A running library of statistics on AI and the software development lifecycle — including coding agents, cost management, code quality, security, adoption, productivity, and the labour market.
Click any statistic to open its commentary and source.
KPMG's 2026 CEO Outlook covered more than 1,300 chief executives globally, including 400 in the United States. Only 9% of US respondents described agents as transformational, while 39% reported significant efficiency or growth improvements. These are executive assessments of organisational impact.
IBM commissioned Censuswide to survey 3,221 executives at organisations with at least 1,000 employees across eight European and Middle Eastern countries on 14–24 September 2026. Eight in ten reported an escalated AI incident. Incidents included agent failures, data exposure, decision errors and security breaches; this measures executive reports, not independently audited incident rates.
Sonatype recorded 9,658 audited Hugging Face model requests in H1 2026 and found that 6.98% were blocked or rejected. The policy-telemetry sample is smaller and separate from its repository-adoption sample. A block reflects an organisation's policy decision and does not by itself establish that a model is vulnerable or malicious.
JetBrains' 2026 CI/CD Tools Adoption Survey covered 598 respondents, including 576 using at least one CI/CD tool. Its combined MCP-or-agent-skills measure reached 28%. The measure combines overlapping responses to two interface options; adding their separate percentages would overcount people who selected both.
IBM's survey of 1,500 CFOs and equivalent finance leaders across 33 geographies and 26 industries ran from February to April 2026. Only 6% placed finance in its transformation-ready stage, defined by consistent AI integration into workflows and decision-making at scale. This is reported maturity within finance, not an enterprise-wide deployment rate.
A Gartner survey of 297 cybersecurity leaders in Q2 2026 found that 54% of organisations either had no defined approach to restricting AI-agent access or relied on predefined human access. The combined response captures an access-control gap; it does not mean all of those organisations have no controls whatsoever.
Salesforce's fourth State of Commerce report surveyed 3,450 commerce professionals across 20 countries in April–June 2026, including 100 in Singapore. The Singapore release reports 28% current agentic-AI adoption. This is a small country subsample of the global survey and should not be presented as a worldwide adoption rate.
Gartner surveyed 243 senior leaders globally at organisations with at least $500 million in annual revenue from 11 November to 18 December 2025. Its September 2026 release reports that 83% had spent at least $3 million on planning automation. Spending includes AI and other automation, so the entire amount cannot be attributed to AI alone.
IBM and Oxford Economics surveyed 1,500 CHROs and 8,800 employees in April–June 2026. Eight in ten CHROs said adoption creates extra work such as checking recommendations, correcting errors and supplying context. The result records leaders' perceptions of work that can be missed by simple time-saving measures.
Capgemini's September 2026 government report draws on a global survey covering 600 public-sector organisations and interviews with government leaders. Only 28% reported scaling AI successfully. The result measures reported organisational progress, rather than the percentage of individual government services already automated.
Deloitte estimated workers' annual personal spending on GenAI tools used for work at £958 million. Its inaugural UK GenAI Workforce Survey collected responses from 25,000 workers between 7 May and 10 June 2026. The figure is a survey-based annual estimate, rather than an observed total of subscription transactions.
Gartner polled 161 chief audit executives in May 2026 and found that 60% had no formal AI strategy, despite widespread AI use. Because respondents came from a webinar poll, the finding describes that participating group and should not be read as a representative estimate for every audit function.
Salesforce's CFO Priorities Report surveyed 865 finance leaders in France, Germany, Japan, the UK and US on 4–15 May 2026 using a double-blind design. Among those already using AI, 90% reported positive ROI. The figure excludes non-users and reflects leaders' assessments, rather than independently verified financial returns.
ServiceNow and ThoughtLab's 2026 Enterprise AI Maturity Index surveyed 4,500 senior leaders, including 350 in India. The India findings report that only 22% had processes covering AI testing, auditing and risk assessment. This measures reported governance practices in that national subsample, rather than a global control-coverage rate.
Deloitte and Docusign's 2026 Capitalizing on AI study gathered responses from more than 1,100 leaders. Surveyed businesses reported average efficiency gains of 36%, encompassing time savings and shorter cycle times across agreement workflows. This is a reported workflow outcome, not a controlled experiment or a guaranteed benefit from a particular platform.
Capgemini's September 2026 digital-sovereignty research surveyed 1,300 business and technology executives. Its separate index analysis covered 866 organisations in the US, Europe and APAC and found significant exposure to foreign or externally controlled supply chains in 86%. The percentage applies to the index population and its five-dimension framework, not all surveyed executives.
Salesforce surveyed 2,025 agentic-AI decision makers across 20 countries on 14–28 May 2026. Among the 30% who had fully deployed agents, reported time to meaningful ROI averaged about eight months. The timing is self-reported and conditional on deployment; it is not a guaranteed payback period for a new project.
Gartner surveyed 1,303 respondents between January and April 2026 at organisations with at least $50 million in annual revenue. Only 22% reported either successful scaling across multiple business units or an AI-first approach. The percentage combines those two maturity outcomes and is based on respondent reports.
PwC's 2026 AI Jobs Barometer analysed more than a billion job advertisements across 27 countries. It reported a 62% premium for AI skills, compared with 57% in the previous edition. The observational comparison describes labour-market associations and does not establish that acquiring an AI skill alone causes that pay increase.
Gartner surveyed 3,566 B2B and B2C customers in February and March 2026. Most considered a human contact option essential when companies use GenAI for service. This is a customer preference about escalation and service design, rather than a measured failure rate for automated support.
CrowdStrike's 2026 Threat Hunting Report records 131 poisoned AI-framework packages associated with STARDUST CHOLLIMA. The report draws on Counter Adversary Operations intelligence from 1 July 2025 to 30 June 2026. This is a count from observed threat activity, not an estimate of the share of AI packages that are malicious.
Harness found that two thirds of organisations already pass $250,000 in monthly AI spending. Sapio Research fielded the survey in May and June 2026 among 700 engineering leaders and practitioners at organisations actively using AI or LLM services across the US, UK, France, Germany and India.
