AI and automation

The government's own progress report on the AI Opportunities Action Plan, published on 29 January 2026, leads with five designated AI Growth Zones, £28.2 billion of investment and more than 15,000 jobs. Further down the same document is a smaller claim that tells you considerably more: about one third of NHS chest X-rays, some 2.4 million scans, are now read with AI assistance.

Those two figures are not the same kind of thing. One is announced capital, which is a promise. The other is a workflow that changed, which is a fact about how a radiology department spends its afternoon.

This section is about the second kind. The test applied to a story here is whether anything reached a person doing a job, and whether the organisation can say what it now costs to keep running.

The interesting question stopped being capability

Between 2023 and 2025 the argument in most British boardrooms was whether the models were good enough. That argument is over for a large class of tasks, and it was replaced by a duller and more expensive one about operations. A pilot is cheap because it runs for six weeks on somebody's enthusiasm. A production system is expensive because it needs evaluation that survives a model version change, a review path for the cases it gets wrong, a retention policy for what it was shown, and somebody whose job it is to notice when its accuracy drifts.

The organisations that got this right treated it as a service management problem from the first week. The ones that did not are now in year two, discovering that the invoice grew while the pilot's benefit case stayed where it was. The full arithmetic is in what automation costs to run.

Much of what is sold as new here is not. A decade of robotic process automation taught British organisations what happens when brittle software is pointed at a process nobody has documented, and the lessons transferred more or less intact. That is the argument in what a decade of RPA taught us about agents.

There is a second reason the cost question is hard to answer honestly, and it is definitional. The word "agent" is currently doing service for three different things: a model that drafts something a person then approves, a model wired into systems it can change without approval, and a scheduled script with a language model somewhere in the middle. Only the second raises genuinely new control questions, and only the first is what most regulated British organisations have actually deployed. Reporting that collapses the three produces both the excitement and the panic in roughly equal measure.

Procurement makes the same mistake in a more expensive way. Buying a model is a licensing decision with a fairly predictable bill. Buying an outcome, which is what most organisations think they are doing, means buying an integration, an evaluation regime and a support model, and those are the lines that grow. The build or buy question has a boring answer, and the boring answer usually depends on whether the process being automated is one your organisation has a genuine opinion about.

Regulation is arriving sideways

Britain has not passed a cross-sector artificial intelligence statute, and on current evidence it is not about to. What it has instead became visible on 28 January 2026, when the technology and business secretaries wrote to sector regulators asking each to publish, by May 2026, a plan setting out how it will enable safe AI-powered innovation, then report annually on what that approach achieved.

For a head of digital that is a more consequential document than any white paper. It means the rule that binds your AI deployment will be written by the regulator that already inspects you, in the language it already uses, and it will arrive as guidance rather than as legislation. Organisations that operate across several regulated markets should expect several of those plans, and should expect them to disagree.

What gets rejected here

Benchmark results, on their own. A model scoring higher on a public evaluation is not evidence that a claims process got faster, that a waiting list moved, or that a council answered more calls. Vendor-sponsored surveys of adoption, which measure how many people said yes to a salesperson. And the productivity percentage with no denominator, which is the most common number in this field and the least checkable.

Most of what goes wrong in these programmes goes wrong a layer below the model, in the records being fed to it, which is why this section leans on data and analytics more than on model releases. Public sector deployments are covered from the other direction, through procurement and audit, in public sector technology.