A program of EffectiveSC
How We Use AISources checked October 29, 2026
Where things stand / What people use AI for / Use cases by industry
What hospitals, law firms, schools, banks, governments, retailers, technology firms, manufacturers and newsrooms use AI for, with trials and organizations' own results.
Summarizing research is doctors' most common use of AI, and hospitals' fastest-growing uses are billing and scheduling. Trials find modest time savings but no better patient outcomes yet, and one study links routine AI use to weaker skills without it.
nearly four in ten
In the AMA's survey of US physicians in January and February 2026, summaries of medical research and standards of care were the most common AI use in practice, used by nearly four in ten physicians who answered the question.
Who was asked: 1,692 US physicians
Self-reported, in a voluntary survey.
36% to 61%
Among US non-federal acute care hospitals that used predictive AI, the share using it to simplify or automate billing rose from 36% to 61% between 2023 and 2024, and the share using it to facilitate scheduling from 51% to 67%, the fastest-growing uses.
Who was asked: US non-federal acute care hospitals that used any predictive AI
Predictive models, such as forecasting no-shows, not chatbots or generative AI.
9.5%
In a randomized trial of 238 outpatient physicians at UCLA Health from November 2024 to January 2025, those given the Nabla AI scribe cut their time writing each note by 9.5% compared with usual care, while the 1.7% drop for those given Microsoft's DAX Copilot was not significant.
Who was asked: 238 outpatient physicians at UCLA Health
The health-record log misses time spent editing drafts inside the scribe apps, so savings may be overstated.
2.2%
In a cluster-randomized trial in 16 primary-care clinics in Kenya from April to July 2025 (9,691 patients, 103 clinical officers), an AI assistant built into the medical record didn't reduce treatment failure within 14 days: 2.2% of patients with it against 2.0% without, not a significant difference.
Who was asked: 9,691 patients seen by 103 clinical officers
The trial could detect only a large effect; clinicians' notes and treatment plans did improve.
Disclosure: Penda Health ran the clinics, and two of the authors hold its stock options; OpenAI gave in-kind support for the tool.
28.4% to 22.4%
At four endoscopy centres in Poland, the share of standard colonoscopies, done without AI, in which doctors found at least one adenoma (a precancerous polyp) fell from 28.4% to 22.4% between the three months before AI tools arrived at the end of 2021 and the three months after (1,443 patients; an observational study).
Who was asked: 1,443 patients who had standard, non-AI-assisted colonoscopy
Observational, before and after: an association, not proof that AI caused the drop.
Lawyers who use AI use it most for research, document review and summaries. In a randomized trial it made legal work much faster, while courts keep ruling on filings that cite cases AI invented.
80%
In a Thomson Reuters survey of professionals in 27 countries in October and November 2025, the most common regular uses of generative AI among legal professionals who use it were legal research (80%), document review (74%) and document summarization (73%).
Who was asked: 1,514 professionals in legal, tax and accounting, corporate risk and fraud
Drawn from the publisher's own lists and screened for familiarity with AI.
Disclosure: Thomson Reuters sells legal AI tools.
34% to 140%
In a randomized trial with 127 upper-level law students in October 2024, AI raised productivity, measured as quality points per minute, significantly on five of six realistic legal tasks: by approximately 38% to 115% with the legal research tool Vincent AI and 34% to 140% with OpenAI's o1-preview reasoning model.
Who was asked: 127 upper-level law students at the University of Minnesota and University of Michigan
Volunteer law students, not practicing lawyers; work done with the reasoning model contained more hallucinations.
2241
As of October 29, 2026, Damien Charlotin's AI Hallucination Cases Database had identified 2241 cases worldwide in which a court or tribunal found or implied that a party relied on AI-hallucinated content, typically fake citations.
Who was asked: Legal decisions worldwide
A hand-built list that grows weekly; it counts court decisions, not every fake citation.
Disclosure: The database's author also offers a reference checker that detects AI hallucinations.
Many teens now do schoolwork with chatbots, and preparing to teach is teachers' most common use. Trials show design matters: a tutor built to teach raised learning, while an open chatbot left students worse off.
One-in-ten
One-in-ten US teens ages 13 to 17 told Pew in September and October 2025 that they do all or most of their schoolwork with chatbots' help; 21% said some, 23% a little, and 45% hadn't used chatbots this way.
Who was asked: 1,458 US teens ages 13 to 17
six in 10
In the 2024-25 school year, six in 10 US public K-12 teachers reported using an AI tool for their work, most often to prepare to teach (37% at least monthly), make worksheets or activities (33%) and modify materials to meet student needs (28%).