One in five respondents to Harness's 2026 survey said their organisation had crossed $1 million in monthly AI expenditure. The sample comprised 700 engineering, DevOps, IT operations and executive respondents whose employers were already using AI or LLM services.
Harness recorded broad recent cost growth rather than an isolated rise in model fees: 80% of respondents said overall AI spend had increased. Separate questions found increases across services, software, data, infrastructure, models, cybersecurity and application-development platforms.
Nearly a third of the organisations in Harness's five-country survey said AI already accounted for over 25% of cloud expenditure. The measure covers organisations actively using AI or LLM services and places AI costs within the wider cloud bill rather than treating model invoices alone as total spend.
Harness found AI-cost accountability divided among platform and DevOps, FinOps, finance and engineering teams. No function held a majority, and 52% of the 700 respondents said there was no clear dedicated owner for the total bill.
Nearly three quarters of Harness respondents had encountered at least one surprise AI bill in the preceding 12 months, and roughly one in three had been caught out multiple times. The result is self-reported incidence among organisations already operating AI or LLM services.
Asked how long diagnosis would take if AI spend doubled overnight, 40% of Harness respondents chose a full day, 32% a week and 8% longer or never. Only 20% said they could identify the cause within hours.
More than half of the engineering leaders and practitioners surveyed by Harness said anticipating AI expenditure depended on guesswork. The report links that uncertainty to fragmented billing across infrastructure, software, models and services, but the percentage itself records respondents' assessment of their budgeting process.
Harness asked about 'tokenmaxxing': incentives to increase AI use without tying consumption to an outcome. A majority said their organisation encouraged that behaviour, a useful measure of the tension between adoption targets and cost control rather than a measured productivity result.
Respondents to Harness's survey estimated that 26% of their AI expenditure was wasted, with similar reported rates across software, models, infrastructure and services. This is an organisational estimate rather than an audit of invoices; at a $1 million monthly budget, the same share would equal $260,000.
Engineering leaders in Harness's survey estimated that fewer than half of their engineers understood the running cost of what they built. In a separate question, 56% said teams did not have cost in mind while building AI features, pointing to a missing cost signal at design and implementation time.
Only about one in five organisations in Harness's study rated their AI cost-management practice fully mature across the company. The report separately found that 73% had AI cost policies and 47% fully enforced them, so the maturity measure is stricter than the presence of policy alone.
Veracode's 2026 longitudinal report applied standardised code-generation tasks with no security-specific prompting across four testing snapshots. The average security pass rate was 56%, while the same models produced compilable code at a near-universal rate.
GPT-5.5 led Veracode's Summer 2026 secure-coding leaderboard at 68%. Six of the 11 models tested clustered between 50% and 53%, so even the leader failed almost one security task in three under the benchmark's unprompted conditions.
Models built specifically for coding averaged a 51% security pass rate in Veracode's raw-model tests, compared with 52% for general-purpose models. The test excluded agent tooling, guardrails and human review, isolating the models' default code-generation choices.
Reasoning models averaged 56% on Veracode's secure-code tasks, five points above non-reasoning models at 51%. It was the only architectural factor in the report associated with a consistent security advantage, though both groups still failed a large share of tasks.
Veracode found little security difference by model size: systems above 100 billion parameters averaged 53%, while medium and small models each averaged 51%. Bigger models did not supply a meaningful secure-coding advantage under the same task conditions.
Language choice produced a 33-point gap in Veracode's 2026 data. Python achieved the highest security pass rate at 63%; Java remained last at 30%, despite being the only language with a clear upward trend over the preceding year.
BCG's 2026 Digital Government Citizen Survey covered 48,500 people in 44 countries. Weekly AI use reached 64%, up 26 percentage points over two years. This measures reported use among citizens in the surveyed markets; it is not a workplace-only or developer-specific adoption rate.
McKinsey's May 2026 Enterprise AI FinOps survey found spending rose almost fourfold between isolated implementations and enterprise-scale adoption. The survey included 120 enterprise participants, of whom 75 qualified, across five major industries.
Budget overrun was nearly universal among the qualified respondents in McKinsey's enterprise AI FinOps survey. The result comes from a relatively small cross-industry sample of 75 qualified participants and is self-reported, but it directly measures experience rather than a forecast.
Almost two thirds of qualified participants in McKinsey's five-industry survey said their organisation was actively deploying AI rather than remaining at the experiment stage. McKinsey reports this adoption level alongside sharply rising spend and widespread budget overruns.
McKinsey found that organisations with mature AI-demand forecasting saved 10% more on AI expenditure than peers on average. This is a cross-sectional survey relationship, not proof that forecasting alone caused the difference, and the underlying qualified sample contained 75 enterprises.
LeadDev and Harness found that fewer than half of surveyed organisations had controls designed specifically for AI-written code. The 2026 report surveyed engineers across company sizes and industries; vulnerability detection was used by 44% of large enterprises but only 16% of smaller companies.
More than half of respondents to the LeadDev and Harness release survey selected unclear measurement as their leading AI-code problem. Only 29% said their organisation measured AI tools' impact on teams, even as 58% reported running more experiments than before.
Google Cloud surveyed more than 1,400 senior IT leaders and found most believed their current infrastructure required upgrades for production agentic systems. The result measures perceived readiness, but it identifies a large installed-base constraint as agents add repeated model calls, memory and orchestration demands.
Nearly two thirds of the senior IT leaders in Google Cloud's infrastructure study reported material cost overhead beyond model calls. Google grouped data-transfer fees, excess storage and underused accelerators into an 'inference tax', so the percentage reflects leaders seeing that combined burden rather than one billing line.
Workday analysed more than 25,000 job listings from over 300 companies using Workday Recruiting between July 2025 and April 2026. Technical AI skills appeared in 9% of technology-sector postings in April, up from 7.4% in September. This measures the share of postings in its customer sample, rather than the share of existing technology workers with AI skills.
