AI and Public Services – What should you know? What should you be thinking about?
AI will be the most consequential disruption today’s public service leaders encounter in their careers.
We have reached a point in AI development where it cannot be assumed that someone else is checking on our behalf that it’s all ok to use. The technologists pushing the frontiers do not have public services front of mind!
Everyone reading this will be using some form of AI, personally as well as within their organisations. But before jumping too far in on where it can be used, there is a need to properly understand what kind of thing it is, what it can and cannot do, what happens when it begins to act rather than merely advise and what must remain distinctively human in public services.
I’ve been studying, reading and thinking about AI for the last five years and I am excited and worried in equal measure. Over those five years, the Mutual Ventures team has worked with councils and social enterprises on AI strategy and has produced a guidance framework on behalf of the Co-operative Councils Innovation Network.
This piece kicks off a series of Mutual Ventures articles, webinars and #RadicalReformers podcasts aimed at equipping public service leaders with the knowledge and understanding they need to ask the right questions and make the right decisions about the use of AI in public services.
FYI - I've always written using dashes - so please don't think this is all AI generated, which would be ironic!
Created, discovered or grown?
Human beings discovered fire - we didn’t invent it. We learned how to harness it. It transformed our existence. We could cook food, heat homes, forge materials and eventually power industries...But we can also lose control of fire. It can burn a building down without us telling it to.
A surprisingly large number of people working in the AI field think AI is something we have discovered rather than invented.
That isn't merely philosophical. Ok, it is a bit! but not entirely…
Modern frontier models (think the various Chat GPT or Claude models) aren't conventional software in the sense that somebody specifies every behaviour line-by-line. Training creates a model whose capabilities are subsequently tested, discovered and grown, often without human input. “Grown rather than built” is therefore a useful metaphor.
That creates the first critical leadership question - What does responsible stewardship look like when even the people creating the technology don’t know what it will be capable of?
What exactly is AI?
Artificial Intelligence itself is not new. Public services have been using forms of AI and machine learning for years, often without describing them as such. What has changed dramatically is what these systems can do. Think about the old-fashioned chatbot on a website (best example is the bin collection day automated response). It might look like you are having a conversation, but underneath it is essentially following a set of rules e.g. if the person asks this, give them that answer. Go outside the questions and answers the designers anticipated and the limitations quickly become obvious.
Generative AI is fundamentally different. Rather than selecting from a set of predetermined responses, models such as ChatGPT and Claude have been trained on enormous amounts of information and have learned patterns and relationships within that data. They can use those patterns to generate a new response to a question they may never have encountered before. Nobody has written a rule telling the system exactly what to say. The response is generated there and then. That is the big leap which has made our recent experience of AI feel so different.
But we are already moving beyond generative AI. The next important development for public-service leaders to understand is agentic AI. A generative AI system can help produce a report, analyse some information or suggest what I might do next. An AI agent can be given an objective and work out a series of steps to achieve it. This can involve using other software, retrieving information, communicating with the real world and potentially taking actions in the real world.
So we are moving from AI that can predict, to AI that can generate - and now increasingly to AI that can act. That final transition is particularly significant for public services. An AI that drafts a recommendation for a social worker or housing officer raises important questions. An AI that can autonomously act on that recommendation, updating systems, contacting people, initiating processes or instructing other agents, raises a different order of questions about judgement, accountability, control and ultimately who, or what, is exercising agency on behalf of the state.
How worried should we actually be?
The first thing to say is that this world is coming whether you choose to engage with it or not.
Individual public service leaders can decide not to use AI themselves. Organisations can decide to proceed cautiously. But they cannot decide that AI will not affect them. Staff are already using it, suppliers are embedding it into products, citizens will increasingly use it when interacting with public services. Some of your colleagues will already have been subject to AI astro-turfing (i.e. the generation of fake grass-roots campaigns!).
In any case, the capabilities of the technology will continue to develop outside the control of any individual council, NHS organisation or central government department.
