[00:00] James Kuht: There's a real moment here to think differently about how not just how we're training, you know, our generalist cadre and our existing senior leadership, but actually how do you take technical experts and teach them the skills of leadership and decision-making, and generalise those as well so that you get—and even bring in from outside—a cadre of senior leaders who, when they say, "We are committing to this," they know what the risks are. It's training, but it's critically getting people across that imagination gap.

And if you can just do those two things, you know, as a starting point, you can have really clear senior, you know, sort of buy-in and vision on what we're trying to achieve here. Is this about productivity gains or is it about achieving some big goal around, you know, increasing the throughput of planning applications or whatever it might be? And then you can empower people with the right tools to actually achieve that, and the right enablement to help them get over the hump of, "Oh my god, I've suddenly got unlimited access to PhD-level intelligence in almost every subject that occasionally lies." You know, if you can get those two things right, I think we generally see in government that the magic can really start to happen.

[01:14] Joe Owen: Hi, welcome to The Policy Fix by Nesta, the research and innovation foundation. I'm Joe Owen. Every episode, we take a policy problem and try and identify ways to fix it. If you enjoy this episode, please do like and subscribe wherever you get your podcasts, and share with the policy nerds in your life.

Artificial intelligence looks set to be one of the most important technological advances in our lifetime. That is what the AI labs think, it's what the investors think, it's what a lot of the companies think, and it's definitely what a lot of governments think. For the UK government, we've had AI action plans, we've had AI safety summits, and we've got huge projections of potential savings for the way that government works. But what is holding us back? What do we need to do in order to take advantage of the opportunities? And in particular, what needs to happen to make the most of AI for public services?

That's the question that we are going to discuss today. Joining me are two people who live in the space between big ideas and delivery. We've got James Kuht, who was one of the founding members of the Prime Minister's Data Science team in Number 10 Downing Street, has also worked in cyber in government, and has now founded a company called Pair, which helps deliver AI-native workforces. And we've got Mallory Durran, who, like James, also spent time in Number 10 working in the PM's Data Science team, was also leading the Incubator for AI team in Cabinet Office, and now runs Nesta's Applied Research and Methods practice.

James, Mallerie, welcome. So to start us off, I want to know where you guys sit on the AI hype spectrum. There's obviously loads of noise about AI, from the sort of "it will drive double-digit growth" to possibly "the end of the world", or maybe "it's all just a sort of bubble waiting to burst." So I want to ask you where you guys sit on the sort of two imaginary axes of: AI good or bad, AI big or small. Mallerie?

[03:06] Mallory Durran: Yeah, of course. Like a true data scientist, there's a huge range of uncertainty, and it would be unwise—and anyone who says they're sure of one thing or the other isn't considering the options enough. That being said, I think it is definitely true that AI will drive economic growth at at least a moderate scale. I don't think that it is certain that it will drive economic growth for everyone, or in a way that is distributed across the population that might drive the kind of social and individual benefits we'd want to see. And similarly, I think generally public value creation, making public services better, you know, better citizen experiences, are not guaranteed because of really good general-purpose technology. So I'm pretty sure the private sector is going to benefit. I'm pretty sure that a whole lot of people are going to get really rich, but a whole lot of people is not everyone. And I think there's a lot of really hard work to do to make the public benefit more certain and bigger.

[04:05] Joe Owen: James, what do you think?

[04:13] James Kuht: Yeah, if I was going to be brave, perhaps overly brave, I'd go big, but it will be slower than Silicon Valley would have us believe. And I'm going to be optimistic and hope for good on the second axis. But, you know, acknowledging Mallerie's point, there's a huge degree of uncertainty. But the key thing is we have a degree of agency as to where we end up on these spectrums. I think the worst position to take on these axes is to just pick your position and wait for this to happen to us. I think as leaders and as listeners to this podcast, we all have a role to play in making sure that we end up as far up and to the right on the good, and well-distributed, and big—and also not leading to rampant inequality—as we can. I think that agency point's key.

[05:08] Joe Owen: I want to drill down into the question around the public sector and public services, and we'll get into it in much more detail. But to keep us at the general vibe-check level as for the first question, what are your views on the genuine size of the prize for public services? Having been inside the machine—you've worked in government, you know what it takes to get stuff done, you've got a sense of how quickly government moves—how big do you think the prize is for the public sector, and how confident are you in the ability to seize it at a top level? James, I'll come to you first on this one and then Mallerie.

[05:46] James Kuht: Well, we're both somewhat biased here because obviously both of us have worked in Number 10 in the Data Science team, right? So let's just caveat that first. We're obviously both pro-AI and think it's a very exciting technology, and we've also been very fortunate to have a front-row seat to seeing data science really making its way into the centre of government. Actually, we all crossed over in that sort of time in Number 10.

So I'm very optimistic, and to answer your question: what's the size of the prize? I think it's enormous. Just to bring it down to something very tactical and precise, now we can already see some departments making significant gains with AI right now. Unfortunately, as Mallerie says, not all departments are doing the same. But I think if you look at departments—and perhaps a front-runner at the minute that's in the press a reasonable amount might be the Ministry of Justice—very vocal about AI adoption, very much driving forward with it across their workforce and starting to see some of the results of that. I think that is just a small slice of what the potential is, but it shows that government departments can move fast even if they've got 100,000 people. So yeah, that'd be how I'd set it out.

[07:05] Joe Owen: And I presume you agree on the scale of the prize, Mallerie. How optimistic or pessimistic are you about the government's ability to grab it?

[07:10] Mallory Durran: Yeah, definitely agree on the hype, like this could change everything. And I think it's right for people to be enthusiastic, loud, shout about what we could be doing here. I think the hard work sits in actually making that happen. The Ministry of Justice is such a good example, where they were also front-runners in general digital transformation. So their cloud analytics services are the envy of most departments, and they really figured out how to do that digital and cloud delivery well, well before the rest of government was really there from a maturity perspective. So they've gone from strength to strength on that.

I think whether or not a department is capable of grabbing and harnessing the AI opportunity sort of rides or dies on that general baseline capability. So part of what I see now as the size of the prize on AI is sometimes about more like base automation, using data really well, better technology services. So I think this is yet another opportunity on the road to go back and fix some of the things we maybe didn't grab in other digital transformations. So it's not just what the capabilities specific to AI can do for departments, it's also about the stuff we have missed in good digital infrastructure, joined-up data, getting the governance around that right—all the boring stuff that people glaze over on, but that are absolutely critical enablers to safe, secure, reliable ability to fundamentally change citizens' experiences of public services.

[08:47] Joe Owen: You mentioned MoJ, both of you. To move us into the more specific then, from the general views across government, I wanted to come back to you, Mallerie, and get your take on a good, specific recent example of where AI has been used well in government. Because we can project out to the future, but this stuff is happening now, people are doing it now. What are the stories of those, and what does it tell us about the potential right now?

[09:16] Mallory Durran: Yeah, because my passion is digital infrastructure, I maybe want to call out some work that MHCLG has done with the Incubator for AI team in Cabinet Office. They started thinking about: how do we achieve the big priorities for the country? One of those is housing—making sure that there is good quality housing at the right scale for the growth that we see.

Actually, the first layer of getting there—speeding up planning processes, making it better and faster to build, but also to extend—is actually about the data layer. Teams worked together on an AI product now known as Extract, that takes all of the information about planning restrictions that might be in place for a very specific geography, which, believe it or not, sit on sheets of paper in lots of cases! You know that there are so many of them in a particular local area that they go floors and floors deep in some of the old bomb shelters and bunkers. But we've got handwritten notes from 1843 about endangered newts on a particular parcel of land, and that's the level of depth that a local planning officer has to go to to understand whether they can tick a box for a new development to happen.