Flexera's separate 2026 IT asset-management survey found AI software was visible to fewer than a third of organisations. The study sampled 512 technology professionals with ITAM responsibilities worldwide through an independently maintained panel in early 2026.
AI had the largest reported rise in waste across the software categories in Flexera's ITAM survey: 59% saw waste increase, 31% said it was unchanged and 7% saw a decrease. The measure captures practitioners' reported direction of change, not the percentage of AI spend wasted.
More than half of the IT asset-management practitioners in Flexera's global survey said their teams supported AI-spend visibility. That responsibility sits alongside cloud and SaaS licensing work, even though only 31% of organisations reported accurate visibility into AI software itself.
Almost half of organisations surveyed for Flexera's State of ITAM said they tracked AI usage and adjusted contract size. Flexera notes a gap between that claimed activity and the 31% reporting visibility into AI software, so the result should not be treated as evidence of complete cost coverage.
Flexera found that 45% of organisations had a practice for finding AI tools used outside approved channels. The result links software governance with cost and security: unidentified tools can escape contract optimisation, data controls and the organisation's main inventory.
AI application tracking and adoption formed the top combined challenge in Flexera's 512-person ITAM survey. The high response sits beside limited AI-software visibility and increasing reported waste, showing that the difficulty covers both inventory control and financial management.
GitLab's 2026 AI Accountability Report found that multi-tool use is already the norm: 91% of surveyed organisations had two or more AI coding tools active. The figure comes from a Harris Poll survey of 1,528 developers and technology buyers in six countries.
More than half of organisations in GitLab's six-country survey reported three or more AI coding tools in active use. The result measures deployed tools rather than future plans and sits within a 1,528-person Harris Poll sample of developers and technology buyers.
GitLab's respondents widely perceived a speed gain from AI coding tools: 78% said developers were writing and committing code faster. This is a self-reported organisational outcome from 1,528 developers and technology buyers, not a measured change in delivery lead time.
Nearly three quarters of GitLab's survey respondents reported better overall code quality after AI adoption. The finding records respondent assessment rather than repository telemetry, which makes it useful evidence of perceived outcomes rather than an objective defect-rate measure.
GitLab found broad agreement that faster generation moves work downstream: 85% said review and validation had become the bottleneck. The result comes from a Harris Poll survey of 1,528 developers and technology buyers across six countries.
GitLab's survey found a basic provenance gap in 43% of organisations: respondents could not reliably tell AI-written code from human-written code in their own repositories. The same study covered 1,528 developers and technology buyers in six countries.
Among organisations that had experienced an incident in the previous year, 34% could not establish whether AI-generated code played a part. GitLab contrasts that observed failure with 87% of respondents who felt confident their team could answer the question within 24 hours.
Nearly three quarters of GitLab's 1,528 developers and technology buyers expressed concern about the maintainability of AI-generated code. The Harris Poll ran the study across six countries for GitLab's 2026 AI Accountability Report.
GitLab's six-country survey found broad concern that generated code is accumulating faster than organisations can govern it. The 82% result comes from 1,528 developers and technology buyers surveyed by The Harris Poll.
Most respondents agreed that AI increased individual developer productivity while the end-to-end delivery process failed to accelerate at the same pace. GitLab calls this the AI paradox; the figure reflects agreement from its 1,528-person Harris Poll survey.
For 84% of GitLab respondents, the central difficulty was controlling what happens to AI-generated code after it is written. The result shifts attention from generation tools to provenance, ownership, review and production accountability.
Fragmented development tools were the second most common structural barrier to control and traceability in GitLab's study, selected by 40% of respondents. Only difficulty separating AI code from human code ranked higher.
Thirty-nine percent of GitLab's survey respondents identified missing code-origin records as a barrier to AI traceability. Without provenance data, teams cannot reliably reconstruct whether generated code contributed to an incident or who approved it.
Fewer than three in ten respondents described their software-development lifecycle tools as fully integrated with common data and workflows. GitLab presents this integration gap as a practical constraint on tracing and governing AI-generated code.
CrowdStrike's 2026 State of Cloud Detection and Response Survey found that 94% of participating organisations reported cloud intrusions resulting in exposure or exfiltration. This is a survey result about those organisations, not a telemetry-derived estimate of breach prevalence across all cloud users.
Anthropic classified roughly 400,000 privacy-preserved Claude Code sessions from about 235,000 people. Writing accounted for 25%, fixing 26%, and testing or orchestration 5%, for a combined 56% of sessions devoted directly to producing or maintaining code.
Across Anthropic's October 2025 to April 2026 session sample, fixing broken code fell from 33% to 19% of Claude Code sessions. Operating software and non-code work grew in its place, indicating a shift in task mix rather than proof that debugging itself became faster.
Anthropic estimated a 27% rise in average session value by matching tasks to freelance-marketplace postings. The company calls the dollar estimates coarse and uses them for relative comparison, so the statistic is directional evidence that session scope increased, not a realised economic return.
New Relic's 2026 survey asked 200 US technology decision-makers at upper-mid-market and enterprise organisations already using generative or agentic AI in software engineering. Two thirds placed AI's weekly code share in the 51%-75% band.
Among 200 US technology leaders at organisations using AI in engineering, 74% said at least 25% of AI-generated code required substantial post-deployment rework. Respondents attributed the rework to missing context, incomplete data, or flawed system assumptions.
New Relic's Hanover Research survey found 82% of its 200 US technology-decision-maker respondents had experienced a major AI-code production failure in the prior six months. The sample is restricted to larger organisations already using generative or agentic AI in engineering.
Software Improvement Group measured AI-generated code at 1.9% of enterprise production code in its 2026 benchmark. The result comes from analysed systems rather than developer estimates; SIG's wider benchmark spans more than 30,000 systems and 400 billion lines of code.
In Software Improvement Group's testing, AI-generated code carried roughly twice the security-risk violations of human-written code and also scored lower on maintainability. The 2026 report draws on SIG's portfolio benchmark of more than 30,000 systems.
