At the Public Service Data.AI Conference on 15th October, Nimble CTAIO, Nigel Garner, joined public sector leaders on the main stage for “Making AI work: Moving from pilots to practice”. The discussion explored a critical question facing the government: how do we move beyond AI experimentation to embed AI into everyday public services, delivering lasting improvements for citizens and the people who serve them?
Nigel’s central message was clear: the biggest opportunity isn’t simply giving people better AI tools. It’s rethinking how we organise work, build trust, and turn innovation into sustainable operational change.
From individual productivity to human-digital teams
Across government, there are encouraging examples of AI adoption, from GOV.UK Chat and HMRC’s early rollout work to initiatives helping departmental teams reduce operational friction and make more time for human judgement.
Tools such as AI assistants are already helping people draft documents, summarise information, and complete everyday tasks more efficiently. These are valuable first steps, improving productivity and building AI awareness. But there is a risk of mistaking individual productivity gains for meaningful transformation of public services.
The next step is to move from one person using AI to a human-digital team: a coordinated workforce in which people and specialist AI agents work together.
In this model, humans set the direction, define boundaries, and retain accountability. Digital agents take on defined responsibilities, from research and analysis to testing and risk assessment. For example, in cyber defence, agents could continuously assess threats, analyse evidence, and prepare reports, allowing human security analysts to focus on critical decisions and strategic judgement.
As Nigel highlighted, the opportunity is to look beyond staff augmentation and process improvements towards operating-model reform. During the discussion, it was questioned whether government is ready for the increase in pace AI could bring. But by redesigning the operating model, we can ensure people don’t become the bottleneck as AI accelerates delivery.
The opportunity is not to replace people, but to multiply their capabilities. The difference lies less in the AI model itself and more in how we design the work around it.
Three foundations for AI that works
If government is to realise this potential, three things need to develop together.
1. Data foundations
AI agents need more than access to data. They need to understand the facts, relationships, and meaning behind it. That means moving towards agent-ready data, supported by clear governance over what each agent can access and under whose authority.
Legacy systems make this challenging, but getting the foundations right is essential for safe, effective deployment.
2. Engineered context
Much of the knowledge that makes a department work isn’t formally documented. It lives in people’s heads: why a previous approach failed, how a policy has evolved, or what lessons have been learned from past projects.
AI needs this context just as a new human colleague would. Capturing, maintaining, and sharing relevant knowledge helps agents reason more effectively, supports better decisions, and can enable smaller, more cost-effective models to deliver useful results.
3. Cultural adoption
Technology alone cannot create trust. Every agent needs a defined role, a named owner, appropriate data access, and clear boundaries on what it can do. Organisations also need the governance and oversight to manage these digital colleagues responsibly.
Evaluation is equally important. Pippa Fernandes, Deputy Director of Strategy at the UK Government’s Incubator for AI (i.AI), highlighted the need for robust ways to assure ourselves that AI systems are working as intended. Finding Evaluation Engineers is challenging, but government is exploring how people in adjacent roles, such as software testing, can bring their critical assessment skills into this emerging profession.
Data foundations without context create systems that cannot act effectively. Context without trust creates infrastructure that people won’t use. All three foundations must progress together.
Why successful pilots don’t always scale
Innovation teams can demonstrate what’s possible, but transferring that success into mainstream policy and operational teams is much harder.
The problem isn’t necessarily the pilot. It’s the gap between the team developing it and the team expected to run it. Innovation specialists can dedicate themselves to emerging technology; frontline and policy teams already have demanding day jobs and cannot be expected to maintain the same level of specialist knowledge.
Bridging this gap requires collaboration from the outset. In our work with the Department for Education, we focus on co-designing and co-authoring solutions with departmental colleagues from day one. This helps ensure that the receiving team understands the context, owns the capability, and is prepared to operate it when the solution goes live.
It also means redefining what “done” looks like. A successful pilot isn’t simply one where the technology works in a sandbox, but one where the receiving team has the governance, knowledge, skills, and support to sustain the intended outcomes.
That requires funding the last mile of adoption, building effective feedback loops between digital and policy teams, and measuring real-world outcomes rather than technical milestones alone. As Pippa noted, AI can also enable teams to experiment more quickly, learning from failure while still progressing faster than traditional approaches.
Richard Henderson, Assistant Deputy Minister and Corporate Chief Information Officer for the Government of Alberta, Canada, offered a complementary perspective:
“Solve things 5 people at a time.”
His point was that incremental improvements, often more achievable than wholesale transformation, can collectively deliver significant impact. Starting small, learning what works, and building on those gains can help government move from successful pilots towards sustainable change.
Invest in context, not just technology
With rising demand and constrained resources, governments must be deliberate about where they invest in AI. Nigel’s recommendation was to look beyond the models themselves: these are increasingly interchangeable.
The more strategic investment is in departmental knowledge, data foundations, and the context that makes intelligent systems useful. Embedding this knowledge in external platforms without considering ownership and control risks creating long-term dependencies and undermining departmental sovereignty.
Attracting and retaining the right talent is another part of the equation. Pippa highlighted that government cannot always compete with the private sector on salary, so it must lean into its mission, offering meaningful work and opportunities to make a difference. It also needs to recruit proactively, recognise transferable skills, and accept that even shorter periods of service can be mutually beneficial, bringing valuable private-sector experience into government.
Moving AI from pilots to practice means treating it as a workforce and operating-model transformation, not simply a technology rollout.
The public sector has already demonstrated what’s possible. The next challenge is to make those gains repeatable, trusted, and sustainable, so that AI improves not just individual tasks, but the services government delivers every day.m, and giving teams the time and structure to explore ideas can be a powerful way to start.