If we constantly keep sending people down the staircase to find the information, we're never going to speed these things up. That has an impact not just on the average homeowner trying to put in an extra bedroom, but on our ability to expand housing at the right rate. So they've really started at that base level of: can we cut down on the 50 years it would take in human time to go through and digitise that information, and use AI capability to extract that up so that then you have open, standard, available data not just for local authorities and planning services, but also so that the sector can innovate? There are some really good pieces of software, companies that have done automation or streamlining of application processes—the private sector can benefit from this as well. So we can fuel innovation by getting that data layer right.

It's a really good example of a major public sector priority that will change citizens' lives if we have enough good quality housing across the country. A really good example of a niche, deep application of AI that makes a system better to fuel the private sector and economic growth as well.

[11:52] Joe Owen: James, I want to ask you about barriers. What do you think are the big barriers for adoption in public services and government? Is it a question of vision and ambition? Is it some of the hard yards on data and digital infrastructure as Mallerie was pointing out? Is it skills? What do you see as the biggest challenges for adoption in government and public services?

[12:11] James Kuht: It's obviously a mix, but I'll pick a couple where I think there's quite tangible examples of what best practice looks like. I think one of the reasons that MoJ and MHCLG and others have succeeded is where they've had a clear vision of the ultimate goal they're trying to support—whether that be improving citizen services, increasing the number of planning applications approved, or whatever it might be. They've been very clear about the goal, and senior leadership have been clear how they're going to invest in AI to achieve that goal. So I think that's really important: to have a senior leader who is clear about the goal they're trying to drive towards, and then empowers people to get on with it and invests in that.

Once you've got a clear vision and goal, and you've got buy-in from top-down, it can come down to things as simple as: do people have access to the right AI tools to actually achieve this? Do they have a Co-pilot licence, or a ChatGPT licence, or a Gemini licence, or whatever is relevant for them to do the task they've been asked of? And are they enabled with it? Is it just slung over the fence and then they've got this licence and they're sort of stuck looking at it like, "Hey, what do I do? It can do my meeting transcription, what else do I do with it?" Or are they properly enabled with this?

The enablement question is obviously multifaceted. Part of it is ideally seeing your senior leader model the behaviour they want to see—seeing their senior leadership using AI, making mistakes with AI, and showing the level of risk tolerance they're willing to accept or not accept. Part of it's the senior leaders, part of it is a training question.

Just to be really specific on this: this is not teaching people to write in the box and click send. Everyone knows how to use chat logs now. It's addressing the imagination gap between, "I know how to do a search with ChatGPT, Co-pilot, or Gemini" (or insert your tool here), and "I actually know how to transform elements of my job with AI." So it's training, but it's critically getting people across that imagination gap.

If you can just do those two things as a starting point—you can have really clear senior buy-in and vision on what we're trying to achieve here (is this about productivity gains, or is it about achieving some big goal around increasing the throughput of planning applications, whatever it might be?), and then you can empower people with the right tools to actually achieve that, and the right enablement to help them get over the hump of, "Oh my god, I've suddenly got unlimited access to PhD-level intelligence in almost every subject that occasionally lies"—if you can get those two things right, I think we generally see in government that the magic can really start to happen. Those would be the two barriers I'd say most commonly happen.

[15:00] Joe Owen: The leadership one is a really interesting point you made. I was thinking about this in the context of: we've got a pretty new Cabinet Secretary, there's been quite a few turnovers of Permanent Secretaries, and if that was happening in private sector organisations of the same size as departments or the civil service, you would imagine those new leaders in an organisation of that size, with that level of complexity, to be obsessed about technology. But we sort of have a system in government by which for Permanent Secretaries or policy professionals, the tech side of government can seem at arm's length and not core. But what I'm hearing from this is if you want to gain the benefits for public services and for policy from AI, that sort of weird arm's-length approach needs to end.

[15:47] Mallory Durran: Maybe a more uncomfortable message is that I think this is a different era of senior leadership. People are endlessly capable of changing how they work and learning new things, but when James talked about that senior buy-in and real commitment to doing it, including taking risks, I think it's a much bigger ask for a senior leader who is not a technology native, who might not be even sure what the risks are. So that when the first bad thing goes wrong, you can really see a loss of confidence or a wobbling. If you're sort of tangoing "we're doing it, oh wait hold back, oh we're doing it, oh wait hold back", you start to get hesitation that trickles down the rest of the system.

So I think there's a real moment here to think differently about not just how we're training our generalist cadre and our existing senior leadership, but actually how do you take technical experts and teach them the skills of leadership and decision-making, and generalise those as well so that you get—and even bring in from outside—a cadre of senior leaders who, when they say, "We are committing to this," they know what the risks are, and they know what it's going to feel like to own when things don't work out, or to invest in something that doesn't come to the promise of the potential growth or opportunity that you wanted it to create. I think that that is absolutely critical to actually harnessing the opportunity.

[17:19] Joe Owen: There's loads in that that I want to come back to. But the risk tolerance thing in particular—where it seems to me, to go back to where we started, which is reasonable confidence that the private sector will do well out of AI, less confident on the public sector—I see a lot of that as a result of risk tolerance. In that there's huge incentives obviously if you're a private company to take a risk to try and get a first-mover advantage, or at the very least definitely not get left behind. Whereas in government, the approach is: how do you fully understand the risk, how do you minimise the risk? A lot of this is imagined, because government's not very good at necessarily doing this all the time, but how do we try and push this risk as close to zero as we possibly can? The nature of this technology, and what we know about it and what we don't know about it, means that that's just not possible.

So how should civil servants listening to this, or senior leaders in the civil service, start to think about taking good risk and calculated risk? I might start on this one, and Mallerie, I'd be fascinated to hear your thoughts.

[18:31] James Kuht: I think it's really important to distinguish real risks here from perceived risks, because I think you're absolutely right: there are some things which are just harder and more risky to do in the public sector. I remember speaking to Tom Read, the former CEO of Government Digital Service a while ago, and he was like, "All of our services need to cater for absolutely everyone." It's a very famous startup mentality that you should fire some of your customers if they're not a good fit for you, fire the customers and focus on the customers who are the best fit for you and the most profitable. You can't do that in government; you have to cater for absolutely everyone. So I think it's first worth just acknowledging you can't be a total techno-maximalist here and say, "Oh, government should just take more risks," because there are some risks which are just totally intolerable to government which the private sector don't have to worry about.

Now we've got that out of the way, a lot of the chats I have with senior government leaders and civil servants, they are citing perceived risks which are not really risks. For example, I'll still very commonly speak to civil servants who will say things like, "I don't want to write something into Co-pilot because then Microsoft is training their models on the data I'm putting in and it's not secure." When of course—in Microsoft's case, and other providers are available—if you're using Co-pilot on your corporate tenant, it has the same level of cyber security as Microsoft Word in the cloud. As soon as you tackle that with them, they're like, "Oh my god, massive unlock."

So if we could just solve that bit—that people are actually informed of the true level of risks of using these AI tools day-to-day, that it's not some black hole where all citizens' data is going to go and be used to train evil algorithms in Silicon Valley or whatever people's perceptions might be—if we could tackle the perceived risk bit, I think we could have enormous unlocks. But we do still need to acknowledge that some level of risk is going to be intolerable in the public sector.

[20:36] Mallory Durran: Yeah, I don't think that this is a new problem or set of questions. We used to talk about how you get better evaluation in government all the time, and the same thing happens, right? You spend money evaluating a policy, well then you've both spent that money, and you might find out that it didn't work the way that you wanted it to. And the short-term nature of a parliamentary term makes that feel really scary. So this isn't just a technology problem, I think it's a muscle of: there is more public good to be done overall when we make mistakes and learn from them, as opposed to never do anything until we're sure it's going to work. So this isn't even a specifically AI problem, it's expectations in a political sphere and a public sector sphere.