SIG found 86% of code below its recommended maintainability threshold. The finding is drawn from enterprise benchmark data across more than 30,000 systems and 400 billion lines, giving it a different evidence base from developer surveys about perceived quality.
Software Improvement Group reports that stronger architecture cuts issue-resolution time by 30%. Half of the code in its benchmark fell below the recommended architecture-quality score, linking a structural quality measure to operational maintenance speed.
Software Improvement Group places the cost of technical debt between 21% and 40% of IT spending. Its 2026 report combines benchmark data from more than 30,000 systems and over 400 billion lines of code, giving the estimate a measured software-portfolio base.
SIG estimates annual developer-time savings of €870,000 for each system when code-level technical debt is reduced. The calculation sits within a portfolio analysis covering more than 30,000 systems, rather than a survey of perceived productivity.
SIG's benchmark analysis rated 71% of code as having weak security controls. The State of Software dataset covers more than 400 billion lines across over 30,000 systems and 300 technologies, so this is a code-analysis result rather than developer sentiment.
Half of the code in SIG's 2026 benchmark fell below its architecture threshold. SIG separates architectural debt between components from file-level maintainability, linking stronger architecture elsewhere in the report with 30% faster issue resolution.
Checkmarx's report released on 9 June 2026 surveyed 2,350 CISOs, application-security managers and developers across 14 countries through Censuswide. Three quarters reported knowingly deploying vulnerable code. Although titled the 2027 Industry Outlook, this is a 2026 publication reporting survey responses, rather than a prediction of 2027 behaviour.
BCG's fourth annual AI at Work survey covered nearly 12,000 employees, managers and leaders across 14 markets. Among frontline regular AI users, 42% reported saving at least eight hours weekly. The denominator is regular users on the frontline, and the time savings are self-reported rather than independently timed.
OpenAI reported more than five million weekly active Codex users by June 2026, over six times the level following the February desktop-app launch. This is first-party product usage data and includes developers and a growing group of knowledge workers.
Developers remained Codex's largest user group in June 2026, but OpenAI reported that knowledge workers already represented about 20% of users. That group was growing more than three times as fast as the developer user base.
KPMG's Q2 2026 AI Pulse recorded a weighted average of $202 million in planned AI investment for the coming year. The figure reflects surveyed US business leaders' plans, not realised expenditure, and sits within a report focused on the economics of operating AI at scale.
Only about a quarter of organisations in KPMG's Q2 AI Pulse had real-time AI cost information, despite 66% using dashboards and 61% requiring approvals. The result measures timely cost visibility across scaled deployments rather than whether any spending report exists.
More than a third of the leaders in KPMG's Q2 survey selected AI cost management and economic literacy as a constraint. KPMG explicitly includes understanding token and inference costs in that category, distinguishing skills and interpretation from a lack of billing data alone.
KPMG found agent deployment had reached a majority of surveyed organisations in Q2 2026. A narrower measure of orchestration across workflows doubled from 9% to 18%, showing that using at least one agent remained much more common than coordinating several agents as a system.
KPMG's Q2 2026 pulse found organisations paying between 6% and 15% more for employees with strong AI skills. The same survey reported 65% investing in upskilling and reskilling, framing specialised labour as part of the full cost of scaling AI rather than a model or infrastructure charge.
Stack Overflow's April pulse survey of 1,100 developers and working professionals found workplace agent use had risen from 31% in its 2025 Developer Survey. The 59% total includes daily, weekly and less frequent use.
Most technologists in Stack Overflow's 1,100-person pulse survey kept a human in the loop instead of letting agents run on autopilot. The result qualifies the same survey's rapid growth in workplace adoption: use has spread faster than trust in full autonomy.
More than two thirds of Stack Overflow's pulse-survey respondents favoured a single, predictable agent over a more complex multi-agent arrangement. The preference provides a user-side evaluation of workflow design, not a benchmark of which setup completes tasks more accurately.
Six in ten respondents to Stack Overflow's April survey prohibited agents from changing systems without approval. The control was common even as agent adoption rose, indicating that practical deployments still depend on explicit human authorisation.
Accenture's research included executives at 2,000 companies across 15 countries and nine industries. Only 7% met its data-reinventor classification for the capabilities needed to scale advanced AI. The proportion depends on Accenture's maturity framework and should not be read as the share of companies with any usable data.
Verizon's 19th Data Breach Investigations Report analysed 2025 incident data and was published in May 2026. Vulnerability exploitation accounted for 31% of breach entry points, surpassing stolen credentials. This is a finding in the contributed breach dataset, not an estimate that 31% of organisations were breached or that AI caused every exploitation.
Electricity price was the leading location factor for 55% of respondents to Flexential's third annual AI-infrastructure survey. The study covered more than 350 director-level or senior IT decision-makers at companies with over $100 million in annual revenue.
Almost three quarters of the enterprise IT leaders in Flexential's survey were concerned about changing electricity prices feeding into AI running costs. The result separates price volatility from the distinct constraint of grid availability, which 89% said influenced deployment decisions.
Cost reduction and operating efficiency became the most common AI-success measure in Flexential's 2026 survey, ahead of revenue growth at 42%. The statistic records which outcome leaders use to judge AI, rather than the share already achieving a cost reduction.
The share of enterprise IT leaders expecting an AI return within 12 months fell 15 percentage points in Flexential's year-on-year survey. Another 20% said returns were already visible, a proportion that remained broadly flat as infrastructure expenses and deployment complexity rose.
Seven in ten organisations in Flexential's survey devoted at least 10% of IT budgets to AI software, hardware, networking and related initiatives. The threshold covers the wider AI programme rather than inference costs alone and comes from companies with at least $100 million in annual revenue.
Four in ten Flexential respondents said tariffs had caused their organisation to postpone or scale back AI-infrastructure procurement. A separate 54% increased reliance on domestic suppliers, showing policy and supply-chain costs changing both timing and sourcing decisions.
