The interview was conducted by Ula Rutkowska (Senior Researcher, Apolitical) and edited by Christina Obolenskaya (MSc in International History, LSE and Communications Intern, Apolitical). D. Laura Gilbert OBE was named on Apolitical’s Government AI 100 2025.

How do you convince people to adopt AI when their first reaction is scepticism? For Laura Gilbert, Head of AI for Government, Ellison Institute of Technology, Oxford and former Director of the Incubator for AI at 10 Downing Street, the key is to lead with the problem, not the technology. People don’t want AI for the sake of it — they want solutions that make their jobs easier and improve public services.

AI isn’t about replacing the human element in government; it’s about boosting efficiency, cutting costs and helping public servants focus on what matters. In this interview, Gilbert discusses how to design AI solutions that work, navigate challenges in adoption and integrate AI into government services effectively.

Q: How has your professional experience shaped your understanding of AI’s potential to enhance frontline public services?

There’s been much discussion about AI from every angle — the good, the bad and the outright strange. But in government, what really matters is hearing from those on the frontline. The technology itself is something we broadly understand. Large language models (LLMs) may feel new, but AI in various forms has been around for decades. Even years ago, we were using early versions of LLMs to analyse factors driving maternal mortality in hospitals.

So, in some ways, this isn’t entirely new work. What’s crucial, though, is how users interact with AI. Problems arise when people build tools and simply hand them over with a, “Here you go, use this” approach — without fully understanding frontline workflows. Public servants are busy, and if a tool doesn’t fit seamlessly into their day-to-day reality, it won’t be used effectively. A lot of our work has focused on engaging directly with those who will be using AI tools, ensuring that what we build actually supports their needs.

Q: What are the steps you take to build a new AI product?

The first step is scoping — figuring out the problem we’re trying to solve. That includes assessing the return on investment, but more importantly, understanding how we would actually deliver the solution. Who will use it? How do we engage them? Do they even want this product? Will it fit into their workflow? And crucially, how do we make it useful, engaging and accessible to the public? Following that, we go through the usual software development lifecycle, incorporating product design specialists and user researchers alongside our skilled engineers to take a product from prototype to alpha (internal testing) to beta (user testing) to full deployment.

Most importantly, we involve frontline workers — the people delivering public services — early in the process. They explain their daily challenges, share their perspectives and work alongside teams to ensure AI solutions genuinely address their needs. Co-building products with the people who need to use them is critical to success. A big part of our team’s work also involves hackathons, but ours work a little differently from most. We run a programme called Evidence House, which has over a thousand public servants signed up. Some have coding experience, while others bring expertise in policy, branding or service delivery. We bring them together, give them a problem and relevant data, and build diverse teams based on their skills. Sometimes real product prototypes come out of these, but there is also huge benefit in involving people in how you’d turn an idea into a workable data-driven tech solution.We’ve also taught thousands of people the basics of how to code in these sessions.

Q: What is an example of successfully building out a new product in this way?

One example is Caddy, a tool we developed for Citizens Advice, which we’re now rolling out across other government services where a human advisor interacts with someone in need. In Citizens Advice, many clients are in vulnerable situations, and they’ve come specifically to speak with a person — not an AI. So, rather than replacing that human connection, we designed Caddy to support advisors in doing their jobs more effectively.

Many advisors are new or volunteers and traditionally, they rely on supervisors, Google searches, or stacks of reference materials to verify their guidance. By understanding how they actually work, we were able to design AI that fits seamlessly into their process.

Caddy can be integrated directly into Google Meet, Microsoft Teams or Slack, functioning like a colleague. An advisor can simply ask, “@Caddy, what advice should I give in this situation?”—without needing to install a new tool or disrupt their workflow. Supervisors are tagged in automatically, so they can oversee and verify responses, but the process is much faster.

The impact has been significant. Our evaluations have shown that advisors using Caddy are 2.5 times more confident in the advise they have given, and members of the public report a 60% higher resolution rate. The tool continuously evolves through built-in evaluation, allowing us to refine and improve it based on real user feedback.

Government provides unique access to the people who deliver public services, and when you treat their expertise, relationships and care for their work as central to product design, you create solutions that are both effective and widely adopted.

Q: How did Caddy evolve through multiple rounds of evaluation?

