Universities are moving quickly to understand where AI can deliver meaningful value.
From research and teaching to student support and professional services, experimentation is happening across institutions, often faster than central technology teams can keep track of it.
But as AI adoption accelerates, a more fundamental question is emerging: are the systems, data and operating models underneath it ready?
During NashTech’s November higher education webinar, NashTech’s Senior Technology Consultant, Chris Weston, University of Birmingham Interim CIO, Kuldip Sandhu, Henley Business School and The NextPath Device Consortium, Research Associate and Co-Director, Freddie Quek and PDC’s Director of Technology Transformation, Sharon Penfold, alongside NashTech’s Higher Education Account Director, host Laura Bailey, explored what needs to change if universities are to move from isolated AI experimentation towards meaningful transformation.
The discussion covered integration, legacy technology, data, cost, governance and responsible AI, and one theme connected them all: AI cannot compensate for weak foundations.
As Kuldip Sandhu put it:
“AI is magnifying our data problems, but it’s also exacerbating our cyber posture.”
NashTech’s wider research paints a similar picture. Among education technology leaders, 42% cite integration with existing or legacy systems as a challenge when implementing custom software, while 41% struggle to adapt solutions as requirements evolve and 37% cite limited internal expertise to maintain or scale them.
The opportunity presented by AI is significant. But before universities ask how quickly they can deploy it, they may need to ask a more uncomfortable question: what exactly are they plugging it into?
University technology estates are rarely simple.
Student records, learning platforms, research systems, finance, HR, admissions and countless specialist applications have often developed independently over many years. Integrations have been added over time, workarounds have become business processes and valuable institutional knowledge can sit with a surprisingly small number of people.
For Kuldip Sandhu, solving that problem requires universities to stop thinking of integration purely as a technical exercise.
“Integration isn’t just about connecting applications. It’s about understanding the data and the information. Does it have clear meaning? Can it be trusted? Is somebody accountable for it, and how does it move from system to system?”
That distinction becomes increasingly important with AI.
Connecting an AI model to fragmented, unreliable or poorly governed information does not resolve the underlying problem; it potentially allows that problem to operate faster and at a greater scale.
Sharon Penfold shared one particularly powerful example. At one institution, a transgender student’s name repeatedly reverted to their birth name as their information moved between systems.
“It kept resetting to their birth name. There was no mapping of integrations, no documentation, and there were no regimes or controls around integration.”
It is an example that turns what can sound like an abstract integration problem into a very human one.
There is an operational impact too. Every failure pulls teams into troubleshooting, testing and tracing information across systems.
Penfold explained:
“You’re taking away a huge amount of time from people who should be doing their operational roles. That’s on top of their day jobs.”
Legacy technology can create similar risks.
Kuldip Sandhu described an institution with a 25-year-old internally developed student records platform containing critical student data, but very limited resources remaining to support it.
His assessment was simple:
“Some people call it a single point of success. I’ll call it a single point of failure.”
Moving from individual AI experiments to institution-wide adoption is particularly challenging in a university, where different groups are using the technology in different ways and moving at different speeds.
Researchers, students, academics, professional services teams and central technology functions all approach the technology from different perspectives.
Freddie Quek summed up the challenge:
“As soon as you use the word AI, which lens are you looking through? Everybody has different answers.”
A researcher may already be experimenting with emerging models at the edge of their discipline. A student may routinely use public generative AI tools. Professional services may see an opportunity to automate repetitive processes, while the technology team is considering security, architecture and data privacy.
All of those conversations may be happening simultaneously. The challenge is connecting them.
“They move at different paces,” Freddie Quek explained. “If we want to find that golden thread across all those groups, it is difficult.”
This is where AI experimentation can easily become fragmented.
Individual pilots can prove that a technology works without answering whether it can operate securely, integrate with existing systems or deliver measurable value at institutional scale.
There is also a danger at the other end of the spectrum.
Sharon Penfold described universities facing the possibility of ‘AI paralysis’, in which concerns about security, budgets, governance, and organisational change prevent institutions from progressing at all.
Institutions need environments where people can test AI safely, understand what works and turn successful experimentation into repeatable approaches.
Kuldip Sandhu described a model built around a secure ‘walled garden’ that gives different university communities access to large language models within an institution-controlled environment.
It enables experimentation, but with greater visibility over data, access and usage than organisations may have when employees or students independently use public AI services.