Gartner's 2026 CMO Spend Survey covered 401 marketing leaders in North America, the UK and Europe during January–March 2026. Respondents allocated an average 15.3% of marketing budgets to AI. Most represented companies with more than $1 billion in annual revenue, so the figure principally describes large organisations.
Microsoft's 2026 Work Trend Index surveyed 20,000 AI-using knowledge workers across ten markets. Only 26% reported clear, consistent leadership alignment on AI. This is an employee assessment of organisational conditions from the survey component, rather than a finding from Microsoft 365 product telemetry.
BCG's inaugural Split Decisions survey covered 351 CEOs and 274 board members at companies with at least $100 million in annual revenue. More than a third of CEOs believed their boards overestimated AI's replacement potential. The finding captures CEO perceptions of board expectations, rather than measured automation capability.
IBM and Oxford Economics surveyed 2,000 CEOs and equivalent leaders across 33 geographies and 21 industries from February to April 2026. Respondents reported a Chief AI Officer at 76% of organisations, compared with 26% in the previous edition. These are separate annual survey results, rather than tracking the same firms' hiring decisions.
Workday commissioned The Harris Poll to survey 6,100 active workplace AI users in finance, HR, IT and operations on 2–24 March 2026. Among respondents whose organisations embedded AI in core systems, 60% reported time savings of at least 25%. This cross-sectional association does not establish that integration alone caused the difference.
Qodo surveyed 500 US IT engineers and engineering leaders about AI-generated code. Eighty-nine percent said their organisation had experienced at least one related production incident, a broad measure that includes incidents short of a full outage.
Among Qodo respondents working at organisations with more than 10,000 employees, 40% reported a production outage caused by AI-generated code. It was the highest outage rate of any company-size bracket in the 500-person US survey.
Qodo found that 41% of developers were spending more time on manual code review than they had before AI coding tools. The result captures the uneven distribution of productivity gains: faster generation can transfer work to verification rather than remove it.
Ninety-five percent of developers in Qodo's survey said they scrutinised AI-generated code more closely. The near-universal caution sits alongside the report's incident findings and indicates that higher review effort has become a normal part of AI-assisted delivery.
Atlassian's double-blind State of Teams survey found broad AI use but much thinner workflow integration. It collected responses from 12,035 knowledge workers and 173 Fortune 1000 executives worldwide in January and February 2026.
Atlassian found a wide gap between local speed and enterprise results. The executive findings come from 173 Fortune 1000 leaders surveyed alongside 12,035 global knowledge workers, with 25 additional Fortune 500 executive interviews.
Atlassian identified 14% of teams as its highest-performing AI cohort. Those teams were up to 9.4 times more likely to report that AI increased collaboration, a relative survey comparison rather than a measured multiplication in collaborative output.
Datadog analysed anonymised LLM usage from thousands of customers running AI in production for its 2026 engineering report. Around 5% of model requests failed, with capacity limits responsible for nearly 60% of failures. These are request-level observations in Datadog's customer population, rather than a failure rate for every AI application.
PwC surveyed 1,217 senior executives across 25 sectors and analysed revenue and efficiency gains attributed to AI, adjusted against industry medians. The top 20% of organisations accounted for 74% of the study's measured AI value. This is a concentration within PwC's surveyed companies, not a census of the global economy.
JetBrains' January 2026 AI Pulse survey found regular workplace AI use among nine in ten professional developers. The survey covered more than 10,000 developers worldwide, was localised into eight languages, and asked specifically about coding and development work.
JetBrains separated general chatbots from developer-specific products and found 74% of professional developers had adopted a specialised AI development tool by January 2026. The global AI Pulse sample included more than 10,000 respondents across eight survey languages.
GitHub Copilot remained the most-used specialised AI development tool in JetBrains' January 2026 survey, with 29% global workplace adoption. Adoption reached 40% among developers in organisations with more than 5,000 employees.
JetBrains measured Claude Code workplace adoption at 18% in January 2026, up from roughly 3% in its April-June 2025 wave. The comparison uses the same global survey programme and shows adoption growth rather than a product usage count.
Among products measured in JetBrains' January 2026 AI Pulse survey, Claude Code recorded 91% customer satisfaction and a Net Promoter Score of 54. The figures describe sentiment among users, not market share or independently measured code quality.
JetBrains measured Codex workplace adoption at 3% in January 2026, when 27% of developers had heard of it. The fieldwork predates the Codex desktop launch and ChatGPT promotion, making this a useful pre-launch baseline rather than a current market-share estimate.
Faros analysed two years of workflow telemetry from 22,000 developers across more than 4,000 teams. In 80% of those teams, AI-tool weekly active use exceeded 50%, a behavioural adoption measure rather than a purchased-seat count.
Faros telemetry shows the acceptance rate for AI-generated code rising from 20% to 60% as adoption deepened. The report compares each organisation's lowest- and highest-adoption periods across two years of data from 22,000 developers.
Faros found epics completed per developer increased 66.2% between the same organisations' low- and high-AI-adoption periods. The telemetry covers 22,000 developers and more than 4,000 teams, measuring completed roadmap units rather than self-reported time saved.
Task throughput per developer rose 33.7% in Faros's comparison of low- and high-adoption periods within the same organisations. The underlying telemetry spans two years, 22,000 developers, and more than 4,000 teams.
Faros measured a 16.2% increase in pull-request merge rate per developer when organisations moved from low to high AI adoption. This is workflow telemetry, but it measures merged volume rather than whether the code remained valuable after release.
Faros measured an 861% increase in the ratio of deleted to added lines for merged code under high AI adoption. The company notes several possible causes, including rework and productive refactoring, so the result signals instability without proving that every deleted line was waste.
Across Faros's telemetry, the ratio of production incidents to merged pull requests rose 242.7% between low- and high-adoption periods. Faros explicitly describes this as a ratio, not proof that each incident was caused by the latest PR.
