructions, assigning appropriate human oversight, monitoring operation and retaining logs under their control where required.

Following the 2026 amendment to the AI Act, obligations for the Annex III high-risk systems relevant to education are scheduled to apply from 2 December 2027.

For university CIOs, the priority is not to classify every AI procurement as high-risk. It is to build a consistent route from use-case classification to proportionate governance.

GDPR and AI governance in higher education

Asking whether an AI tool is “GDPR compliant” rarely gets a university far. The answer depends on the service, the data involved, the purpose of the processing, the chosen configuration, the contractual arrangements and the way the institution intends to use it.

A lecturer using generative AI to brainstorm public course material presents a very different data-protection scenario from a system processing identifiable student records. Universities need enough information to understand that difference.

Relevant questions include what personal data is processed and why, where it is handled, how long it is retained, which subprocessors are involved, whether information may be used for model training or service improvement, and whether international transfers take place. These issues sit at the intersection of AI privacy and deployment decisions, procurement, legal review and data protection.

When faculties adopt tools independently, gathering reliable answers becomes much harder.

A DPIA is not automatically required for every AI tool

Universities should avoid turning AI governance into a blanket compliance exercise. Under GDPR, a Data Protection Impact Assessment is required where processing is likely to result in a high risk to the rights and freedoms of individuals; the involvement of AI may be relevant, but it does not mean every application needs its own DPIA.

A low-risk administrative use should not necessarily travel through the same review process as a system that handles identifiable student information or contributes to consequential decisions. A consistent screening process can send higher-risk cases toward deeper assessment without creating unnecessary friction everywhere else.

Good governance applies scrutiny where it matters.

International research makes vendor transparency essential

Research-intensive universities face an additional complication: international collaboration is part of everyday academic work. Researchers rely on institutions, datasets, platforms and technology providers across borders, so an AI service that transfers personal data outside the EU requires the university to understand the applicable transfer mechanism and safeguards.

Basic vendor facts suddenly carry considerable weight. Where is the service hosted? Which subprocessors are involved? How long is information retained? Can submitted data be used for model training? What security and contractual commitments has the provider made?

Without reliable answers, the institution is being asked to approve a service with only part of the picture. AI governance for universities is therefore becoming inseparable from software procurement and vendor management.

AI governance is also a budget problem

CIOs have a more ordinary concern alongside privacy and regulation: who is paying for all of this?

AI does not always fit traditional software licensing. One service charges per user, another uses credits, a research team consumes an API, and a faculty buys subscriptions of its own. Employees may expense premium accounts while students pay for AI services personally.

Each choice may be reasonable. Collectively, they can create a fragmented cost base that no one sees in full.

The useful questions are practical: what is the university already paying for, which products overlap, which licenses are actively used, where demand is growing, and which costs should sit centrally rather than within a faculty? Better visibility over AI tools for universities can answer those questions without dictating which model every academic should use.

Equal access is becoming part of AI governance

When individuals are left to arrange their own AI access, purchasing power starts to influence capability. Students who can afford premium subscriptions may receive better models, higher usage limits or functions that free services do not offer.

As generative AI in higher education becomes part of everyday study and research, institutions will have to decide what baseline of capability belongs in the digital environment they provide. That does not imply giving every student every model, and different disciplines will continue to need different tools.

It does, however, make equal access an AI governance question as much as a licensing decision. A centrally managed baseline can coexist with specialist services for faculties and research teams, while the wider environment still supports universities, students and staff with appropriate digital access.

AI literacy belongs inside the governance model

Access is only part of the job. People also need to understand what they are using.

AI literacy varies widely across a university. Some researchers know model limitations, hallucinations, evaluation methods and data risks in detail; other users may have little experience beyond asking a chatbot a question and accepting the answer.

The EU AI Act’s AI literacy provisions have applied since 2 February 2025, requiring providers and deployers to take measures that support an appropriate level of AI literacy among relevant staff and other people dealing with AI systems on their behalf. The European Commission also frames AI literacy and digital skills as a broader policy priority.

For universities, AI literacy frameworks should develop alongside access rather than appear as a training campaign after deployment. Users need practical guidance on appropriate use, sensitive information, output verification, bias, hallucination, institutional policy and the situations in which human judgment remains essential.

A well-governed environment still depends on informed people inside it. Practical examples of AI use in education can help connect those principles with daily work.

