What comes next for AI?

The short version

The measured capability of AI rose sharply through 2025 and 2026. By one careful measure, the length of software task an AI agent can complete half the time has doubled roughly every seven months, and the UK AI Security Institute finds a similar rate for cyber security tasks. On real freelance projects, the best agent on one evaluator's leaderboard in September 2026 matched a professional's work about one time in five.

Measured business impact has moved far less. About a third of UK firms with ten or more staff say they use AI, which the ONS describes as relatively shallow adoption, and nine in ten executives surveyed in four countries, including the UK, reported no effect on their firm's productivity or employment over three years. Estimates of what AI will add to productivity growth vary widely: the Office for Budget Responsibility's central case for the UK is around 0.2 percentage points by about 2030, while outside estimates for various countries and periods run from 0.1 to 3.4.

Informed views of the next few years range from gradual diffusion over decades to rapid automation within a few years, and the Bank of England and the IMF both warn of a possible correction after an investment boom that projections cited by the IMF put at $3.4 trillion through 2029. The evidence cannot yet say which future is arriving. The decisions worth taking now are the ones that pay off in each: building on the organisation's own processes and data, keeping integrations and contracts open to a change of supplier, measuring a baseline, and training staff to check output, whatever product they use.

What has measurably changed

One careful measure of AI progress is the length of task a system can complete. The research group METR times human experts on most of its tasks, estimating the rest, then finds the length at which an AI agent succeeds half the time. Across its full series of models, that length has doubled roughly every seven months, and a revised set of tasks published in January 2026 shows the same rate. The UK AI Security Institute, measuring a different domain, finds the length of cyber security tasks models can complete unassisted doubling roughly every eight months.

The caveats come from the evaluators themselves. Succeeding half the time is not the same as completing work dependably. METR's tasks are in software, machine learning and cyber security, and it says most jobs involve other people and outcomes that cannot be scored automatically. It says its estimates above 16 hours are unreliable, and in March 2026 it reported that fixing a modelling mistake had lowered recent models' results by up to 20 per cent.

The Security Institute notes that performance in controlled tests may not reflect effectiveness in the real world, and in April 2026 it reported the first model to complete its 32-step simulated attack on a corporate network, in three of ten attempts, on test networks with no active defenders.

Closer to ordinary work, the evidence is thinner and more modest. The Remote Labor Index, run by Scale AI and the Center for AI Safety, gives AI agents real projects taken from freelance platforms and has people grade the results against the work a professional delivered. On its leaderboard in September 2026, the best agent listed matched the professional standard on about one project in five. Projects that need interaction with people, physical work or a long wait to judge the result are excluded, and Scale AI sells evaluation services to AI developers.

Stanford's AI Index 2026 describes the pattern as jagged: frontier models match or exceed human performance on PhD-level science questions and competition mathematics, while robots succeed at only 12 per cent of real household tasks.

What has not changed yet

Measured effects on businesses have moved far less than measured capability. In June 2026 about 35 per cent of UK firms with ten or more employees told the Office for National Statistics they used at least one AI technology, up from about 12 per cent in late 2023, but the ONS describes that adoption as relatively shallow: the average number of AI technologies per adopting firm has risen only from about 1.4 to 1.6.

Around half of firms using AI say it has not changed their headcount. The survey excludes public administration and defence, publicly provided education and health, and finance and insurance, so it says nothing about AI in those areas.

A survey of about 6,000 executives in the UK, the US, Germany and Australia, published in 2026 by the National Bureau of Economic Research, found that 69 per cent of their firms actively used AI, yet nine in ten reported no effect on their firm's productivity or employment over the previous three years. They expected AI to raise productivity at their firms by an average of 1.4 per cent and to cut employment by 0.7 per cent over the next three, which are forecasts, not findings.

In September 2026 the Bank of England's agents, who talk to firms across the country, reported that automation and AI were influencing role design and replacement hiring but that "quantified productivity gains remain rare", with limited evidence of broad AI-driven job losses.

A large Danish study found no measurable effect of AI chatbots on workers' earnings or recorded hours two years after ChatGPT's launch, ruling out effects larger than 2 per cent, even among daily users and in firms that encouraged use. Its data end in late 2024, before the current generation of agents.

Why capability and measured impact have diverged is not yet established. Adoption takes time, surveys do not separate light use from embedded use, the outcome data lag, and laboratory tasks may transfer poorly to messy, interactive jobs. Each explanation has support. None has been shown to dominate, and which one is right matters a great deal for what happens next.

What it will cost

The price of a given level of AI capability has fallen fast. The research group Epoch AI found in 2025 that the price of reaching a fixed level of performance had been falling by between 9 and 900 times a year, depending on the task, and cautioned that the fastest falls might not persist. Its estimate in February 2026 was "very roughly" a 5 to 10 times reduction each year.

The cost of the work AI is asked to do is a different matter. Reasoning models work through problems at length before answering, and Epoch found in April 2025 that their answers were growing about five times longer each year, against 2.2 times for other models. More computation per task can outweigh a lower price for each unit of it.

