AI Engineer Europe 2026
Rewiring the State
Loading slides…
Read the talk
Rewiring the State
Inside Number 10’s experiment with technical fellowships, embedded engineers and reusable AI tools—and the institutional changes needed to make their work scale.
From a talk by Eoin Mulgrew
What can a small team do about public-service backlogs?
How can a small technical team at the center of government help a public-service system struggling with waiting lists, court backlogs and slow planning decisions? Eoin Mulgrew leads cross-government transformation and the fellowship program in the Number 10 Data Science team, 10DS. Established during the pandemic, partly in response to it, the team’s core responsibility is to bring the best available evidence to consequential national decisions. It is now expanding its AI engineering capability, both inside Number 10 and across strategically important parts of the state.
The scale of the problem needs careful denominators. Mulgrew opens with an NHS waiting-list figure of 7.25 million; the contemporaneous NHS England release counts waiting-list entries, representing approximately 6.13 million unique patients in England in January 2026. He also cites roughly 350,000 backlogged court cases, without specifying the jurisdiction or court types. His description of only one in five planning decisions arriving on time similarly needs a narrower scope: England’s July–September 2025 statistics report 19% of major applications decided within the statutory 13 weeks, rising to 90% when agreed extensions are included. That is not a rate for all planning applications.
Behind these delays sits a public-sector productivity problem, worsened since the pandemic. Mulgrew cites a Tony Blair Institute estimate of £40 billion in potential annual productivity gains from AI in government. This is an estimate of opportunity, not savings already delivered. His broader framing is useful: government behaves more like a complex industry than a single organization. His shorthand of 400,000 people understates the civil service measured in the March 2025 statistics, which record 549,660 people and exclude NHS employees. The organizational challenge remains one of many different services, professions and operating environments.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The constraints are organizational as well as technical
Government has struggled to build and retain high-performing technical teams. Pay makes recruitment and retention difficult; hierarchy and bureaucracy can make delivery slow. Some barriers are perceived, while others are real. Some also exist for good reasons: public bodies are accountable to the public and Parliament, and their safeguards cannot simply be treated as unnecessary friction. Together, however, these conditions can repel precisely the impatient, capable engineers a transformation program wants to attract.
Changing the whole system exceeds the remit of a small central team. Mulgrew assigns that wider reform agenda to the Chief AI Officer, Kal Beer, and the Department for Science and Technology. For 10DS, the immediate opening is narrower: political leaders want new technology to improve services, so what can a small team do with that backing?
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
An exceptional team inside the system
The answer is an insurgency model: a small unit at the center with different operating conditions. Its design combines several permissions that normally sit apart.
- Political mandate: Number 10 backing to enter departments and work on consequential problems.
- Viable compensation: market rates within reason, without promising compensation comparable to Meta.
- Autonomy: freedom to pursue opportunities discovered while working inside a department.
- Technical selection: a recruitment process designed around technical ability rather than the standard civil-service process.
- Outside recruitment: fellows come exclusively from outside government, bringing experience the system would otherwise struggle to acquire.
Mulgrew reports an applicant success rate of about 0.7–0.8%. The longer-term benefit is not confined to a fellow’s initial placement: some stay in government and establish teams of their own.
Political backing creates room to work; it does not make implementation easy. A ministerial mandate cannot, by itself, dissolve data silos or make departments cooperate. If authority alone were sufficient, this model would already be commonplace.
The recruitment offer has nevertheless attracted people from AI labs, big technology companies and research institutes, alongside Y Combinator founders and serial entrepreneurs. The combination matters: economically viable pay, consequential decisions and an environment where technical people can do effective work. Mulgrew describes the desired recruits as “missionaries, not mercenaries.” Public purpose must sustain them when the work becomes difficult; compensation alone will not.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Start with the workflow, then choose the deployment
Legacy workflows offer many relatively simple opportunities for AI. For those, 10DS usually builds directly. Mulgrew describes Number 10’s first forward-deployed engineers working alongside policy advisers, lawyers, communications staff and pollsters. Their procedure starts with the people doing the work:
- Observe the existing workflow and its pain points.
- Co-design a solution with the team that will use it.
- Take the idea through implementation and put the capability into users’ hands.
Mulgrew describes some simple use cases taking days, with this embedded route typically moving from idea to implementation in a couple of weeks.