While it's true that you can't just leave a group out because they're hard to cater for, it's got to be true that you can put investment into an application. We weren't sure at the outset of the Extract programme whether we would be able to, with sufficient fidelity, actually extract information from enough of different types of documents to do it, but it required an initial upfront investment. There was a chance that we were going to put money into that that wasn't going to result in an outcome. I think it was great bravery on the part of Directors General, Permanent Secretaries, Ministers to say, "Yeah, we're going to have to take that mentality." And if we maybe do it for AI because it feels like this big, shiny opportunity, maybe we get better at doing it in other kinds of policy areas too. But I think it's been a big barrier to policy impact in general that we're not good at failing.

[22:24] James Kuht: Can I add one more point on this one? I think this is such a rich seam. One of the best ways to manage some of these risks with AI is actually getting people to just experiment with the tools. I think two great things come from this:

One, you start to work out the type of tasks where AI is good and where it's pretty naff. That's actually really important—just that curation of where I use ChatGPT versus where I don't at the individual level. Obviously at the more strategic level you might have more top-down, exquisite tools, but that's actually really important and a very valuable skill.

The second thing is: if you invest in people's skills and you continuously reinvest in them, you also buy out this challenge where the frontier of what AI can achieve is constantly moving. So people need to evaluate and re-evaluate constantly what's within the art of the possible with AI, because chances are in three months, the frontier changes. So I just wanted to make that point: that a good way to manage risk is actually by investing in AI skills in your workforce.

[23:31] Joe Owen: You both made the point around experimentation being key. When we find something that works, what's the journey then—the scaling question? James, I want to come to you on this: how straightforward is it in your experience to scale innovation in government around technology? What breaks or what's at risk of breaking?

[23:52] James Kuht: It's a great question, and we've seen lots of things fail and some things succeed as well. It's worth acknowledging before I get into the answer to this question that it's good that there is a high bar for things to scale in government. By very definition, when you run 10 experiments of a new technology, despite everyone's motives for their pilot to succeed, there should be a survival of the fittest and only the best ones should get through.

Specific to AI, what we see when we see pilots scale is that there's a degree of IT professionalisation that often has to come in. Getting to an 80% prototype is often relatively straightforward and exciting, and everyone's like, "Oh my god, we built this agent, it does the thing!" The final 20% is pretty hard, and there's loads of stuff that goes into it—change management and stuff. But I think often the first piece is IT getting involved, working out: how do we scale this thing from an infrastructure perspective? Does it make sense from token costs? Are we going to move it onto some separate bit of infrastructure? Should we build this or should we just buy something off the shelf? I think a lot of things go into the IT bit.

Once you've worked out the feasibility of scaling this agent or this new AI workflow, or building AI into a new part of your planning process, it's then a question of change management. How do you actually embed this in a way which actually gets used by people and doesn't just sit on the shelf as something shiny, and actually gets evaluated on whether it has actually shifted the dial on the metric you're aiming for? To bring it back to senior leaders: it being crucially important you set the goals that are measurable that you're aiming for.

[25:49] Mallory Durran: I might jump off that evaluation point and say the 80% can often feel really easy, and then when you run an evaluation, you realise that there are missing capabilities or instances in which you're not covering that are actually critical failure points. So designing that evaluation in is key.

I also think there's a structural funding problem. Government runs on these four-year Spending Reviews. What it takes for a product team to run an MVP, get a pilot going, even maybe to integrate that into some IT systems, is actually relatively different than what it takes to maintain that service, and indeed respond to the ways in which a really good piece of tech will change ways of working and then need adaptation to features.

Setting product teams up that scale is not easy within the funding structures of government. If you're a tech company, you start with a really lean team, you have the minimum viable number of people, probably working crazy hours—which also happens in government! But there comes a point where Jane, the amazing engineer working overnight a couple of times a week, just doesn't cut it anymore. The way that you get past that in the private sector is raising capital, so you get these outside funding sources and you can kind of go from there.

In government, even where you've got ring-fenced pots of transformation funding, it's actually quite hard to unlock that for headcount, and even harder to unlock it for investment in compute or base infrastructure, just the way that those funding rules are set up. So harnessing technology scale for government is probably going to require a rethink of some of those funding rules, and/or getting creative about how outside capital can fuel that, whether it's joint innovation partnerships. But just thinking differently about how you fund teams to grow as a product grows, rather than saying, "Well, we think there's probably something here at point of Spending Review or budget," and guessing that you need 70 people a year from now. Do you contract that in? Do you just bring it in? But then you've got people sitting there not ready. It's a really weird funding picture, and one that getting to scale consistently is going to require some serious thought on.

[28:07] Joe Owen: Mallerie, you were talking about raising funds in order to bring people in. There's a question about whether one of the perpetual things is: can the government compete with the private sector for the skills and the talent that it needs in order to drive the transformation that you're talking about? Can the Cabinet Office really compete with the compensation packages that people could get elsewhere? You've worked in government with these teams, what is your take on that question?

[28:32] Mallory Durran: Yeah, so I think there have been successes, right? James worked there, absolutely nailing it. But yeah, there are barriers, and I think it's really tricky to look at a group of incredible professionals and say, "There are some other people who aren't you whose skills are more valuable." The reality is it's a market valuation, right? Maybe AI coding agent assistants will get so good that we won't need software developers at some point in the future—although objectively, software developers prompt AI for better software, but a different tale for a different day!

So I think there have been real successes in pushing beyond the bounds of typical public sector contract and compensation packages to get there. There are heaps of people with incredible technical skill who also—maybe it's like this elsewhere, but in the UK in particular—have just an unbelievable public-spirited mindset, and have taken pay cuts of more than 50% to come in and say, "I'm going to use the best skills I have at this once-in-a-generation opportunity." So I think it's possible.

I think the AI Safety Institute has done some really good things to get top-line researchers in, because it is a genuine national and international priority to do that. But I think there's a retention thing: how long is someone willing to keep turning down daily, weekly offers from new AI companies or the frontier labs? I think it's about compensation, but it's also about giving those people an environment in which they actually feel they're using their skills to best interest. So when we talk about having leadership that are willing to commit and take on the genuine risks of delivery, and invest—some of that sort of "sludge" that gets in the way of delivery is just as important to talent retention. If you're going to bring the people in, getting the most out of them actually means giving them the space to run and do their best work as well.

[30:31] James Kuht: Yeah, I'd back that up. I think the value proposition of working in government is just fundamentally different to working in the private sector for a lot of people. Mallerie's teams have done this well in the past, as well as the teams I've worked for: they lean into that. They really lean into the mission, which sounds a little bit trite, but honestly, the access you get in government and the levers you can pull—whether that's in one of the frontline departments or more centrally—they're extraordinary. You chat to your friends in the private sector, and I don't really look at many of them with any degree of envy. I'm now in the private sector socially, but I think the mission is fab, and the learning rate you can have as well in career progression I think is fab too.

So I think we need to lean into that value proposition. When it comes to technical skills, the real deep technical experts, I think I'd just be at pains to remind people that's the exception, not the norm. These people who are getting hired, prepared to go off to frontier labs and things, is probably a unique situation to just some of these real front-end AI teams, like the AI Safety Institute. This isn't for your average software engineer in DWP, and that's absolutely no slight on them, but they're not getting calls from Anthropic to go for £200k. So I think we should just be really careful that we don't overstate the salary piece. We've got a fundamentally different value proposition. The salaries are generally pretty good with the Digital, Data and Technology (DDaT) pay framework that came in, and there's just this exception pool for exquisite AI talent. The places Mallerie's mainly played in I think is an important exception, but it's not the norm.