Who was asked: 2,232 US public K-12 teachers from the RAND American Teacher Panel
17%
In a randomized trial with nearly 1,000 math students at one high school in Turkey in fall 2023, those who practiced with an unrestricted GPT-4 chat tool scored 17% worse than those without AI on the exam that followed, taken without it; a tutor version with teaching guardrails removed the harm but didn't raise scores.
Who was asked: Nearly 1,000 students in about fifty 9th to 11th-grade classes
One school, with each exam taken right after the practice session.
over double
In a randomized crossover trial in Harvard's introductory physics course in fall 2023, students' median learning gains from a lesson with a GPT-4-based AI tutor at home were over double those from the same lesson taught in class with active learning.
Who was asked: 194 eligible students in Harvard's introductory physics course for life sciences
Two lessons, each tested right after.
Disclosure: The authors built the tutor they tested and teach the course.
Internal processes, cybersecurity and fraud detection are the most common uses of AI in UK financial firms. Banks and exchanges report faster checks and less fraud.
41%
In the Bank of England and FCA's 2024 survey of 118 UK financial services firms, the most common uses of AI were optimising internal processes (41% of respondents), cybersecurity (37%) and fraud detection (33%).
Who was asked: 118 UK financial services firms
Counts all kinds of AI, including machine learning, not only generative AI.
30%
Coinbase reported in a 2026 blog series that layering an LLM risk agent onto the existing fraud model and rules for its Onramp service cut fraudulent transactions by 30% and fraud value by 22% in an online A/B experiment.
Who was asked: Coinbase Onramp transactions in an online A/B experiment
The company's own A/B test, reported without sample size, period or method.
Disclosure: Coinbase reporting on its own fraud system.
more than double
In its 2025 annual report, JPMorgan Chase says AI has let its Commercial & Investment Bank review more than double the volume in transaction screening while halving the number of manual operator checks.
Who was asked: JPMorgan Chase Commercial & Investment Bank transaction screening operations
The bank's own claim, with no baseline, period or error rates.
Disclosure: JPMorgan Chase reporting its own results to shareholders.
US federal agencies report thousands of AI uses, and State Department staff use its chatbot most often to draft. The department reports large savings from automating document work, while a small UK test found some tasks faster and better and others faster but worse.
3,611
In their 2025 inventories, compiled by OMB as of April 13, 2026, US federal agencies individually reported 3,611 AI use cases at all stages of development, 445 of them high-impact.
Who was asked: 41 federal agency submissions of individually reported AI use cases
Off-the-shelf tools are listed separately and defense uses are excluded; not comparable with earlier years.
approximately 51 percent
In the State Department's opt-in staff survey of August 26 to September 26, 2025, drafting communications was the most common reported use of its StateChat tool, selected by approximately 51 percent of users, with time savings averaging 1.6 hours per week.
Who was asked: 2,706 State Department employees giving substantial answers to a voluntary opt-in survey
Opt-in survey with self-reported savings; the department says it can't support statistical inference.
Disclosure: The State Department evaluating its own tool.
$6M
The State Department says its AIRMOD pilot used generative AI to digitize, translate and extract key data from over 115K documents, saving 76.5K hours and avoiding $6M in costs (reported in its July 2026 playbook).
Who was asked: The State Department's AI for Reports Modernization (AIRMOD) pilot
The department's own figures; it doesn't say how hours or costs were counted.
Disclosure: The State Department reporting on its own project.
over 7 minutes
In observed tasks during the UK Department for Business and Trade's Microsoft 365 Copilot pilot (October to December 2024), Copilot users made PowerPoint slides over 7 minutes faster on average than non-users but of worse quality and accuracy, and did Excel data analysis more slowly and worse; their report summaries and emails were faster and better.
Who was asked: UK Department for Business and Trade staff
Tiny samples: three Copilot users and three non-users for the slides and Excel tasks.
Shoppers who use AI most often compare prices and narrow their choices, and fewer let it buy for them. Retailers say assistant users buy more, but the controlled experiments come from a single platform.
50%
In Menlo Ventures' July 2026 survey of US adults, 50% of shoppers who use AI had used it in the past six months to compare prices and 43% to narrow a shortlist; 37% made a final purchase decision based on its recommendation, and 26% let AI complete a purchase.
