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

The UK government's Number 10 Data Science team is building an insurgent unit model that recruits exceptional technical talent from outside government to rap...

By Sean Weldon

Insurgent Units and Forward-Deployed Engineers: Accelerating AI Deployment in Government Through Organizational Innovation

Abstract

The UK government's Number 10 Data Science team has implemented an insurgent unit model to address critical public sector productivity challenges through rapid deployment of artificial intelligence solutions. Operating with unprecedented political mandate and autonomy, the team recruits exceptional technical talent from industry and research institutions using highly selective processes (0.7-0.8% acceptance rates) and deploys them as forward-embedded engineers working directly with operational teams. Early implementations demonstrate deployment cycles measured in weeks rather than years, including policy simulation tools, AI-powered planning application digitization using Gemini, and criminal justice system improvements. This analysis examines the structural innovations enabling these outcomes, including custom recruitment mechanisms, partnership models for cross-government challenges, and horizontal scaling strategies. The findings suggest that small, elite technical teams with appropriate organizational positioning can rapidly deliver transformative capabilities within large bureaucratic institutions when freed from traditional constraints.

1. Introduction

Public sector productivity in the United Kingdom faces a crisis of significant magnitude. Operational metrics reveal systemic dysfunction across critical government services: 7.25 million individuals await NHS treatment, 350,000 court cases remain in backlog, and only 20% of planning applications receive decisions within prescribed timeframes. These failures represent not merely administrative inefficiencies but fundamental barriers to effective governance that compound over time, eroding public trust and economic productivity.

The UK government estimates that artificial intelligence applications could generate £40 billion in annual productivity gains across public sector operations. However, traditional civil service structures encounter systematic difficulties in recruiting and deploying technical talent capable of realizing these opportunities. Compensation disparities with private sector employers create retention challenges, while hierarchical organizational frameworks and risk-averse regulatory environments impose operational friction incompatible with rapid technical iteration. Traditional recruitment processes optimize for generalist capabilities and procedural compliance rather than exceptional technical specialization, proving inadequate for building teams capable of developing sophisticated AI systems.

The Number 10 Data Science team represents a deliberate organizational experiment designed to circumvent these structural constraints. Operating as an insurgent unit—a small, autonomous team with exceptional political backing positioned at the center of government—the initiative demonstrates that alternative organizational models can achieve disproportionate impact. This analysis examines the mechanisms enabling accelerated AI deployment within government contexts, including recruitment innovations, the forward-deployed engineer model, partnership frameworks for complex challenges, and strategies for horizontal scaling across the 400,000-person civil service.

2. Background and Related Work

2.1 Structural Impediments to Technical Excellence in Government

Government organizations face systematic disadvantages in competing for technical talent beyond simple compensation differentials. The hierarchical decision-making structures characteristic of civil service organizations create approval chains that extend deployment timelines from weeks to months or years. Regulatory safeguards and procurement requirements, while serving legitimate governance functions, impose procedural overhead that high-performing technical professionals accustomed to rapid iteration cycles find operationally prohibitive.

Furthermore, traditional civil service recruitment optimizes for different capabilities than those required for advanced technical work. Standardized processes designed to ensure fairness and reduce bias through uniform evaluation criteria prove ill-suited for identifying exceptional technical talent, where relevant experience may manifest in non-traditional credentials such as open-source contributions, research publications, or startup founding experience. The resulting talent pool, while competent for many government functions, lacks the specialized expertise required for frontier AI development and deployment.

2.2 The Forward-Deployed Engineer Paradigm

The forward-deployed engineer (FDE) model embeds technical personnel directly within operational contexts rather than isolating them in centralized development teams. This approach, refined in technology companies serving complex enterprise customers, enables engineers to observe workflows firsthand, identify inefficiencies through direct user interaction, and co-design solutions with domain experts. The methodology reduces communication overhead and misalignment between technical capabilities and operational requirements, accelerating the feedback loops essential for effective system design.

Application of this model within government contexts represents a significant departure from traditional IT procurement and development practices, which typically separate technical development from operational deployment through formal requirements documentation and vendor relationships. The insurgent unit model provides the organizational flexibility necessary to implement forward-deployed engineering at scale within government.

3. Core Analysis

3.1 Recruitment Mechanisms for Exceptional Technical Talent

The Number 10 Data Science team employs a custom recruitment process fundamentally distinct from traditional civil service hiring. The process achieves acceptance rates of 0.7-0.8%, comparable to highly selective technology companies and research institutions, through laser-focused evaluation of technical capabilities rather than generalist competencies. Recruitment targets individuals from AI research laboratories, major technology companies, research institutes, and Y Combinator-backed startups—populations with demonstrated technical excellence but limited prior government experience.