Infrastructure should follow the governance model

Centralizing AI can sound like a major infrastructure project, although the technical requirements vary considerably. A university that relies mainly on commercial SaaS products faces different choices from one running self-hosted models or providing large-scale research compute; many institutions will end up with a mixture of commercial services, APIs, managed gateways, specialist research tools and locally hosted models.

The common requirement is not a particular architecture. It is the ability to support institutional governance through identity, entitlements, software distribution, usage visibility, integrations, security controls and trustworthy vendor information.

Decide what must be governed first. Then design the architecture around it.

A practical AI governance framework for universities

Universities do not need a final answer to every AI question before making progress. They do need a repeatable way to decide.

Six foundations offer a practical starting point:

1. Define decision rights

   Decide who can approve an AI vendor, assess data protection, own the budget, judge whether a teaching or research use case is acceptable, and authorize exceptions. Those responsibilities do not all belong with central IT, but every important decision needs a named owner.
   
   

2.  

   Classify AI use cases by risk

   An assistant used for administrative brainstorming should not automatically follow the same process as an AI system involved in student assessment. Different risk levels need different review paths, allowing the institution to move quickly where consequences are limited and apply deeper scrutiny where the stakes are higher.
   
   

3. Centralize vendor evidence

   Maintain reusable information about contracts, privacy, security, hosting, subprocessors, retention and AI-specific terms. Faculties can then focus on whether a tool suits their use case instead of repeatedly investigating the same supplier.
   
   

4. Connect AI access to institutional identity

   Where appropriate, manage AI entitlements through existing identity, onboarding and offboarding processes. Access becomes easier to grant, change and remove, while the university gains a clearer view of who can use institutionally provided services.
   
   

5. Make AI costs visible

   Bring subscriptions and relevant usage information into a common view. Procurement and IT can then identify duplication, understand demand and make better licensing decisions without interfering unnecessarily with academic choice.
   
   

6. Build AI literacy alongside access

   Whenever a new AI service becomes available, users should also understand the conditions under which it can be used. Access, governance and literacy are far more effective when they develop together.

Centralize accountability, not every AI decision

Universities depend on experimentation. Researchers need to test emerging models, faculties need space to explore teaching methods, and students need opportunities to develop practical AI skills. Trying to control every AI decision centrally would work against that mission.

Complete decentralization creates the opposite problem: visibility over access, contracts, costs, data practices and risk begins to disappear.

A workable model sits between those extremes. Universities can centralize the areas for which the institution ultimately carries accountability (identity, entitlements, procurement evidence, compliance information, cost visibility and baseline safeguards) while preserving flexibility around model choice, research experimentation, disciplinary requirements and teaching practice.

That is a more useful definition of AI governance in higher education than placing every tool behind one platform.

Where Academic Software fits

Academic Software already helps higher-education institutions manage a similar challenge across their wider software environments: giving students and staff access to the tools they need while supporting centralized licensing, distribution and onboarding. AI extends that challenge.

Through centralized software access and solutions such as Mila, universities can create a more manageable access layer for AI and other digital tools while retaining flexibility for different users and faculties.

Technology does not replace an institutional AI governance framework. It should not be presented as if it does.

Its value is practical: helping universities translate governance decisions into the way software and AI services are actually made available. Depending on the institution’s setup, that can include access, entitlements, licensing, onboarding, usage visibility, cost management and access to multiple AI models.

For CIOs and procurement teams, the benefit is not simply another AI platform. It is a clearer connection between institutional decisions and the digital tools people actually use, supported by the wider range of Academic Software services where relevant.

Conclusion: AI governance starts with accountability

AI is already part of university life. Students use it, researchers experiment with it, and staff are finding places where it removes repetitive work.

AI adoption in higher education is moving faster than the governance around it.

Closing that gap will take more than a policy or a decision to standardize on one model. Universities need clear decision rights, reliable vendor information, proportionate safeguards, visibility over access and costs, and users who understand how to work with AI responsibly. Just as importantly, they need to preserve the autonomy that makes universities productive places for research and experimentation.

The principle remains straightforward:

Centralize assurance, not necessarily choice.

Institutions that find the right balance will not be those that exert the most control over AI. They will be the ones that know where control is necessary, where flexibility is valuable, and who is accountable for the difference.