AI automation that survives an HE audit committee

Audit and risk committees in UK universities exist to protect the institution from decisions it cannot reverse, explain or evidence. An AI automation proposal that reaches the room as a pitch — features, efficiency percentages, vendor optimism — invites the only answer a careful committee can give: not yet.

The proposals that pass arrive differently. They treat the committee's scrutiny as a specification: every question internal audit is likely to raise has a designed answer, and the evidence for it exists before the meeting starts.

Human-in-the-loop on consequential actions. The simplest way to satisfy "who approved this" is to make consequential actions — anything affecting a student's record, fee status or progression — require explicit human approval. Automation does the gathering, checking and drafting; a person signs. Committees approve that division of labour because they can see themselves in it.

Data residency inside the institution's tenant. Running models and queues inside the university's own Azure tenant — with no student data leaving for third-party APIs — converts the hardest residency question into a short answer. Our localised LLM deployments exist precisely for this constraint.

Logging as a first-class deliverable. Every automated proposal, approval and execution written to an append-only store, exportable on demand. When the auditor asks for a sample, you hand them a query, not a promise.

Arrive with the artefacts, not just the narrative: the DPIA (completed, not drafted); the one-page data flow diagram; the risk register with a named owner per risk; the human-approval matrix showing which actions can never be fully automated; and a rollback plan that proves the service can revert to manual without data loss.

For anything touching special category data — disability records, extenuating circumstances, disciplinary cases — UK GDPR Article 9 conditions and the OfS's expectations on ongoing conditions of registration belong in the pack from day one, not as a supplementary paper. Our note on AI governance for UK higher education data covers the detail.

The person carrying the proposal into the room is rarely the person who built it. If the champion cannot explain the architecture in their own words under pressure, the proposal fails regardless of quality. Part of any serious engagement is preparing the internal case: briefing the champion, pre-answering procurement and risk questions, and staying available while the decision works through governance. We make that explicit in how we work — consensus and commitment is a named stage, not an afterthought.