Faros found bugs per developer 54% higher as AI adoption deepened, compared with a 9% rise in its prior report. The 2026 result uses two years of engineering-system telemetry from 22,000 developers and more than 4,000 teams.
Median time in review rose 441.5% in Faros's within-organisation comparison of low- and high-adoption periods. Time to first review increased 156.6% and average review time 199.6%, showing pressure at several points in the review queue.
Faros recorded a 31.3% increase in pull requests merged without any review as organisations moved to high AI adoption. The telemetry does not establish why review was skipped, although rising volume and longer review queues are consistent with capacity strain.
Respondents to Flexera's 2026 State of the Cloud survey estimated that 29% of their infrastructure- and platform-cloud expenditure was wasted. The annual global study covered 753 cloud decision-makers and users; the figure is a respondent estimate rather than a provider-billing audit.
Cloud-cost control was cited by 85% of Flexera's 753 respondents, ahead of security at 82% and software-licence management at 78%. The result comes as estimated waste rose and AI introduced more variable services and pricing models into cloud estates.
Flexera found that monthly cloud expenditure exceeded $5 million at more than three quarters of large-enterprise respondents. The threshold covers overall cloud spend rather than AI alone, but establishes the financial base on which growing AI workloads and a 29% estimated waste rate operate.
Cost optimisation remained the most common priority in Flexera's 2026 global cloud survey, selected by 68% of respondents. The report distinguishes this initiative ranking from its separate finding that organisations increasingly judge cloud progress using business-value and unit-economic measures.
Almost two thirds of organisations in Flexera's State of the Cloud survey reported an established FinOps team. The wording includes teams that advise, manage or execute optimisation, so the percentage reflects organisational presence rather than a single maturity level or operating model.
Nearly half of Flexera respondents used unit metrics such as cost per service, up from 40% in 2025. The survey records adoption of unit economics, not the quality or coverage of each implementation, but it shows more teams relating spend to output rather than tracking the bill alone.
Three in ten respondents to Flexera's cloud survey selected unpredictable cost as a leading generative-AI scaling problem. The report attributes the forecasting difficulty to variable workloads, new pricing units and the need to balance AI-resource rightsizing against performance.
GitGuardian retested its historical set of exposed credentials for the 2026 State of Secrets Sprawl report. More than 64% of secrets confirmed valid in 2022 still worked in January 2026. This is a longitudinal result for that tracked credential set, rather than the proportion of all leaked secrets that remain active.
GitHub reported 60 million Copilot code reviews between the feature's April 2025 launch and 5 March 2026. This is first-party product telemetry rather than a survey estimate, and GitHub says usage grew tenfold over that period.
GitHub says Copilot now accounts for more than one in five code reviews on its platform. More than 12,000 organisations had also configured it to run automatically on every pull request by March 2026.
Palo Alto Networks' 2026 Unit 42 Global Incident Response Report analysed more than 750 incidents. Identity weaknesses were exploited in 89% of investigations. This describes cases handled by its response team, which can differ from the wider population of organisations, and does not establish that AI was responsible for each incident.
Sonar's survey found 64% of professional developers had used AI agents in development: 39% were experimenting and 25% used them regularly in daily work. The result distinguishes goal-driven, multi-step agents from simpler assistant features.
Snyk's 2026 State of Agentic AI Adoption report uses anonymised AI-BOM telemetry from more than 500 early adopters of Evo by Snyk. It identifies autonomous systems in one quarter of observed organisations. The selected customer population is an early-adopter sample and should not be treated as a representative enterprise adoption rate.
OpenAI said overall Codex usage doubled between GPT-5.2-Codex's mid-December launch and the Codex desktop app announcement on 2 February 2026. More than one million developers had used Codex in the preceding month, providing a dated pre-app adoption baseline.
The FinOps Foundation's sixth annual survey found that AI costs had entered the remit of nearly every respondent, compared with 63% in 2025 and 31% in 2024. The 2026 study collected 1,192 responses from a global community representing more than $83 billion in annual cloud spend.
SaaS entered the present or near-term scope of nine in ten respondents to the FinOps Foundation survey, up from 65% in 2025. The question measures both current management and stated 12-month plans, so it should not be read as 90% already operating a mature SaaS cost practice.
Licensing appeared in the current or planned remit of 64% of State of FinOps respondents, 15 percentage points above 2025. The result sits within a multi-select question on technology-spend categories and documents FinOps expanding beyond public-cloud infrastructure.
Private cloud was in scope or planned for 57% of the global FinOps community surveyed in 2026, up 18 points from the previous year's 39%. The finding covers cost-management responsibility, not the share of organisations running private-cloud workloads.
Almost half of State of FinOps respondents included data-centre expenditure in their current or near-term remit, compared with 36% in 2025. The Foundation's 1,192-response survey shows cost management following AI infrastructure into environments outside the public cloud.
The FinOps Foundation found that more than a quarter of respondents were adding or planning to add labour expenditure to their cost view. This is an early measure of scope, not a claim that teams can already allocate people costs reliably to individual AI or cloud services.
CrowdStrike's 2026 Global Threat Report records average eCrime breakout time of 29 minutes, with the fastest observed case at 27 seconds. Breakout time describes movement from an initially compromised host to another host. The measure concerns observed intrusions and should not be interpreted as time to discover, patch or fully recover from a vulnerability.
Across the ten leading AI coding tools in Sonar's survey, 35% of access occurred through personal accounts. The rate rose to 52% for ChatGPT, exposing a governance gap that tool-provisioning policies alone cannot see.
Zscaler's 2026 AI Security Report analysed 989.3 billion AI/ML transactions from roughly 9,000 organisations during 2025. Transfers to AI/ML applications rose 93% to 18,033 terabytes. The figure measures traffic in Zscaler's platform population; it does not mean that all transferred data was sensitive or leaked.
Sonar's global survey of 1,149 professional developers found near-universal reservations about functional correctness: 96% did not fully trust AI-generated code. The measure is about complete trust, so it should not be read as 96% believing all AI code is wrong.