Choosing not to engage with AI is therefore not really a viable strategy. However, acknowledging that AI is coming is very different from arguing that we should embrace everything it offers.
There are two equally unhelpful positions here. One is to dismiss the concerns as another round of technological panic. At the other extreme are predictions that increasingly capable AI could escape meaningful human control and ultimately represent an existential threat. I don't think public-service leaders need to become experts in that debate or decide which side is right. The important point is that there is a sufficiently wide range of credible outcomes that we should take the uncertainty seriously.
This is where I find the analogy with fire useful. Harnessing fire was a huge leap forward for humanity. It allowed us to cook food, keep warm and eventually transform how we made things - but can also burn your house down. Our response wasn't to stop using fire. Nor was it to say that because fire creates enormous benefits we shouldn't worry too much about the risks. We learned to contain it, developed rules about where and how it could be used and built protections around it. After all, an entire public service was created around managing the risks of fire…
AI presents an additional complication. We pretty much understand fire – it hasn’t evolved in any real sense. With AI, its full capabilities often only become apparent once increasingly large and sophisticated models have been created and tested. It is constantly evolving.
Therefore, a key question for public-service leaders is: How do we capture what could be immense public value from this technology without surrendering judgement, control and responsibility along the way?
That is where the conversation about risk needs to begin.
Public services cannot adopt Silicon Valley's risk model
This brings us to regulation and to an uncomfortable question about who we are trusting to determine how quickly this technology develops.
The frontier labs themselves (Anthropic and Open AI primarily) broadly accept that industry self-regulation is only a first step and that government intervention will be needed to establish standards and ensure compliance. But there is a fairly obvious problem. These companies are also engaged in an intense commercial race. They are competing for investment, talent and customers. Even if the leadership of one company genuinely believed that development should slow while particular risks were better understood (they are increasingly stating this publicly), what happens if its competitors don't slow down too?
This isn't simply speculation from people standing outside the industry throwing stones. In July this year (2026), nearly 1,400 employees from these companies signed a statement making this point. They warned that their companies are under intense competitive pressure not to slow down unilaterally and called for mechanisms that would allow society to deliberately slow the pace of AI development.
Yikes! That is quite an scary position for us to be in...
This should matter enormously to government - but at the minute we have a US President who is instinctively against anything that will slow America’s speed and position in the great AI race.
The extension of this view is that we should avoid regulating AI too heavily while the technology is developing. Allow innovation to flourish, rely wherever possible on existing law - privacy, discrimination, customer protection, negligence etc. and intervene when identifiable harms emerge.
Obviously, an approach of “allow it to develop and deal with the harms afterwards” is not remotely sufficient for public services!
We don't regulate medicines that way. Imagine a pharmaceutical company developing an entirely new drug and saying: “We think this could have enormous benefits. We have tested it ourselves and we're pretty confident it is safe. Let's give it to several million people and, if it causes serious harm, existing law allows people to sue us.”
We wouldn't accept that.
The same principle applies to food safety, aviation and numerous other areas where society has concluded that some risks are simply too consequential.
That distinction becomes particularly important when AI moves into public services because government and public services are not simply another consumer of technology. They exercise power over people. They decide whether someone receives care or support. They determine eligibility for housing and benefits and safeguard children. In delivering these functions, public services holds some of the most sensitive information imaginable about people’s lives. We have seen some of this play out publicly with regard to Palantir’s contract with the NHS and concerns over that data.
This doesn't mean freezing innovation or subjecting every AI tool used by public services to something resembling a clinical trial. The level of assurance should obviously reflect the potential consequences. Using generative AI to help somebody draft an internal meeting note is very different from allowing an AI agent to make or enact decisions affecting a vulnerable person's care.
The more useful question for public service leaders is: What level of evidence, independent assurance and human control should be required for different levels of AI risk?
