Research
The AI Evidence Index
The case against enterprise AI runs on about a dozen numbers, and almost nobody quotes them correctly. This is what each study actually measured, how big its sample was, and the way it gets restated wrong.
Version 1.0 · 56 studies · 51 organisations · 2024 to 2026 · CC BY 4.0
56 of 56
entries carry a written caveat
21 of 56
report a sample size
17
findings we had written wrong
4
citations with no study behind them
The third and fourth numbers are about us, not about the publishers. We re-checked every entry against its primary document, and where our own phrasing had drifted from what the study said, the corrected version is printed with the original still visible. Four entries turned out to have no study behind them at all, which is a citation problem of ours, and they are marked rather than quietly deleted.
The index
2026
American Bar Association
Legal Technology Survey
Survey
Law-firm AI adoption tripled from 11% to 30% in one year, reaching 46% at large firms.
Sample: 512 attorneys in private practice in the United States, fielded 2024. Distributed to a sample of ABA members with voluntary response, so the risk is nonresponse bias rather than self-selection. No response rate is published, and per-band counts are not published either.
Corrected. The finding above did not match the source.
American Bar Association, 2024 Legal Technology Survey Report, released March 2025: 30% of responding attorneys reported using AI tools, up from 11% in 2023. By firm size, 46% at firms with 100 or more attorneys, 30% at firms with 10 to 49 attorneys, and 18% at solo practices.
What it actually measured
There is no 2026 ABA Legal Technology Survey Report. The most recent edition with AI figures broken out by firm size is the 2024 report, released 3 March 2025, based on 512 United States attorneys in private practice. Response was voluntary and the ABA publishes no response rate, so the respondents are not a random cross-section of the profession. The report does not define "artificial intelligence", a gap Robert Ambrogi flagged in his review, so it is not clear whether respondents counted generative tools only or any AI-based feature. That ambiguity also reaches the jump from 11% in 2023 to 30% in 2024. Per-band respondent counts are not published, so the solo and largest-firm figures rest on small cells.
Usually restated as: Nearly half of law firms are now using AI.
2026
BizBuySell
Small Business AI Adoption Survey
Method not recorded
Small-business AI adoption reached 63-88% depending on survey, with productivity gains the #1 cited motivation at 78%.
Sample: not published
What it actually measured
Adoption range spans multiple surveys with different definitions of 'using AI'; cite as a range, never a single number.
2026
Cloudera / Harvard Business Review Analytic Services
Enterprise data readiness for AI
Survey
Only 7% of enterprises say their data is completely ready for AI; 73% say preparing data for AI is challenging; top obstacles are siloed data (56%) and no clear data strategy (44%).
Sample: 231 members of the Harvard Business Review audience involved in their organization's AI data decisions, online survey fielded October 2025. The full report is free and ungated and publishes its methodology and participant profile, including organization size and industry.
What it actually measured
It measured self-reported readiness of data, not of the enterprise. Respondents rated their own organization's data for AI adoption on a five option scale and 7% picked the top option, completely ready, against 15% almost completely ready, 51% somewhat ready and 27% not very or not at all ready. The base is 231 members of the Harvard Business Review audience, surveyed online in October 2025, all involved in their organization's data decisions and all at organizations already practising, piloting or exploring AI, so the sample is screened toward AI adopters. HBR publishes the organization size, industry, region and seniority splits in a free report, though nothing at country level. Cloudera sponsored the report and sells a hybrid data platform, so a data readiness gap is the sponsor's own sales case, and the usual misquote drops both the word data and the word say.
Usually restated as: Only 7% of enterprises are AI-ready, according to Harvard.
2026
Construction Owners
Construction AI adoption doubles in 2026
Survey
38% of contractors report measurable business impact from AI, up from 17% a year earlier.
Sample: 235 general and trade contractors in the United States, surveyed September and October 2025
Corrected. The finding above did not match the source.
Dodge Construction Network, SmartMarket Brief "AI for Contractors", released 5 December 2025 in partnership with CMiC. 235 US general and trade contractors surveyed in September and October 2025. 87 percent expect AI to transform construction. 19 percent have actually adapted a workflow.
What it actually measured
The research is not by Construction Owners Club, the site that carried it. constructionowners.com is a construction industry news and marketing site with no disclosed corporate entity and no named editorial staff, and its write-up does credit the original: it names Dodge Construction Network and CMiC, quotes Dodge's Steve Jones, and links to Construction Dive's earlier reporting. The study itself is Dodge Construction Network's SmartMarket Brief "AI for Contractors", released 5 December 2025 with CMiC, a construction financial and project management software vendor, as partner. It surveyed 235 US general and trade contractors in September and October 2025, so it measured contractors rather than owner organizations, and the year is 2025 rather than 2026. Nearly every headline figure is an expectation about the future: 87 percent expect transformation against 19 percent who have changed a workflow.
Usually restated as: 85% of construction firms say AI has cut the time they spend on repetitive tasks.
2026
Deloitte
State of AI in the Enterprise 2026
Survey
Only about 1 in 5 companies has a mature governance model for autonomous AI agents.