Despite the widespread trust gap, only 48% of developers in Sonar's survey said they always checked AI-generated or assisted code before a commit. The question measures consistent developer verification, not whether automated checks run elsewhere in the pipeline.
More than a third of developers in Sonar's 1,149-person survey said AI output required greater review effort than human-written code. The finding supports the reported verification bottleneck, while remaining a perception measure rather than timed review telemetry.
Sonar found 61% of surveyed professional developers agreed that AI tools often generate plausible-looking but unreliable code. The result captures developer experience of hidden defects and explains why surface readability is a poor substitute for functional verification.
More than half of the 648 senior US technology professionals surveyed by KPMG said the cost of remediating technical debt was crowding out new programmes. Every respondent worked at a company with annual revenue above $100 million.
KPMG found weekly glitches in foundational enterprise IT at four in ten surveyed US companies. The rate had fallen by 18 percentage points from the prior year's survey, but remained a recurring operational cost of older systems.
KPMG's survey found 62% of US technology leaders aim to move as quickly and cheaply as possible even when doing so harms security, scalability or data standardisation. The comparison was 69% globally, making the pressure widespread rather than US-specific.
Only one in ten US organisations in KPMG's survey placed their emerging-technology implementations at the highest maturity stage, down from 25% the year before. The result comes from 648 senior technology professionals at companies above $100 million in revenue.
Dynatrace's Pulse of Agentic AI 2026 surveyed 919 senior leaders responsible for agents at enterprises with at least $100 million in revenue. Y2 Analytics conducted the research in November–December 2025. Leaders reported human verification for 69% of agent-powered decisions; the figure is a survey estimate rather than inspection of agent execution logs.
CyberArk commissioned Ponemon Institute to study nearly 2,000 IT and security practitioners globally for its January 2026 PKI report. More than half reported unplanned outages linked to expired or misconfigured certificates. The result identifies a reported operational consequence of certificate-management failures, rather than the frequency of outages per organisation.
PwC's 29th Global CEO Survey collected responses from 4,454 CEOs in 95 countries and territories. More than half reported no significant financial benefit from AI so far. This describes CEOs' assessments at the time of the survey; it does not imply that their companies had no AI deployments or operational gains.
BCG's 2026 AI Radar surveyed 2,360 executives, including 640 CEOs, across 16 markets. Nearly three quarters of CEOs said they were their organisation's main AI decision maker, twice the share in the previous edition. The result concerns reported executive responsibility and compares annual survey editions.
Capgemini's multi-year AI study surveyed 1,505 director-level and more senior executives in November 2025 and published its findings in January 2026. All participating organisations had more than $1 billion in annual revenue and some AI deployment. The 38% result therefore describes an AI-adopting large-enterprise sample.
Capgemini surveyed 500 C-suite executives, including 100 CEOs, at companies with more than $10 billion in revenue across 16 countries and 13 industries. Fieldwork took place in August–September 2025, with publication in January 2026. Just 11% said they disclosed or planned to disclose AI use in decisions; the result concerns stated communication practice.
Workday's Beyond Productivity study surveyed 3,200 full-time employees who actively use AI at organisations with at least $100 million in annual revenue. Hanover Research fielded it in November 2025; the results were published in January 2026. Nearly 40% of reported saved time went back into checking, rewriting and correcting AI output.
Sonar's 2026 State of Code survey found daily use among 72% of developers with AI coding-tool experience. The survey included 1,149 professional developers globally; this denominator excludes people who had never tried the tools.
Professional developers in Sonar's 1,149-person global survey estimated that 42% of their committed code was generated or assisted by AI. This is a respondent estimate of code share, distinct from repository-level authorship telemetry.
Sonar found AI use had moved well beyond prototypes: 58% of surveyed professional developers used it for business-critical or mission-critical services. The same global survey recorded use in customer-facing applications at 73%.
In Sonar's survey, 88% of 1,136 developers selected at least one negative effect of AI-assisted or generated code on their team's technical debt. The options covered unreliable, duplicated and hard-to-maintain code, integration problems, vulnerabilities and weak documentation.
The most common negative effect in Sonar's 1,136-developer technical-debt question was code that looked correct but was not reliable. Sonar notes that this failure mode can evade casual review because plausibility creates false confidence.
Four in ten respondents to Sonar's technical-debt question said AI had added unnecessary or duplicative code. This was the second most frequently selected negative impact among the 1,136 developers answering that part of the survey.
Twenty-nine percent of the developers in Sonar's technical-debt analysis selected unreliable or buggy generated code as a negative impact. The result is separate from the larger group reporting code that merely looked correct while being unreliable.
More than a quarter of Sonar's 1,136 respondents linked AI-assisted code with maintainability or refactoring difficulty. The survey treated this as one of several ways generated code can increase a team's technical-debt burden.
One in five developers answering Sonar's technical-debt question reported integration or compatibility issues from AI-generated or AI-assisted code. The finding places system fit, rather than code generation alone, among the recurring sources of AI-related debt.
Seventeen percent of the 1,136 developers in Sonar's technical-debt section said AI-assisted or generated code had increased vulnerabilities. Respondents could select multiple effects, so this figure measures reported exposure rather than a share of all vulnerabilities attributable to AI.
Poor or missing documentation was selected by 14% of developers as a negative AI effect on technical debt. It was the least common of Sonar's seven listed problems, but still affected roughly one in seven respondents to the 1,136-person question.
Managing technical debt, including refactoring complex or outdated code, was the most frequently chosen source of core-development toil in Sonar's 1,149-developer survey. It ranked ahead of debugging legacy code, finding system information and correcting AI-written code.
Nearly one third of the 1,149 developers surveyed by Sonar placed legacy or poorly documented code in their five biggest sources of productivity-sapping work. It was second only to managing technical debt among the core-development tasks listed.
One in five of Sonar's 1,149 respondents put reworking output from AI coding tools in their five largest sources of development toil. The result quantifies review and repair work that is absent from simple measures of generation speed.
