What the latest research tells us about AI adoption

AI adoption is rising quickly. Returns are rising more slowly and unevenly. The most credible recent research suggests the gap has less to do with the technology than with how organisations manage it.

This article summarises the strongest recent evidence on how organisations and employees use AI, where returns are falling short and why governance is associated with better results. Each finding carries a rating, so you can see how much weight it can bear.

How to read AI research

Much of what is published about AI adoption is based on surveys of opinion, often by organisations with something to sell. That does not make it wrong, but it does mean it should be read with care. We have rated each source on three questions: who was asked and how many, whether it measures behaviour or opinion, and how independent it is.

• Strong. What it means: Official statistics, administrative data or controlled studies. Measures what happened, with a transparent method.
• Moderate. What it means: Large, clearly described surveys. Useful, but based on what people report rather than what can be measured.
• Indicative. What it means: Small, selective or opaque samples. Worth knowing as a signal, not as a precise figure.

1. Adoption is rising, but it is shallow

Office for National Statistics, July 2026. About 35% of UK businesses with ten or more employees use at least one AI technology, almost three times the figure in late 2023. Among businesses with 250 or more employees, the figure is 49%. Use is not deep, though: only 10% of adopting businesses report extensive use, and the average number of AI technologies per adopter has risen only from about 1.4 to 1.6.

Evidence rating: Strong. Official statistics from 38,637 business responses in June 2026. The survey excludes several sectors, including finance and insurance, and public administration, education and health.

UK government AI Adoption Research, January 2026. Of 3,500 UK businesses surveyed in spring 2025, 16% used at least one AI technology, ranging from 36% of large businesses to 14% of micro businesses. Among adopters, 75% reported improved workforce productivity, but 77% had seen no change in revenue yet.

Evidence rating: Strong. Government-commissioned, large and weighted. Fieldwork predates the ONS figure, and respondents were usually technology leads, so the report notes it does not capture shadow AI.

The two UK figures differ because they were collected at different times, from different populations and with different definitions of AI. The direction is the useful part: adoption is spreading, larger firms lead and most use remains light.

McKinsey, State of AI, August 2026. Nearly nine in ten respondents report regular use of AI in at least one business function.

Evidence rating: Moderate. 1,719 participants in 97 countries, surveyed May to June 2026. A large sample, but self-reported and drawn from organisations engaged enough to respond.

2. Employees are ahead of their employers

KPMG and the University of Melbourne, 2025. 65% of UK workers say they intentionally use AI at work. 38% admit using it in inappropriate ways, and only 27% of UK respondents say they have had any AI education or training.

Evidence rating: Moderate. 1,029 UK respondents, including 617 workers, surveyed November 2024 to January 2025, as part of a 47-country academic study. Self-reported.

IBM, Cost of a Data Breach, July 2026. Shadow AI, meaning staff use of unapproved tools, was involved in 43% of the security incidents studied, up from 20% a year earlier.

Evidence rating: Moderate. 602 breached organisations across 16 countries. Figures are taken from independent summaries, as the full report requires registration.

MIT Project NANDA, 2025. Workers at over 90% of companies regularly use personal AI tools for work, while only 40% of companies say they have bought an official subscription.

Evidence rating: Indicative. Based on 52 interviews, 153 conference survey responses and a review of public AI initiatives. Commentators have questioned how its headline figures were calculated.

3. Where returns are falling short

McKinsey, August 2026. Only 37% of respondents attribute any EBIT impact to AI, and about 6% qualify as high performers, attributing at least 5% of EBIT to AI. Both figures are essentially unchanged from the previous year.

Evidence rating: Moderate. As above.

Humlum and Vestergaard, National Bureau of Economic Research, 2025. Linking surveys of AI chatbot use to Danish administrative records, the authors found precise null effects on earnings and recorded hours, ruling out effects larger than 2% two years after ChatGPT’s launch. They also found that new tasks, including AI oversight and integration, were widespread.

Evidence rating: Strong. Administrative data rather than opinion. The limits are that it covers Denmark, chatbots specifically and the early period of adoption.

METR, July 2025. In a randomised controlled trial, 16 experienced software developers took 19% longer to complete 246 real tasks when allowed to use AI tools. Before the study they expected to be 24% faster, and afterwards they still believed they had been about 20% faster.

