The State of Applied AI

In early 2026, 81% of U.S. physicians reported using AI tools in their practice — up from 38% in 2023, more than doubling in three years.

The FDA had authorised 1,524 AI-enabled medical devices as of 30 March 2026 (about 76% of them in radiology) — up from 1,451 at end-2025 and 1,247 in mid-2025 — with the list growing each quarter (a further update followed in July 2026).

Kaiser Permanente (The Permanente Medical Group): an ambient AI scribe listens to the visit, with patient consent, and drafts the clinical note into the record. A peer-reviewed one-year evaluation (NEJM Catalyst, 2026) reported 7,260 physicians using it across 2,576,627 encounters (Oct 2023-Dec 2024), saving an estimated 15,791 documentation-hours — one of the largest real-world ambient-AI deployments to date.

In a cluster-randomised trial published in Nature Medicine (2026), an LLM clinical decision-support copilot ('AI Consult') across 16 Kenyan primary-care clinics (>9,600 patients) significantly improved documentation and treatment-planning quality and lowered antibiotic-related costs, but produced no significant change in patient outcomes — 14-day treatment failure was 2.2% vs 2.0%, with no difference in hospitalisation, death or satisfaction. The first RCT-grade evidence that an LLM copilot can improve clinician process without (yet) moving hard patient outcomes.

UpDoc became the first Software as a Medical Device cleared by the FDA with a patient-facing large-language-model interface (510(k) K253281, cleared 23 December 2025) — a prescription tool for insulin management in adults with type 2 diabetes, in which an LLM front-end sits over deterministic dosing logic. It retires the common 'no generative-AI/LLM medical device has been authorised' claim, though reviewers note its 'AI doctor' marketing overstates what is a bounded, single-purpose device.

ECRI ranked AI-driven diagnostic error as the #1 patient-safety concern for 2026 in its annual Top 10 list (published March 2026), citing inconsistent real-world performance — the first time AI has topped the list from a major patient-safety authority.

A cluster-randomised trial (Scientific Reports, February 2026; 10,422 inpatient visits on an 85-bed cardiology/cardiac-surgery ward) found that a passive AI risk-trajectory display, plus staff education, did not improve patient outcomes — trial-grade evidence that a deployed AI-monitoring tool can fail to help.

As of the Bank of England and FCA's 2024 survey, 75% of UK financial services firms reported already using AI, up from 58% in 2022, with a further 10% planning to adopt it within three years (118 firms surveyed).

Machine-learning AI helped the US Treasury recover US$1 billion by speeding up detection of Treasury check fraud in fiscal year 2024, part of a record US$4 billion in fraud and improper payments the department prevented and recovered through its wider technology- and data-driven approach, up from US$652.7 million the previous year.

JPMorgan Chase built LLM Suite, its own secure generative-AI platform, and onboarded 200,000 employees within eight months of its summer-2024 launch (American Banker 2025 Innovation of the Year). By 2026 LLM Suite reaches over 230,000 employees globally, used firm-wide for drafting, summarising and idea generation that saves users several hours a week.

Commonwealth Bank of Australia deployed an in-house 'agentic' AI fraud-detection system (built in about three months) that reduced customer fraud and scam losses by more than 20% in H1 2026 versus H1 2025; the AI now develops or updates around three-quarters of the bank's card-fraud rules under human-in-the-loop approval, across more than 80 million signals a day.

The UK Financial Conduct Authority launched 'AI Live Testing' — a voluntary supervisory service (~5-10 firms per cohort) letting firms trial live AI proofs-of-concept with the regulator (first cohort October 2025, second launching April 2026), part of the FCA's AI Lab.

The US Treasury's Do Not Pay service prevented, detected or recovered US$11.7 billion in fraud and improper payments in FY2025 — up 63% from US$7.2 billion in FY2024, a record for the programme (614 million data-source queries in FY2025). A companion to the FY2024 machine-learning check-fraud figure; FY2025 reporting does not isolate the ML-only line.

Generative AI usage in the legal sector nearly doubled year-over-year, from 14% in 2024 to 26% in 2025, according to Thomson Reuters' survey of nearly 1,800 professionals across legal, tax, risk and government.

