Twelve months into the AI Opportunities Action Plan, the scorecard reads like a win. Thirty-eight of fifty commitments met. A pledge to expand compute twentyfold by 2030, with Isambard-AI already running in Bristol. £28.2bn of private investment routed through five AI Growth Zones, more than 15,000 jobs attached. A Sovereign AI Unit with up to £500m behind it. On every input a government can buy, order or announce, Britain has moved. And yet the opportunity the country keeps being told it can seize — the one The Economist set out this month in its case for what Britain must do next — is not an input. It is an outcome, and outcomes do not arrive on a press release.
The prize is a dividend, not a data centre
The reason any of this matters is a single number the inputs are meant to produce. The OECD’s 2025 estimate is that AI adoption could lift UK productivity growth by 0.4 to 1.3 percentage points, worth somewhere between £55bn and £140bn of gross value added by 2030. That is the prize. It is also the most demand-side figure imaginable: it materialises only if firms across the economy, not the labs, change how they work.
This is the distinction Britain’s strategy keeps blurring. Compute, investment and Growth Zones are the supply of AI capability. The productivity dividend is the demand for it, realised inside ordinary businesses and public services. The first is something the state can fund. The second is something the state can only encourage, and the gap between the two is exactly where the opportunity is currently going missing.
Strategic Reality: Britain has spent a year proving it can build the supply side of AI. The £55bn–£140bn prize sits entirely on the demand side, and almost none of the year’s headline wins move that number directly.
The numbers that frame the choice
| Measure | Where it stands | Source |
|---|---|---|
| Action Plan commitments met in year one | 38 of 50 (76%) | DSIT, One Year On |
| Private investment via AI Growth Zones | £28.2bn, 15,000+ jobs | DSIT, One Year On |
| Productivity prize from AI adoption | +0.4 to +1.3 ppts → £55bn–£140bn GVA by 2030 | OECD, 2025 |
| AI use, whole economy | ~25% of firms (Dec 2025) | ONS / DSIT |
| AI use, professional and business services | 43.4% (up from 31.4% a year earlier) | DSIT sector data |
| Firms not yet ready on core enablers | ~75% (data, orchestration, monitoring) | DSIT / sector survey |
The supply-side column is full. The demand-side column is where the story turns: a quarter of the economy using AI at all, the most AI-ready sector still under half, and three in four firms not ready on the plumbing that turns a pilot into production.
What’s really happening underneath the announcements
Britain’s instinct is to admire its assets and assume diffusion follows. The assets are genuine. Google DeepMind is headquartered in London. The research base across Oxford, Cambridge, Imperial and UCL is world-class. The regulatory posture is lighter than Brussels, the language is the internet’s default, and the professional-services economy is deep. The country invents, hosts and regulates AI at a level most nations would envy.
What it does not yet do is use it widely. The adoption figures tell a two-speed story. Professional and business services have crossed 43%, but the whole-economy number sits around a quarter, and the distribution underneath that average is brutal — large firms experimenting freely, the long tail of SMEs barely started. More telling still is the readiness gap inside the firms that have begun. Around three-quarters report they are not ready on the core enablers — clean data, orchestration, monitoring — and most have made little progress on redesigning the processes that AI is supposed to transform. The bottleneck is not access to models. It is the unglamorous work of moving from a chatbot trial to a production system that actually changes a cost line.
Critical Context: The adoption barrier in 2026 is no longer technology. Firms have the tools. What they lack is the data foundations, the process redesign and the operational discipline to run AI in production rather than in a pilot that quietly never scales.
This is the British disease in a new costume. The country has always been better at the breakthrough than the rollout — strong on invention, weak on diffusion. The Action Plan’s first year, for all its merits, repeats the pattern. It is overwhelmingly a supply-side document, and its most quantified wins are the ones easiest to count: gigawatts, pounds invested, zones designated. The harder, slower, less photogenic work of getting a mid-sized manufacturer in the Midlands to re-engineer a workflow around AI does not announce well, so it gets less of the oxygen.
The human factor decides the dividend
The productivity dividend is, in the end, a management and skills story rather than an infrastructure one. AI raises output only when someone redesigns the job around it, and that someone is usually a middle manager with no slack and no playbook. The government has noticed: over a million free AI courses delivered, a target to upskill ten million workers by 2030, £187m into TechFirst. These are real and worth backing. But training individuals is necessary, not sufficient. The binding constraint is organisational capability — whether a firm can absorb the skill once a worker has it.
The public sector is where the diffusion test is most visible, and most encouraging. The NHS now AI-assists 2.4 million scans and has halved stroke treatment time in places. School trials are reaching up to 450,000 children on free school meals. A planning tool goes to every council by spring 2026. These are demand-side wins — AI doing work, not AI being procured — and they are the template the rest of the economy needs.