Who was asked: 5,067 US adults, online survey with Morning Consult
Disclosure: Menlo Ventures invests in consumer AI companies.
up to 16.3%
In randomized field experiments on a leading cross-border online retail platform in 2023 and 2024, generative AI raised sales by up to 16.3% with a pre-sale service chatbot and typically by 2–3% with search-query refinement and product descriptions, with no detectable effect in advertising.
Who was asked: Seven randomized field experiments on a leading cross-border online retail platform
Short-term effects on one platform.
Disclosure: One author worked for the platform during the project.
250 million
Amazon said in its third-quarter 2025 results that 250 million customers had used its AI shopping assistant Rufus that year, and that shoppers using Rufus were 60% more likely to complete a purchase.
Who was asked: Amazon store customers (company data)
Compares Rufus users with other shoppers; it doesn't show Rufus caused the purchases.
Disclosure: Amazon describing its own product.
Technology firms use AI first for research and development, and they lead in putting coding agents to work. Company reports of AI-written code use their own definitions, so they can't be compared.
42.51%
Among EU information and communication enterprises using AI, the main purpose was research and development or innovation (42.51%) in 2025.
Who was asked: EU enterprises with 10 or more employees and self-employed persons
The sector also covers publishing, broadcasting and telecoms, not only software.
31%
In McKinsey's May and June 2026 survey, 31% of respondents at technology companies said their organizations had reached the scaling phase with AI agents in software engineering, the highest share for any industry and function in its exhibit; across all industries, 10% said so.
Who was asked: 1,719 participants at all levels in 97 nations
Respondents describing their own organizations, not a count of companies.
Disclosure: McKinsey sells consulting on AI.
nearly 60%
Airbnb says nearly 60% of the code its engineers produced in Q1 2026 was coauthored with AI, roughly twice the industry average by its own estimate.
Who was asked: Airbnb's own engineers
Airbnb doesn't define coauthored, so the share can't be compared with other firms'.
Disclosure: Airbnb describing its own operations to investors.
See also: Software: big claims, mixed trials
Manufacturers' most common uses of AI are office work: marketing, administration and finding information. A chipmaker reports faster design work.
30.41%
In 2025, EU manufacturers that used AI used it most often for marketing and sales (30.41%) and for organising business administration or management (27.05%); production processes ranked lower, according to Eurostat.
Who was asked: EU enterprises with 10 or more employees and self-employed persons
50 to 60 percent
In the New York Fed's August 2025 surveys of firms in the New York-Northern New Jersey region, 50 to 60 percent of both manufacturers and service firms that used AI in the past six months used it for marketing and advertising, and more than 40 percent of those manufacturers used it to search for information.
Who was asked: Firms in the New York-Northern New Jersey region
Firms that used AI only to search for information weren't counted as users.
approximately 50%
Samsung said at NVIDIA's GTC conference in March 2026 that using AI, including reinforcement learning and genetic algorithms, to automate transistor sizing and predict design-rule violations had cut design turnaround time for its analog circuits by approximately 50%, enabling speeds of up to 13 Gbps per pin in its HBM4 memory.
Who was asked: Samsung's own analog design workflow for high-bandwidth memory
Samsung's own account, with no baseline, period or method.
Disclosure: Samsung describing its own engineering; partners NVIDIA and Synopsys sell the tools involved.
News leaders rate behind-the-scenes automation as AI's most important use, and few call its impact transformational. A streaming company reports using generative AI in hundreds of titles, most of it in post-production.
64%
In the Reuters Institute's survey of 280 news leaders in 51 countries and territories (November and December 2025), back-end automation such as transcription, copyediting help and automated metadata was most often rated a 'very important' AI use (64%), followed by coding and product development (44%), commercial uses (33%) and newsgathering (29%).
Who was asked: A strategic (non-random) sample of 280 news and digital leaders
Importance ratings from an invited, non-random group, not measured use.
Disclosure: Published with support from the Google News Initiative.
roughly 300
Netflix said in July 2026 that generative AI workflows had been used in roughly 300 of its titles so far that year, with the largest concentration of work in post-production.
Who was asked: Netflix titles in 2026
Netflix doesn't say how much of each title; use could be a few shots.
Disclosure: Netflix describing its own operations to investors.
13%
In the Reuters Institute's survey of 280 news leaders in 51 countries and territories (November and December 2025), 13% described the impact of their current AI initiatives as transformational, 44% as promising and 42% as limited.
Who was asked: A strategic (non-random) sample of 280 news and digital leaders
An invited, non-random group of news leaders.
Disclosure: Published with support from the Google News Initiative.
The AI at Work Benchmark
Score your organization's practices, measure your people with an anonymous pulse, and see both against the published evidence.