Sample: 3,235 business and IT leaders, director through C-suite, split equally between IT and line of business, all with direct involvement in their organization's AI work, across 24 countries in the Americas, Asia Pacific, Europe and the Middle East, and 6 industries. Fielded August and September 2025. Published 21 January 2026 at Davos.
Corrected. The finding above did not match the source.
21% of the companies that plan to deploy AI agents within two years say they have a mature agent governance model. That group is close to three-quarters of the 3,235 leaders surveyed, so the figure is about 15% of all respondents, not 1 in 5 of all organizations.
What it actually measured
Deloitte publishes this figure two different ways and the gap matters. Its press releases say 21% of the companies planning to deploy AI agents within two years report a mature agent governance model, and that group is close to three-quarters of the 3,235 leaders surveyed. Its Deloitte Insights article drops the restriction and says 21% of all respondents, adding that roughly 80% of organisations surveyed lack mature governance capabilities. Cite whichever base you use and say which. Either way the finding is about governance that is not yet mature, measured on agent decision boundaries, real-time monitoring and audit trails, so "no governance in place" overstates it. Maturity is self-assessed by the respondent and never audited, and Deloitte sells AI governance consulting. Fieldwork ran in August and September 2025, not 2026.
Usually restated as: Deloitte found that 80% of companies deploying AI agents have no governance in place.
2026
IDC
Agentic AI pilot-to-production rates
Method not recorded
88% of agentic AI pilots never reach production.
Sample: not published
What it actually measured
Reported via press coverage of IDC research rather than the primary report. Use it as the reason we gate on evals and scope, not as a precise industry benchmark.
2026
JAMA
Ambient AI scribes and clinician documentation burden
Method not recorded
Clinicians using ambient AI scribes saved 16 minutes of documentation time and spent 13 fewer minutes in the medical record per 8 hours of patient care, and added roughly one extra patient visit every two weeks.
Sample: 1,800 clinicians across 5 academic medical centres, 2023 to 2025
What it actually measured
The denominator is per 8 hours of patient care, NOT per encounter, and the study characterizes the savings as modest against vendor claims. It found no significant effect on after-hours EHR time. Never restate this as a per-visit number.
2026
JPMorgan Chase
2026 US Business Leaders Outlook
Method not recorded
No study behind this
89% of midsize businesses plan to implement AI in 2026.
Sample: 1,469 business leaders, November 2025
Corrected. The finding above did not match the source.
No stat was supplied and no study exists to correct against. If a JPMorgan figure is cited at all, attribute it as a company statement. Dimon said in an October 2025 Bloomberg interview that roughly $2 billion of AI expense returned roughly $2 billion of benefit.
What it actually measured
JPMorgan has published no study behind these numbers. The figures are executive statements and trade coverage, not research with a sample or a method. In an October 2025 Bloomberg interview Jamie Dimon said that for about $2 billion of AI expense the bank saw about $2 billion of benefit, which is a wash rather than a net saving. The 230,000 employees on its internal LLM Suite, the 450 production use cases and the 30% to 40% efficiency figures come from company statements with no published measurement method and no external audit. Attribute all of it to JPMorgan as a company claim, never as measured evidence.
Usually restated as: JPMorgan's AI deployment saves $2 billion a year and delivers 30% to 40% productivity gains across 230,000 employees.
2026
Recon Analytics
AI Choice 2026: Why Licenses Don't Equal Adoption
Method not recorded
Only 35.8% of workers with paid Copilot access actually adopt it, vs 83.1% for ChatGPT.
Sample: US worker survey, July 2025 to January 2026
What it actually measured
The companion '64% of Copilot licenses unused' figure comes from vendor adoption analyses, not an independent study; treat it as directional and cite the 35.8% instead.
2026
Wolters Kluwer
Future Ready Lawyer Survey 2026
Method not recorded
Over 90% of legal professionals report using at least one AI tool, while firm-level adoption sits far below individual use.
Sample: not published
What it actually measured
Individual tool use is not firm deployment. Cite this as the gap between personal use and governed firm systems, never as firm-level adoption.
2025
AIxEnergy
What US utilities are actually doing with AI
Method not recorded
Energy companies report roughly 170% average ROI on deployed AI despite below-average adoption rates.
Sample: not published
What it actually measured
Industry-analysis estimate; treat as directional.
2025
American Medical Association
AI scribes save 15,000 hours and restore the human side of medicine
Method not recorded
The Permanente Medical Group's ambient AI scribes saved physicians an estimated 15,791 hours of documentation time.
Sample: not published
What it actually measured
A 2026 multi-center study (STAT) found more modest average savings, roughly 16 minutes per 8 hours of patient care. Cite the Permanente figure as one health system's estimate, not a universal result.
2025
Association of Corporate Counsel and Everlaw
In-house counsel on outside counsel AI use
Method not recorded
Nearly 60% of in-house counsel reported no noticeable savings from their outside counsel's use of AI. Of the 40% who did see a benefit, only 13% pointed to fewer billable hours.
Sample: not published
What it actually measured
Measures what clients perceived, not what firms saved. The honest reading is that efficiency gains are not reaching clients, which is a pricing question before it is a technology one.
2025
BCG
The 10-20-70 rule of AI transformation
Vendor analysis
Success in AI deployment is 10% algorithms, 20% technology and data, and 70% people and processes.