Caddy was one of the rare cases where we got it right early on. By then, we had already learned some key lessons and we worked closely with Citizens Advice throughout development. The core design — integrating directly into Teams and Google to fit existing workflows — remained the same. Most of the changes focused on improving the quality of advice and fine-tuning responses to be more accurate and useful.

That said, we’ve definitely built products that changed dramatically over time. A great example is Redbox, our first major project, which started as a hackathon idea to help ministerial private offices manage a minister’s red box. Ministers often need to form a view on complex 200-page regulatory documents without much context, so their private office has to read everything, extract the key points and provide advice. Our goal was to automate parts of that process, making it faster and easier to get relevant insights.

The first version, though, didn’t meet user needs at all. We rolled it out to a few private offices and, while they liked the concept, the execution wasn’t quite right. It took about six months — and a major reset — to transform it into something people actually wanted to use. At one point, we had to take a step back and admit we hadn’t fully understood what users needed, leading us to rethink and rebuild large parts of the tool.

Now, Redbox is thriving, particularly in the Cabinet Office, where it has spread organically through word of mouth. Its net promoter score (NPS) is about 57 — comparable to Apple (61) and far ahead of Microsoft (38) — meaning that people are recommending it strongly to colleagues. That shift from an underwhelming prototype to a widely adopted, high-scoring tool, is a testament to the importance of listening to users and being willing to pivot when necessary.

Q: What’s the biggest lesson you’ve learned when rolling out AI-enabled services?

The most important lesson is understanding how people will actually use the service — and caring about that. If you don’t design for real-world workflows, adoption won’t happen. A great example of this is our early work on AI for prescriptions.

In the UK, studies suggest that up to 22,000 people die each year due to harmful prescription interactions. As people age, they tend to accumulate more medications, sometimes up to 19 prescriptions at once — creating serious risks. While existing systems prevent GPs from prescribing two directly conflicting drugs at the same time, they don’t flag dangerous interactions when adding new prescriptions to an existing list. The result is not only harm to patients but also an estimated £1 billion in costs, with a quarter of that coming from wasted medicine and the rest from treating adverse effects.

Our first instinct was to offer GPs a tool to prevent these harmful interactions. But when we approached them, their response was clear: “Please go away.” They were already overwhelmed with pop-ups and alerts — adding another interruption wasn’t an option.

Instead, we turned to pharmacists, who were much better suited for this kind of intervention. They already conduct structured medication reviews, and our AI could help by automating parts of that process. It could triage patients, prioritising those at highest risk, and suggest ways to optimise prescriptions safely and efficiently. For pharmacists, this wasn’t just another alert — it was a valuable tool that fits naturally into their workflow. The solution is currently in testing and evaluation.

The takeaway? You have to find the right users for your product. If the intended audience won’t — or can’t — use it, the technology doesn’t matter. Success comes from understanding user needs, fitting into their existing processes and ensuring the tool provides meaningful value.

Q: How do you think one should approach balancing the push for AI-driven solutions with the need to ensure that they're ethical and equitable?

We often overlook the reason that AI products are not ethical — unless you're a marketing giant or social media platform, arguably most unethical AI is not on purpose. No one sits down and thinks, for example, "You know what we're going to do is we're going to make an AI solution that checks your passport photo, but we'll make sure that it rejects black women twice as often as everybody else." It's an accidental professional failing when that happens not a deliberate attempt to make unfair technology.

When people have (for example in the USA) built parole decision algorithms where the outcome was to essentially to keep all the black people in prison and let all the white people out on probation, they didn't sit down and consciously do that. They fundamentally failed to understand the statistics. They failed to understand biases in the data and they failed to build the tool that they should have built. When it comes to the ethics sort of thing for me the key thing is ensuring that you are hiring very well-qualified, suitably experienced professional experts who understand the statistics and the data and care about outcomes. You need people who understand how to build the tool and test it for bias, a system in place that you can red team and permit external people to try it and see if they can make it become biased or not. Then you sort of have an iterative process to fix that and you have statistical measures to make sure that you are not producing a biased outcome.

If you don't do all of those things, you're not fulfilling the basic function of your job. For any single product you build, you need to have those checks in place. And I think it's professional standards that are the most significant problem as far as that goes.