Critically, the model addresses compensation constraints not through matching private sector salaries entirely but by offering "market rates within reason" that make roles economically viable while recruiting what the team characterizes as "missionaries, not mercenaries." This framing acknowledges that individuals motivated primarily by financial compensation will find superior opportunities elsewhere, while those attracted to high-impact problem domains may accept moderate compensation discounts when working on challenges of genuine societal significance. The ability to offer competitive compensation represents a departure from standard civil service pay scales, enabled by the unit's exceptional political mandate.

The exclusive recruitment of external talent rather than internal civil servants represents a deliberate strategic choice. While potentially creating organizational friction, this approach ensures the team possesses technical capabilities and operational norms aligned with private sector standards rather than government conventions. The model explicitly trades internal political capital for technical excellence and deployment velocity.

3.2 Rapid Deployment Through Organizational Autonomy

The insurgent unit operates with "unusually high levels of autonomy and opportunism," enabling deployment cycles measured in weeks rather than the year-plus timelines typical of government IT projects. Concrete examples demonstrate this acceleration: the team published two public-facing government delivery dashboards within two months, developed a policy simulation tool enabling policy teams to model universal credit scenarios before implementation, and launched a new public service reaching millions of users in 2.5 weeks—a timeline that would typically require year-long discovery phases under standard procurement processes.

The statute book analysis project illustrates both technical capability and economic efficiency. The team replaced a £1.5 million external legal firm contract with an in-house engineer embedding for two weeks, developing an AI-powered tool capable of continuous updates at the pace of new legislation rather than requiring periodic expensive re-engagement with external vendors. Similarly, a delivery red teaming tool analyzes departmental reporting to flag optimism bias and risk rating patterns, providing ministers with independent assessment of project health beyond self-reported metrics.

This velocity derives not from reduced quality standards but from organizational positioning that enables rapid decision-making. The mandate from Number 10 and high-level political backing allows the team to bypass approval chains that would otherwise extend timelines, while the small team size reduces coordination overhead. The model demonstrates that appropriately positioned small teams can achieve disproportionate impact relative to their size.

3.3 Forward-Deployed Engineers and Operational Embedding

The team deploys engineers directly into operational contexts across government, representing the first forward-deployed engineers in Number 10 history. These personnel embed with policy teams, operational units, legal departments, communications teams, and polling operations, observing workflows to identify pain points and co-designing solutions with domain experts. The approach achieves capability deployment from initial concept to implementation within weeks.

The Just AI initiative in the Ministry of Justice exemplifies this model's application to complex operational challenges. Forward-deployed engineers embed with parole officers and prison wardens, working directly in facilities such as HMP Wandsworth to develop AI solutions addressing drug flow reduction and security improvements. One fellow—a Harvard dropout and Y Combinator founder—embedded in a prison during his second week, illustrating the model's emphasis on immediate operational engagement rather than extended orientation periods.

This embedding approach addresses a fundamental challenge in government technology deployment: the gap between technical capability and operational context. Traditional models rely on formal requirements documentation to bridge this gap, but such documentation often fails to capture tacit knowledge and workflow nuances critical for effective system design. Forward deployment enables engineers to acquire this contextual understanding directly, reducing misalignment between technical solutions and operational needs.

3.4 Partnership Models for Cross-Government Challenges

For challenges requiring sustained engagement beyond the core team's capacity, the initiative employs a partnership model deploying fellows into other government entities while maintaining coordination. The AI Safety Institute received deployed fellows including Dr. Harry Coppock, who leads cybersecurity workstreams and development of Inspect—a CFI isolated environment for testing AI agent autonomy and tool usage. The Incubator for AI, a spin-out program in the Department for Science and Technology, was founded primarily by fellows from the core team.

The Extract tool represents a significant partnership with DeepMind, leveraging Gemini to digitize planning applications including handwritten documents and hand-drawn maps. The system is rolling out to every local authority in England, addressing the planning application backlog that currently sees only 20% of applications decided on time. This partnership model enables the core team to access frontier AI capabilities while maintaining deployment velocity through embedded fellows who understand both the technology and operational context.

Education initiatives focus on AI tutors, with the team developing safeguards and evaluating frontier models against benchmarks including cognitive load metrics—a sophisticated evaluation approach that considers not merely answer correctness but the pedagogical quality of AI-generated explanations. These partnerships demonstrate how a small central team can achieve broad impact through strategic deployment of personnel to high-leverage initiatives across government.