Sample: No sample is published for the 10-20-70 rule itself. The report it appears in, BCG AI Radar 2025, surveyed 1,803 C-level executives across 19 markets and 12 industries, fieldwork September to December 2024, at companies above 500 million dollars revenue in the US, Europe, Japan and Australia and above 100 million dollars elsewhere.
Corrected. The finding above did not match the source.
BCG's 10-20-70 rule advises companies to put 10 percent of their AI effort into algorithms, 20 percent into data and technology, and 70 percent into people and process change. It is guidance from BCG's client case work, not a measurement of where value came from.
What it actually measured
The 10, 20 and 70 are a BCG rule of thumb from its own consulting case work, which BCG describes as hundreds of client engagements. No sample frame, no measurement method and no underlying data are published, so nobody outside BCG can check it. BCG states it two ways. Its case-work language says about 10 percent of AI value comes from the algorithms, 20 percent from the technology, and 70 percent from rethinking people and process. The 2025 AI Radar states it as an effort split that top performers follow, which is a claim about what leaders do rather than a measurement of where value came from. The rule is not new to that report. BCG's Sylvain Duranton was giving the same three numbers in a TED talk published in January 2020. BCG sells AI transformation work, so read the 70 percent as guidance from an interested party.
Usually restated as: BCG research shows 70% of AI value comes from people and process, not technology.
2025
Beeontrade
Leveraging AI to transform logistics in 2025 and beyond
Method not recorded
AI-driven freight operations report up to 20% lower logistics spend.
Sample: not published
What it actually measured
Industry-blog estimate; treat as directional, phrase as 'up to'.
2025
C.H. Robinson
AI in freight brokerage (via Tank Transport)
Method not recorded
C.H. Robinson runs 30+ AI agents managing more than 3 million shipment tasks, including quoting, booking, appointment scheduling and tracking, trained on 37 million annual shipments.
Sample: not published
What it actually measured
Company-reported figures via trade press.
2025
Capgemini
AI agents in insurance (as cited by Digiqt)
Method not recorded
82% of insurance executives plan to integrate AI agents within three years.
Sample: not published
What it actually measured
Capgemini finding as cited by a third-party industry blog.
2025
CarLotAI
AI BDC platforms guide
Method not recorded
Dealerships typically miss around 60% of inbound calls, and 85% of missed callers do not leave a voicemail; they call the next dealer.
Sample: not published
What it actually measured
Industry-guide figures; directional, not an audited study.
2025
CenterPoint Energy (via T&D World)
LiDAR and AI reshape vegetation risk management
Method not recorded
CenterPoint Energy delivered a 50% year-over-year reduction in vegetation-related outage minutes in 2025 using a LiDAR-based AI digital twin.
Sample: not published
What it actually measured
Utility-reported outcome via trade press.
2025
Cohere Health
National survey on AI prior authorization
Method not recorded
Cohere Health's AI auto-approves up to 90% of prior-authorization requests for covered health-plan members.
Sample: not published
What it actually measured
Vendor-reported; the company sells the platform it measures.
2025
CPA.com
2025 AI in Accounting Report
Vendor analysis
Accounting firm AI adoption jumped from 9% to 41% in 2025.
Sample: not published
Corrected. The finding above did not match the source.
CPA.com and AICPA, 2025 AI in Accounting Report, June 2025, produced with the consultancy be radical. It is a 19 page trend narrative, not a survey, and it publishes no sample. Its two most quotable numbers are attributed inside the report to vendors: time savings of 30 to 70 percent on common workflow tasks, and over 80 percent automation of individual return preparation at unnamed firms.
What it actually measured
The 2025 AI in Accounting Report is not a survey. CPA.com states that it synthesizes the AICPA and CPA.com AI Symposium, ecosystem research, vendor and practitioner interviews, and firsthand observations, and it publishes no respondent count, no sampling frame and no field dates. Its figures are vendor claims by its own wording: vendors report time savings of 30 to 70 percent, and over 80 percent automation of return preparation is credited to unnamed "some firms". CPA.com is the AICPA's commercial technology subsidiary, and the report's own footnote reads "Some solution providers listed within this report are CPA.com Preferred Partner Solutions".
Usually restated as: A 2025 CPA.com study found AI saves accounting firms 30 to 70 percent of their time.
2025
Deloitte
AI budget allocation (technology vs workforce)
Method not recorded
Companies spend 93% of AI budgets on technology and only 7% on the people who are supposed to use it.
Sample: not published
What it actually measured
As cited by General Assembly's AI Academy page.
2025
EliseAI
Multifamily AI operations survey
Method not recorded
85% of AI-using property operators saw measurable lead-to-lease conversion improvement; 78% report losing business to AI-enabled competitors.
Sample: not published
What it actually measured
Vendor-published industry survey; treat as directional.
2025
EY
AI productivity and talent strategy survey
Survey
Companies forfeit up to 40% of AI's potential productivity gains through weak talent strategy; only 12% of employees say they get sufficient AI training.
Sample: 15,000 employees and 1,500 employers across 29 countries and 19 sectors, organizations with 1,000+ global employees, fieldwork August 2025
Corrected. The finding above did not match the source.
