Right now, almost every public sector leader I talk to is being told they need an AI strategy. There’s a lot of buzz among ministers and the top of the shop in Whitehall on the potential for AI to drastically change how the civil service operates. But enthusiasm isn’t a strategy. And in my recent work looking at how organisations across government are adopting AI, the space between the two is where value is slipping away.
The pattern repeats itself almost everywhere. Leaders see the promise, move quickly to put something in place — usually a copilot licence for everyone — and wait for the productivity gains to arrive. A year on, the tools are installed, a handful of people use them well, and the organisation can do no more than it could before.
But it doesn’t have to be this way. In my work on AI adoption in government, I’ve seen that the organisations getting real, lasting value from AI tools aren’t the ones that deployed it fastest. They’re the ones that started with the problem they were trying to solve, and then invested in the specific data and AI that can help solve that problem.
It’s the same lesson I’ve written about with data strategy. A generic “let’s adopt AI” fails for the same reason a generic “let’s treat data as an asset” does: it isn’t tied to anything anyone actually cares about. The organisations that get it right don’t buy a strategy off the shelf. They make a few clear choices about where AI will earn its keep, and then they resource the things that make those choices real.
So what does a good AI strategy actually put its money and attention into? In my experience, four things. Only one of them is the technology.
Clarity of intent
Start with the problem, not the product. Search for the most sticky, time-sapping activities where automation could really add value. Or the things you just can’t do today because the process is too complex or you don’t have the right inputs. Make clear choices about what tools you’ll build and what you’ll buy to tackle the problem – understanding the true cost and risks of each and how they’ll integrate into your wider digital ecosystem. Off-the-shelf tools are quick and comparatively cheap to begin with; they speed up what you already do, but costs can build up later if inefficient token use racks up sky high or they don’t line up with your other systems. Only AI built on your own data gives you a capability you didn’t have before. Both, done the right way, can be the right call. What doesn’t work is reaching for the quick win while expecting the transformational one.
Data that’s AI-ready
This is the one that everything else rests on. Government usually holds the data it needs; it rarely holds it in a state AI can read or learn from. It sits in ageing systems, unclassified, scattered across teams, and not machine-readable. GDS has published some helpful guidance on Making government datasets ready for AI. This is unglamorous work – cleaning it, linking it, labelling it – but it’s what turns raw data into something AI can actually use, and it has to come first. I’ve seen so many organisations buy clever tools while the data underneath them isn’t ready to feed them. There’s no shortcut here, and no amount of budget on the tool makes up for skipping it. This is where an AI ambition and a data strategy become the same conversation.
The right specialists
Getting the best out of AI is a bit like competing in Formula 1. The driver, the aerodynamicist and the race engineer each master a completely different craft. None can do another’s job, but the team needs to come together to win. Building AI, integrating it into how you work, and buying it well call for different, scarce specialists too, and they’re rarely already sitting in the building. The organisations that pull ahead treat this as capability-building, not a one-off software rollout. They grow and hold on to the people who can do the work, rather than assuming a licence and a lunchtime webinar will get them there. In government, these specialists are especially scarce, with the public sector competing for them against employers who can pay a good deal more. That makes it all the more important to use the ones you have wisely: point them at the problems only they can solve and surround them with others who do the prep work. Don’t make your driver fuel then engine or change the tyres – they’ll get bored and leave for a job that will have them racing around the track in no time.
Guardrails, not hurdles
Good governance isn’t red tape. Done well, it’s the thing that lets you move quickly and safely. But a lot of organisations are running AI through security and impact assessments built for a pre-AI world: slow, exhaustive, and designed for a different set of solutions. Add the cautious culture and low AI literacy we have in the public sector on top, and you get governance far more restrictive than the risk warrants, smothering experiments before they can prove themselves. The answer isn’t less oversight. It’s proportionate guardrails, clear ownership, and a real mechanism for backing the use cases worth pursuing. Too little governance invites risk; too much invites paralysis. The organisations that move fastest are usually the ones that gave their people a safe space to experiment and a clear route from “this works” to “let’s scale it”.
Buying AI tooling is the easy part
Look again at those four, and notice what’s missing: the AI itself.
That’s the whole point. None of the things that make AI work are really about the technology. They’re about intent, data, people, and the permission to try — the same foundations that decide whether any ambitious change succeeds in the public sector. The departments and bodies moving ahead aren’t the ones with the shiniest tools. They did the plumbing first, and then gave their people room to experiment once it was in.
None of this is an argument for going slow. AI’s potential in government is real, and so is the cost of standing still while others move. But there’s a difference between moving fast and simply moving first. The organisations that will get the most out of AI are the ones treating it as a capability to build, not a product to buy.
So before the next licence is signed, the more useful question isn’t “which AI tool should we adopt?” It’s “are the foundations underneath it ready, and if not, what’s our plan to get them there?” That’s a harder conversation. It’s also the one that separates the organisations transforming public services from the ones left holding a very expensive chatbot.
This is one part of the wider data and AI strategy work I do with public sector organisations. If it’s set off a thought, I’d be glad to talk it through — find me on LinkedIn or at contact@kompassconsulting.co.uk.
For more detailed analysis on AI adoption in government read Kompass Consulting’s report on Building AI in Government.