WEEKEND READING: AI strategy has the wrong address

This blog was kindly authored by Michal Bobula, Senior Lecturer in Accounting and Finance at University of the West of England, Bristol.

Most universities now have an artificial intelligence (AI) strategy or are writing one. Almost all of it speaks at the institutional level: principles, goals, generic guidance on AI literacy. A recent HEPI review found no shared approach across the sector: two in five institutions have nothing publicly findable, and much of what exists polices misuse rather than guides teaching. None of it reaches the place where teaching actually changes. That is not an argument against policy; it is a gap. And meaningful adoption may need to start where the gap sits: in the module.

The case here is deliberately narrow. Every module leader finds the single best use of generative AI in their module, the one they could genuinely defend in their discipline and lets it run for a year, with students and colleagues, before refining it for the next.

With students, not on them

Letting it run means exactly that. The use case is introduced to the cohort, discussed with them, and shaped by their feedback as it goes; colleagues get a say too. Students practise judging AI against the standards of a given discipline. Staff experiment and find out what the technology is actually good for in their subject. Sometimes the answer is that AI has no useful place in the module at all, and that is important learning too, because it was discovered through practice rather than assumed from a distance. This is what makes the module the right place to take the risk: the stakes stay small and visible, stakeholders are directly involved, and every outcome feeds next year’s version, and, in time, how the module assesses.

Answering ‘what is the one good use in my module?’ does not demand expertise upfront; instead, it should be a process. Start small, think it over, try something, refine it. Working through the question builds the module leader’s and their team’s own understanding of AI while drawing on their expertise in the field. Staff learn AI the same way we want students to learn it: through the discipline, one real use at a time. It does mean watching our own biases, ironically the very failings we accuse AI of: a no driven by distaste is no more considered than a yes driven by hype. That honest judgement is exactly what we want graduates to leave with, and a considered no counts as an answer.

Shape, don’t adopt

There is a bigger prize here. So far the sector has mostly been adopting: taking the tools as they arrive and looking to industry for cues on how to use them. But there is nothing settled out there to copy. Industry has taken up generative AI at speed, yet Gallup finds that only around one employee in ten at organisations using AI says the technology has changed how work actually gets done. Everyone is using it; few are using it well.

Universities can do better than adopt: they can shape what good use means. A module has a cohort that tests an idea every year, and academics are trained to examine claims rather than accept them. That is exactly the machinery needed to work out what good, honest, properly considered AI use looks like in each field. Do that, and universities make an impact on how AI is used rather than just keeping pace with it. Graduates arrive at work able to judge AI use in their subject and to shape it, and they take that into the organisations they join. And underneath it all sits agency. Academics already decide what good work looks like in their subject. Deciding what good AI use looks like is the same job, and students learn it from the people who do it.

The scaffolding pays off

For students, this adds up, and it does not need every module to say yes. Even if only a third or half of module leaders find a use they can defend, a student still graduates having worked through a series of considered uses of AI, each embedded in a discipline’s way of thinking, asking the same questions: where does the tool genuinely enhance the work, what are its downsides, and what stays human, and why? And each is there because someone decided it belonged, not because a policy demanded one. Students want this. In In this year’s HEPI Kortext student survey, over two-thirds of undergraduates said AI skills are essential to their futures, yet fewer than half felt teaching staff were helping them build those skills.

Tis compounding does not happen by accident though, and no single mechanism solves coordination. But the minimum is visibility: someone keeping a simple map of who is doing what, so that uses can build on one another rather than repeat. Without even that, six modules produce six disconnected encounters rather than a curriculum.

Get the sequencing right and the result is AI competence that no bolt-on literacy module can produce. Generic literacy is a unit students forget: some prompting tips, a slide on hallucination, a quiz on the policy, all gone by next month. Literacy built through the discipline is a skill they keep. An accountancy graduate who has checked AI-extracted figures against source documents forty times knows output must be verified. A law graduate who has tested AI case summaries against the judgments knows exactly where the tool cuts corners. Neither needs a certificate saying they can use AI. Their degree is the evidence.

The support that makes it real

None of this is free, and pretending otherwise is how initiatives die. One model could be an AI lead in each department: someone whose job is to save colleagues time by knowing which tools are licensed and where the traps are, and who keeps the map. What cannot vary are two conditions:

•  The work must be properly incentivised, as a named part of the workload model rather than goodwill.
• Experimentation needs guidance and support in place from the start, with checks that match the size of the experiment rather than the full machinery designed for enterprise systems.

The Monday-morning ask

Heads of department, put the question to every module leader: what is the one good use in yours? Then protect the time to answer it. Some will conclude, honestly, that the answer is none, and that is a finding too. Get the momentum going and delegate: the lead keeps the map, so uses build across levels rather than collide. And university leaders: resource the time, and keep it going year after year, not as a one-off.

And in a few years, the measure of success is that nobody is talking about it. AI will simply be part of how each discipline teaches, because the work happened where it always belonged. One module. One good use. One academic who can defend it. Start there.