Hi Reader,

In 2022, before any of this was fashionable, Andy Burnham said something that caught my eye.

“Technology for technology’s sake doesn’t change lives.”

He was right. And as the prime minister in waiting, he is about to get the chance to act on it from Number 10.

Here my hot take (we are in a UK heatwave after all) for the incoming Prime Minister. The job is not to win the AI race. It is not to publish another strategy, stand up another institute, or be photographed next to a datacentre that turns out to be a scaffolding yard. The job is to use technology to solve the hard thing, not the easy thing. To assure it for outcomes, not paperwork. And to refuse the siege that a handful of vendors have already laid against the British state.

None of that requires him to become a technologist. It requires him to trust the instincts he already has, and apply them somewhere he hasn’t yet looked.

AI is an ingredient, not a meal

The government’s flagship programme is called the AI Opportunities Action Plan. It's named after the technology.

That is the whole problem in four words. When AI is the meal, it gets its own plan, its own institute, its own budget line and its own press release. When it is an ingredient, it disappears into how the country is run, and the only thing anyone notices is that something got better. A waiting list got shorter. A benefit got paid correctly and on time. A planning decision came back in weeks instead of years.

The test of success is not an AI dashboard. It is a citizen describing a better outcome, with AI somewhere in the background as a contributing factor. If the programme is still called an AI programme when it’s finished, it hasn’t worked.

The easy thing is a giveaway. The hard thing is the job.

The easy thing looks like progress and changes nothing. Give every civil servant a Copilot licence. Pay a Big Four firm around £4 million to build citizens an “AI Skills Hub”, then ship a directory of other people’s online courses, most of them paid for. Announce “AI for everyone” and do your best Oprah impression. You get a chatbot, you get a chatbot, everybody gets a chatbot.

Being surrounded by AI is not the same as getting anything from it. You can drive a Fiat Uno around an oil refinery, with millions of litres of fuel all around you, and still run out of petrol because none of it is in the tank. Licences are not adoption. A badge is not a skill. Ubiquity is not benefit.

The hard thing is to point technology at the problems the country cares about, and most of those problems are messy. Which brings us to the most important sentence in this piece.

Edge cases are the business of government

A company can choose its customers. It can serve the easy, profitable middle of the market and decline everyone else. Government cannot. It is obliged to serve the whole population, and the people who need the state most are almost always the people with the most complicated situations. The edge cases are not a rounding error for government. They are the business of government.

Now the awkward bit.

The edge case may be where Gen AI is least reliable.

These systems are not reasoning. They are predicting the next most plausible word, and they are most accurate where the training data is densest, which is the well-trodden middle. Out at the fringes, where the unusual benefit claim and the atypical medical presentation and the complicated housing case live, the model gets, in plain terms, a bit bullshitty. Worse, it tells you so with the same confidence it uses when it’s right. Old software failed by breaking. AI fails by agreeing. It hands a fluent, wrong answer to a real person and nothing in the output flags which kind of answer you just got.

It may be at its weakest where its impact is most destructive to vulnerable people’s lives. A confidently wrong decision about someone’s benefits, their care, their immigration status isn’t an admin error. It’s a failure of statecraft.

The government has just renamed the AI Safety Institute the AI Security Institute, dropping algorithmic bias and ethics from its remit to focus on national security. The security case is real. But the state has moved its watchdog away from the harm that lands on a citizen.

We assure AI for conformance: were the right boxes ticked? The right principles published? We don’t assure it for outcomes: does it behave fairly in the decision that changes someone’s life.

The DWP is already running machine-learning fraud detection that the National Audit Office and campaign groups have flagged for bias. That is not a future risk. It is a live one, and it is the template for what gets worse if nobody changes the question from “did we follow the framework?” to “is this fair to the person on the receiving end?”

The fix isn’t to ban AI from public services. It’s to assure it for outcomes, match the scrutiny to what’s at stake for the person affected, and design the decision so a human expert holds it rather than signing off whatever the model produces. The hard cases should reach real judgement by design, not by luck.

It’s not a moat. It’s a siege.

Now the part that should alarm Burnham who built his career on the idea that London extracts and the regions get left holding the bill.

The same dynamic is playing out between a few large vendors and the British state. When one company embeds itself deep enough in critical public infrastructure that swapping it out becomes unthinkable, that is not a moat protecting good service. It is a siege. It is expensive, hard to escape, and it captures the institution it claims to serve.

Burnham’s first real AI decision is already on his desk. The Palantir contract for the NHS Federated Data Platform, worth up to £330 million, has a break clause. Its first term ends in March 2027, which means the Department of Health has to actively choose to extend it, or it lapses. The British Medical Association has already told doctors to limit their engagement with it. NHS users have reportedly described it as awful to use.

Is this just a Palantir problem? No. The people advising government on AI are too often the people selling it. That is the structural problem, and it is the one nobody at the top is naming: the body with the independence to make the call is the only body not in the room when the case is made.

The decision about whether to trigger that break clause is the cleanest test the new Prime Minister will get of whether he means any of this. It is the siege, made concrete, with a date attached.

And it isn’t only his decision. Every board in a regulated sector is about to face its own version of it. Most of them are being advised on it by the people who would lose the contract.

The future of work is a choice, not a forecast

One last thing, and it’s the one closest to his heart. The dominant story about AI and work is a fatalist one: the machines arrive, the jobs go, and the rest of us brace for impact. The tech bros rather like that story, because if the outcome is inevitable, nobody has to answer for it.

It isn’t inevitable. In May, Pope Leo argued that AI is never neutral, that it takes on the character of those who finance and control it, and that its gains must not be concentrated in a few hands while everyone else is shut out. I agree. What AI does to working lives is a matter of choice, and government has far more agency over it than it has chosen to use. Whether the gains are shared or hoarded, whether someone starting out still has a way into a trade, whether a town keeps work worth having: these are political decisions, not the outputs of a model.

On the current course, a handful of tech bros take the upside and the downside lands on everyone else. Is that levelling down? Burnham has built his career arguing to remove regional inequality. This is right in his lane. And it starts by leading the argument for future jobs in an AI world. Citizens will only be the victims of this technology if their government decides to let them. I’m advocating for debate and design rather than fatalistic acceptance.

What I’d say to him

If I had the audience, it comes down to three lines.

AI is an ingredient, not a meal.

It’s not a moat, it’s a siege.

Edge cases are the business of government.

He already has the instinct. He said it himself in 2022. Now he has to govern like he meant it.

I've built an AI confidence snapshot, if you want to be a beta tester, reply to this message.

And share with a friend, in the hope that someone passes this on to Andy B.

Until next time!

Helen