EY's 2025 Work Reimagined Survey found 12% of employees received 81 or more hours of AI training in the past 12 months. EY separately says AI can unlock up to 40% more productivity when talent foundations are strong. That is an upside ceiling from a segment comparison, not a measured 40% of gains forfeited.
What it actually measured
EY surveyed 15,000 employees and 1,500 employers across 29 countries in August 2025. The 12% is measured and real, but it has a specific threshold behind it: 12% of employees reported 81 or more hours of AI training in the past 12 months. The 40% is not a measured loss. EY's own body text says AI can unlock up to 40% more productivity when it lands on strong talent foundations, which is a ceiling drawn from comparing survey segments, and EY sells the workforce consulting that closes the gap it describes.
Usually restated as: EY found that companies lose 40% of their AI gains because only 12% of employees get proper training.
2025
Gallup (via Engageli)
AI in education statistics
Method not recorded
Teachers who use AI weekly save an average of 5.9 hours per week, about six weeks per school year.
Sample: not published
What it actually measured
Gallup finding as cited by an education-platform roundup.
2025
Gartner
Over 40% of Agentic AI Projects Will Be Canceled by End of 2027
Method not recorded
More than 40% of agentic AI projects are predicted to be canceled by end of 2027 due to escalating costs, unclear business value, or inadequate risk controls.
Sample: not published
What it actually measured
A prediction, not an observed outcome.
2025
Gartner
Agentic AI vendor analysis (agent washing)
Prediction
Only an estimated ~130 of the thousands of vendors claiming agentic AI are genuine; the rest is rebranded RPA, chatbots, and assistants.
Sample: No sample is published for either headline claim. The only sample in the release is a January 2025 Gartner poll of 3,412 webinar attendees, which measured investment posture: 19% significant investment, 42% conservative investment, 8% no investment, 31% waiting or unsure. Webinar attendees are self-selected. No geography, industry or job title breakdown is published.
What it actually measured
Neither number is a measurement. The 40% is a forecast about 2027, and Gartner published no model and no sample for it in public, because the report behind it is sold to clients. The 130 vendor figure is an analyst estimate with no published denominator and no published method, so "thousands" is Gartner's word rather than a count anyone can check. The only sample in the release is a January 2025 poll of 3,412 self-selected webinar attendees, and that poll asked how much organizations had invested in agentic AI, not whether anything was canceled.
Usually restated as: Gartner found that over 40% of agentic AI projects fail.
2025
Gartner
HR survey on employee AI use
Prediction
No study behind this
37% of employees don't use available AI because their co-workers aren't using it.
Sample: 2,986 employees, July 2025
Corrected. The finding above did not match the source.
No Gartner publication titled "AI as a coworker" exists. If a Gartner agent prediction is wanted, use: Gartner predicts that by 2028, at least 15% of day-to-day work decisions will be made autonomously through agentic AI, up from 0% in 2024, and that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024. Both come from the Top 10 Strategic Technology Trends for 2025, announced October 2024.
What it actually measured
Gartner has published no document titled "AI as a coworker". The nearest real Gartner publications on AI in the workplace are a 24 April 2025 press release predicting that one third of finance staff will occupy "shared jobs" with AI by 2029, and an 11 November 2025 release on four scenarios for human and AI collaboration at work. The most quoted Gartner agent number, that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024, comes from the Top 10 Strategic Technology Trends for 2025, announced 21 October 2024. It is a forecast rather than a measurement, and Gartner publishes no sample or confidence interval for it. Gartner separately forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027, so quote both halves or neither.
Usually restated as: By 2028, 15% of all work will be done autonomously by AI agents.
2025
GE (via UITG industry roundup)
AI predictive maintenance outcomes in manufacturing
Method not recorded
GE reports a 20% reduction in unplanned downtime from AI predictive maintenance; industry benchmarks for the category run 15-30% downtime reduction and 10-25% maintenance cost savings.
Sample: not published
What it actually measured
Operator-reported outcomes compiled by an industry roundup, not an audited study.
2025
HeroHunt
AI adoption in recruiting: 2025 year in review
Method not recorded
In 2025, 41% of talent-acquisition teams piloted AI interview scheduling and 23% standardized it; 99% of hiring managers report using AI somewhere in hiring.
Sample: not published
What it actually measured
Vendor roundup of industry surveys; treat the 99% figure as 'somewhere in the process', not full automation.
2025
Hostie
Voice AI restaurant reservations: 2025 adoption, accuracy, revenue impact
Method not recorded
Restaurants deploying voice AI report 87% fewer missed calls and $3,000-$18,000 per month in recovered revenue per location.
Sample: not published
What it actually measured
Vendor-reported; the company sells the voice agents it measures.
2025
HSMAI Asia
AI revolution in hotel revenue management
Method not recorded
86%+ of hoteliers now depend on AI for demand forecasting; advanced revenue-management systems are associated with 10-15% RevPAR lift.
Sample: not published
What it actually measured
Industry-association reporting; RevPAR lift is a typical range, not a guarantee.
2025
Iowa State / University of Arkansas
See & Spray field research
Field trial
University field trials measured 43.9-90.6% herbicide reduction (average 76%) and roughly $15.70 per acre savings, with a +2 bu/A soybean yield effect.