The values governing public services are different from those governing a technology company. Public service leaders should be leading this conversation rather than waiting for the technology industry to tell us what they intend to do.
For many services, a “human in the loop” isn't enough
One of the phrases we hear repeatedly in discussions about responsible AI is that there should always be a “human in the loop”. It sounds reassuring. The AI may be doing more, but ultimately there is still a person involved. The problem is that the phrase can conceal as much as it reveals. There is a considerable difference between a human genuinely making a decision supported by AI and a human simply sitting somewhere in a process that is, in reality, being driven by AI.
It is useful to think about three different models.
With a human in the loop, the AI supports a process but a person remains directly involved in consequential decisions, reviewing what the AI has produced, applying their own judgement and deciding whether to act.
With a human on the loop, the AI has greater autonomy. It can make decisions or take actions within defined boundaries while a human monitors what it is doing and can intervene when necessary.
With a human out of the loop, humans set the initial parameters but the system can make and enact decisions without requiring human involvement in the individual decision at all.
This distinction becomes much more important as we move from generative to agentic AI.
If AI simply drafts a report, a human can read it before anything happens. But imagine an AI agent that can access personal data, interpret information, decide what to do next, communicate with other organisations and initiate actions. A person might technically remain “on the loop”, supervising the system and retaining the ability to intervene. But if the agent is initiating multiple actions, potentially at a speed we can’t keep up with, how meaningful is that supervision in practice? A human who theoretically has the power to intervene, but cannot realistically understand what the system is doing quickly enough to do so, is not exercising meaningful control.
That is why I think the UK Government's Artificial Intelligence Playbook points us towards a much more useful principle i.e. meaningful human control at the appropriate stages of an AI system's lifecycle and use. The important word is meaningful. Human oversight should not become a governance comfort blanket.
For public-service leaders, therefore, the key questions are:
What judgement are we asking the human to exercise?
What information do they have to challenge the AI?
Do they have the time, authority and confidence to disagree with it?
Can they understand why it has reached its recommendation?
What happens when they override it?
And which decisions are sufficiently consequential that we should never allow the human to move from in the loop to merely on it?
The thing technologists may underestimate: relationships!
There is another aspect of this debate which I think those developing AI can sometimes underestimate - the importance of human relationships!
I was struck by this listening to Ezra Klein interviewing Bill Gates about AI. Klein raised the possibility that, even in a world in which AI could provide human level levels of expertise, people might still value some services precisely because they involved another human being. Gates initially seemed to interpret the point as being about whether we enjoy everyday human interactions - something akin to preferring to have a conversation with a taxi driver rather than travelling in an autonomous car. But Klein was getting at something deeper - whether there are circumstances in which the relationship itself is part of what creates value. That distinction matters enormously for public services but it is not front of mind for those driving the AI bus at breakneck speed!
Much of the way we think about technology starts by breaking work down into tasks. Which of those tasks could AI perform more quickly, cheaply or accurately? Viewed through that lens, as AI becomes better at more of those tasks, the role of the human inevitably appears to shrink. But that is not how people experience many public services especially those which are there to support the most vulnerable people.
A good social worker does not simply gather information, assess it and produce the correct intervention. A good community nurse does not simply diagnose a problem and prescribe the appropriate response. They develop relationships and notice things that haven't been said. They understand the history behind the information in front of them. They challenge people and are challenged in return. And, over time, they understand something about a person's family, relationships, ambitions, fears and the place in which they live. I’m teaching you to suck eggs at this point I know! This sits at the heart of much of Mutual Ventures’ work on place-based relational services.
There is a danger that AI leads us down a path towards making a siloed model really efficient. No need for deep reform, folks!
Thanks - but we want more than that…
AI doesn't replace relational working. It creates the possibility of relational working at a level of insight and capacity that our fragmented systems have rarely allowed.
