Universities are moving rapidly from debating artificial intelligence to buying it. That shift is understandable. Students already use generative AI extensively, employers increasingly expect graduates to understand it, and leaving access to the best systems dependent on who can afford which commercial subscriptions creates obvious problems of imbalance.

In the United Kingdom, Universities UK has now called for AI tools to be accessible to every undergraduate student. In Australia, UNSW Sydney has announced one of the largest higher education deployments in the Asia-Pacific region, providing ChatGPT Edu to more than 80,000 students and staff.

These initiatives point toward a future in which institutional access to AI becomes as normal as access to a learning management system, academic databases or university email.

Universities do need AI infrastructure. But they do not need AI lock-in.

That distinction is becoming urgent because universities risk treating AI procurement as though they were buying another piece of software. Generative AI is developing into something considerably more consequential: a computational layer through which students learn, academics research and professional staff increasingly perform institutional work.

A procurement decision can gradually become an institutional architecture.

Education’s emerging platform problem

UNESCO put this problem unusually clearly on the international policy agenda in September. At its Digital Learning Week in Paris, education ministers called for AI systems used in education to be auditable, for data to remain portable and for education systems to consider the total costs of ownership before adopting AI. The resulting ministerial statement explicitly favors interoperability, portability, and open-source systems so that education systems retain the ability to change course.

The important principle can be expressed more simply: institutions need to be able to leave.

This is not an argument against commercial AI providers. Universities need access to capable systems, and institutionally managed versions of commercial products can provide significant advantages over consumer accounts in security, privacy, administration and equitable access.

The problem comes when access becomes dependence. Consider what happens over several years after a university selects a preferred AI platform. Staff develop workflows around it. Students learn through it. Bespoke assistants are created inside it. Institutional knowledge accumulates within it. Other university systems are connected to it. Eventually, AI agents may begin initiating actions across those systems.

At that point, changing providers is no longer comparable to replacing a software license. The institution must unwind an ecosystem. This is a familiar feature of platformization: the value of a platform increases as more activity takes place within it, while the cost of leaving rises at the same time.

Higher education has encountered versions of this problem before with learning management systems, cloud infrastructure, academic publishing and research analytics. AI potentially intensifies it because the technology is not simply storing information or providing access to content. Increasingly, it is mediating the work itself.

The relevant procurement question is therefore changing. It is no longer only, “which AI system offers the best capabilities today?” Universities also need to ask, “what architecture preserves our ability to choose differently tomorrow?”

The university should control the layer above the model

UNSW offers an interesting glimpse of one possible answer. Alongside its large OpenAI deployment, the university says it is developing a multi-model AI environment in which students and staff can experiment with different AI systems as the technology evolves. That may ultimately prove more strategically important than the ChatGPT agreement itself.

The objective for universities should be to separate their institutional AI environment from any individual model operating within it.

An institutionally governed layer could establish common rules for identity, data access, security, privacy, and appropriate use while allowing different AI models to operate underneath. A university might use one model for particular research tasks, another for coding, and another for routine administrative work. Those choices could change as technology, cost, and evidence change.

Today’s leading model may not be tomorrow’s. Universities should be positioned to benefit from competition between providers rather than finding themselves locked into the winner of a procurement exercise conducted several years earlier.

This principle matters beyond higher education. Public institutions are under increasing pressure to adopt AI while a small number of companies control many of the most capable models and the cloud infrastructure on which they depend. Will Stronge, chief executive at the Autonomy Institute, has recently pushed the argument further, describing the AI systems through which students increasingly think and research as a form of “cognitive infrastructure” and arguing that questions of ownership and control therefore belong at the heart of educational policy.

Universities do not need to wait for publicly owned foundation models to act on that insight. The immediate policy challenge is to create conditions under which institutions can adopt commercial AI without surrendering technological agency. Interoperability and portability consequently become governance issues rather than technical preferences.

When platforms begin to exercise institutional authority

There is a second reason universities should care about the architecture they are constructing: AI systems increasingly do more than generate text.

Universities are experimenting with AI in student support, research, administration, assessment, recruitment, and the identification of students considered at risk. The development of AI agents will push this further because systems will increasingly be capable not merely of recommending actions but of initiating workflows.

Imagine a university uses an AI-enabled system to identify students at risk of dropping out. Formally, the algorithm may make only a recommendation. A human member of staff retains responsibility for deciding whether intervention is required. But suppose staff accept the recommendation almost every time. Where does the decision actually sit?

The university can truthfully maintain that a human remains “in the loop” while the technical system increasingly determines which students attract institutional attention. Authority can migrate without any formal decision to transfer it.

The same process occurs through more mundane technical choices. A university regulation might give academics discretion over appropriate AI use in assessment, while the rule experienced by students is implemented through settings inside a learning management system: AI permitted here, prohibited there, one application accessible and another blocked.

Institutional policy is translated into permissions, defaults, interfaces, and automated workflows.

Policy becomes code

This is code-mediated governance: institutional authority exercised partly through technical systems rather than exclusively through conventional rules and human decisions. That makes the question of who controls university AI infrastructure inseparable from the question of who governs the university. AI governance needs to move beyond AI principles. Many universities responded to generative AI by producing guidance. That was sensible when the principal institutional problem appeared to be how students and staff should use ChatGPT. It is increasingly inadequate when AI becomes infrastructure. Universities now need governance capable of answering a different set of questions. Which AI systems operate within the institution? What information can they access? Which other systems can they interact with? Which decisions do they inform? Who authorized those uses? Can their outputs be audited? When must a human intervene? What happens when a model changes? And what is the exit plan if a supplier no longer serves the institution’s interests?

These are not questions for an IT department alone. They involve procurement, academic governance, information security, data protection, research ethics, employment relations, student representation, and senior institutional leadership.

Nor can they be answered simply by creating another AI ethics committee. The governance has to reach the technical architecture itself. Procurement requirements can mandate portability. Data architecture can restrict what models can access. Institutional AI gateways can provide multiple models through common controls. Logging can make automated activity auditable. Governance processes can establish where human judgment must remain decisive.

In other words, principles have to become infrastructure.

Universities should also be cautious about allowing the urgency of AI adoption to reverse the appropriate sequence of governance. The pattern too often is: procure a system, distribute access, encourage experimentation, then work out how to govern what follows. The more sustainable sequence is governance architecture first, scalable adoption second.

The question is not which model wins

None of this requires universities to become AI skeptics. The opposite may be true. Institutions with strong governance should be able to adopt AI more confidently because they know where responsibility lies and have reduced the consequences of any individual technology failing.

Universal access to capable AI could become an important part of higher education. Students should graduate understanding how to use these systems critically and professionally. Researchers should be able to exploit rapidly improving tools. Universities should seek productivity improvements in administration rather than preserving inefficient processes simply because they are familiar.

But technological ambition and institutional autonomy need not be opposites. The most important AI decision university leaders make over the next few years may therefore not be whether to choose ChatGPT, Claude, Gemini, Copilot or whatever comes next; it will be whether they build an institutional architecture in which those remain choices.

The real test of a university’s AI strategy is not which model it buys in 2026. It is whether it will still be free to choose in 2030.