Evidence rating: Strong design, narrow scope. METR itself describes the result as a snapshot of early-2025 tools in one setting, and its later attempt to repeat the study produced data it judged unreliable. The lasting lesson is the gap between perceived and measured productivity.

BetterUp Labs and Stanford Social Media Lab, September 2025. 40% of US workers surveyed had received AI-generated “workslop”, meaning AI-generated work that looks finished but has to be fixed, queried or redone, in the past month. Each instance took nearly two hours to deal with, at an estimated cost of about US$186 per employee per month.

Evidence rating: Moderate to indicative. Over 1,100 US full-time workers. Time and cost figures are self-reported estimates.

Read together, these studies point to five recurring leaks: pilots that never scale, productivity that is perceived but not measured, time saved that is absorbed by new tasks, output that has to be fixed and use that happens outside any programme anyone is measuring.

4. The picture is improving for some

BCG, Applied AI Index, September 2026. 7.5% of companies are “future-built” and generating value from AI at scale, and a further 41% are scaling AI and outperforming, so that nearly half are creating real value. Future-built companies report 2.4 times the revenue growth and 2.8 times the EBITDA growth of laggards. BCG’s 2025 study, a different survey, had put the leading group at about 5%, with around 60% gaining no material value.

Evidence rating: Moderate. 1,330 senior executives. A large sample, but self-reported, and from a consultancy with a commercial interest in AI transformation.

The improvement is real but concentrated. The organisations pulling ahead are not using different technology. According to BCG, they run AI as an enterprise-wide programme, redesign workflows end to end, invest in governance and the workforce alongside the technology, and measure the impact in their profit and loss.

5. Why governance improves the odds

Across the studies, the practices most associated with returns are governance practices in the full sense of the word: ownership, measurement, workflow design, controls and training.

• Leadership and measurement. McKinsey’s 2026 high performers are twice as likely to report visible senior leadership commitment and defined processes for measuring AI’s impact, and nearly three-quarters have fundamentally redesigned workflows, against a quarter of others. In its 2025 survey, CEO oversight of AI governance was one of the elements most correlated with bottom-line impact.
• Risk controls. The same high performers are much more likely to be mitigating technical vulnerabilities and unintended actions by AI. EY’s 2025 survey of 975 senior leaders found that real-time monitoring and oversight committees had statistically significant, positive effects on revenue growth and cost savings.
• Readiness for agents. BCG found that 42% of companies expect to give AI agents decision-making authority by 2030, but only 5% currently have the full set of controls needed to do so safely.

These are correlations, and the organisations that govern well may also be better run in other ways. But the mechanism is plausible, because each governance practice closes one of the leaks described above.

• People think AI is helping but nobody measures it. The governance response: Set baselines and measure outcomes, not opinions. What the research shows: Developers in the METR trial believed they were faster when they were slower. McKinsey’s high performers are twice as likely to have defined processes for measuring AI impact.
• Pilots never scale. The governance response: Give AI a senior owner and an enterprise-wide plan. What the research shows: CEO oversight of AI governance was among the factors most correlated with bottom-line impact in McKinsey’s 2025 survey.
• Time saved is spent fixing poor output. The governance response: Set quality standards, review rules and training. What the research shows: Recipients of “workslop” report spending nearly two hours dealing with each instance.
• Use happens outside approved tools. The governance response: Find out what people actually use and offer a better approved route. What the research shows: Shadow AI was involved in 43% of the incidents in IBM’s 2026 study.
• Agents are given authority before controls exist. The governance response: Put oversight, audit and rollback in place before autonomy. What the research shows: Only 5% of companies in BCG’s 2026 index have the full set of controls, while 42% expect to give agents decision-making authority by 2030.

