Better models or better use? Where AI's next gains will come from
AI models are still improving fast, but only about a quarter of US firms and under a third of UK firms use AI at all, and the measured payoff so far is small. The evidence on whether we need smarter models, or businesses that use today's models well.
AI can already write, summarise, answer customers and fix code. So where does the next gain come from: from models that are smarter still, or from businesses actually putting today's models to work? This article puts the two sides of the evidence next to each other: how fast the models are improving, and how far businesses have got in using them.
The short version
- The models are improving fast. The length of task that leading AI agents can complete has been doubling roughly every four months since 2023, according to METR, an independent research group.
- Most businesses are not using them yet. About a quarter of US firms and under a third of UK firms used AI in the latest official surveys. In the EU it was one in five firms with 10 or more employees in 2025.
- Where AI is used, the measured gains are real but narrow. Studies of specific tasks find clear gains; studies of whole jobs and whole economies find small ones so far.
- The Bank of England's reading: AI's contribution to growth so far comes from building AI, "rather than broader gains from their adoption elsewhere in the economy".
How fast are the models still improving?
METR measures AI agents by the length of tasks they can finish, in terms of how long the same task takes a skilled person. Its findings:
- The length of task an agent can complete with 50% success has doubled every 130.8 days since 2023, and every 88.6 days since 2024 (January 2026).
- In February to March 2026 the leading public models reached about 12 hours at 50% success. The most capable unreleased models, which METR says run about two months ahead, were likely at 16 hours or more.
- METR's own test suite "can't reliably measure time horizons above 16 hours". The yardstick has run out of room at the top.
- Reliability is much lower. At 80% success, the leading public models managed tasks of about 1.5 hours, not 12.
Epoch AI, which tracks the inputs, finds the computing power used to train leading models has been "growing at 5× per year since 2020", and the cost of training them by "3.5×/year". In April 2026 it reported that "Three of four metrics show strong evidence of acceleration, driven by reasoning models" - models that work through a problem step by step before answering.
So the evidence does not show models hitting a wall. If anything, the pace picked up in 2024-2026.
How far business has got
Official surveys in the US, UK and EU all ask businesses whether they use AI. Use has roughly tripled in the UK in three years:
UK businesses using AI
- All businesses
- Businesses with 10 or more employees
Show the numbers
| Values (%) | All businesses | Businesses with 10 or more employees |
|---|---|---|
| Sep '23 | 9.4 | 11.9 |
| Dec '23 | 9.9 | 11.8 |
| Mar '24 | 13.8 | 13.8 |
| Jun '24 | 12.7 | 14.8 |
| Sep '24 | 14.7 | 17.8 |
| Dec '24 | 16.1 | 18.2 |
| Mar '25 | 17.8 | 20.6 |
| Jun '25 | 20.5 | 25.1 |
| Sep '25 | 23.5 | 27.2 |
| Dec '25 | 24.7 | 28.7 |
| Mar '26 | 25.9 | 32.1 |
| Jun '26 | 28.9 | 34.9 |
| Sep '26 | 29.7 | 38.1 |
But in every place measured, most businesses still do not use AI at all:
Share of businesses using AI, latest figures
Show the numbers
| Values (%) | Using AI |
|---|---|
| United States, all firms (two weeks to 1 Nov 2026) | 24.2 |
| United Kingdom, all firms (Sep 2026) | 29.7 |
| United Kingdom, 10+ employees (Sep 2026) | 38.1 |
| EU, 10+ employees (2025) | 20 |
| Denmark, 10+ employees (2025) | 42 |
| Germany, 10+ employees (2025) | 26 |
| France, 10+ employees (2025) | 18.2 |
- United States: 24.2% of employer firms used AI in any business function in the two weeks to 1 November 2026, up from 17.3% when the current question began in November 2025. 27.6% expect to use it within six months.
