Britain gave its AI minister a place at the Cabinet table, then heard a much smaller number for AI’s economic contribution. Kanishka Narayan now estimates £55 billion to £140 billion by 2030, against a £400 billion estimate from a Google-led report that officials had previously touted. The top of his range is £260 billion lower.

Key takeaways

  • Kanishka Narayan forecast that AI will add £55 billion to £140 billion to the UK economy by 2030.
  • UK businesses with at least 10 employees reporting AI use rose from 12% in September 2023 to about 35%, while average tool use rose from 1.4 to 1.6.
  • Only 11% of 396 surveyed businesses said they could forecast their AI spending.
  • Across more than 6,800 tasks, Microsoft found lower-priced models cost more than higher-priced models in 32% of cases.
  • Tussell recorded £573 million in UK government AI-related contracts by July 15, 2025, versus £468 million during all of 2024.

A promotional forecast could help make AI an economic priority without explaining how the gain would be delivered. Once ministers make AI part of a growth strategy, they inherit a different obligation: to show how business adoption, net productivity and implementation costs turn a supplier’s modeled upside into growth that public institutions can plan around.

The Cabinet seat gave a borrowed estimate an owner

In July, the UK government elevated Narayan to attend Cabinet while he continued to oversee the AI Security Institute. Narayan could now bring AI policy to Cabinet discussions about public services and economic strategy. By November, Narayan’s lower estimate exposed a distinction that matters in that setting: officials had touted the £400 billion Google-led figure, but the available coverage does not establish that the Treasury had adopted it as a budget assumption.

Nor does the smaller range prove the larger estimate impossible. Researchers cited slow UK adoption in explaining Narayan’s forecast; the coverage does not provide enough of the two models’ assumptions to reconcile them line by line. What ministers can no longer leave implicit is the route from a possible technological gain to a national one. Cabinet attendance, granted to the minister in July, makes that route a question for government rather than for a report sponsor alone.

The Treasury Committee had already shown how quickly AI claims become questions of public responsibility. In January, it warned that the government, Bank of England and Financial Conduct Authority’s wait-and-see approach to AI risks in finance could expose the UK to serious harm. That warning concerned downside risk, not a GDP forecast. In both cases, the institution making decisions needs evidence it can inspect before the consequences arrive.

Business adoption counts reach, not redesigned work

In the coverage record, UK government data show AI use among businesses with at least 10 staff rising from 12% in September 2023 to about 35%. Average tool use rose from 1.4 to 1.6. The first change is substantial evidence of diffusion. The second gives a narrower view of how deeply tools have entered firms’ work.

A national forecast has to cross the distance between those measures. A firm may give staff a tool without changing who checks its output, which decisions it informs, or what happens to time saved. Prediction can become cheaper while the judgment and workflow around it remain the same. Enterprise AI adoption creates more economic value when firms change those surrounding systems, not merely when another employee gains access.

There is evidence that firms are making that harder transition. SAP, Capgemini, Sopra Steria and OVHcloud reported stronger AI demand as European customers moved from experiments toward deployment. The general-purpose-technology J-curve offers a reason costs might precede visible gains: organizations must invest in complementary changes before new tools alter measured output. That is a reason to track implementation, not to treat every delayed return as a return already earned.

Firms cannot yet price the unit a national model adds up

Even a company that has redesigned a workflow needs to know what the new workflow costs. In a survey of 396 businesses, only 11% could forecast AI spending. Microsoft found that lower-priced models cost more than higher-priced ones on 32% of more than 6,800 tasks. A model’s posted price did not reliably settle its cost per useful task; the task and the work needed to complete it mattered too.

Returns are similarly difficult to establish. An IBM survey of 2,000 CEOs found that 25% of AI initiatives had delivered their expected return on investment and 16% had scaled across the enterprise. Researchers have also identified “workslop”: passable AI-generated material that saves its creator effort by giving someone else more to check or repair. A firm counting drafts produced without counting recipients’ rework can mistake transferred labor for saved labor.

OpenAI has described productivity benefits from its tools, while launching a 12-month study in July 2025 to assess their broader impact. The study’s existence does not negate those benefits. It marks the distance between a useful tool in a worker’s hands and a defensible estimate of its contribution across an economy.

Investment reaches the ledger before productivity does

The Office for National Statistics reported more than £11 billion of UK digital-infrastructure investment in 2025 amid AI data-centre growth. Separately, Tussell recorded £573 million in UK government AI-related contracts by July 15, 2025, compared with £468 million during all of 2024. These measures cover different spending; neither is a measure of the productivity that spending ultimately yields.

Suppliers are paying for the organizational work as well as the machines. AWS committed $1 billion in resources to an AI-focused forward-deployed-engineer organization, and OpenAI launched a deployment company backed by more than $4 billion to help customers build systems. Those commitments give implementation a budget and a workforce. Alongside contracted compute capacity, they also make the expenditure visible well before anyone can isolate its contribution to national output.

A growth print is evidence, not the whole return

The ONS cited AI as a major driver of Britain’s summer growth spurt, when GDP grew 0.4% in July 2026, above expectations. That July reading is genuine counter-evidence to any claim that AI’s macroeconomic effects cannot yet be observed. It cannot, by itself, establish how much of Narayan’s projected 2030 contribution will come from lasting productivity improvements rather than investment or a short period of stronger activity.

Frequently asked questions

When would OpenAI’s 12-month productivity study have been expected to reach its planned end point?

OpenAI launched the study in July 2025, so its 12-month period would have reached July 2026. The evidence provided does not identify any published findings from it.

Can the UK’s £11 billion digital-infrastructure investment and £573 million in AI-related government contracts be added together as a single AI-spending total?

No. The figures cover different categories of spending, and the evidence does not establish whether they overlap. Neither figure measures the productivity ultimately produced by that expenditure.

How much did average AI-tool use rise among UK businesses covered by the government data?

It increased by 0.2 tools per business, from 1.4 to 1.6—about a 14% rise. That measure records tool use, not whether firms redesigned workflows or achieved a measured productivity gain.

Competing estimates of AI’s UK economic contribution by 2030

Estimate sourceProjected contribution to UK economyGap versus £400B figure
Kanishka Narayan — low end£55B£345B lower
Kanishka Narayan — high end£140B£260B lower
Google-led report£400B—

A cross-party committee of MPs has questioned the government’s plans to improve public services with AI, citing outdated technology, poor data and insufficient digital skills. Those are identifiable obstacles against which a public project can be tested. Ministers can count a contract when they sign it and a data centre when it is built; they have to follow the changed service or business process to find out whether either made work more productive. The £400 billion figure belonged on a presentation slide. Narayan’s smaller range has to survive contact with the ledger.