Don’t rely on generative AI to do your novel thinking
There is a danger that because generative AI communicates so convincingly, we assume it is thinking in the same way that we do. I’ve been doing an Oxford University course on Generative and Agentic AI and one of the professors, Matthias Holweg (along with Teppo Felin), has set out an important challenge to that assumption. Their argument is that generative AI is pretty special at identifying patterns in existing knowledge and re-cutting and combining what humanity already knows in potentially new ways. However, that this is fundamentally different from the way humans develop theories, think about cause and effect and imagine explanations and ideas that are not contained within existing data. AI can produce things that appear novel, but that does not necessarily mean it is creating genuinely new knowledge in the way that humans can.
That distinction matters when public-service leaders start using AI not simply to summarise information, but to develop strategy. This is happening a lot.
If we ask an AI trained on the accumulated knowledge and experience of existing public services how to reform public services, there is an obvious risk. It is well equipped to tell us what we already know. At its worst, AI could become a champion for the assumptions of the existing system, at the very point when we should be questioning them.
As a side note, we should acknowledge that Agentic AI complicates this argument.
What happens when AI no longer simply processes the accumulated knowledge of the past, but can act, observe what happens, learn from the result and try again?
Experimentation is one of the ways humans create new knowledge. We test, we learn, and we grow!
If autonomous AI systems can test hypotheses in digital or real-world environments (this is known as Recursive self-improvement – RSI) the distinction between recombining and re-cutting existing knowledge and discovering something new starts to become blurred. RSI is how some of the more recent high profile “jail breaks” have occurred, such as the Open AI model breaking out of it’s secure sandbox (essentially a safe space for experimentation) and hacking another application (Hugging Face).
But we don’t use agents in that way when writing reports and strategies…yet.
The main point here is you should not rely on generative AI (Chat GPT, Claude, Co-pilot) to do your new strategic thinking for you - unless of course you are happy to follow what has gone before.
So what could/should we do with it?
Having spent much of this article talking about risk, I don't want to leave the impression that I am pessimistic about AI. Quite the opposite. I think the potential for AI to help us create better public services is very real but only if we are deliberate enough and ambitious enough about what we use it for.
The easiest thing for any of us to do will be to apply AI to the public-service system we already have. We can process an assessment faster. Write follow up notes faster. Determine eligibility more quickly. And the rest…
There will be real value in some of these things, particularly where they free up humans from repetitive administrative work. But there is also a danger. We could make every individual transaction significantly more efficient while leaving the experience of the person using public services almost completely unchanged.
Someone could still have to tell their story five times to different services. They could still be assessed separately by health, housing, social care and employment services. They could still fall into the spaces between organisational silos. You could still spend enormous amounts of public money responding to crises which might have been prevented - but we’d be doing it really efficiently!
This is where AI connects with the wider re-wiring the state agenda and our wider public service reform work on relational place-based public services. People don't live their lives in public-service silos, and any technology solution shouldn't either.
AI will give us the ability to make sense of information across organisational boundaries, understand patterns of demand, identify where problems are escalating and give frontline professionals a much richer understanding of people and communities.
Imagine neighbourhood teams where practitioners spend far less of their time searching systems, writing notes, completing repetitive assessments and navigating organisational processes. Imagine instead that AI does more of that work in the background - bringing together relevant information, spotting patterns, reminding practitioners what has happened before and helping them understand the wider context. The professional arrives at the conversation well informed and prepared - but then does the human/relational/trust building things the AI is not capable of doing and doesn’t value.
That is the sweet spot I am increasingly interested in.
Of course, none of this happens simply because we implement one or another AI solution. It requires organisations to collaborate around people and places rather than simply optimise their own part of the system. AI cannot solve the fragmentation my colleagues and I have written about elsewhere - but it might make some of the excuses for maintaining it increasingly difficult to sustain.
That feels like an optimistic place to end this part of the discussion. Thanks for sticking with it!
Watch this space for more on our series or articles, webinars and podcasts.
If you are interested in discussing any of this, I would love to hear from you – andrew@mutualventures.co.uk
Comments