30% of US lawyers reported using AI-based technology tools in 2024, up from 11% in 2023, with adoption highest at large firms (46% at firms of 100+ attorneys).

61% of UK lawyers were using generative AI in their day-to-day work by September 2025, up from 46% in January 2025.

Ashurst (global law firm): In firm-wide generative AI trials (November 2023 to March 2024) involving 411 partners, lawyers and staff across 23 offices in 14 countries, Ashurst measured first-draft time savings of about 80% on UK corporate filings, 59% on sector research reports and 45% on legal briefings. The firm stressed that human review remained essential (its lawyers' outputs still out-scored AI on accuracy) and framed the gains as freeing lawyers to focus on higher-value work.

As of late 2025, 54% of U.S. teens aged 13–17 say they have used AI chatbots to get help with their schoolwork.

Three in ten (30%) U.S. public K–12 teachers report using AI tools at least weekly, and those who do estimate they save 5.9 hours a week — about six weeks across a school year.

In 2026, 95% of UK undergraduates report using AI in at least one way and 94% use generative AI for assessed work — up from 66% overall use in 2024.

World Bank / Edo State, Nigeria: in a six-week after-school randomised trial, about 800 senior-secondary students used a generative-AI tutor (Microsoft Copilot / GPT-4) under teacher supervision and improved by ~0.31 standard deviations overall (0.23 SD on English) — gains the World Bank framed as close to two years of typical learning — and did better on their regular end-of-year exams. Now documented in a formal working paper.

In a randomised controlled trial of about 1,000 Turkish high-school maths students (Bastani et al., PNAS, 2025), students given an unguarded GPT-4 assistant scored about 17% worse on subsequent AI-free exams than peers who never had AI access — while a guardrailed 'tutor' version that withheld answers eliminated the harm. The strongest causal evidence that unscaffolded AI use can degrade learning.

In the 2026 Gallup/Walton 'Teaching for Tomorrow' survey (n=2,069 US K-12 teachers, Feb-Mar 2026), only 18% reported receiving any formal guidance on using AI, and fewer than 1 in 10 had guidance for any specific task (tutoring, grading, feedback) — with higher-needs schools lagging wealthier ones. Adoption is running well ahead of policy.

A 2026 peer-reviewed evaluation of AI-text detectors (International Journal for Educational Integrity) found false positives were rare, but most detectors substantially under-detected AI-generated and 'humanised' text — only one tool performed reliably — leading the authors to conclude detectors are unfit as sole evidence of academic misconduct. The real risk is false negatives, not false accusations.

The OECD's flagship Digital Education Outlook 2026 (January 2026) foregrounds the tension between AI's short-term task success and its effect on learning retention, and compiles national data on AI adoption in schools across member countries — the authoritative policy frame for education AI in 2026.

In 2025, 84% of developers were using or planning to use AI tools in their development process (up from 76% in 2024), and 51% of professional developers now use them daily.

In Google Cloud's 2025 DORA report (nearly 5,000 technology professionals, 90% of whom use AI at work), AI adoption was positively associated with software delivery throughput but continued to be negatively associated with software delivery stability.

Developer trust in AI is falling even as adoption rises: in 2025, 46% of developers actively distrusted the accuracy of AI tools versus 33% who trusted them, and favourable sentiment toward AI tools fell to 60% (from over 70% in 2023 and 2024).

Google (Alphabet): AI now generates a large and rising share of new code at Google, with human engineers reviewing and approving it. CEO Sundar Pichai's stated figure has climbed from 'more than a quarter' (Q3 2024) to about 50% (late 2025) to 75% of all new code at Google as of Google Cloud Next in April 2026 — a vendor self-claim with a loose definition of 'AI-generated', but Google's own on-record statement.

The US Federal Trade Commission issued 6(b) study orders to seven companies — OpenAI, Character Technologies (Character.AI), Meta, Instagram, Snap, xAI and Alphabet — on 11 September 2025 (3-0 vote), examining harms to children and teenagers from AI 'companion' chatbots, amid reports of safety guardrails weakening in long emotional conversations. It is the defining AI-safety incident cluster of 2025-26, running alongside wrongful-death litigation (Raine v. OpenAI, filed August 2025).