[32:10] Joe Owen: One of the ways the government has managed the talent thing in the past is just use outsourcing, or contractors, or suppliers. Is there a risk that we sort of reheat that, whether it's to do with PFI initiatives or outsourcing initiatives, we take the same approach with AI and start relying on a small number of firms to do the government's work for it?

[32:33] Mallory Durran: Yeah, I mean, I think one of the downsides of that is that you don't build capability in-house. I think a certain amount of relying on outside skill is really intelligent, especially when it comes to the foundational work to set systems up to succeed, and to move existing workforces into a level of comfort to be able to operate themselves. So I think when you get these sort of insource-outsource combinations, or more partnership working between contractors or the private sector and government teams, you get the best of both worlds: people who really understand how services are being delivered, the challenges that live in that space, and the technical skill to achieve it.

Outsourcing in a "we're just going to throw this over the wall, go deliver it, we'll be stuck in this contract forever" way is a travesty. But actually really creatively setting up partnerships that leave behind capability, that bring the workforce along with them, and that then have access to the nuts and bolts of how things are really working so that those technical skills can be applied differently in a way that really creates a solution—I think it's not that outsourcing is or isn't the answer, but thinking about how you do that in an integrated way that leaves a legacy behind is really important.

[33:57] Joe Owen: So I am going to ask for the last few minutes or so a few more quickfire-style questions. The first one is: what are the two or three things that you would be looking for over the next year or so to show that the government is really serious about AI adoption? What are the signals through the noise where you're like, "Right, this government is really getting on with it"? James, I'm going to pick on you first.

[34:23] James Kuht: I'd love to see them clearly articulate the measurable goals that they are trying to drive with AI adoption. I think it's all very well and good—I think they've made some AI training freely available to 10 or 12 million people in the UK with a load of partners, and obviously there's a big push internally too, and that's great as a starting point. But unless it's actually tied to outcomes, we're never going to get the chance to see whether this has actually been impactful or not, and whether it needs more investment or a different approach or not.

I think this is just so important. It's absolutely undeniable that AI adoption across government and outside in the UK can drive massive productivity gains, but also allow us to do fundamentally more valuable tasks, and also make work more fulfilling. I worry at the minute that government is throwing over the fence access to these AI training courses, not measuring any outcomes whatsoever, and is going to wonder why we don't get the benefits of AI across the spread of the population when we made this training available to 12 million people. That would be my one.

[35:39] Joe Owen: Mallerie?

[35:45] Mallory Durran: I might go the opposite end of the spectrum and think about a real laser focus and commitment on a couple of really big bets, probably in some of the places where it's hardest or riskiest to achieve change, where you probably have lots of different kinds of products available being procured in different ways, but where the state—and only the state—has the authority and responsibility to ensure that that really delivers well.

So unified approaches to AI for unlocking healthcare outcomes; how AI will change and shape and equalise access to education or educational attainment. Actually, it's not going to be a one-off six-month programme bit of pre-training and off you go—there's going to be repeated testing and failures and things not meeting the mark. But I think real commitment to some of the places where the impact for the average citizen is going to be outsized. Actually just willingness to take those on and say, "Whatever it takes, we are going to achieve progress in those areas." I think there's been some good signals to that with leadership positions both in Number 10 and departments trying to create that focus, but I think it's about sustained commitment, both financially and otherwise, to big bets.

[37:00] Joe Owen: Okay, we've got big bets, we've got some targets and some ambition. What's one very practical and unglamorous reform—a sort of plumbing fix—that you would be looking for? Is it something to do with procurement, data sharing? Is there a particular thing on your wish list, Mallerie?

[37:20] Mallory Durran: I think I hit a few with how we do external funding of things, who leaders are and how we train them, but probably the thing I will spend the rest of my life on is better digital public infrastructure. The ways in which we set standards for how data is structured and held, and the technology that allows linkage and use of that data. It's a real barrier already to great use cases and AI applications that are being developed by civil servants, and is also a barrier to private sector innovation using great public sector data. It's got to be secure, it's got to work for everyone, can't be locked behind big, shiny, expensive walls—so really, implementation of open standards and technology.

[38:11] Joe Owen: Do you have an unglamorous reform, James?

[38:17] James Kuht: It's not a reform as such, but I think it comes back to the funding question. I think people have to be really clear on what's going to be replaced by the thing they're doing, because the civil service can't just keep growing—things have got to give. So I think people have got to be really clear why their AI initiative, and their top-down or bottom-up AI programme, is going to deliver better outcomes for the citizen or increase productivity, and what is going to be turned off to make space.

Obviously the Nudge Unit bit, now part of Nesta, is a great example of this: there should be sunset clauses around this sort of stuff where if things aren't working—which statistically, lots of things end up not working way more than we like to think—things should be turned off and we should move on to the next thing. We're in this incredible era of experimentation where we're all working this out, and we need to find space to do that, probably at the expense of other things, because getting private capital in is going to be hard. If things don't work, we should move on to the next thing.

[39:20] Joe Owen: And then my final question: if there's anyone listening to this who works in a government department or a local authority who thinks all of this sounds slightly intimidating, maybe exciting, what's the one thing that I could do tomorrow that would move me up the AI curve? What would your advice be, James first and then Mallerie?

[39:40] James Kuht: This has got a little bit of a bar because you're going to have to pay £20 for it if you don't already have it: I would have a go at trying to build an agent which automates a significant part of your job. That might require a ChatGPT licence or a Claude licence or whatever, and try and build an agent to do it. It's not too difficult. You need to use Claude Cowork Enterprise and do it within the bounds of data protection and cyber security, but I think in just doing that exercise for a few hours and building an agentic workflow, you learn so much stuff:

One, you learn exactly what the models and the tools are capable of, and you suddenly become a lot better at cutting through the hype.

Two, you become much better at procuring these sort of solutions because you know what you can do out of the box, so you're less prone to being sold snake oil, which is a real danger in government.

And three, you crucially work out what the things can't do, and therefore what's squarely in the realm of being a uniquely human skill, or something that you're going to have to keep testing with the frontier.

So yeah, take half a day, you'll have a ton of fun doing it, get some sort of agentic solution if you don't have it as part of your corporate IT already, and be really careful with data—Claude Cowork I suggest or ChatGPT—and try and build an agent which automates something really hard that you do. You might not succeed, but you will definitely learn a lot in doing that. There's nothing that beats doing stuff.

[41:12] Mallory Durran: So I would say if you heard that and you're a bit overwhelmed by it, you're like, "Well, what if I mess something up?" Or if your organisation doesn't have policies in place that allow you to access that stuff, actually I think you can gain some of the same learning, and then the bravery to go and ask for that or bring it into your work, if you just pick a problem that you really understand.

I spend heaps of time on trains, for example. Train times change, platforms change, there's loads of different places you can go to to find those times. I've set up an agent that at the times of day I usually take a train sends me a push notification to do it. Pick a problem in your life that you really understand and set something up that works for that—whether it's an agent or an automated workflow. The great thing about these AI tools is if you don't know where to start, ask them! Put that chat in. So what James said, if that felt overwhelming, just pick a problem that you really understand and do something with it.

[42:14] Joe Owen: James, Mallerie, thank you so much.

[42:16] Mallory Durran: Thanks for having us.

[42:17] Joe Owen: If you enjoyed this episode, please do like, share and subscribe wherever you get your podcasts. As a reminder, Nesta is a research and innovation foundation, and we design, test and scale solutions to society's biggest challenges. We are funded by a charitable endowment, and we are politically neutral. If you would like any more information, please go to nesta.org.uk.