Sample: Iowa State: 5 conventionally managed soybean fields using conventional or vertical tillage, 415 acres, Boone and Story counties, Iowa, 2024 season, unreplicated field-scale demonstration. Arkansas: 3 year soybean trial from 2022, Northeast Research and Extension Center, Keiser, Arkansas. Peer reviewed Arkansas paper: 4 site years, 2021 to 2022, Keiser AR and Greenville MS.
Corrected. The finding above did not match the source.
Two separate programmes, not one joint study. Iowa State Extension, 2024: a 415 acre unreplicated demonstration on 5 soybean fields with a John Deere See and Spray Ultimate. Product savings by field of 43.9, 71.2, 87.2, 87.6 and 90.6 percent, averaging 76 percent, worth $15.70 per acre. University of Arkansas, reported March 2025 from a trial begun in 2022: 43 to 59 percent reduction in postemergence herbicide at the high sensitivity setting, worth $30.49 per acre, breaking even at 819 acres.
What it actually measured
These are two separate programmes, not one study. Iowa State ran a field scale demonstration in 2024, not 2025: five conventionally managed soybean fields, 415 acres, Boone and Story counties, using a John Deere See and Spray Ultimate. Product savings ran from 43.9 percent to 90.6 percent by field and averaged 76 percent, the fields were not replicated, and Iowa State states that savings are a function of both the technology and the underlying weed pressure, so a single average travels badly. Arkansas is a separate plot trial reporting 43 to 59 percent postemergence reduction at the high sensitivity setting, and its peer reviewed paper in Weed Technology was funded by Blue River Technology and Deere and Company, which also supplied the machine, with three Blue River employees as co-authors who disclose that employment.
Usually restated as: Iowa State found targeted spray technology cuts herbicide use by 76%.
2025
John Deere
See & Spray herbicide savings
Method not recorded
In 2025, See & Spray ran across 5 million acres and saved more than 31 million gallons of herbicide mix, roughly a 50% average reduction.
Sample: not published
What it actually measured
Vendor-reported acreage and savings; see the university research entry for independent measurement.
2025
Klarna
AI assistant performance (via CX Dive)
Method not recorded
Klarna says its AI assistant performs the work of 853 full-time agents, handling roughly two-thirds of customer-service chats.
Sample: not published
What it actually measured
Company-reported figure via trade press.
2025
KPMG / Melbourne Business School
Trust, attitudes and use of AI (global study)
Survey
Almost half of employees admit using AI in ways that violate company policy, including uploading sensitive company information to free public tools; 57% hide their AI use and present AI-generated work as their own.
Sample: 48,340 adults across 47 countries, fielded November 2024 to mid January 2025 by online panel (Dynata), nationally representative on age, gender and regional distribution, country samples of 1,001 to 1,098. 32,352 respondents were working full or part time and answered the workplace questions.
Corrected. The finding above did not match the source.
57% of employees who use AI at work say they have used it in non-transparent ways, meaning presenting AI-generated content as their own or avoiding revealing that they used AI. The figure counts anyone who did either behavior, including rarely.
What it actually measured
It measured two self-reported behaviors rolled into one number. The report says 57% of employees admit using AI in non-transparent ways, meaning presenting AI-generated content as their own or avoiding revealing that they used AI. Treat the number as soft. The report prints no base line for it, and the two underlying chart rows sit at 61% and 55% for doing it rarely or more often, so 57% is a composite the report never explains. A nearby sentence scopes a related finding to employees who use AI, while the headline sentence, the executive summary and the press release all say employees, so the base is genuinely ambiguous. KPMG funded the work and sells AI advisory, and its press release restates the finding as employees who hide their use of AI and present AI-generated work as their own, which turns an either/or into a both. The research itself was designed and run independently by University of Melbourne academics, and the authors flag social desirability bias, meaning the true rate is probably higher.
Usually restated as: 57% of workers hide their AI use from their bosses.
2025
Kyndryl
People Readiness Report
Survey
50% of organizations say they lack the skilled talent to manage AI; 45% of CEOs say most of their employees are resistant or openly hostile to AI.
Sample: 3,700 senior leaders and decision makers across 21 countries, fielded by Edelman DXI for Kyndryl between June 20 and August 15, 2025, by online survey and telephone interview. 50% C-suite, 50% senior directors and business unit leaders. Half represented companies with $1 billion or more in revenue. Paired with aggregated telemetry from 1,200 companies on Kyndryl Bridge, dated August 2025.
Corrected. The finding above did not match the source.
The 2025 Kyndryl Readiness Report reports 54% of leaders seeing a positive return on their AI investments, up 12 points year over year, against 36% who say their AI is completely ready to manage future risks and 62% who say they are still in the experimentation phase.
What it actually measured
All of these are executive perceptions collected by Edelman DXI for Kyndryl, not measured returns, and Kyndryl sells the IT modernization work that a readiness gap implies. The 2025 report surveyed 3,700 senior leaders in 21 countries between June 20 and August 15, 2025. It gives 54% seeing a positive return on AI investments, up 12 points year over year, 36% saying their AI is completely ready to manage future risks, up 7 points, and 62% saying they are still in the experimentation phase. The 29% often quoted as the AI readiness number is the 2024 wave of the same annual report, republished by Kyndryl as an AI Readiness Report. A separate 29%, workforce readiness, comes from Kyndryl's People Readiness Report, a different survey of 1,100 leaders in 8 markets.