What this means for your business

1. Measure, don’t ask. Before rolling out a tool, record how long the task takes and how good the output is. Compare afterwards. Staff impressions are useful, but the METR trial shows they can point the wrong way.
2. Count the hidden work. Include the time spent checking, correcting and redoing AI output when you calculate return.
3. Bring shadow use into the light. Find out which tools people actually use and why. If the approved option is worse, they are likely to keep using the unapproved one.
4. Redesign one workflow properly. The evidence favours changing one process end to end over adding AI to many processes at the edges.
5. Give it an owner and a number. A senior owner, and a measure that appears in the management accounts, turn AI from an experiment into a programme.
6. Put controls before autonomy. Before giving an agent the authority to act, make sure you can see what it does, stop it and reverse it.
7. Read research critically. Ask who was asked, how many, whether it measured behaviour or opinion, and who paid for it.

The research does not say AI fails to deliver. It says AI delivers for organisations that manage it deliberately, and that the number of those organisations appears to be growing.

Sources checked October 2026. Evidence ratings are Turma’s editorial assessment, based on each publisher’s description of its own method. Survey findings are not directly comparable across studies, and several findings are correlations rather than evidence of cause.

This article provides general information and does not constitute investment, financial or professional advice.

References

UK official and government research

• Office for National Statistics, Artificial intelligence in UK businesses: 2023 to 2026, 20 July 2026: https://cy.ons.gov.uk/businessindustryandtrade/business/businessservices/articles/artificialintelligenceinukbusinesses/2023to2026
• Department for Science, Innovation and Technology (IFF Research and Technopolis Group), AI Adoption Research, January 2026: https://www.gov.uk/government/publications/ai-adoption-research/ai-adoption-research

Surveys of organisations and workers

• McKinsey & Company, The state of AI, 25 August 2026: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
• McKinsey & Company, “The state of AI: how organizations are rewiring to capture value”, 12 March 2025: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-how-organizations-are-rewiring-to-capture-value
• BCG, “AI starting to pay off as companies generate value” (Applied AI Index 2026), 30 September 2026: https://www.bcg.com/press/30september2026-ai-starting-to-pay-off-companies-generate-value
• Entrepreneur India, “Nearly half of companies now generate real value from AI: BCG report”, 1 October 2026: https://india.entrepreneur.com/?p=100173
• Computer Weekly, “Companies reap the rewards of being AI-first” (BCG 2025 study): https://www.computerweekly.com/news/366631942/Companies-reap-the-rewards-of-being-AI-first
• KPMG UK, “UK attitudes to AI” (KPMG and University of Melbourne, Trust, attitudes and use of AI: a global study 2025): https://kpmg.com/uk/en/insights/ai/uk-attitudes-to-ai.html
• EY, “How can responsible AI bridge the gap between investment and impact?”, 8 October 2025: https://www.ey.com/en_ch/insights/ai/how-can-responsible-ai-bridge-the-gap-between-investment-and-impact
• Virtualization Review, “MIT report finds most AI business investments fail, reveals GenAI divide”, 19 August 2025: https://virtualizationreview.com/articles/2025/08/19/mit-report-finds-most-ai-business-investments-fail-reveals-genai-divide.aspx
• BigDATAwire, “MIT report flags 95% GenAI failure rate, but critics say it oversimplifies”, 4 September 2025: https://www.hpcwire.com/bigdatawire/2025/09/04/mit-report-flags-95-genai-failure-rate-but-critics-say-it-oversimplifies/
• Allwork.Space, “AI workslop is upsetting coworkers and costing companies millions”, September 2025: https://allwork.space/2025/09/ai-workslop-is-upsetting-coworkers-and-costing-companies-millions/

Measured and controlled studies

• Humlum and Vestergaard, NBER Working Paper 33777, 2025: https://ideas.repec.org/p/nbr/nberwo/33777.html
• METR, “Measuring the impact of early-2025 AI on experienced open-source developer productivity”, 10 July 2025: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
• METR, “We are changing our developer productivity experiment design”, February 2026 (republished by Conffab): https://conffab.com/?p=10163

Security and governance

• Baker Donelson, “Ten takeaways from IBM’s 2026 Cost of a Data Breach Report”, 30 July 2026: https://www.bakerdonelson.com/ten-takeaways-from-ibms-2026-cost-of-a-data-breach-report
• Sentra, blog on shadow AI in IBM’s Cost of a Data Breach 2026 report, 13 August 2026: https://sentra.io/blog/ibm-cost-of-a-data-breach-2026-shadow-ai