Veracode's 2026 GenAI Code Security Report found that roughly 44% of AI code-generation tasks introduced a security vulnerability, with an average security 'pass rate' of about 56% — essentially flat on the prior year (55%). AI-generated-code security has not improved even as the volume of AI-written code surges.

Google Cloud's 2026 DORA research ('The ROI of AI-Assisted Software Development') describes a 'J-curve' of value from AI adoption — an early productivity dip on rollout before gains — and reaffirms that AI amplifies an organisation's existing strengths and weaknesses rather than uniformly boosting performance.

In a randomised controlled trial by METR (16 experienced open-source developers, 246 real tasks, using Cursor Pro with Claude 3.5/3.7 Sonnet), developers were about 19% slower when allowed early-2025 AI tools — even though they had forecast being 24% faster and believed afterward they had been 20% faster. The canonical rigorous rebuttal to self-reported AI productivity gains.

78% of surveyed organisations reported using AI in 2024, up from 55% in 2023.

Across Dec 2025–May 2026, overall AI use among U.S. businesses ran 17–20%, rising to 37% among firms with at least 250 employees.

DBS Bank: DBS Bank (Singapore) scaled AI/ML and data-analytics across more than 430 use cases powered by over 2,000 models, spanning risk management, controls and back-office productivity. In 2025 these initiatives generated approximately S$1 billion in economic value, up from S$370 million in FY2023.

43% of organisations now use AI in HR tasks, up from 26% in 2024, according to SHRM's survey of 2,040 HR professionals.

37% of organisations are actively integrating or experimenting with generative AI in hiring, up from 27% a year earlier.

Recruiters who use generative AI in hiring report saving about 20% of their workweek — roughly a full workday.

Chipotle Mexican Grill (recruiting AI built by Paradox): Chipotle deployed Paradox's conversational AI hiring assistant, 'Ava Cado,' across its 3,500-plus US restaurants to answer applicant questions, collect details, schedule interviews and send offers. Chipotle reported the tool nearly doubled applications and cut the time from application to starting the job from 12 days to four — reducing time-to-hire by up to 75% — while freeing restaurant managers to focus on operations as it scaled hiring for peak season.

As of CMI/MarketingProfs' 2026 research cycle (fieldwork mid-2025, n=1,015 B2B marketers), 89% of B2B marketers use AI to generate marketing copy and 95% say their organisation uses AI-powered applications — up from 81% team gen-AI use the year before. (The metric wording shifted from 'teams use generative AI' to 'org uses AI apps'; the 89% copy-generation figure is the closest like-for-like successor.)

Nearly 90% of marketers report having used generative AI tools at work, per a September 2024 survey of over 1,000 marketers.

US 'AI advertising' spending — ads placed in and around AI-powered results such as Google's AI Overviews and chatbots — is forecast to more than double from US$32.03 billion in 2026 to US$68.25 billion in 2030 (about 2.1x), with over 80% flowing to ads next to AI content; in-chatbot ads stay under US$1 billion in 2026, rising to roughly US$5 billion by 2030.

Klarna: Klarna used generative AI image tools (Midjourney, DALL-E, Adobe Firefly) plus an internal 'Copy Assistant' that now produces about 80% of its marketing copy. The company reported cutting image-production costs by around $6 million, compressing the image-production cycle from six weeks to seven days, and attributing roughly $10 million in annualised marketing savings to AI while reducing external marketing-agency spend by 25% (company-reported figures, May 2024). Klarna's widely-reported 2025 rebalancing of AI (rehiring humans after 'going too far') concerned its customer-service agents, not marketing — these marketing figures were not retracted.

Google integrated its Veo video model into Google Ads (Asset Studio and Demand Gen, rolled out March 2026), generating video ads from a static image and a text prompt without leaving the ad console — bringing AI-generated video creative natively into the platform where ads are bought.

New York's Synthetic Performer Law takes effect 9 June 2026, requiring a conspicuous disclosure when an advertisement knowingly uses an AI-generated 'synthetic performer' (penalties $1,000 / $5,000) — the first US labelling law directly governing AI creative in advertising, championed by SAG-AFTRA alongside its June 2026 contract terms restricting synthetics.