Usually restated as: Kyndryl's 2025 survey of 3,700 executives found only 29% say their AI is ready to manage future risks.
2025
LayerX
Enterprise AI & SaaS Data Security Report 2025
Method not recorded
About 18% of enterprise employees paste data into GenAI tools, and over half of those paste events include corporate information; 72-82% of GenAI access happens through unmanaged personal accounts.
Sample: not published
What it actually measured
LayerX's separate 77% figure is ChatGPT's share of LLM traffic, not the share of employees leaking data. Use the 18% / over-half-of-pastes framing.
2025
McKinsey
The state of AI in 2025: Agents, innovation, and transformation
Method not recorded
88% of organizations report regular AI use in at least one business function, up from 78% a year earlier. About 6% qualify as AI high performers, the group attributing more than 5% of EBIT to AI.
Sample: not published
What it actually measured
The 6% is McKinsey's 'AI high performer' segment, which requires both significant reported value and more than 5% of EBIT attributed to AI. Do not phrase it as '6% see any return': 39% report enterprise-level EBIT impact, most of it below 5%.
2025
McKinsey
Superagency in the Workplace
Survey
48% of employees rank training as the most important factor for AI adoption, yet nearly half receive minimal or no training.
Sample: 3,613 employees (managers and independent contributors) and 238 C-level executives, fieldwork October to November 2024, 81 percent United States with the remainder in Australia, India, New Zealand, Singapore and the United Kingdom
What it actually measured
McKinsey surveyed 3,613 employees and 238 C-level executives in October and November 2024. It did not score readiness on any scale. It measured what employees said they already do against what leaders guessed employees do, and the two groups answered different questions. Leaders estimated 4 percent of employees use gen AI for at least 30 percent of daily work, employees self-reported 13 percent, and that 3x gap is the headline. McKinsey's conclusion that employees are ready is an inference from that gap plus what employees said they want, such as the 48 percent asking for formal training. The reported findings cover US workplaces only, so quoting this as a global figure is wrong.
Usually restated as: McKinsey found employees are 3x more ready for AI than their leaders believe.
2025
MIT NANDA
The GenAI Divide: State of AI in Business 2025
Method not recorded
95% of enterprise GenAI pilots show no measurable P&L impact; externally purchased tools succeed roughly three times as often as internal builds.
Sample: 150 interviews, 350-employee survey, 300 deployments analyzed
What it actually measured
Measures no measurable P&L impact, not total project failure. Phrase as 'show no measurable return', never 'fail completely'.
2025
Numa
AI for dealership service departments
Method not recorded
Numa serves 1,200+ dealerships; one multi-location GM group recovered over $853,000 in missed service revenue, handling roughly 16,500 calls per store per year.
Sample: not published
What it actually measured
Vendor-reported case figures.
2025
OpenAI
The State of Enterprise AI 2025
Method not recorded
Daily AI use reaches only about 10% of the US workforce; power users get roughly 6x the productivity of typical employees on the same tools.
Sample: not published
What it actually measured
Vendor-published report; the company sells the tools it measures.
2025
PayPal
Beyond Efficiency: Small Businesses Look to AI for Competitive Edge
Survey
74% of SMB AI 'Explorers' would adopt with clearer ROI evidence; 78% of small-business AI users feel pressure to adopt AI to keep up with competitors; 66% say adoption is essential to staying competitive.
Sample: 498 US merchants (166 small under $3M revenue, 143 mid-market $3M to $20M, 189 large enterprise over $20M), plus 7 qualitative interviews, fielded February 23 to March 3, 2026
Corrected. The finding above did not match the source.
No stat was supplied, so nothing can be confirmed or corrected. A defensible figure from the primary report: 52% of large enterprises have integrated AI into regular operations, and roughly 1 in 5 merchants have 80% or more of their product catalog available as structured, machine-readable data.
What it actually measured
PayPal's Agentic Commerce Pulse is an online survey of 498 US merchants, fielded between February 23 and March 3, 2026 by Stripe Partners, an independent London research consultancy that is not Stripe the payments company. Seven interviews ran alongside it. It measures merchant awareness, readiness, trust and perceived risk. It does not measure sales, conversion or revenue. The 99% figure for large enterprises is familiarity with a term PayPal defined for respondents inside the survey, not adoption. Cite the primary PDF and not PayPal's article page, which prints the mid-market sample as 243 and makes the segments sum to 598 against a stated total of 498. PayPal commissioned the research while selling agentic commerce services.
Usually restated as: PayPal found that 99% of large enterprises are already adopting agentic commerce.
2025
Phenom
Companies using AI recruiting platforms (Mastercard case)
Method not recorded
Mastercard cut interview scheduling time by more than 85% and scheduled 88% of interviews within 24 hours of request.
Sample: not published
What it actually measured
Vendor case study published by the platform provider.