Large public-service backlogs require a different commitment. They are not just a collection of quick automation tasks, so 10DS deploys people into partner teams or departments, sometimes for prolonged periods.
| Work | Deployment approach |
|---|---|
| Bounded internal workflow | Build directly with its users |
| Complex departmental problem | Embed engineers through a sustained partnership |
The distinction is the depth and duration of the engagement, not whether engineers talk to users: both approaches depend on working inside the operational setting.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Model policy choices and make legal analysis reusable
Much of Number 10’s internal workflow automation is sensitive, but the policy simulation example shows the basic interaction. A policy team can explore a proposed change before making a decision. The demonstrated case concerns Universal Credit and its effects on household finances. The tool supports human analysis rather than necessarily replacing it; Mulgrew reports that more decisions can receive modeling support, and receive it faster. The useful capability is access to analysis during policy development, while choices are still open.
The next example changes the economics of repeated analysis. Mulgrew reports that the Cabinet Office was preparing to spend £1.5 million on external lawyers to analyze the entire UK statute book. He describes the volume of legal text as a stack four African elephants high. Instead, one engineer embedded with the in-house lawyers for a couple of weeks to build a tool they could use themselves.
The benefit was not only avoiding a proposed commission. The external analysis was expected to take longer than the pace at which new laws and regulations appeared. Its output would therefore age during production and eventually need to be repeated.
| Proposed commission | Internal tool |
|---|---|
| Analysis delivered after an external engagement | Analysis rerun by the legal team |
| New legislation makes the output stale | New legislation can prompt another run |
| Repeated work requires another engagement | Capability remains with the users |
The durable output is the ability to repeat the analysis. Mulgrew also raises the possibility of open sourcing the tool and sharing it across government; those are potential next steps, not completed releases.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Challenge delivery reports and expose progress
Number 10 receives reports on major projects and manifesto commitments across government. A delivery red-teaming tool gives its delivery teams something like a project management office in their pockets. Mulgrew says the team built it a couple of weeks earlier and that it is already used daily. Beyond interrogating an individual report, it questions the reporting organization’s judgment:
- Optimism bias: does this team or department habitually present an overly positive picture?
- Risk ratings: does it disproportionately classify risks as amber?
- Mitigations: have its proposed mitigations usually worked?
This adds a second judgment about the source of a report, rather than treating each report as an isolated account.
In-house engineering also supports public transparency. Mulgrew reports publishing two public-facing delivery dashboards in two months, describing them as a departure from previous practice. The named example is the AI Opportunities Action Plan dashboard, tracking commitments from Matt Clifford’s plan, including compute rollout and AI adoption. Its official landing page confirms publication on 29 January 2026; it does not establish the broader claim that government had never previously published a public delivery dashboard.
A further example is deliberately unnamed: a new public service awaiting a ministerial launch. Mulgrew says its idea originated two months earlier, with launch expected two and a half weeks after the talk and millions of prospective users. He contrasts that timetable with a conventional government discovery phase that might last a year or more. At this point in the recording, launch and public adoption are still expected outcomes.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Build capability through partner teams
The partnership model extends beyond Number 10 through three examples: the AI Safety Institute, the Incubator for AI and Justice AI. Mulgrew describes AISI as the world’s first government body for evaluating frontier models and says early fellows helped establish its cybersecurity workstream. He credits fellow Dr. Harry Coppock, placed there from the beginning, with leading work on Inspect.
Inspect supports evaluation of models and agents, including prompts, tools and multi-turn interactions. Mulgrew introduces it through the need to test what agents actually do when given autonomy and tools in an isolated environment. The technical distinction is that Inspect is an evaluation framework with sandboxing support, not an automatic isolation boundary around every tool call. Its sandboxing documentation distinguishes code running in the evaluation process from work explicitly routed through a sandbox interface.
The Incubator for AI, now within DSIT, illustrates a second form of institutional growth. Mulgrew describes it as a spin-out of the fellowship program, with fellows making up much of its original technical team. Its role is to incubate AI solutions for use across the public sector; the relationship with 10DS continues as those solutions scale.
Extract is one such collaboration. Built with DeepMind on Gemini, it digitizes planning materials, including handwriting and hand-drawn maps. It was unveiled at London Tech Week in June 2025, and Mulgrew describes rollout to every local authority in England as underway. Digitization addresses a bottleneck in the planning process; automatic AI decisions on planning applications are a future aspiration he raises, not the tool’s stated current capability.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Evaluate tutoring before broad adoption
AI tutoring offers the possibility of making strong educational support available regardless of a child’s socioeconomic background. But access alone is not enough: classroom use requires safeguards and evaluations suited to children. The team is evaluating frontier models against benchmarks for both safe interaction and educational behavior.