- United Kingdom: 29.7% of businesses use at least one AI technology, and 38.1% of those with 10 or more employees (late September 2026).
- European Union: 20.0% of enterprises with 10 or more employees used AI in 2025, up from 8.1% in 2023. Denmark led at 42.0%.
One change to note in the US figures: since 19 October 2026 the Census Bureau's question asks about "Super Intelligence or SI (formerly referred to as Artificial Intelligence or AI)". In that first round, the share of firms answering "Do not know" rose to 13.9%, from 9.6% two weeks earlier.
Where AI pays off, and where it does not
Studies that measure real work, not opinions, find gains on specific tasks:
- Customer support: in a study of 5,172 support agents, AI assistance raised the number of issues resolved per hour by 15% on average. Less experienced workers gained the most; the most experienced saw "small gains in speed and small declines in quality" (Quarterly Journal of Economics, 2025).
- Professional writing: in an experiment with 453 professionals, ChatGPT cut the time taken by 40% and raised quality by 18% (Science, 2023).
But the picture changes when the task gets harder or the view gets wider:
- Experienced software developers working on their own projects took 19% longer with AI tools in METR's 2025 study. Beforehand they expected a 24% speed-up, and afterwards they "still believed AI had sped them up by 20%". METR has since said the result reflects early-2025 tools.
- Whole jobs in Denmark: linking surveys to official records, economists found "precise null effects on earnings and recorded hours", ruling out effects larger than 2% two years after ChatGPT launched. "What moves is the structure of work."
What the economy shows so far
- Bank of England (November 2026): firms estimate that "AI has increased productivity by around 0.5% over the past three years", which the Bank reads as about 0.1 to 0.2 percentage points a year since early 2024. It adds that AI's contribution to growth "appears to reflect activity in the upstream development and implementation of AI technologies rather than broader gains from their adoption elsewhere in the economy", and that firms "report close to zero impact of AI on employment so far".
- Executives in seven countries (nearly 9,000 surveyed in summer 2026 for an NBER working paper): AI "raised annual labor productivity growth by 0.2% over the last 3 years", and they "expect a substantially larger 0.6% per annum boost over the next 3 years". In that survey 78% of firms use AI, a much higher share than the official statistics above; the two ask different questions of different samples.
In plain terms: the money and the growth so far are in building AI. The payoff from using it is only starting to show.
What holds businesses back
The UK survey asks businesses what has stopped or delayed them adopting AI in the last three months. In the latest round (late September 2026), across all businesses:
- Not enough expertise in AI: 9.3%.
- Lack of trust in the safety or transparency of AI products: 9.2%.
- Cost: 7.7%.
- Security concerns: 6.8%.
- Difficulty identifying a use for it: 6.2%.
- 49.0% said nothing had stopped them, and 21.8% had not tried to use AI at all.
Each obstacle is named by fewer than one business in ten. The largest groups report no obstacle at all, or have simply not tried. The Danish study adds that where AI is adopted, "employers absorb AI through task reorganization" - the change shows up first in how work is organised, not in pay or hours.
So do we need better models?
The case for better models:
- Reliability. The gap between what agents can do half the time (about 12 hours of work) and what they can do four times out of five (about 1.5 hours) is wide. A business handing over a whole job needs the second number.
- Harder work. The gains in the studies above are largest on well-defined tasks and for less experienced workers. Experienced developers on complex projects were slower, not faster, in early 2025.
- Automation needs more than help. In Anthropic's own usage research, about three-quarters of business use through its API was classed as automation, against 45% on its Claude.ai app, where more of the use was people working alongside the model. Automating a task asks more of a model than assisting a person.
The case for better use:
- Most businesses have not adopted the models that already exist. Even among UK firms with 10 or more employees, more than 60% do not report using AI.
- The barriers businesses name are mostly about skills, trust, cost and finding a use.
- Where AI has been studied carefully, today's models already deliver gains on the right tasks. The economy-wide payoff is small because use is still narrow, not because the tools do nothing.