Who the diffusion gap lands on
| Stakeholder | What the demand-side gap means in practice |
|---|---|
| SMEs and the long tail | Falling behind larger competitors who can absorb AI; the productivity dividend bypasses them entirely without hands-on support |
| Professional services | Furthest ahead on adoption, but readiness gaps mean much of the activity is experimentation that has yet to reach a cost line |
| Public sector | Best-placed to demonstrate diffusion at scale; early wins in health and education are the proof of concept for everyone else |
| Government | Can fund supply indefinitely, but cannot book the £55bn–£140bn until firms operationalise; the metric it is judged on is the one it least controls |
| Workers | Skills programmes help individuals, but the dividend depends on employers redesigning roles, which most have not started |
What Britain should actually do next
The supply-side work is largely done or under way. The next phase has to tilt hard towards demand, and towards the firms least able to cross the gap on their own.
For government and policy. Rebalance the scorecard. The Action Plan’s year-two metrics should lead with diffusion — share of firms running AI in production, not merely “using” it — and weight support towards the SME long tail where the gap is widest and the market will not close it unaided. BridgeAI’s expansion to “thousands of businesses” is the right instrument; it should be the headline, not a footnote behind the Growth Zones. Treat public-sector deployments as the national demonstration estate: every NHS or council win is a reusable playbook, and publishing the how, not just the outcome, is the cheapest diffusion lever available.
For boards and executives. Stop measuring AI by tools adopted and start measuring it by processes redesigned. A pilot that never reaches production is a cost, not a capability. Pick one or two workflows on the critical path, fix the data and monitoring underneath them, and run them in production end to end before widening. The firms that win the dividend are not the ones with the most pilots; they are the ones that finished the first one properly.
Take Action: Audit your AI activity by stage, not by count. Sort every initiative into experiment, pilot or production. If almost nothing sits in the production column, the problem is operationalisation — data, process redesign, monitoring — not model access. That is where the next pound should go.
For SMEs. The gap is real but the entry point is smaller than it looks. One well-chosen process — quoting, scheduling, document handling — taken properly into production beats a scattergun of trials. The support exists; the constraint is usually capacity to do the integration work, which is exactly what diffusion programmes are meant to underwrite. Use them.
Four challenges the headline figures hide
Averages flatter, distributions bite. A 25% economy-wide adoption rate hides a chasm between AI-fluent large firms and an SME tail that has barely started. National figures will keep rising on the back of the leaders whilst the laggards fall further behind, and the productivity dividend will concentrate rather than spread unless support is aimed deliberately downward.
“Using AI” is not “running AI in production.” Adoption surveys count firms that have touched a tool. The dividend requires firms that have rebuilt a process around one. The distance between those two states is where most current AI activity is stranded, and no headline adoption number captures it.
Skills without organisational slack go to waste. Upskilling ten million workers does nothing if their employers have no capacity to redesign work around the new skill. The binding constraint is management bandwidth and process change, neither of which a training course supplies.
Supply-side wins are easy to count, so they crowd out demand. Gigawatts and investment pledges make better announcements than the slow grind of SME integration. The political incentive favours the metrics that move fastest, which are precisely the ones least connected to the £55bn–£140bn the strategy is ultimately chasing.
The strategic takeaway
Britain has answered the easy half of its AI question well. It can build, attract and host the capability, and the first-year scorecard proves it. The half The Economist and the OECD both point at — turning that capability into a measurable productivity dividend — is a demand-side problem the state can fund but not deliver, and it is where the opportunity is currently leaking away.
Three things decide whether Britain banks the dividend or just builds the infrastructure for someone else’s:
- Measuring production, not adoption. The metric that matters is the share of firms running AI in production and changing a cost line, not the share that have tried a tool. Year two should be judged on diffusion, not inputs.
- Aiming support at the long tail. The dividend concentrates in firms that can already help themselves. Spreading it means funding the integration work for the SMEs that cannot, where the market will not.
- Using the public sector as the demonstrator. Health, education and local government are proving diffusion at scale. Published, reusable playbooks from those wins are the cheapest way to pull the rest of the economy across the gap.
For UK leaders, the immediate checklist is short:
- Sort your AI initiatives into experiment, pilot and production — and be honest about how little is in the last column.
- Fix the data, process and monitoring under one critical workflow and take it fully live before starting another.
- Treat AI as an operating-model change owned by the business, not a tooling decision owned by IT.
We have argued before that Britain’s regulatory advantage over Europe is real but unearned, and that hosting American compute is not the same as controlling it. This is the same argument from the demand end. The rulebook and the infrastructure decide whether Britain can deploy AI. Diffusion decides whether deploying it is worth anything.
Sources and attribution
This analysis takes its framing from “What Britain needs to do to grasp its big opportunities in AI”, published by The Economist on 17 June 2026. The article sits behind a subscription paywall; the strategic interpretation and all figures here are drawn from public sources rather than reproduced from it.
The first-year figures — 38 of 50 commitments met, the twentyfold compute pledge, £28.2bn via the AI Growth Zones, the Sovereign AI Unit, the skills targets and the public-sector deployments — come from the Department for Science, Innovation and Technology’s AI Opportunities Action Plan: One Year On (January 2026). The productivity range is the OECD’s 2025 estimate; the economy-wide and sector adoption figures are from DSIT and ONS reporting current to December 2025. The Resultsense interpretation — that Britain’s strategy is strong on supply and weak on diffusion — is our own.