A consumer backlash against AI-generated advertising has taken hold: brands including Heineken, Polaroid and Cadbury now market their creative as 'human-made' or mock AI outright, while H&M and Guess drew criticism for using AI brand ambassadors. Authenticity concerns are strong enough that major brands advertise their non-use of AI.

Meta is targeting fully end-to-end AI-generated advertising by the end of 2026: an advertiser supplies a product image and a budget, and Meta's AI produces the imagery, video and text, sets the targeting and allocates the budget, with real-time per-user personalisation. An announced aim (Zuckerberg; Wall Street Journal), not yet fully deployed.

Sellers who effectively partner with AI tools are 3.7 times more likely to meet their sales quota than those who do not, according to Gartner's survey of 1,026 B2B sellers conducted January to March 2024.

Bain & Company reports that early AI deployments in sales are delivering conversion gains that add up to 'more than a 30% increase in win rates', and notes that sellers today spend only about 25% of their time actually selling — work AI 'could double' by absorbing surrounding tasks.

AI saves sellers an average of 4.8 hours a week, yet 72% of sales organisations fail to reinvest that freed-up time in high-value selling (Gartner survey of 210 chief sales officers and senior leaders, early 2026). The sales-AI story has shifted from 'does it help?' to a management gap in capturing the time it frees.

69% of B2B buyers turn to a human sales representative to validate AI-generated insights (Gartner, May 2026) — evidence that even as AI research spreads, buyers still rely on people as the trust and quality check.

In Bain's B2B Growth Agenda 2026 (1,125 sales and marketing leaders, 40 countries), AI 'knowledge assistants' delivered a 25-35% increase in effective selling time, and leading adopters achieved about 2x the AI-attributed revenue growth and 1.8x the cost efficiency of laggards — grounding earlier 'double the selling time' aspirations in a measured figure.

Gartner predicts that by 2028 AI sales agents will outnumber human sellers by 10 to 1, yet fewer than 40% of sellers will report that AI agents improved their productivity (Gartner, 18 Nov 2025) — an honest counterweight to AI-SDR hype.

Sales organisations that provide AI-enabled 'next best actions' to sellers are 2.6x more likely to achieve commercial growth, and Gartner forecasts 95% of seller research workflows will begin with AI by 2027 (from under 20% in 2024) — Gartner, May 2026.

As of the 2026 AI Index (2025 data), 88% of surveyed organisations reported using AI — up from 78% in 2024 — and 70% used generative AI in at least one business function.

In PwC's 29th Global CEO Survey (January 2026; 4,454 CEOs across 95 countries), 56% of chief executives said their companies had realised neither higher revenue nor lower costs from their AI investments over the previous 12 months.

In McKinsey's The State of AI (2025), 28% of respondents whose organisations use AI reported that their CEO is responsible for overseeing AI governance.

Moderna: Moderna's leadership deployed OpenAI's ChatGPT Enterprise company-wide and enabled staff to build their own AI assistants; within about two months employees had created more than 750 custom GPTs across legal, research, manufacturing and commercial functions, with the legal team reaching 100% adoption.

77% of PR professionals report using AI tools such as ChatGPT in their work (2025).

60% of communications professionals say AI will benefit the PR profession, while 28% expect its impact to be negative.

PR teams most often use AI for social media content (43%), followed by research and analytics (36%) and preparing press materials (36%).

UK Government Communication Service (GCS): The Cabinet Office's Government Communications team built 'Assist' in-house — a bespoke generative AI tool that drafts and reviews communications content for UK government communicators. Since its November 2023 pilot it has been used by over 50% of government communicators across more than 200 government organisations and has saved thousands of hours of work (as of May 2025).

87% of newsroom leaders say their newsrooms are being fully or somewhat transformed by generative AI, with just 13% saying not so much or not at all (survey of 326 digital news leaders across 51 countries and territories, Nov–Dec 2024).

Researchers' overall use of AI tools surged from 57% in 2024 to 84% in 2025 (survey of 2,430 researchers, August 2025).

85% of researchers surveyed said AI has improved their efficiency, and close to three-quarters said it has enhanced both the quantity and quality of their work (Wiley survey of 2,430 researchers, August 2025).