The technology, however, is only half the challenge. Even when approved tools are available, people need the confidence and knowledge to use them.
“They are fearful of using it. They don’t want to use it. They do not know how to use it,” Freddie Quek said.
Successful AI transformation therefore requires more than deploying technology. Universities need AI literacy, practical guidance and leadership capable of understanding both where AI can create value and where it should not be used.
| Understand more about what universities need to solve as they move from AI pilots towards institution-wide adoption in From AI pilots to campus-wide impact. |
As AI experimentation expands, governance inevitably becomes part of the conversation. But governance does not have to mean slowing everything down.
Sharon Penfold argued that universities need to reconsider the role risk plays in transformation.
“One thing is trying to look at risk management as a benefit, as delivering return on investment, instead of just getting in the way.”
That becomes particularly important because different AI use cases create very different levels of risk.
An AI assistant helping someone find basic information may require a very different level of assurance from a model processing sensitive student information or contributing to decisions affecting an individual.
As Kuldip Sandhu explained:
“Each use case should have an outcome, and it should have a risk profile depending on what type of AI you’re using.”
That moves governance away from a blanket question of whether AI is allowed and towards a much more useful conversation about how an individual use case should be governed.
Who owns it? What information is being used? What outcome is expected? What would happen if it produced the wrong answer? Who remains accountable?
Done well, governance can actually accelerate adoption because teams no longer need to answer those questions from scratch every time they want to test something new.
The principle extends to the role of people, too.
Freddie Quek argued that the industry needs to move beyond simply talking about keeping a ‘human in the loop’.
“It’s about human, not in the loop, but human in the lead.”
That means technology can carry out more of the work, but people remain responsible for defining the outcome, challenging the output and making decisions where human judgement matters.
For a wider look at the governance and execution challenges organisations face as they scale AI, explore NashTech’s ‘Closing the gap between AI ambition and execution’.
The economics of AI present universities with another emerging challenge.
Generative AI does not necessarily behave like familiar enterprise software with a predictable annual licence cost. Usage can vary substantially depending on the model, task, volume of information and number of users.
Chris Weston offered a memorable description of the problem:
“We had cloud shock once upon a time. Now we’re getting token shock.”
It is particularly pertinent for universities, where potentially thousands of students, academics, researchers and professional services staff could eventually be consuming AI services.
Kuldip Sandhu described an approach in which different groups receive defined allowances, while additional usage can be approved where there is a legitimate need.
This gives institutions visibility into where AI is being consumed and creates an opportunity to understand where additional spending is justified by the outcome.
As adoption increases, the conversation therefore needs to evolve from “How much AI are we using?” to “Where is AI actually creating enough value to justify its cost?”
That distinction will become increasingly important as experimentation moves into operational services and AI expenditure becomes a meaningful part of university technology budgets.
Perhaps the most important takeaway from the conversation was also the simplest. Universities should resist the temptation to start transformation with the technology.
Kuldip Sandhu described seeing organisations invest significantly in sophisticated data infrastructure before deciding exactly what they wanted that infrastructure to achieve.
“Before you give a lot of capital and operational expenditure to a technology function to implement the infrastructure to support the data world, make sure you as an institution know what you want from it.”
The same principle applies to AI. If the goal is to improve student support, start there.
If it is reducing administrative pressure, improving retention, accelerating research or strengthening financial sustainability, make that the starting point.
Then determine what data, integration, governance and AI capabilities are required to deliver it.
That reverses a pattern technology leaders will recognise all too well: implementing a platform first, then searching for sufficient use cases to justify it.
Kuldip Sandhu captured the alternative in five words:
“Start top down, not bottom up.”
Universities do not need perfect technology estates before they can make progress with AI. Nor should legacy complexity become another reason to wait.
But they do need to understand the foundations on which AI will depend.
Because ultimately, the institutions that gain the most from AI may not be those experimenting with the greatest number of models or launching the largest number of pilots.
They will be the ones that can connect AI to trusted data, resilient systems, proportionate governance and, most importantly, a clearly defined institutional outcome.
Did you miss the live event or want to revisit the discussion? Access the webinar on demand here.
If your institution is considering how to modernise its technology estate while creating the foundations for AI, explore how NashTech supports higher education organisations across data, integration, legacy modernisation and digital transformation.