4. Technical Insights

The technical implementations reveal several patterns relevant to AI deployment in complex organizational contexts. The Inspect tool developed for the AI Safety Institute provides isolated testing environments for evaluating AI agent autonomy and tool usage, addressing fundamental safety concerns around deployed AI systems. This infrastructure investment enables more confident deployment of AI capabilities across government by providing standardized evaluation frameworks.

The Extract tool's use of Gemini for planning application digitization demonstrates effective application of frontier model capabilities to domain-specific challenges. The system handles not only typed text but handwritten documents and hand-drawn maps—formats that previous digitization efforts struggled to process. This capability directly addresses a bottleneck in the planning approval process, where manual digitization of legacy documents creates substantial delays.

Policy simulation tools enable ex-ante evaluation of policy decisions, modeling impacts on household finances and other metrics before implementation. This represents a shift from purely retrospective policy evaluation to prospective scenario analysis, potentially reducing costly policy failures. The delivery red teaming tool applies similar analytical approaches to departmental reporting, identifying patterns indicative of optimism bias and tracking mitigation effectiveness—essentially applying adversarial analysis to internal government communications.

Implementation trade-offs merit consideration. The model's reliance on external recruitment creates knowledge transfer challenges when fellows complete their terms, potentially creating dependencies on individuals rather than building institutional capabilities. The emphasis on rapid deployment may also create technical debt if systems lack adequate documentation or maintainability considerations. However, the team explicitly frames current operations as a "pilot" or "hack" designed to prove concept viability before institutionalization, suggesting awareness of sustainability challenges.

5. Discussion

The insurgent unit model demonstrates that organizational innovation can overcome structural barriers to technical excellence within large bureaucratic institutions. The critical enabling factors include exceptional political mandate, compensation flexibility, recruitment process customization, and operational autonomy. These elements combine to create an environment where high-performing technical talent can operate effectively despite the constraints typical of government contexts.

The forward-deployed engineer model addresses a fundamental challenge in government technology: the separation between technical development and operational deployment. By embedding engineers directly in operational contexts, the approach reduces the communication overhead and misalignment that extend deployment timelines in traditional models. The effectiveness of this approach suggests broader applicability beyond government to any large organization struggling to deploy technical capabilities rapidly.

Scaling represents the critical challenge for the initiative's next phase. The current model operates as a small elite team, but the civil service encompasses 400,000 employees, predominantly in operational roles such as call center operators, prison wardens, and nurses rather than policy functions. The team's stated focus for the next 12-24 months on "horizontal work targeting processes applicable en masse"—including transcription and call center operations—reflects recognition that demonstration projects must evolve into systematic capability deployment. This transition from insurgent unit to institutionalized practice requires strategic intervention to modify how the broader government operates, a challenge qualitatively different from the initial technical deployment focus.

International interest from the United States government, Singapore, and Norway suggests the model's potential generalizability beyond UK-specific contexts. However, replication requires not merely technical knowledge transfer but organizational conditions enabling autonomy and political mandate—factors that may prove more challenging to reproduce than technical methodologies.

6. Conclusion

The Number 10 Data Science team's insurgent unit model provides empirical evidence that small, elite technical teams with appropriate organizational positioning can rapidly deploy transformative AI capabilities within large bureaucratic institutions. Key innovations include highly selective recruitment of external technical talent, forward-deployed engineering embedding personnel in operational contexts, partnership models for cross-government challenges, and organizational autonomy enabling deployment cycles measured in weeks rather than years.

Early outcomes demonstrate practical impact: policy simulation tools enabling ex-ante decision evaluation, AI-powered planning application digitization addressing critical backlogs, criminal justice system improvements through embedded engineers, and substantial cost reductions through in-house capability development. These results suggest the model's viability as an approach to accelerating technical capability deployment in government contexts.

The critical question for the initiative's next phase concerns scalability: whether practices proven effective in a small insurgent unit can become "business-as-usual" across a 400,000-person civil service. This transition requires not merely additional deployment but strategic intervention to modify institutional practices and incentive structures. Success would demonstrate a pathway for other large organizations—governmental or private—to overcome structural barriers to technical excellence through targeted organizational innovation. The model's emphasis on horizontal process automation and international collaboration suggests recognition that sustainable impact requires moving beyond demonstration projects to systematic capability transformation.

Sources

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub