We Are Building AI Infrastructure. But Are We Building the Capacity to Use It?

Why judgement, adaptation and organisational memory deserve the same strategic attention as compute

By Louize Clark

Nobody seriously disputes that Britain needs to invest in the technical foundations of AI. Data centres. Compute capacity. Sovereign models. Chips, energy, talent, the infrastructure that sits underneath all of it. The scale of the investment is unusual, but the logic behind it isn't: a country that wants a serious AI capability has to build the physical and computational base that capability requires. That much is settled, and rightly so.

It's worth pausing on that agreement for a moment, because it's genuine. The organisations, investors and policymakers making the case for technical infrastructure are not wrong. Compute matters. Sovereign capability matters. Nothing here argues against building it.

But sitting underneath that agreement is a quieter assumption, one that rarely needs to be stated out loud because it rarely feels like an assumption at all: that once the technical infrastructure exists, the organisations and people meant to use it will more or less follow. Adoption gets treated as the easy part, a training session, a change-management plan, a rollout schedule. The hard part, the expensive part, the part that attracts the investment and the headlines, is assumed to be the technology itself.

Follow that assumption into an actual organisation, though, and it starts to strain almost immediately.

Picture one employee, in one team, doing a job they've done competently for years. An AI tool is introduced to help with part of it, drafting, summarising, flagging anomalies, whatever the task requires. For the first few weeks, they check its output carefully, the way anyone would with something new. Then, gradually, because the tool is usually right and checking takes time nobody has been given back, the checking gets lighter. Not through negligence. Through the ordinary, reasonable accumulation of trust that any competent person extends to a tool that keeps performing well.

Six months in, they are still nominally "reviewing" the output. In practice, review has become a formality.

Multiply that one person across a team, and it becomes a habit rather than an individual's drift, the kind of thing nobody decided, exactly, but that has become how the work gets done. Multiply the team across an organisation, and something closer to institutional memory begins to change hands. Knowledge that once lived in experienced people starts to live instead in the prompts, defaults and shortcuts that only a handful of people fully understand in exactly the pattern explored in the previous piece in this series: nobody decided this should happen, and yet, eighteen months on, it has.

Multiply the organisation across a country trying to build sovereign AI capability, and the same gap simply scales: a national strategy that assumes a workforce capable of operating that capability well, exercising judgement over it, catching its failures, adapting as it changes, without ever quite funding, or measuring, whether that capability actually exists.

Technical capability determines what becomes possible. It does not determine what an organisation becomes capable of doing well. The technology is not the only constraint in that chain - availability, integration, cost and legacy systems can all genuinely limit what's possible. But increasingly, once the technology is in place, what determines whether it's used well is something else entirely: the human and organisational capacity surrounding it.

Give the same platform to two people doing the same job, or two teams inside the same organisation, and behaviour will diverge within months. Some keep sharpening their judgement around the tool, learning precisely where it's strong and where it fails. Others start deferring to it by default, because nobody was ever equipped, or given the standing, to do otherwise. Most organisations contain both patterns at once, often in the same department, and sometimes in the same person, depending on workload, confidence and the attention a task happens to get that week. The technology is identical throughout. The outcome is not.

This is worth naming carefully rather than dramatically, because the temptation with a pattern like this is to reach for alarm it doesn't yet deserve. Emerging research into how sustained AI assistance affects critical thinking and cognitive effort is still early, and the conclusions aren't settled. But the underlying question, what capability a person retains as more stages of analysis, drafting or judgement are delegated to a system, deserves the same seriousness as any question asked about compute.

It's tempting to file all of this under governance, or under HR, because those are the categories that already exist to hold organisational concerns like this. Neither quite fits.

Governance can establish responsibilities, controls and boundaries around how AI is used and a thorough governance framework is genuinely valuable. But those structures still depend on people retaining the judgement and confidence required to actually enact them, day to day, under normal pressure; a policy is only as good as the review culture built underneath it. Training programmes, similarly, describe activity, attendance, completion rates, engagement scores rather than the quieter question of whether the judgement they were meant to build is still there six months later, when it's actually needed.

The language available for this is thin. "Training" describes an event with an end date. "Change management" describes a project that gets marked complete.

What actually sits underneath adoption, the ongoing capacity to exercise judgement, adapt as tools keep shifting, and carry knowledge forward as the people around a system change starts to look like something else.

Not a programme that finishes. Not a department that already exists under another name. A form of infrastructure in its own right: human infrastructure, the human counterpart to the compute and data centres currently receiving so much of the investment and one that deteriorates quietly if left unmaintained, in much the same way a technical system does, except that unlike compute, it cannot simply be purchased and installed.

Nobody publishes a figure for a workforce's collective judgement the way figures get published for compute capacity. That absence is itself informative. It means the gap between the two what's being built and what's being asked to carry it goes largely unmeasured and almost entirely unfunded, right alongside a national strategy that quietly assumes the gap isn't there.

Compute can be built in a data centre. Judgement can only be built in people. One is being funded at national scale. The other is still being treated as though it will arrive by itself.

AI Policies UK helps organisations see the AI infrastructure they're already standing on, chosen, inherited and embedded — before decisions like this one have to be made under pressure. Get in touch: louize@aipolicies.uk