[00:00] James Kuht: There's a real moment here to think differently about how not just how we're training, you know, our generalist cadre and our existing senior leadership, but actually how do you take technical experts and teach them the skills of leadership and decision-making, and generalise those as well so that you get—and even bring in from outside—a cadre of senior leaders who, when they say, "We are committing to this," they know what the risks are. It's training, but it's critically getting people across that imagination gap.

And if you can just do those two things, you know, as a starting point, you can have really clear senior, you know, sort of buy-in and vision on what we're trying to achieve here. Is this about productivity gains or is it about achieving some big goal around, you know, increasing the throughput of planning applications or whatever it might be? And then you can empower people with the right tools to actually achieve that, and the right enablement to help them get over the hump of, "Oh my god, I've suddenly got unlimited access to PhD-level intelligence in almost every subject that occasionally lies." You know, if you can get those two things right, I think we generally see in government that the magic can really start to happen.

[01:14] Joe Owen: Hi, welcome to The Policy Fix by Nesta, the research and innovation foundation. I'm Joe Owen. Every episode, we take a policy problem and try and identify ways to fix it. If you enjoy this episode, please do like and subscribe wherever you get your podcasts, and share with the policy nerds in your life.

Artificial intelligence looks set to be one of the most important technological advances in our lifetime. That is what the AI labs think, it's what the investors think, it's what a lot of the companies think, and it's definitely what a lot of governments think. For the UK government, we've had AI action plans, we've had AI safety summits, and we've got huge projections of potential savings for the way that government works. But what is holding us back? What do we need to do in order to take advantage of the opportunities? And in particular, what needs to happen to make the most of AI for public services?

That's the question that we are going to discuss today. Joining me are two people who live in the space between big ideas and delivery. We've got James Kuht, who was one of the founding members of the Prime Minister's Data Science team in Number 10 Downing Street, has also worked in cyber in government, and has now founded a company called Pair, which helps deliver AI-native workforces. And we've got Mallory Durran, who, like James, also spent time in Number 10 working in the PM's Data Science team, was also leading the Incubator for AI team in Cabinet Office, and now runs Nesta's Applied Research and Methods practice.

James, Mallerie, welcome. So to start us off, I want to know where you guys sit on the AI hype spectrum. There's obviously loads of noise about AI, from the sort of "it will drive double-digit growth" to possibly "the end of the world", or maybe "it's all just a sort of bubble waiting to burst." So I want to ask you where you guys sit on the sort of two imaginary axes of: AI good or bad, AI big or small. Mallerie?

[03:06] Mallory Durran: Yeah, of course. Like a true data scientist, there's a huge range of uncertainty, and it would be unwise—and anyone who says they're sure of one thing or the other isn't considering the options enough. That being said, I think it is definitely true that AI will drive economic growth at at least a moderate scale. I don't think that it is certain that it will drive economic growth for everyone, or in a way that is distributed across the population that might drive the kind of social and individual benefits we'd want to see. And similarly, I think generally public value creation, making public services better, you know, better citizen experiences, are not guaranteed because of really good general-purpose technology. So I'm pretty sure the private sector is going to benefit. I'm pretty sure that a whole lot of people are going to get really rich, but a whole lot of people is not everyone. And I think there's a lot of really hard work to do to make the public benefit more certain and bigger.

[04:05] Joe Owen: James, what do you think?

[04:13] James Kuht: Yeah, if I was going to be brave, perhaps overly brave, I'd go big, but it will be slower than Silicon Valley would have us believe. And I'm going to be optimistic and hope for good on the second axis. But, you know, acknowledging Mallerie's point, there's a huge degree of uncertainty. But the key thing is we have a degree of agency as to where we end up on these spectrums. I think the worst position to take on these axes is to just pick your position and wait for this to happen to us. I think as leaders and as listeners to this podcast, we all have a role to play in making sure that we end up as far up and to the right on the good, and well-distributed, and big—and also not leading to rampant inequality—as we can. I think that agency point's key.

[05:08] Joe Owen: I want to drill down into the question around the public sector and public services, and we'll get into it in much more detail. But to keep us at the general vibe-check level as for the first question, what are your views on the genuine size of the prize for public services? Having been inside the machine—you've worked in government, you know what it takes to get stuff done, you've got a sense of how quickly government moves—how big do you think the prize is for the public sector, and how confident are you in the ability to seize it at a top level? James, I'll come to you first on this one and then Mallerie.

[05:46] James Kuht: Well, we're both somewhat biased here because obviously both of us have worked in Number 10 in the Data Science team, right? So let's just caveat that first. We're obviously both pro-AI and think it's a very exciting technology, and we've also been very fortunate to have a front-row seat to seeing data science really making its way into the centre of government. Actually, we all crossed over in that sort of time in Number 10.

So I'm very optimistic, and to answer your question: what's the size of the prize? I think it's enormous. Just to bring it down to something very tactical and precise, now we can already see some departments making significant gains with AI right now. Unfortunately, as Mallerie says, not all departments are doing the same. But I think if you look at departments—and perhaps a front-runner at the minute that's in the press a reasonable amount might be the Ministry of Justice—very vocal about AI adoption, very much driving forward with it across their workforce and starting to see some of the results of that. I think that is just a small slice of what the potential is, but it shows that government departments can move fast even if they've got 100,000 people. So yeah, that'd be how I'd set it out.

[07:05] Joe Owen: And I presume you agree on the scale of the prize, Mallerie. How optimistic or pessimistic are you about the government's ability to grab it?

[07:10] Mallory Durran: Yeah, definitely agree on the hype, like this could change everything. And I think it's right for people to be enthusiastic, loud, shout about what we could be doing here. I think the hard work sits in actually making that happen. The Ministry of Justice is such a good example, where they were also front-runners in general digital transformation. So their cloud analytics services are the envy of most departments, and they really figured out how to do that digital and cloud delivery well, well before the rest of government was really there from a maturity perspective. So they've gone from strength to strength on that.

I think whether or not a department is capable of grabbing and harnessing the AI opportunity sort of rides or dies on that general baseline capability. So part of what I see now as the size of the prize on AI is sometimes about more like base automation, using data really well, better technology services. So I think this is yet another opportunity on the road to go back and fix some of the things we maybe didn't grab in other digital transformations. So it's not just what the capabilities specific to AI can do for departments, it's also about the stuff we have missed in good digital infrastructure, joined-up data, getting the governance around that right—all the boring stuff that people glaze over on, but that are absolutely critical enablers to safe, secure, reliable ability to fundamentally change citizens' experiences of public services.

[08:47] Joe Owen: You mentioned MoJ, both of you. To move us into the more specific then, from the general views across government, I wanted to come back to you, Mallerie, and get your take on a good, specific recent example of where AI has been used well in government. Because we can project out to the future, but this stuff is happening now, people are doing it now. What are the stories of those, and what does it tell us about the potential right now?

[09:16] Mallory Durran: Yeah, because my passion is digital infrastructure, I maybe want to call out some work that MHCLG has done with the Incubator for AI team in Cabinet Office. They started thinking about: how do we achieve the big priorities for the country? One of those is housing—making sure that there is good quality housing at the right scale for the growth that we see.

Actually, the first layer of getting there—speeding up planning processes, making it better and faster to build, but also to extend—is actually about the data layer. Teams worked together on an AI product now known as Extract, that takes all of the information about planning restrictions that might be in place for a very specific geography, which, believe it or not, sit on sheets of paper in lots of cases! You know that there are so many of them in a particular local area that they go floors and floors deep in some of the old bomb shelters and bunkers. But we've got handwritten notes from 1843 about endangered newts on a particular parcel of land, and that's the level of depth that a local planning officer has to go to to understand whether they can tick a box for a new development to happen.