In 2025 one AI coding-tool supplier moved its customers from a price per request to usage-based allowances, saying that new models "can spend more tokens per request on longer-horizon tasks", and promised to refund unexpected charges. The Bank of England noted in July 2026 that the rapid expansion of AI capacity was pushing up the price of inputs such as microchips.

For a buyer, the practical consequence is to budget for tasks rather than seats, to expect pricing models to change during a contract, and to route routine work to smaller, cheaper models, a choice our earlier report, What an answer costs, also discusses.

More than one future

Informed views of the next few years differ widely, partly because they describe different things: when AI could do a job, and how fast its effects spread. The AI Futures Project, whose authors wrote the widely read AI 2027 scenario, updated its forecasts in April 2026: the median dates for AI that an AI company would rather use than employ human software engineers moved to mid-2028 for Daniel Kokotajlo and mid-2030 for Eli Lifland. The authors have stressed that they were never confident about 2027, and their August 2026 update said their timelines "haven't changed much".

The chief executive of one frontier developer wrote in January 2026 that powerful AI "could be as little as 1–2 years away, although it could also be considerably further out", and called the continuation of the exponential trend "not certain".

Arvind Narayanan and Sayash Kapoor, whose essay AI as Normal Technology is a widely cited counterweight, expect transformative economic and social effects to be "slow (on the timescale of decades)", even if capabilities keep advancing quickly, because institutions and the rules they work under take time to adapt. In November 2025 the two groups published a statement of common ground. They agreed that before strong AGI, AI will be a normal technology, that it will be a "big deal" either way, and that it is possible that even by 2029 AI will not be usable in settings where failure is unacceptable.

The official economic estimates show the same spread. The Office for Budget Responsibility's central estimate is that AI will gradually add around 0.2 percentage points to UK productivity growth by the end of its five-year forecast, with an optimistic case nearer 0.5, and it thinks a slow start is more likely than early gains. The outside estimates it reviewed, for various countries and periods, run from 0.1 to 3.4 percentage points a year, and the OECD puts the range for high-exposure economies such as the UK at 0.4 to 1.3 points a year over a decade. The differences come mainly from assumptions about how fast AI is adopted and how many tasks it can take on.

In July 2026 the Governor of the Bank of England said "we are between two major cycles" of innovation showing up in the statistics, and a Bank staff blog found some evidence that AI-producing industries are adding more to productivity growth, and a tentative, correlational link with AI adoption elsewhere.

A third possibility runs alongside the other two. The Bank of England's Financial Policy Committee said in July 2026 that the risk of a sharp correction in equity markets "remains high", with valuations more stretched on some metrics and AI-related companies' borrowing accelerating rapidly, though it found little evidence yet that this was crowding out other borrowers. Projections cited by the IMF put AI-related capital spending at $3.4 trillion through 2029, about 70 per cent of it by the largest cloud companies, and the IMF warns that a pullback could bring a synchronised fall in revenue across the whole AI supply chain.

The second-quarter 2026 results of Alphabet, Meta and Amazon showed no cuts: Alphabet raised $49.6 billion in new shares partly to fund AI infrastructure, and Amazon's free cash flow over twelve months turned negative.

Decisions that hold up either way

The evidence cannot yet say which future is arriving, so the decisions worth taking now are the ones that pay off in more than one of them. None requires a bet on a date.

• Build around the organisation's own workflows and data, which it keeps whatever happens to a supplier or a model.
• Keep integrations on open, documented interfaces, so a model or supplier can be changed without rebuilding the work around it.
• Write the way out into the contract: the return of data, notice of price and model changes, and a route to leave.
• Keep an alternative in view. On Epoch AI's index, the best openly available models trailed the best closed ones by about four months in 2026.
• Train people to check and challenge AI output, whichever product they use.
• Measure a baseline and set review dates, so that the next investment decision rests on evidence.

There is official backing for most of this. The UK government's AI Playbook tells public bodies to "consider strategies to avoid vendor lock-in". The EU's rules on switching between cloud providers have applied since September 2025. The Model Context Protocol, a widely used standard for connecting AI tools to other systems, moved to a neutral Linux Foundation body in December 2025, although the maintainers already stewarding it still take the decisions. And the capability gap between open and closed models does not make switching easy: on one large platform, closed models took nearly 80 per cent of usage, according to research reported by MIT Sloan in January 2026.

A note from the founder

Leaders tend to underestimate what these systems can already do, and to overestimate how much of it runs unattended. The capability is real and it compresses work that used to take days. It is still a tool, and it still needs someone accountable for what it produces.

What I have changed my mind about in the last year is the pace. I now expect this to reach further into ordinary work, and sooner, than most plans assume.

My bet for the next two years is that the organisations which do this well will be measurably quicker and cheaper than their competitors at the same quality. What would prove me wrong is a run of projects that improve quality without changing cost, or cut cost without holding quality.

The commitment I would avoid is a general no-code or subscription product that cannot follow your own process as you grow. A tool you cannot shape eventually becomes the limit. I have an interest in saying that, because we build bespoke software, so weigh it against what you already run well.