When it comes to the way we choose to build tools though, certainly the current raft of generative AI is still very new. There will be job losses and interestingly we're seeing them in places that we didn't expect. The creative industry is obviously very early in government. One of our very first use cases was actually to replace analysts who do government consultations. It wasone of the first tools that we built and it does save money that you can then spend on cancer treatments.

We're certainly not targeting job loss. We are not building tools for the most part that are designed to replace people and we're never trying to replace somebody's human interaction with government. We don't think AI should make decisions for people. We do think AI can get the right decision to the right person much more quickly. We should be using AI to make government more human. Your interaction with government should be more human, not less.

One of my favourite examples of this is in the Department for Work and Pensions, where people write letters in. The people that write physical letters, not emails, are probably not that tech enabled. Historically, it could take 50 weeks for someone to read that letter, and the people that write in — they are probably vulnerable. As a result, some of them may not even survive 50 weeks, so by the time you read the letter, they're not there anymore. What DWP have done brilliantly is get an AI to read all the letters immediately and do a sentiment analysis where it figures out if they sound desperate and vulnerable. If they are, someone will phone them. That's a great way to use AI to make government much more human.

When we're building these algorithms, we are trying to pick the right use cases, trying to replace paperwork or speed up decisions or give people better information on which to decide — not trying to replace human contact. Nobody really wants to replace human decision-making — we're not at that stage in society or in tech readiness where that's an appropriate thing to do most of the time, but we do also need to realise that government is very expensive. The health service is incredibly expensive. We don't have enough taxes to pay for all the things we need to do.

If the result of really good implementation of AI is fewer civil servants, that is better for the country to some extent. There is something where it will replace jobs and hopefully we'll be able to reskill people and put them into sectors where there’s massive growth from AI. It will be economically a plus side, but there is a threat there to jobs realistically and particularly jobs like mine — software engineering and data science are actually really automatable in large parts!

Q: How critical are skills to the successful adoption of AI in government?

Having the right skills is essential. Without them, we risk building tools that exacerbate inequality, produce poor outcomes and even endanger people’s well-being — completely counter to what AI should be used for.

It’s a tricky challenge because there aren’t many highly technical professionals in the civil service, and hiring them can be difficult for generalists. A CV might be filled with the right buzzwords, but that doesn’t guarantee competence. That’s why high-quality recruitment is so important. In our team, we use a technical test — candidates get four hours, some data and a problem to solve. You need professionals to hire professionals.

Upskilling is just as crucial. Public servants need basic AI literacy, from prompt engineering to understanding what tools are available. Without that, AI adoption won’t scale effectively.

Procurement is another major factor. The government spends billions on technology, but if there aren’t technical professionals in the room, departments risk overpaying for solutions that don’t meet their needs. Sometimes, people don’t even realise they’re asking for the wrong thing because it’s not their area of expertise.

To save money, deliver AI well and avoid algorithmic failures, we need a highly skilled workforce. The AI Action Plan is focused on bringing in those professionals and ensuring government is equipped to use AI responsibly and effectively.

Q: What advice would you offer public sector leaders facing pushback or anxiety around AI adoption?

This is a classic leadership challenge, and the key is to start with the problem, not the technology.

When introducing an AI solution, I never begin with, “We’re using AI to improve prescriptions.” That doesn’t resonate. Instead, I say:

"Did you know that up to 22,000 people die every year due to harmful prescription interactions? And that this costs the NHS £1 billion — money that could be spent on cancer treatment? We think we have a solution. We’re testing it with patients, and here’s how it works."

Framing AI in terms of real-world impact makes it more relatable and urgent. If people don’t see a problem, they won’t see the need for a solution.

Engage early adopters first. The diffusion of innovation theory suggests that about 16% of people are innovators or early adopters — target them. Get them excited, give them incentives to share their experiences and let enthusiasm spread organically. Trying to win over sceptics too soon can be a waste of time.

Make training hands-on. Watching a video or reading about AI won’t change behaviour. People need to try the tool themselves, in a practical setting, with guidance. If they still resist after that, it might not be the right tool — and that’s valuable feedback. AI adoption isn’t about forcing change; it’s about finding the right solutions that genuinely improve people’s work.

And at the end of the day remember why we all do public sector work — to make people's lives better.