Public-sector AI deployment is outrunning assurance, transparency and evaluation capacity

Humphrey tools are rolling out across Whitehall and 2.4m NHS chest X-rays a year are AI-assisted, but the Public Accounts Committee (March 2025) found 28% of government IT systems are end-of-life legacy, ~50% of civil-service digital roles advertised in 2024 went unfilled, 21 of the 72 highest-risk legacy systems lack remediation funding, and only 33 Algorithmic Transparency Recording Standard records had been published despite the standard being mandatory for central government. No independent body audits whether deployed tools actually work; the NAO flagged fragmented accountability between DSIT and Cabinet Office.

Government is simultaneously the UK's biggest AI adopter and its least assured. A high-profile public-sector AI failure (benefits, policing, health) without transparency or audit trails would set adoption back years and corrode trust exactly when institutional capacity matters most.

A statutory ATRS publication duty with enforcement; an independent public-sector AI evaluation and audit function (NAO-linked, or CLTR's 'three lines' model separating risk ownership, oversight and audit) publishing before/after performance of tools like Consult; ring-fenced remediation funding for the unfunded high-risk legacy systems.

// State-led: Instrument: statutory ATRS publication duty plus NAO-linked audit function and ring-fenced funding.

Government is the biggest AI adopter yet its mandatory transparency standard is barely used and nothing audits whether tools work; one visible failure could set adoption back years.