The verdict
This is our reading of the evidence, not a forecast and not a view on any company's shares.
For now, the bigger gap is on the business side. The models are already well ahead of how most businesses use them, and the payoff will depend more on adoption than on the next leap in intelligence. Where better models matter most is reliability, not raw ability.
The models keep improving on every independent measure, and the pace has picked up rather than slowed. But about three in four businesses in the US and seven in ten in the UK do not report using AI at all, the barriers they name are mostly about people and cost, and the measured economic payoff so far is a fraction of a percentage point a year. A smarter model does not help a business that has not started. What would help those that have is a model they can trust to finish long tasks without supervision: the 80% horizon, not the 50% one.
What would change this reading:
- Adoption jumps - for example, UK use passing half of all businesses, or the EU's 2026 figure (due 15 December 2026) rising much faster than before.
- Productivity statistics move - official figures showing gains well beyond 0.1 to 0.2 points a year.
- Reliability catches up - agents completing long tasks four times out of five, which would make full automation, not just assistance, practical.
What to watch
- The US Census Bureau's business survey, every two weeks.
- The UK's next survey wave with the AI questions.
- Eurostat's 2026 figures for AI use in EU enterprises, due 15 December 2026.
- Productivity releases and the Bank of England's next Monetary Policy Report.
- METR and Epoch AI's next measurements of model progress.
What these numbers do and do not show
Every figure here comes from a government statistics office, a central bank, a peer-reviewed or working paper, or an independent research group, and each one is listed under Sources.
- The surveys differ. The US counts employer firms of every size and asks about the last two weeks; the UK counts all trading businesses; the EU counts enterprises with 10 or more employees over a year and leaves out the financial sector. The bar chart is not a like-for-like ranking.
- The UK shares are worked out from the ONS table as 100% minus businesses that do not use AI and those not sure, so they can differ from ONS's own rounded figures by a point.
- The US question changed in November 2025 (to "any business function") and again in wording in October
- We use only the series since November 2025.
- Studies measure specific tasks and periods. The software study used early-2025 tools; the writing and customer-support studies used 2023-era tools.
- METR's data is quoted, not charted: its site does not allow republishing it.
- Anthropic's usage research is published by Anthropic about its own products.
Disclosure: this article was drafted with help from Claude, an AI model made by Anthropic, whose research and models are among those discussed above. Anthropic has no part in what we publish; every fact here is taken from the sources listed under Sources.
Drafted with help from an AI tool, then checked and edited by a person. Sources are listed below.
Sources (14)
- METR, Time Horizon 1.1 (29 January 2026) - checked
- METR, Frontier Risk Report (19 May 2026) - checked
- METR, Measuring the impact of early-2025 AI on experienced open-source developer productivity (10 July 2025) - checked
Show all 14 sources
- Epoch AI, Trends in artificial intelligence - checked
- Epoch AI, Have AI capabilities accelerated? (16 April 2026) - checked
- Brynjolfsson, Li and Raymond, Generative AI at Work, Quarterly Journal of Economics (2025) - checked
- Noy and Zhang, Experimental evidence on the productivity effects of generative artificial intelligence, Science (2023) - checked
- Humlum and Vestergaard, NBER Working Paper 33777 (revised March 2026) - checked
- Yotzov, Barrero, Bloom, Davis and others, Firm Data on AI, NBER Working Paper 34836 (revised October 2026) - checked
- Bank of England, Monetary Policy Report, November 2026 - checked
- Office for National Statistics, Business insights and impact on the UK economy, wave 165 (Open Government Licence v3.0) - checked
- US Census Bureau, Business Trends and Outlook Survey, national data downloads - checked
- Anthropic, Anthropic Economic Index report (January 2026) - checked
- Eurostat, Use of artificial intelligence in enterprises (isoc_eb_ai, CC BY 4.0) - checked