Rewiring the State — Eoin Mulgrew, No. 10 (Downing Street)

Eoin Mulgrew, from the Number 10 Downing Street data science team (10DS), presents the UK government's insurgent approach to deploying AI engineering talent at the center of state to tackle chronic public service delivery failures. He opens by citing stark public-sector statistics: 7.25 million people on NHS waiting lists, 350,000 court cases backlogged, and only one in five planning applications decided on time. He references a Tony Blair Institute figure estimating a £40 billion annual productivity prize from AI adoption in government.

The core of the talk is 10DS's fellowship model — an explicitly insurgent unit operating under direct political mandate from Number 10. Key structural features: market-rate pay (not Meta-level, but competitive enough), a 0.7–0.8% acceptance rate from a technically rigorous selection process, and a deliberate policy of recruiting exclusively from outside the civil service — pulling candidates from AI labs, big tech, top research institutes, and Y Combinator founders. The recruitment pitch is explicitly missionary: people take pay cuts because the problems are consequential.

Mulgrew walks through several concrete shipped projects. First, a policy simulation tool allowing Number 10 teams to model the household financial impact of policy decisions like Universal Credit changes before they are made — described as accelerating the rate at which modeling informs ministerial decisions. Second, a legal analysis tool built in two weeks by one engineer embedded with the in-house legal team, replacing a planned £1.5 million contract with an outside law firm to analyze the UK statute book — with the added advantage that the tool runs continuously as new laws are made, whereas the external contract would have become stale. Third, a delivery red-teaming PMO tool used daily inside Number 10 that interrogates departmental delivery reports, flags optimism bias by department, and assesses whether mitigations are historically effective. Fourth, two public-facing dashboards published within two months — including the AI Opportunities Action Plan tracker unveiled by Matt Clifford — the first time the government had ever published delivery dashboards for public scrutiny. Fifth, a new public service conceptualized two months prior and set to launch in 2.5 weeks, serving millions of citizens — a project that Mulgrew notes would normally spend a year or more in discovery under standard government processes.

Beyond Number 10 itself, 10DS has seeded capability elsewhere. Fellows helped stand up the AI Safety Institute (AISI) from day one, including Dr. Harry Coppock who led development of the Inspect tool — an isolated environment for evaluating what AI agents actually do when given autonomy and tools. The Incubator for AI (now in DSIT) was effectively spun out of the fellowship program, and produced Extract — a Gemini-powered tool built in collaboration with DeepMind that digitizes planning application documents including hand-drawn maps, unveiled by the Prime Minister at London Tech Week and now being rolled out to every local authority in England. A third partner, Just AI, sits in the Ministry of Justice and deploys forward-deployed engineers into prisons and the criminal justice system, using AI to intercept drug smuggling and improve prison security.

On AI tutors, 10DS is working to establish safety benchmarks and evaluation frameworks — including testing frontier models against cognitive load metrics with 70 teachers role-playing pupils — rather than building competing products. In the Q&A, Mulgrew addresses sycophancy risks in policy simulation tools (red-teaming models before deployment, user coaching), scaling challenges beyond the insurgent model (needs systemic civil service reform, not just elite teams), and future ambitions to pursue horizontal use cases — transcription for police, large call centers in DWP and HMRC — that can be applied en masse across the 400,000-person civil service.