What we could not establish

Why measured capability and measured business impact have diverged is not established. The candidate explanations, slow adoption, measurement that cannot separate light from embedded use, lagging data and poor transfer from laboratory tasks, all have some support.

We found no independent measurement of how reliable AI agents are in live business use, and no primary evidence of how an investment correction would reach organisations that buy AI services; the effects described in the scenarios are our inference.

Whether the new UK government keeps the AI Opportunities Action Plan, and what replaces it if not, had not been announced by the time of writing.

How it was researched

Every figure is traced to its primary source and dated. Forecasts are labelled as forecasts, with who made them and when, and views are quoted from their authors' own publications, including where the authors disagree. Capability claims come from independent evaluators, not developers' own announcements, and no model or supplier is ranked. An independent check then tested every claim against its source before publication. The report does not predict a date for any capability or outcome.

Sources

Links open the primary source. Forecasts and views are dated because they change.

Capability

1. METR, Time Horizon 1.1, 29 January 2026
2. METR, Task-completion time horizons of frontier AI models, updated May 2026
3. METR, Impact of modelling assumptions on time horizon results, 20 March 2026
4. UK AI Security Institute, Frontier AI Trends Report, 2025
5. UK AI Security Institute, evaluation of a frontier model's cyber capabilities, 13 April 2026
6. Stanford HAI, The 2026 AI Index Report, April 2026
7. Stanford HAI, Inside the AI Index: 12 takeaways from the 2026 report, 13 April 2026
8. Scale AI and Center for AI Safety, Remote Labor Index leaderboard, read 19 September 2026

Adoption and impact

1. Office for National Statistics, Artificial intelligence in UK businesses: 2023 to 2026, 20 July 2026
2. Yotzov, Barrero, Bloom and others, Firm Data on AI, NBER Working Paper 34836, 2026
3. Bank of England, Agents' summary of business conditions, September 2026
4. Humlum and Vestergaard, NBER Working Paper 33777, revised March 2026

Cost

1. Epoch AI, LLM inference price trends, 12 March 2025
2. Epoch AI, How persistent is the inference cost burden?, 16 February 2026
3. Epoch AI, LLM responses to benchmark questions are getting longer over time, 17 April 2025
4. Cursor, Clarifying our pricing, 4 July 2025
5. Bank of England, Monetary Policy Report, July 2026

Views of the future

1. AI Futures Project, Q1 2026 timelines update, 2 April 2026
2. AI Futures Project, Q2.5 2026 timelines update, 16 August 2026
3. AI Futures Project, Clarifying how our AI timelines forecasts have changed since AI 2027, 27 January 2026
4. Narayanan and Kapoor, AI as Normal Technology, Knight First Amendment Institute, 15 April 2025
5. Narayanan and Kapoor, A guide to understanding AI as normal technology, 9 September 2025
6. Narayanan and Kapoor, Why AI hasn't replaced software engineers, and won't, 11 June 2026
7. Asterisk, Common ground between AI 2027 and AI as Normal Technology, 12 November 2025
8. Dario Amodei, The Adolescence of Technology, January 2026

Economic estimates

1. Office for Budget Responsibility, Briefing paper No. 9: Forecasting productivity, November 2025
2. OECD, Macroeconomic productivity gains from artificial intelligence in G7 economies, June 2025
3. Acemoglu, The Simple Macroeconomics of AI, NBER Working Paper 32487
4. Andrew Bailey, Mansion House speech, July 2026
5. Bank Underground, Is artificial intelligence making us more productive?, 6 August 2026
6. IMF, Global economic and financial implications of AI: lessons from a scenario-planning exercise, 2026

Investment and correction risk

1. Bank of England, Financial Policy Committee Record, October 2025
2. Bank of England, Financial Policy Committee Record, July 2026
3. Bank of England, Financial Stability Report, July 2026
4. IMF, Global Financial Stability Report, April 2026, chapter 1
5. European Central Bank blog, The AI boom: rational enthusiasm or the next dot-com bubble?, 17 August 2026
6. Alphabet, second quarter 2026 results, 22 July 2026
7. Meta Platforms, second quarter 2026 results, 29 July 2026
8. Amazon.com, second quarter 2026 results, 30 July 2026

Hedges and policy

1. Epoch AI, Open models lag state-of-the-art closed models by 4 months, 2026
2. MIT Sloan, AI open models have benefits. So why aren't they more widely used?, 20 January 2026
3. European Commission, AI Act: regulatory framework, updated 3 August 2026
4. Linux Foundation, announcement of the Agentic AI Foundation, 9 December 2025
5. Model Context Protocol blog, MCP joins the Agentic AI Foundation, 9 December 2025
6. Government Digital Service, AI Playbook for the UK Government
7. European Commission, Data Act
8. Competition and Markets Authority, Cloud services market investigation
9. UK Parliament, written statement HLWS298, Machinery of Government Changes, 21 July 2026
10. GOV.UK, AI Security Institute
11. AI Opportunities Action Plan: One Year On, 29 January 2026