The illustrated criterion is cognitive load placed on the student. The slide pairs evaluation criteria with a student–AI tutor conversation, making the interaction itself the object of assessment. The question is not merely whether the model can produce an answer, but how demanding its tutoring behavior is for the learner.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Embedded engineers on the frontline
Justice AI takes the embedded model into the Ministry of Justice. Its founder, Dan James, is a former fellow, though Mulgrew explicitly declines to call the team a fellowship spin-out. Instead of working beside policy advisers and communications staff, its forward-deployed engineers work alongside parole officers and prison wardens. The examples include using AI to interrupt drug flows into prisons, reduce labor-intensive manual processes and improve prison safety and security. The operational details are not disclosed.
A fellow named Will makes that placement model tangible. A few months earlier he had been in California, having left Harvard and founded a company accepted into Y Combinator. Mulgrew describes a photograph of him outside HMP Wandsworth in the rain, in his second week on the job, holding the prison’s actual keys. The recruitment proposition is unusually concrete: experienced industry practitioners can enter government and receive access to consequential operational work.
Mulgrew calls the program an early experiment. He points to savings, faster public-service delivery, frontline reform and new internal AI capabilities as evidence that small, highly capable teams can accomplish useful work. The prepared talk ends with a hiring invitation through an on-screen QR code and conversation afterward. The questions that follow test what would be needed to make the experiment reliable and scalable.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
What if the policy tool tells users what they want to hear?
The first question returns to policy simulation. A user who already believes a policy is excellent may steer a conversational model toward confirming that belief. Mulgrew illustrates the failure with the idea of cutting income tax to zero and receiving enthusiastic agreement. Sycophancy would turn an analytical aid into a source of apparently authoritative reassurance.
The response has two parts. Before release, the team red-teams models specifically for this behavior. Afterward, it coaches users on the risks of the tools they are receiving. Those users may be lawyers, sociologists or professors with substantial domain expertise but little experience of AI failure modes. Mulgrew says the team has encountered relatively little sycophancy so far and attributes that partly to testing before deployment; he does not dismiss the risk.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make the exception normal, then work across services
The next question asks how the model scales across central and local government, including different political contexts. Fellows staying in government and founding teams help, but Mulgrew judges that organic growth too slow for the urgency of the problems. Scaling requires strategic changes to how the rest of government operates.
The bargain with ministers was to let a small team operate under different rules and use its results as a proof point. The intended destination is for those conditions to become business as usual. Mulgrew describes the present arrangement as a way around the system; institutional reform is needed so effective technical delivery no longer depends on an exception.
The other route to scale is horizontal work: capabilities applicable to repeated processes across services rather than one targeted use case at a time. Mulgrew sets this as an ambition for the next 12–24 months. His underlying distinction is between the visible policy workforce around Westminster and the much larger operational workforce. He ranges across call-center staff, prison wardens and nurses to illustrate frontline public service, although the latter sit outside civil-service statistics. The concrete opportunities are police transcription and the large call centers in DWP and HMRC. A reusable improvement to a common process could reach much further than a tool serving one team.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The remaining gaps: motivation and shared experience
An edtech practitioner challenges the tutoring example from another direction: even an excellent tutor may fail if a child does not want to learn. What would school deployment look like, and how would the government address motivation? Mulgrew clarifies that the current plan centers on benchmarks and guardrails for schools choosing products, rather than building a government tutor to compete with vendors. Student uptake has received little work so far.
Mulgrew says the initial tutor testing involved about 70 teachers role-playing their pupils. That supports a different kind of assessment from observing children’s motivation or sustained use directly. He invites the questioner to share practical experience afterward, leaving adoption as an open area for collaboration rather than a solved part of the program.
The final question comes from a Norwegian attendee asking about international exchange. Mulgrew reports some collaboration, but none with Norway so far. He names Tech Force and parts of the US Digital Service as broadly similar initiatives, and says the team talks frequently with Singapore. There is room to do more, and he welcomes contacts in the Norwegian government: the experiment can learn from other countries as well as from the departments where its engineers are embedded.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The June 2025 announcement explains Extract's use of Gemini and its planned rollout to English councils.
AISI's introduction to its open-source framework for evaluating language models and tool use.
Official access page for tracking progress against the government's AI Action Plan commitments.
Further reading
- AI action plan for justiceArticle
The Ministry of Justice's approach to AI adoption, specialist recruitment, workforce training and partnerships.
Updates since the talk
Current instructions and examples for running evaluation workloads in isolated environments.
Full transcript
Loading transcript…