An eight-hundred-year-old institution might not seem a natural place to integrate AI, a technology advancing at the pace of weeks and months. Cambridge’s operations are shaped by commitments accumulated over centuries, delivered across a network of hundreds of departments and colleges. They combine statutes developed over decades, funding rules from hundreds of funders, and estates that house cutting-edge research machinery in historic buildings. Six months of experimentation through ai@cam’s AI for Operations programme shows how AI can be integrated into this institution. It also illustrates how AI adoption does not follow the path of an IT rollout. Getting the most out of these technologies requires a different approach.
The eight projects funded through ai@cam’s AI for Ops programme – on energy use in laboratories, research grants administration, enquiry responses, library cataloguing, exam standing, regulatory compliance, committee management, and analysis of funder terms – are showing a path to integrating AI into University operations.
How AI shapes our relationship with complexity
A building consuming three quarters of a million pounds of electricity a year contained, in principle, the data needed to understand how to be more energy efficient. In practice, harvesting it from the building management system would mean at least 26 hours of manually clicking through five hundred-plus meters. AI-assisted scripts brought that to twelve minutes, presenting information about where energy is being used and where there may be opportunities for efficiency. A process of checking student exam paper choices against Statutes and Ordinances, which previously took a team of six the best part of two months, now runs in under a minute. Committee minute-writing across thirty meetings of twenty-three committees took, on average, about half as long with AI assistance as it had without, with no measurable drop in quality.
Exam regulations are still detailed and changeable. Funder terms still run to thousands of pages. Buildings still draw their electricity from hundreds of meters. What changes is the surface area of that complexity; how much friction it creates, how navigable it becomes, whether the questions it raises can be asked and answered in time to do something about them.
One of the projects describes its work as not having fixed the plumbing, but having reached a point where the leaks could be dealt with in seconds rather than months. The institution remains as complex as it was before. The practical experience of working with that complexity is beginning to change.
Most AI projects are not really “AI” projects
A recurring observation across the programme was that the technical work of using AI was rarely the binding constraint. The hard parts sat in finding and curating data that nominally existed but could not be reached, in designing workflows where human authority had to be preserved through any new tool, and in deserving the trust of staff whose work was being changed.
Funder terms exist as web pages with expandable boxes and embedded links, with the same requirement referred to differently across documents. Recruitment enquiries had to be classified across nine hiring rounds before any drafting tool could be designed around them. Statutes had to be parsed sentence by sentence into a form a rules engine could act on. The cliché that AI projects are data projects in disguise is, for this kind of operational work, true.
The same is true of workflows. Knowing that AI can produce a serviceable first draft of committee minutes is one finding; understanding how a chair, a clerk, and a secretary should interact with that draft to preserve the authority of the final record is another. Building a tool that flags issues with examination coordination is one thing; producing a trail that lets a non-specialist see where the issue arose, and update the rules in time for next year, is what makes it usable.
In the process, those innovating with AI need to be mindful of the individuals and teams whose work is likely to be affected. The committee management project devoted as much energy to developing a set of principles for use – on data protection, human oversight, transparency, and ensuring AI does not generate work that others then have to read – as it did to testing tools. Those principles are likely to outlast any one tool.
These principles illustrate the vigilance this work requires. When a professional delivers their duties, their reputation is in some sense at stake in the result, and that human stake shapes the quality of attention applied. When an AI does the same piece of work, no such stake exists. The outputs can have a kind of artificial fluency that masks the absence of real care underneath. Several projects observed that AI outputs tend to look more reliable than they are. Catching these hidden failures depends on understanding what is happening under the surface of a tool, and on remaining the person accountable for its outputs.
The technology is the visible part of this work. The data, the workflows, the governance, the trust, and the organisational design around it are where most of the effort lies.
Distributed experimentation with institutional scaffolding
These projects, taken together, make a case against treating AI in operations as a central transformation programme; AI adoption doesn’t look like a traditional IT roll out. Successful projects are being led by teams close enough to a workflow to see where complexity accumulates, where judgement matters, and where automation could help.
A new department designing a recruitment drafting tool spent significant time on its applicant query journey before designing an AI question-answering system. A library team thinking about catalogue cards had to distinguish which kinds of cards were tractable for handwriting recognition, which required specialist curatorial judgement, and how to involve staff whose careers had been built on careful manual work. A committee management team had to think carefully about which kinds of meetings could use AI assistance and which could not.
Local operational knowledge is the condition for useful AI adoption. The kind of central function that helps is the kind that creates conditions for distributed work to happen well – shared principles, secure tooling, training designed for the needs of professional services staff, and communities of practice in which teams can compare notes across departments. The programme has begun to build some of that scaffolding; more is needed. Among the gaps the projects have identified are inconsistencies in which teams can access which tools, the absence of clear institutional guidance on how to handle particular kinds of data, and a training landscape that has so far been designed mostly for researchers rather than for the professional services staff who carry much of this work.
Iterative innovation as technology moves
Traditional technology rollouts value clear scope, defined outputs, and indicators against which delivery can be measured. The teams in this programme reflected, in different ways, that this kind of planning can narrow innovation when the technology is changing every few months.
The right kind of planning here is iterative; setting a direction rather than a fixed output, with checkpoints to reflect on what is and is not working and the willingness to adapt accordingly. Permission to experiment, to learn, to spend time on practices that are not yet measurable as outputs.
Consistent across these innovations is that teams use AI as a tool to support the careful, accountable thinking their roles already require. That stance is what makes the efficiencies trustworthy, and what protects against the kinds of hidden failure these tools are prone to.
Looking ahead
The projects in this programme have delivered real efficiencies, anchored in the careful judgement of the teams who built them. They have also shown the conditions under which such gains are possible, and what is needed to make them durable.
The question is how to scale this kind of work. What these projects point to is a community-led model in which staff who have learned how to use these tools well reinvest some of the time AI saves them into helping colleagues do the same – sharing prompts, running training, surfacing what has and has not worked. The role of the centre is to make that possible: by providing the scaffolding of secure tools, shared principles, and dedicated time.
An eight-hundred-year-old institution is not going to be transformed by a single project. It can, in many places at once, be helped to work better, if the people closest to that work are given the tools, the time, and the permission to keep learning together.
Jessica Montgomery, Director, ai@cam