If we constantly keep sending people down the staircase to find the information, we're never going to speed these things up. That has an impact not just on the average homeowner trying to put in an extra bedroom, but on our ability to expand housing at the right rate. So they've really started at that base level of: can we cut down on the 50 years it would take in human time to go through and digitise that information, and use AI capability to extract that up so that then you have open, standard, available data not just for local authorities and planning services, but also so that the sector can innovate? There are some really good pieces of software, companies that have done automation or streamlining of application processes—the private sector can benefit from this as well. So we can fuel innovation by getting that data layer right.

It's a really good example of a major public sector priority that will change citizens' lives if we have enough good quality housing across the country. A really good example of a niche, deep application of AI that makes a system better to fuel the private sector and economic growth as well.

[11:52] Joe Owen: James, I want to ask you about barriers. What do you think are the big barriers for adoption in public services and government? Is it a question of vision and ambition? Is it some of the hard yards on data and digital infrastructure as Mallerie was pointing out? Is it skills? What do you see as the biggest challenges for adoption in government and public services?

[12:11] James Kuht: It's obviously a mix, but I'll pick a couple where I think there's quite tangible examples of what best practice looks like. I think one of the reasons that MoJ and MHCLG and others have succeeded is where they've had a clear vision of the ultimate goal they're trying to support—whether that be improving citizen services, increasing the number of planning applications approved, or whatever it might be. They've been very clear about the goal, and senior leadership have been clear how they're going to invest in AI to achieve that goal. So I think that's really important: to have a senior leader who is clear about the goal they're trying to drive towards, and then empowers people to get on with it and invests in that.

Once you've got a clear vision and goal, and you've got buy-in from top-down, it can come down to things as simple as: do people have access to the right AI tools to actually achieve this? Do they have a Co-pilot licence, or a ChatGPT licence, or a Gemini licence, or whatever is relevant for them to do the task they've been asked of? And are they enabled with it? Is it just slung over the fence and then they've got this licence and they're sort of stuck looking at it like, "Hey, what do I do? It can do my meeting transcription, what else do I do with it?" Or are they properly enabled with this?

The enablement question is obviously multifaceted. Part of it is ideally seeing your senior leader model the behaviour they want to see—seeing their senior leadership using AI, making mistakes with AI, and showing the level of risk tolerance they're willing to accept or not accept. Part of it's the senior leaders, part of it is a training question.

Just to be really specific on this: this is not teaching people to write in the box and click send. Everyone knows how to use chat logs now. It's addressing the imagination gap between, "I know how to do a search with ChatGPT, Co-pilot, or Gemini" (or insert your tool here), and "I actually know how to transform elements of my job with AI." So it's training, but it's critically getting people across that imagination gap.

If you can just do those two things as a starting point—you can have really clear senior buy-in and vision on what we're trying to achieve here (is this about productivity gains, or is it about achieving some big goal around increasing the throughput of planning applications, whatever it might be?), and then you can empower people with the right tools to actually achieve that, and the right enablement to help them get over the hump of, "Oh my god, I've suddenly got unlimited access to PhD-level intelligence in almost every subject that occasionally lies"—if you can get those two things right, I think we generally see in government that the magic can really start to happen. Those would be the two barriers I'd say most commonly happen.

[15:00] Joe Owen: The leadership one is a really interesting point you made. I was thinking about this in the context of: we've got a pretty new Cabinet Secretary, there's been quite a few turnovers of Permanent Secretaries, and if that was happening in private sector organisations of the same size as departments or the civil service, you would imagine those new leaders in an organisation of that size, with that level of complexity, to be obsessed about technology. But we sort of have a system in government by which for Permanent Secretaries or policy professionals, the tech side of government can seem at arm's length and not core. But what I'm hearing from this is if you want to gain the benefits for public services and for policy from AI, that sort of weird arm's-length approach needs to end.

[15:47] Mallory Durran: Maybe a more uncomfortable message is that I think this is a different era of senior leadership. People are endlessly capable of changing how they work and learning new things, but when James talked about that senior buy-in and real commitment to doing it, including taking risks, I think it's a much bigger ask for a senior leader who is not a technology native, who might not be even sure what the risks are. So that when the first bad thing goes wrong, you can really see a loss of confidence or a wobbling. If you're sort of tangoing "we're doing it, oh wait hold back, oh we're doing it, oh wait hold back", you start to get hesitation that trickles down the rest of the system.

So I think there's a real moment here to think differently about not just how we're training our generalist cadre and our existing senior leadership, but actually how do you take technical experts and teach them the skills of leadership and decision-making, and generalise those as well so that you get—and even bring in from outside—a cadre of senior leaders who, when they say, "We are committing to this," they know what the risks are, and they know what it's going to feel like to own when things don't work out, or to invest in something that doesn't come to the promise of the potential growth or opportunity that you wanted it to create. I think that that is absolutely critical to actually harnessing the opportunity.

[17:19] Joe Owen: There's loads in that that I want to come back to. But the risk tolerance thing in particular—where it seems to me, to go back to where we started, which is reasonable confidence that the private sector will do well out of AI, less confident on the public sector—I see a lot of that as a result of risk tolerance. In that there's huge incentives obviously if you're a private company to take a risk to try and get a first-mover advantage, or at the very least definitely not get left behind. Whereas in government, the approach is: how do you fully understand the risk, how do you minimise the risk? A lot of this is imagined, because government's not very good at necessarily doing this all the time, but how do we try and push this risk as close to zero as we possibly can? The nature of this technology, and what we know about it and what we don't know about it, means that that's just not possible.

So how should civil servants listening to this, or senior leaders in the civil service, start to think about taking good risk and calculated risk? I might start on this one, and Mallerie, I'd be fascinated to hear your thoughts.

[18:31] James Kuht: I think it's really important to distinguish real risks here from perceived risks, because I think you're absolutely right: there are some things which are just harder and more risky to do in the public sector. I remember speaking to Tom Read, the former CEO of Government Digital Service a while ago, and he was like, "All of our services need to cater for absolutely everyone." It's a very famous startup mentality that you should fire some of your customers if they're not a good fit for you, fire the customers and focus on the customers who are the best fit for you and the most profitable. You can't do that in government; you have to cater for absolutely everyone. So I think it's first worth just acknowledging you can't be a total techno-maximalist here and say, "Oh, government should just take more risks," because there are some risks which are just totally intolerable to government which the private sector don't have to worry about.

Now we've got that out of the way, a lot of the chats I have with senior government leaders and civil servants, they are citing perceived risks which are not really risks. For example, I'll still very commonly speak to civil servants who will say things like, "I don't want to write something into Co-pilot because then Microsoft is training their models on the data I'm putting in and it's not secure." When of course—in Microsoft's case, and other providers are available—if you're using Co-pilot on your corporate tenant, it has the same level of cyber security as Microsoft Word in the cloud. As soon as you tackle that with them, they're like, "Oh my god, massive unlock."

So if we could just solve that bit—that people are actually informed of the true level of risks of using these AI tools day-to-day, that it's not some black hole where all citizens' data is going to go and be used to train evil algorithms in Silicon Valley or whatever people's perceptions might be—if we could tackle the perceived risk bit, I think we could have enormous unlocks. But we do still need to acknowledge that some level of risk is going to be intolerable in the public sector.

[20:36] Mallory Durran: Yeah, I don't think that this is a new problem or set of questions. We used to talk about how you get better evaluation in government all the time, and the same thing happens, right? You spend money evaluating a policy, well then you've both spent that money, and you might find out that it didn't work the way that you wanted it to. And the short-term nature of a parliamentary term makes that feel really scary. So this isn't just a technology problem, I think it's a muscle of: there is more public good to be done overall when we make mistakes and learn from them, as opposed to never do anything until we're sure it's going to work. So this isn't even a specifically AI problem, it's expectations in a political sphere and a public sector sphere.

