Creative Use of AI for Research Administration

Summary

Research centres run on administration: meetings, minutes, follow-up, agreements, reports, events, communication and the website. This project explores how AI can be used creatively in that work, with the ambition of letting AI agents take on as much of it as is responsible, so that people spend their time on what needs human judgement. MishMash uses its own administration as the test case.

The core challenge is not any particular agent, but the context agents work in: shared rules, routines and knowledge about how the centre works, written so that different agents and different ways of working can use them, and kept current as the centre changes.

Critical reflection

Using AI in administration is not only a question of efficiency. The project reflects critically on what it does, as it goes:

  • Privacy: which information may be given to which systems, and how personal data about members, partners and staff is protected.
  • Ethics and accountability: people decide, sign and send; it must always be clear who is responsible for what the centre does, and what may not be delegated to a machine at all.
  • Agents and systems: the project compares different AI agents and model services, from the university’s own to commercial ones, on the same tasks and with the same shared context, to learn what each does well, where each fails, and how to avoid depending on any one of them.
  • Models: which models are used, where they run, who controls them, and what that means for openness, dependency and sustainability.
  • Quality and trust: how mistakes are found, how much checking the work needs, and whether it actually saves time.
  • Work and roles: what changes for the people doing the administration, and which skills they need.

Questions

  • What does an agent need to know to do useful administrative work, and how is that knowledge kept up to date?
  • Can the same shared context serve different agents, models and people?
  • Which tasks benefit from AI, which do not, and which should be left to people?
  • What do privacy, ethics and accountability require when AI drafts what people later decide, sign and send?

How it works

The people in the MishMash administration try out agents in their daily work, write down what works and what does not, and revise the shared rules and routines after each round. What is learnt is shared with the centre, and with other research administrations that face the same questions.

Work Package

Cross-work-package initiative, closest to WP7

Duration

2026–ongoing

People

Institutions