Reach plc (UK — publisher of the Mirror, Express and 100+ regional titles): The UK's largest commercial news publisher built Guten, an in-house generative-AI tool that drafts and repurposes stories with a human-in-the-loop review by journalists before publication. It cut the time to publish breaking news from nine minutes to 90 seconds and, according to WAN-IFRA, now supports around a quarter of the publisher's total output.

In RICS's Q1 2025 global survey of more than 2,200 construction professionals, just under 12% reported regular use of AI in specific processes, while around 45% reported no AI implementation in their organisations at all.

In a December 2025 Dodge Construction Network/CMiC survey of 235 US contractors, 86% of large contractors said AI will give them a competitive advantage versus 69% of small and mid-sized firms, and only 40% were allocating a dedicated AI budget.

McKinsey estimates AI could automate 39% of nonphysical work in the construction sector and 50% in the architecture and engineering sectors — a projection of automation potential, not current adoption.

Zachry Construction (using ALICE Technologies): On a $149M Texas highway project (I-10, Graytown to Bexar County Line), Zachry Construction used ALICE Technologies' generative-AI scheduling tool to run about 60 schedule scenarios in roughly three minutes; by resequencing the work and reducing flat-work crews from nine to seven, the team reported cutting 28 days from the schedule. (Outcome as reported by Zachry and ALICE's founder, not independently third-party verified.)

As of 2023, only 27% of U.S. farms and ranches used precision agriculture practices to manage crops or livestock, indicating that adoption remains a minority even in a leading agricultural market.

Guidance and auto-steer systems on tractors and harvesters were used by 52% of midsize and 70% of large-scale U.S. crop-producing farms in 2023, up from single digits in the early 2000s, with adoption rising sharply with farm size.

In a 2025 peer-reviewed field study (four arable fields, 2020-2023), sensor-based variable-rate nitrogen application raised nitrogen-use efficiency by about 15% and cut nitrogen fertiliser use by roughly 18% in the best-responding field, with no significant yield penalty versus uniform application.

PlantVillage Nuru (Penn State University, with the FAO, IITA and CIMMYT): Nuru is a free, offline smartphone app that uses deep-learning computer vision to help African smallholder farmers and agricultural extension agents diagnose crop diseases in the field. In peer-reviewed testing it diagnosed cassava viral-disease symptoms more accurately (65%) than extension agents (40-58%) or farmers (18-31%), and has been scaled out as a public good to smallholder farmers across sub-Saharan Africa.

As of the 2023 OECD Digital Government Index, about 70% of participating countries had used AI to enhance internal government operations, but only 30% had used AI for policymaking — adoption is far more advanced in back-office efficiency than in policy work.

As at autumn 2023, just over a third (37%) of the 87 UK central government bodies surveyed by the National Audit Office had deployed AI, and over two-thirds (70%) were piloting or planning it (report published March 2024).

In a yearlong state-government pilot concluded in March 2025, participating Pennsylvania employees reported saving an average of 95 minutes per day using ChatGPT Enterprise (175 employees across 14 agencies; a self-selected group, figures self-reported).

UK Government — Incubator for AI (i.AI), Department for Science, Innovation & Technology: The government's in-house AI tool 'Consult' grouped more than 50,000 responses to the Independent Water Commission consultation into policy themes in about two hours at a cost of £240 — work that would otherwise take weeks — with human experts needing only 22 hours to review and validate the output. DSIT estimates that using Consult across the 500-plus consultations government runs each year could save around 75,000 days of manual analysis, worth roughly £20 million in staff time.

In 2024, CAL FIRE's ALERTCalifornia AI camera network detected 636 California wildfires before anyone dialled 911 — 38% of the 1,668 fires the cameras spotted that year.

Reported AI use cases nearly doubled across the 11 US federal agencies GAO reviewed — from 571 in 2023 to 1,110 in 2024, with generative-AI uses rising from 32 to 282 — a group that includes the Department of Homeland Security.

CAL FIRE (California Department of Forestry and Fire Protection) with UC San Diego — ALERTCalifornia: CAL FIRE and UC San Diego deployed ALERTCalifornia, a network of more than 1,200 AI-enabled cameras feeding all 21 CAL FIRE command centres that automatically scan for smoke and alert dispatchers. In 2025 the system helped CAL FIRE respond to about 3,600 wildfire incidents, flagging more than half before any 911 call was placed.