2025
S&P Global Market Intelligence
Voice of the Enterprise: AI & Machine Learning 2025
Survey
42% of companies abandoned most of their AI initiatives in 2025, up from 17% a year earlier; the average org scrapped 46% of AI proofs-of-concept before production.
Sample: 1,006 midlevel and senior IT and line-of-business professionals across North America and Europe, online, fielded October 21 to November 25, 2024. Margin of error plus or minus 3 points at the 95% confidence level. The 17% comparison comes from the prior wave: 1,001 respondents, fielded November 14 to December 14, 2023. The abandonment question was asked only of respondents whose organization had AI or ML in production or proof of concept, and that base size is not published.
Corrected. The finding above did not match the source.
42% of respondents said most of their proof of concept AI projects are abandoned before reaching production, up from 17% in the prior wave. The data was collected October 21 to November 25, 2024, not in 2025. The 17% figure was collected November 14 to December 14, 2023. S&P's exact wording is: "The proportion of companies that abandon most of their AI initiatives has increased from 17% to 42%, with the average organization scrapping 46% of its proof-of-concept projects prior to production."
What it actually measured
This measures proof of concept projects, not AI programs. Respondents estimated what share of their organization's AI projects sitting in proof of concept get abandoned before production, and 42% put that share above half. The fieldwork ran October 21 to November 25, 2024, so the 2025 in the survey name is a label and not the collection year. It gets quoted as companies walking away from AI altogether, which is not what the question asked.
Usually restated as: 42% of companies abandoned their AI initiatives in 2025, up from 17% in 2024.
2025
Salesforce
2025 holiday shopping data
Method not recorded
AI and agents influenced $262B, 20% of all global online holiday sales in 2025; retailers operating their own AI agents grew 59% faster (6.2% vs 3.9% YoY).
Sample: not published
What it actually measured
Vendor-published commerce data drawn from Salesforce's own platform traffic.
2025
Spendflo
State of SaaS Buying and Procurement 2025
Method not recorded
No study behind this
Average tool count per company jumped 23% year over year, driven by AI-labeled purchases; the average organization now runs 100+ tools.
Sample: not published
Corrected. The finding above did not match the source.
There is no verifiable Spendflo stat to correct. If the index needs an AI spend benchmark, use SpendHound's AI Spend Report 2026: 46% of finance and procurement leaders exceeded their AI budget in 2025, against 37% who exceeded budget on traditional finance and accounting software. Sample is 172 survey respondents plus spend data from 1,300+ companies.
What it actually measured
This entry carries no verifiable statistic. Spendflo's State of SaaS Procurement 2025 is a gated ebook that discloses no sample size, no field dates and no geography, and describes its inputs only as finance leader responses plus Spendflo's own contract data. Spendflo's State of SaaS Buying Survey 2026 likewise names no sample. Neither discloses enough method to cite as a benchmark. The widely repeated figure of roughly 300 SaaS applications per company with about half of licences unused comes from Zylo's SaaS Management Index, not from Spendflo. For an AI spend benchmark with a disclosed sample, use SpendHound's AI Spend Report 2026, which reports 172 survey respondents plus spend data from more than 1,300 companies, and note that SpendHound sells spend management software.
Usually restated as: Companies manage over 300 SaaS tools and only about half get used, according to Spendflo's 2025 SaaS and AI spend benchmarks.
2025
SQ Magazine
AI in education statistics
Method not recorded
Teacher AI usage roughly doubled year over year to 53-61% in the 2024-25 school year.
Sample: not published
What it actually measured
Aggregated survey range; cite as a range.
2025
Stanford GSB
AI in accounting study
Field trial
AI-using accountants closed monthly statements 7.5 days faster.
Sample: 277 accountants surveyed, plus hundreds of thousands of transaction-level records from 79 small and mid-sized firms covering January 2023 to March 2025, United States, plus a framed field experiment run November 2024 to March 2025
Corrected. The finding above did not match the source.
Choi and Xie, "Human + AI in Accounting: Early Evidence from the Field", Stanford GSB Working Paper 4261 and MIT Sloan Working Paper 7280-25, May 2025, published in the Journal of Accounting Research in 2026: AI use corresponds to a 7.5-day reduction in monthly close time and a 12% increase in general ledger granularity, measured across 79 small and mid-sized firms.
What it actually measured
The 7.5 days is an association, not a causal estimate. It comes from field data on 79 small and mid-sized firms supplied by one AI accounting software vendor, so the sample is that vendor's own customer base and adoption was self-selected rather than assigned. The authors write that AI use "corresponds to" the reduction, and their own framed experiment found accountants sometimes accepted AI classifications that were wrong. Nothing in the study covers large enterprise closes, which is exactly where the number usually gets applied.
Usually restated as: AI cuts monthly financial close time by 7.5 days.
2025
Statista
Barriers to AI adoption survey
Method not recorded
No study behind this
Lack of skilled professionals was the #1 barrier to AI adoption in 2025, cited by 50% of businesses.
Sample: not published
Corrected. The finding above did not match the source.
No corrected phrasing is possible. No study is named and no sample is published, so there is nothing to correct.