While it's true that you can't just leave a group out because they're hard to cater for, it's got to be true that you can put investment into an application. We weren't sure at the outset of the Extract programme whether we would be able to, with sufficient fidelity, actually extract information from enough of different types of documents to do it, but it required an initial upfront investment. There was a chance that we were going to put money into that that wasn't going to result in an outcome. I think it was great bravery on the part of Directors General, Permanent Secretaries, Ministers to say, "Yeah, we're going to have to take that mentality." And if we maybe do it for AI because it feels like this big, shiny opportunity, maybe we get better at doing it in other kinds of policy areas too. But I think it's been a big barrier to policy impact in general that we're not good at failing.

[22:24] James Kuht: Can I add one more point on this one? I think this is such a rich seam. One of the best ways to manage some of these risks with AI is actually getting people to just experiment with the tools. I think two great things come from this:

One, you start to work out the type of tasks where AI is good and where it's pretty naff. That's actually really important—just that curation of where I use ChatGPT versus where I don't at the individual level. Obviously at the more strategic level you might have more top-down, exquisite tools, but that's actually really important and a very valuable skill.

The second thing is: if you invest in people's skills and you continuously reinvest in them, you also buy out this challenge where the frontier of what AI can achieve is constantly moving. So people need to evaluate and re-evaluate constantly what's within the art of the possible with AI, because chances are in three months, the frontier changes. So I just wanted to make that point: that a good way to manage risk is actually by investing in AI skills in your workforce.

[23:31] Joe Owen: You both made the point around experimentation being key. When we find something that works, what's the journey then—the scaling question? James, I want to come to you on this: how straightforward is it in your experience to scale innovation in government around technology? What breaks or what's at risk of breaking?

[23:52] James Kuht: It's a great question, and we've seen lots of things fail and some things succeed as well. It's worth acknowledging before I get into the answer to this question that it's good that there is a high bar for things to scale in government. By very definition, when you run 10 experiments of a new technology, despite everyone's motives for their pilot to succeed, there should be a survival of the fittest and only the best ones should get through.

Specific to AI, what we see when we see pilots scale is that there's a degree of IT professionalisation that often has to come in. Getting to an 80% prototype is often relatively straightforward and exciting, and everyone's like, "Oh my god, we built this agent, it does the thing!" The final 20% is pretty hard, and there's loads of stuff that goes into it—change management and stuff. But I think often the first piece is IT getting involved, working out: how do we scale this thing from an infrastructure perspective? Does it make sense from token costs? Are we going to move it onto some separate bit of infrastructure? Should we build this or should we just buy something off the shelf? I think a lot of things go into the IT bit.

Once you've worked out the feasibility of scaling this agent or this new AI workflow, or building AI into a new part of your planning process, it's then a question of change management. How do you actually embed this in a way which actually gets used by people and doesn't just sit on the shelf as something shiny, and actually gets evaluated on whether it has actually shifted the dial on the metric you're aiming for? To bring it back to senior leaders: it being crucially important you set the goals that are measurable that you're aiming for.

[25:49] Mallory Durran: I might jump off that evaluation point and say the 80% can often feel really easy, and then when you run an evaluation, you realise that there are missing capabilities or instances in which you're not covering that are actually critical failure points. So designing that evaluation in is key.

I also think there's a structural funding problem. Government runs on these four-year Spending Reviews. What it takes for a product team to run an MVP, get a pilot going, even maybe to integrate that into some IT systems, is actually relatively different than what it takes to maintain that service, and indeed respond to the ways in which a really good piece of tech will change ways of working and then need adaptation to features.

Setting product teams up that scale is not easy within the funding structures of government. If you're a tech company, you start with a really lean team, you have the minimum viable number of people, probably working crazy hours—which also happens in government! But there comes a point where Jane, the amazing engineer working overnight a couple of times a week, just doesn't cut it anymore. The way that you get past that in the private sector is raising capital, so you get these outside funding sources and you can kind of go from there.

In government, even where you've got ring-fenced pots of transformation funding, it's actually quite hard to unlock that for headcount, and even harder to unlock it for investment in compute or base infrastructure, just the way that those funding rules are set up. So harnessing technology scale for government is probably going to require a rethink of some of those funding rules, and/or getting creative about how outside capital can fuel that, whether it's joint innovation partnerships. But just thinking differently about how you fund teams to grow as a product grows, rather than saying, "Well, we think there's probably something here at point of Spending Review or budget," and guessing that you need 70 people a year from now. Do you contract that in? Do you just bring it in? But then you've got people sitting there not ready. It's a really weird funding picture, and one that getting to scale consistently is going to require some serious thought on.

[28:07] Joe Owen: Mallerie, you were talking about raising funds in order to bring people in. There's a question about whether one of the perpetual things is: can the government compete with the private sector for the skills and the talent that it needs in order to drive the transformation that you're talking about? Can the Cabinet Office really compete with the compensation packages that people could get elsewhere? You've worked in government with these teams, what is your take on that question?

[28:32] Mallory Durran: Yeah, so I think there have been successes, right? James worked there, absolutely nailing it. But yeah, there are barriers, and I think it's really tricky to look at a group of incredible professionals and say, "There are some other people who aren't you whose skills are more valuable." The reality is it's a market valuation, right? Maybe AI coding agent assistants will get so good that we won't need software developers at some point in the future—although objectively, software developers prompt AI for better software, but a different tale for a different day!

So I think there have been real successes in pushing beyond the bounds of typical public sector contract and compensation packages to get there. There are heaps of people with incredible technical skill who also—maybe it's like this elsewhere, but in the UK in particular—have just an unbelievable public-spirited mindset, and have taken pay cuts of more than 50% to come in and say, "I'm going to use the best skills I have at this once-in-a-generation opportunity." So I think it's possible.

I think the AI Safety Institute has done some really good things to get top-line researchers in, because it is a genuine national and international priority to do that. But I think there's a retention thing: how long is someone willing to keep turning down daily, weekly offers from new AI companies or the frontier labs? I think it's about compensation, but it's also about giving those people an environment in which they actually feel they're using their skills to best interest. So when we talk about having leadership that are willing to commit and take on the genuine risks of delivery, and invest—some of that sort of "sludge" that gets in the way of delivery is just as important to talent retention. If you're going to bring the people in, getting the most out of them actually means giving them the space to run and do their best work as well.

[30:31] James Kuht: Yeah, I'd back that up. I think the value proposition of working in government is just fundamentally different to working in the private sector for a lot of people. Mallerie's teams have done this well in the past, as well as the teams I've worked for: they lean into that. They really lean into the mission, which sounds a little bit trite, but honestly, the access you get in government and the levers you can pull—whether that's in one of the frontline departments or more centrally—they're extraordinary. You chat to your friends in the private sector, and I don't really look at many of them with any degree of envy. I'm now in the private sector socially, but I think the mission is fab, and the learning rate you can have as well in career progression I think is fab too.

So I think we need to lean into that value proposition. When it comes to technical skills, the real deep technical experts, I think I'd just be at pains to remind people that's the exception, not the norm. These people who are getting hired, prepared to go off to frontier labs and things, is probably a unique situation to just some of these real front-end AI teams, like the AI Safety Institute. This isn't for your average software engineer in DWP, and that's absolutely no slight on them, but they're not getting calls from Anthropic to go for £200k. So I think we should just be really careful that we don't overstate the salary piece. We've got a fundamentally different value proposition. The salaries are generally pretty good with the Digital, Data and Technology (DDaT) pay framework that came in, and there's just this exception pool for exquisite AI talent. The places Mallerie's mainly played in I think is an important exception, but it's not the norm.