Just 18% of anti-money laundering and compliance professionals have AI or machine-learning solutions in production, with a further 18% piloting and 25% planning to adopt within 12-18 months.

Active use of generative AI across professional-services organisations (spanning legal, tax, and risk and compliance functions) nearly doubled to 22% in 2025, up from 12% in 2024.

HSBC: HSBC deployed an AI-based anti-money-laundering system, Dynamic Risk Assessment (co-developed with Google Cloud), to screen around 980 million transactions a month; it detects two to four times more financial crime with 60% fewer false-positive alerts than its previous rules-based system, letting compliance investigators focus on genuine risk.

As of April 2025, only between 0.3% and 5.5% of analysed training courses delivered AI content across Australia, Germany, Singapore and the United States, signalling that training supply has yet to scale up to meet demand for AI skills.

59% of the global workforce will need reskilling or upskilling by 2030, and 77% of employers plan to reskill and upskill their workforce to work alongside AI, according to the World Economic Forum's Future of Jobs Report 2025.

In Singapore's national SkillsFuture system, training participation in areas such as artificial intelligence, cybersecurity and digital marketing nearly tripled, from 34,000 in 2023 to 96,000 in 2024.

Skill India Digital / Skill India Assistant — Ministry of Skill Development & Entrepreneurship (MSDE) and NSDC, India: India's national skilling platform applies AI/ML for skill discovery and course recommendation and, from October 2024, added the Skill India Assistant — a multilingual AI chatbot on WhatsApp (built on Meta's Llama, developed with Sarvam AI) that helps learners find courses, skilling centres and jobs 24/7 in English, Hindi and Hinglish. The government reports the platform has onboarded over 1.5 crore (about 15 million) candidates.

OpenAI closed the largest private funding round in history — approximately US$122 billion at an US$852 billion post-money valuation (announced 27 February 2026 at US$110bn / US$730bn pre-money; closed 31 March 2026). Lead participants included Amazon (~US$50bn, partly contingent), Nvidia (US$30bn) and SoftBank (US$30bn). OpenAI filed a confidential S-1 in June 2026, signalling an imminent IPO.

Anthropic raised US$65 billion in a Series H round at a US$965 billion post-money valuation (28 May 2026), led by Altimeter, Dragoneer, Greenoaks and Sequoia — nearly tripling its February-2026 US$380bn valuation and briefly making it the world's most valuable AI start-up. Q2-2026 revenue was reported above US$11.5 billion.

xAI raised US$20 billion in a Series E round at a ~US$230 billion valuation (January 2026); backers included Nvidia, Cisco, the Qatar Investment Authority and MGX, funding data-centre build-out.

MGX (Abu Dhabi) closed a US$49 billion dedicated AI fund in July 2026 — the largest AI-focused investment fund ever raised, above its US$45bn target — chaired by Sheikh Tahnoon bin Zayed; it already holds stakes in OpenAI, Anthropic and xAI, making it the single largest pool of sovereign AI capital.

Governments moved from grants to direct equity in domestic AI: the UK launched a £500 million Sovereign AI Fund (April 2026), and South Korea announced a ~₩20 trillion (~US$13.6 billion) strategic investment fund for AI and semiconductors (July 2026; capital deploying from 2027).

SpaceX (which had absorbed xAI earlier in 2026) acquired Anysphere, the maker of the AI-coding tool Cursor, in an all-stock deal valuing Anysphere at US$60 billion — announced 16 June 2026 and completed 14 August 2026. It is the largest acquisition of a venture-backed startup on record, surpassing Google's US$32 billion purchase of Wiz.

The four largest US AI spenders — Microsoft, Alphabet, Amazon and Meta — guided to roughly US$725 billion in combined 2026 capital expenditure (the sum of their own 2026 guidance: Amazon ~$220bn, Alphabet ~$200bn, Microsoft ~$185bn, Meta ~$135bn), up about 77% on 2025's ~$410 billion. Meta separately pledged at least US$600 billion of US investment through 2028, centred on AI data centres.