Rewiring the State — Eoin Mulgrew, 10 Downing Street
TL;DR
Downing Street is building an AI "insurgency" inside government — Eoin Mulgrew says 10 Downing Street is using a small, politically backed team with market-rate pay, custom hiring, and unusual autonomy to move faster than normal civil service structures allow.
The pitch is simple: give elite outsiders real state power, fast — the fellowship recruits people from DeepMind, frontier labs, big tech, research institutes, YC, and startups with a brutal 0.7-0.8% acceptance rate, then drops them into departments to ship in weeks, not years.
There’s massive low-hanging fruit in UK government operations — Mulgrew cites 7.25 million people on NHS waiting lists, 350,000 backlogged court cases, and only 1 in 5 planning applications decided on time as evidence of a public-sector productivity crisis with a claimed £40 billion annual AI upside.
Some wins are already concrete, not theoretical — one embedded engineer reportedly saved the Cabinet Office £1.5 million by building a legal analysis tool instead of hiring outside lawyers to review the UK statute book, which Mulgrew jokes is "the height of four African elephants."
The model is already spawning new institutions and products — fellows helped launch work at the AI Safety Institute, contributed to Inspect, seeded the Incubator for AI, and supported Extract with DeepMind and Gemini to digitize planning workflows now being rolled out to every local authority in England.
Mulgrew frames this as a pilot for rewiring government itself — he calls the current setup a "hack" around the system, with the longer-term goal of turning these exceptional methods into business-as-usual across the UK state over the next 12 to 24 months.
Summary
The opening grumble: government is broken, and everyone knows it
Eoin Mulgrew opens with a half-joking, half-pained reality check: the UK has 7.25 million people on NHS waiting lists, roughly 350,000 court cases stuck in backlog, and only 1 in 5 planning applications decided on time. He frames government as an "industry" of 400,000 people suffering a deep productivity crisis, with a Tony Blair Institute estimate putting the annual AI upside at £40 billion.
Why government struggles to attract the people it needs
He doesn’t sugarcoat the civil service problems: pay is weak, hierarchy is heavy, bureaucracy is real, and accountability to Parliament means things move slowly for understandable reasons. The result, he says, is an environment that often repels the exact sort of technical people government most needs — ambitious builders who want to leave a mark quickly.
The "insurgency model" inside Number 10
Mulgrew’s answer is to "take the shackles off" and build a small insurgent unit at the center of power with political cover, more autonomy, and permission to recruit differently. Their selection process is intentionally harsh — just a 0.7-0.8% success rate — and they recruit only outsiders, betting that importing talent from labs, big tech, research, YC, and startups can change the system from within.
Missionaries, not mercenaries
He stresses that this isn’t just about paying more, even if market-rate pay helps make the jump into government viable. What they want are "missionaries, not mercenaries" — people who’ve already done well elsewhere but are motivated by the chance to work on decisions crossing ministers’ desks and affecting millions of lives.
Quick wins: simulations, legal tooling, and delivery dashboards
Mulgrew then gets into the fun part: AI tools shipped in days or weeks. He describes policy simulation tools for testing decisions like Universal Credit changes, a legal analysis tool that replaced a planned £1.5 million spend on outside lawyers to review the UK statute book, and a delivery "red teaming" tool that acts like a PMO in officials’ pockets, flagging things like optimism bias and weak risk mitigation.
Shipping public services at startup speed
One of his loudest claims is that a new public service, conceived just two months earlier, is about to launch for millions of people — something he says would normally sit in government discovery for a year or more. He also points to new public-facing dashboards, including one for Matt Clifford’s AI Opportunities Action Plan, as evidence that internal AI capability can improve transparency as well as efficiency.
From Number 10 to the wider state
The fellowship’s influence now extends beyond Downing Street. Mulgrew highlights the AI Safety Institute, where fellow Harry Coppock helped lead work on Inspect; the Incubator for AI, which he describes as effectively spun out of the program; and Extract, a DeepMind collaboration built on Gemini to digitize handwritten planning documents and maps for rollout across England.
The prison story that lands the point
He closes with Just AI in the Ministry of Justice, founded by former fellow Dan James, where engineers embed with parole officers and prison staff to tackle things like drugs entering prisons and manual operational bottlenecks. The image he leaves the room with is perfect: Will, a Harvard dropout and YC founder who was "in California getting a tan" a few months earlier, now standing outside HMP Wandsworth in the rain with the prison keys in his second week — "come join us, and we’ll give you the keys to the state."
Was This Useful?
Share
Keep Reading
Make Alcreon Yours
Tune your feedFive quick questions, and the feed ranks what matters to you first.Or just get notified
The weekly Echo. Signal worth keeping in your inbox.
Every new piece, announced on X.
Read Next
See all
Playbook
The Retirement Email Isn't a Warning
Model retirements now arrive every few weeks; the config-eval-rehearsal loop turns each deprecation email from a fire drill into an afternoon swap.

Playbook
The Cheapest Model That Passes
OpenRouter lists 400 models behind one API. The fix for choosing isn't a better leaderboard, it's a four-step protocol that ends in a real eval.

Playbook
Cheap Models, Hard Tasks
Most agent workflows route every step to the frontier model by default. The bill scales with how chatty the agent gets, even when most steps don't need that brain.