[32:10] Joe Owen: One of the ways the government has managed the talent thing in the past is just use outsourcing, or contractors, or suppliers. Is there a risk that we sort of reheat that, whether it's to do with PFI initiatives or outsourcing initiatives, we take the same approach with AI and start relying on a small number of firms to do the government's work for it?

[32:33] Mallory Durran: Yeah, I mean, I think one of the downsides of that is that you don't build capability in-house. I think a certain amount of relying on outside skill is really intelligent, especially when it comes to the foundational work to set systems up to succeed, and to move existing workforces into a level of comfort to be able to operate themselves. So I think when you get these sort of insource-outsource combinations, or more partnership working between contractors or the private sector and government teams, you get the best of both worlds: people who really understand how services are being delivered, the challenges that live in that space, and the technical skill to achieve it.

Outsourcing in a "we're just going to throw this over the wall, go deliver it, we'll be stuck in this contract forever" way is a travesty. But actually really creatively setting up partnerships that leave behind capability, that bring the workforce along with them, and that then have access to the nuts and bolts of how things are really working so that those technical skills can be applied differently in a way that really creates a solution—I think it's not that outsourcing is or isn't the answer, but thinking about how you do that in an integrated way that leaves a legacy behind is really important.

[33:57] Joe Owen: So I am going to ask for the last few minutes or so a few more quickfire-style questions. The first one is: what are the two or three things that you would be looking for over the next year or so to show that the government is really serious about AI adoption? What are the signals through the noise where you're like, "Right, this government is really getting on with it"? James, I'm going to pick on you first.

[34:23] James Kuht: I'd love to see them clearly articulate the measurable goals that they are trying to drive with AI adoption. I think it's all very well and good—I think they've made some AI training freely available to 10 or 12 million people in the UK with a load of partners, and obviously there's a big push internally too, and that's great as a starting point. But unless it's actually tied to outcomes, we're never going to get the chance to see whether this has actually been impactful or not, and whether it needs more investment or a different approach or not.

I think this is just so important. It's absolutely undeniable that AI adoption across government and outside in the UK can drive massive productivity gains, but also allow us to do fundamentally more valuable tasks, and also make work more fulfilling. I worry at the minute that government is throwing over the fence access to these AI training courses, not measuring any outcomes whatsoever, and is going to wonder why we don't get the benefits of AI across the spread of the population when we made this training available to 12 million people. That would be my one.

[35:39] Joe Owen: Mallerie?

[35:45] Mallory Durran: I might go the opposite end of the spectrum and think about a real laser focus and commitment on a couple of really big bets, probably in some of the places where it's hardest or riskiest to achieve change, where you probably have lots of different kinds of products available being procured in different ways, but where the state—and only the state—has the authority and responsibility to ensure that that really delivers well.

So unified approaches to AI for unlocking healthcare outcomes; how AI will change and shape and equalise access to education or educational attainment. Actually, it's not going to be a one-off six-month programme bit of pre-training and off you go—there's going to be repeated testing and failures and things not meeting the mark. But I think real commitment to some of the places where the impact for the average citizen is going to be outsized. Actually just willingness to take those on and say, "Whatever it takes, we are going to achieve progress in those areas." I think there's been some good signals to that with leadership positions both in Number 10 and departments trying to create that focus, but I think it's about sustained commitment, both financially and otherwise, to big bets.

[37:00] Joe Owen: Okay, we've got big bets, we've got some targets and some ambition. What's one very practical and unglamorous reform—a sort of plumbing fix—that you would be looking for? Is it something to do with procurement, data sharing? Is there a particular thing on your wish list, Mallerie?

[37:20] Mallory Durran: I think I hit a few with how we do external funding of things, who leaders are and how we train them, but probably the thing I will spend the rest of my life on is better digital public infrastructure. The ways in which we set standards for how data is structured and held, and the technology that allows linkage and use of that data. It's a real barrier already to great use cases and AI applications that are being developed by civil servants, and is also a barrier to private sector innovation using great public sector data. It's got to be secure, it's got to work for everyone, can't be locked behind big, shiny, expensive walls—so really, implementation of open standards and technology.

[38:11] Joe Owen: Do you have an unglamorous reform, James?

[38:17] James Kuht: It's not a reform as such, but I think it comes back to the funding question. I think people have to be really clear on what's going to be replaced by the thing they're doing, because the civil service can't just keep growing—things have got to give. So I think people have got to be really clear why their AI initiative, and their top-down or bottom-up AI programme, is going to deliver better outcomes for the citizen or increase productivity, and what is going to be turned off to make space.

Obviously the Nudge Unit bit, now part of Nesta, is a great example of this: there should be sunset clauses around this sort of stuff where if things aren't working—which statistically, lots of things end up not working way more than we like to think—things should be turned off and we should move on to the next thing. We're in this incredible era of experimentation where we're all working this out, and we need to find space to do that, probably at the expense of other things, because getting private capital in is going to be hard. If things don't work, we should move on to the next thing.

[39:20] Joe Owen: And then my final question: if there's anyone listening to this who works in a government department or a local authority who thinks all of this sounds slightly intimidating, maybe exciting, what's the one thing that I could do tomorrow that would move me up the AI curve? What would your advice be, James first and then Mallerie?

[39:40] James Kuht: This has got a little bit of a bar because you're going to have to pay £20 for it if you don't already have it: I would have a go at trying to build an agent which automates a significant part of your job. That might require a ChatGPT licence or a Claude licence or whatever, and try and build an agent to do it. It's not too difficult. You need to use Claude Cowork Enterprise and do it within the bounds of data protection and cyber security, but I think in just doing that exercise for a few hours and building an agentic workflow, you learn so much stuff:

One, you learn exactly what the models and the tools are capable of, and you suddenly become a lot better at cutting through the hype.

Two, you become much better at procuring these sort of solutions because you know what you can do out of the box, so you're less prone to being sold snake oil, which is a real danger in government.

And three, you crucially work out what the things can't do, and therefore what's squarely in the realm of being a uniquely human skill, or something that you're going to have to keep testing with the frontier.

So yeah, take half a day, you'll have a ton of fun doing it, get some sort of agentic solution if you don't have it as part of your corporate IT already, and be really careful with data—Claude Cowork I suggest or ChatGPT—and try and build an agent which automates something really hard that you do. You might not succeed, but you will definitely learn a lot in doing that. There's nothing that beats doing stuff.

[41:12] Mallory Durran: So I would say if you heard that and you're a bit overwhelmed by it, you're like, "Well, what if I mess something up?" Or if your organisation doesn't have policies in place that allow you to access that stuff, actually I think you can gain some of the same learning, and then the bravery to go and ask for that or bring it into your work, if you just pick a problem that you really understand.

I spend heaps of time on trains, for example. Train times change, platforms change, there's loads of different places you can go to to find those times. I've set up an agent that at the times of day I usually take a train sends me a push notification to do it. Pick a problem in your life that you really understand and set something up that works for that—whether it's an agent or an automated workflow. The great thing about these AI tools is if you don't know where to start, ask them! Put that chat in. So what James said, if that felt overwhelming, just pick a problem that you really understand and do something with it.

[42:14] Joe Owen: James, Mallerie, thank you so much.

[42:16] Mallory Durran: Thanks for having us.

[42:17] Joe Owen: If you enjoyed this episode, please do like, share and subscribe wherever you get your podcasts. As a reminder, Nesta is a research and innovation foundation, and we design, test and scale solutions to society's biggest challenges. We are funded by a charitable endowment, and we are politically neutral. If you would like any more information, please go to nesta.org.uk.