What it actually measured
There is no study behind this entry. The link points at the Statista homepage rather than at a statistic, and the stat string carries no number, no sample and no year of fieldwork. Statista both runs its own surveys and republishes other people's, and the source detail usually sits behind a subscription, so a Statista link on its own does not tell a reader who fielded the survey, when, or how many people answered. Name the group that ran it, or drop the entry.
Usually restated as: According to Statista, most organizations report an AI skills shortage.
2025
University of Kansas / Togal.AI
AI takeoff speed study
Method not recorded
A 2025 University of Kansas study found Togal.AI 76% faster than leading takeoff tools; the vendor reports a full architectural takeoff in roughly 12 minutes.
Sample: not published
What it actually measured
The 76% figure is the university study; the 12-minute and 97-98% accuracy claims are vendor-reported.
2025
Voxel51
Visual AI in manufacturing: 2025 landscape
Method not recorded
Computer-vision quality inspection reaches 98-99% defect-detection accuracy with 50-70% inspection labor savings as industry benchmarks.
Sample: not published
What it actually measured
Benchmark ranges across deployments, compiled by a vision-tooling vendor; treat as directional.
2024
Gartner
30% of Generative AI Projects Will Be Abandoned After Proof of Concept by End of 2025
Method not recorded
30% of generative AI projects predicted abandoned after proof-of-concept by end of 2025.
Sample: not published
What it actually measured
A prediction, not an observed outcome.
2024
RAND Corporation
The Root Causes of Failure for Artificial Intelligence Projects
Interviews
More than 80% of AI projects fail, roughly double the failure rate of comparable non-AI IT projects.
Sample: 65 semistructured interviews, August to December 2023. 50 industry practitioners drawn from 379 people contacted, representing more than 50 organizations, plus 15 academics. All had at least 5 years building AI or ML models. Industry participants were recruited through LinkedIn Recruiter and paid a $100 honorarium for 45 minutes. Academics came from a convenience sample at conferences and the research team's own contacts. No geography published.
Corrected. The finding above did not match the source.
RAND cites this figure rather than measuring it, and the hedge matters. The report says: "By some estimates, more than 80 percent of AI projects fail. This is twice the already-high rate of failure in corporate information technology (IT) projects that do not involve AI." The first sentence is footnoted to a 2022 Fortune article, the second to a 2023 Harvard Business Review article. RAND's own finding is five root causes of AI project failure drawn from 65 interviews, four of which are organizational rather than technical.
What it actually measured
RAND measured root causes, not a failure rate. The 80% figure is not RAND's own. It appears twice, once in the summary and once in the introduction, hedged both times as "by some estimates," and the footnote points to a July 2022 Fortune article that names no survey and puts the range at 83% to 92%. That article credits the range to unnamed surveys of business leaders, inside a profile of an AI vendor's CEO, so there is no study anyone can check behind the number the internet now credits to RAND. RAND's own work is 65 interviews that produced five root causes of failure and no failure rate at all. RAND calls it an exploratory analysis and warns its findings may skew toward blaming leadership, because most interviewees were engineers rather than executives.
Usually restated as: RAND found that more than 80% of AI projects fail, twice the rate of non-AI IT projects.
2024
US Federal Trade Commission
Operation AI Comply
Enforcement action
The FTC launched a coordinated crackdown on deceptive AI claims with five simultaneous enforcement actions, followed by roughly a dozen AI-washing cases in 2025.
Sample: 5 enforcement actions against 5 operations, announced 25 September 2024, United States only. No survey, no respondents, no sample. One of the five, FBA Machine, was filed in June 2024 and folded into the September announcement.
What it actually measured
This is a law enforcement sweep, not research. It measures nothing about how often AI claims are exaggerated, and it has no sample. The FTC characterised three of the five targets as business opportunity schemes that marketed AI-powered tools, so only DoNotPay and Rytr sold an actual AI product. No civil penalty was levied in the sweep. DoNotPay agreed to pay 193,000 dollars in consumer redress, which is the only money that changed hands. The FTC then reopened and set aside the Rytr order on 22 December 2025, so anyone citing Rytr as live precedent is citing a vacated order.
Usually restated as: In 2024 the FTC fined five AI companies for overstating what their AI could do.
Method
Every study Advizr cites anywhere on this site is recorded in a typed registry with the source URL, the finding phrased to what was measured, the sample where one is published, and an honesty note. A reference to a study that is not in the registry fails the build, so the site cannot cite something the index does not list.
Each entry was researched against its primary document, then handed to a second reviewer instructed to refute the first and to default to rejection when uncertain. On the most recent pass that reviewer rejected 16 of 21 first drafts, almost always for over-claiming about a named firm's method. The corrected version is what appears here.
The unit of analysis is one published study, report or enforcement action. Entries are selected because they appear in the enterprise AI debate often enough to be worth correcting, not by any sampling frame, so this is a curated index and not a representative sample of the literature.
Method labels separate measurement from forecast. A survey and a prediction can produce a similar looking percentage, and a forecast about 2027 is not evidence about today.
Advizr sells AI systems. That is a commercial interest in this subject matter, and it is why the caveats are written to be usable against Advizr as readily as for it. Several entries here undercut claims the company would find convenient.
Corrections are welcome and will be versioned rather than silently edited. The data is CC BY 4.0, so it can be reused with attribution.