Sonar found positive effects alongside the negative ones: 93% of 1,136 developers selected at least one AI-assisted improvement to documentation, tests, refactoring, maintainability, defect rates or security. The questions allowed multiple answers, so benefits and harms can coexist for the same respondent.
Improved documentation was the most common positive technical-debt effect in Sonar's survey of 1,136 developers. The share rose to 65% among respondents with more than 20 years of experience, according to Sonar's cohort analysis.
More than half of developers answering Sonar's technical-debt question credited AI with better test coverage and debugging. It was the second most common benefit among 1,136 respondents, behind documentation and ahead of refactoring.
Nearly half of the 1,136 developers in Sonar's technical-debt analysis reported that AI had helped with refactoring or optimisation. The finding captures a use of AI for reducing an existing backlog rather than generating net-new code.
Almost a third of respondents to Sonar's 1,136-person technical-debt question reported better maintainability from AI-assisted or generated code. Because the survey also recorded maintainability harms, the result shows that outcomes vary by use and team rather than moving in one direction.
LinkedIn's January labour-market report put global hiring 20% below its pre-pandemic level and job transitions at a ten-year low. The finding comes from LinkedIn's Economic Graph rather than a respondent estimate; the network had more than 1.3 billion members when the report was published.
LinkedIn measured a 70% annual rise in US job postings that ask for AI literacy, including prompt engineering. The report distinguishes these broadly applicable skills from specialist AI-engineering skills, showing demand spreading beyond technical roles.
LinkedIn counted at least 1.3 million new opportunities across roles such as data annotator, AI engineer and forward-deployed engineer. The total is drawn from LinkedIn's global jobs and member data and covers the two years preceding its January 2026 report.
India recorded the strongest hiring level in LinkedIn's country comparison, 40% above its pre-pandemic baseline. LinkedIn calculated a three-month trailing hiring-rate average and compared October 2025 with October 2019.
The UAE's LinkedIn Hiring Rate stood 37% above its pre-pandemic level, second only to India in the report's country comparison. The measure uses members who added a new employer in the month their job began, with a three-month trailing average for October 2025 versus October 2019.
LinkedIn found that hiring across technology, information and media was 27% below its October 2019 baseline by October 2025. That placed the sector among the weaker parts of the global labour market despite rapid growth in AI-specific roles.
Professional-services hiring was 29% below its pre-pandemic level in LinkedIn's cross-industry comparison. The figure compares three-month trailing hiring-rate averages for October 2025 and October 2019 across the network's member data.
LinkedIn classified 3% of US members as AI-skilled in 2025. Its definition requires at least two AI-engineering skills or employment in an AI-engineering occupation, making this a tighter measure than self-reported familiarity with AI tools.
Engineering tied for the highest AI-skill concentration of any job function in LinkedIn's 2025 US data. Members counted only if they listed at least two AI-engineering skills or worked in an AI-engineering occupation.
Product management matched engineering for the highest share of AI-skilled members in LinkedIn's US workforce data. The result shows specialist AI capability extending into product roles, using LinkedIn's two-skill-or-AI-occupation threshold.
LinkedIn found employees at organisations using LinkedIn Learning developed AI skills at 3.4 times the year-on-year rate of employees elsewhere. The comparison comes from behaviour on LinkedIn's platform, though it cannot by itself establish that access to the product caused the difference.
LinkedIn estimates companies can expand their global AI-talent pool by 8.2 times when they search by skills rather than degrees or previous job titles. The estimate uses LinkedIn's member skills and occupation data and describes addressable talent, not completed hires.
LinkedIn reports that its AI-driven hiring products can cut time-to-hire by 30%, including for hard-to-fill AI roles. This is first-party product evidence from LinkedIn rather than a neutral market benchmark, so the statistic is best read as measured customer performance on its own platform.
Applicants already connected on LinkedIn to someone at the company were 3.6 times as likely to be hired as cold applicants. LinkedIn also found the share of pre-connected applicants rose 12% from 2023, making professional networks a stronger filter as application volumes increased.
Seven in ten executives surveyed by LinkedIn said AI tools improve employee efficiency. The sentiment result was collected in September and October 2025 and published in the January 2026 labour-market report; it records leaders' assessment rather than directly observed output.
Workday Government and GovExec reported a June 2026 Market Connections poll of 200 US federal civilian and defence IT employees. Almost all said AI had entered their workflows. The result describes this federal IT sample and should not be extrapolated to the entire federal workforce or interpreted as autonomous agent use.
Deloitte's 2026 State of AI in the Enterprise report surveyed 3,235 senior business and IT leaders across 24 countries in August–September 2025. Only one quarter reported putting at least 40% of their AI pilots into production. The result is an organisational conversion threshold, not a claim that 25% of all pilots succeeded.
Sonatype followed a fixed cohort of enterprise applications continuously monitored from June 2022 to June 2026, keeping evaluation criteria consistent. The average application accumulated 4.31 times as much critical and high-severity risk. The study explicitly does not isolate AI's causal contribution; newly disclosed weaknesses can raise measured risk even without code changes.
CodeRabbit analysed 470 open-source pull requests: 320 identified as AI-co-authored and 150 classified as human-only. AI PRs averaged 10.83 issues each versus 6.45 for the comparison group. The authors caution that authorship was inferred from public signals and could not be guaranteed for every PR.
Logic and correctness findings were 75% more common in CodeRabbit's AI-co-authored PR sample. The category covers business-logic mistakes, incorrect dependencies, flawed control flow and misconfiguration, with findings normalised per 100 pull requests.
Readability showed the largest gap in CodeRabbit's open-source PR study, occurring more than three times as often in the AI-co-authored group. The findings capture departures from local naming, clarity and structural conventions rather than functional defects alone.
CodeRabbit found nearly twice as many error-handling gaps in AI-co-authored PRs. Missing null checks, guard clauses, early returns and complete exception paths were recurring examples, making this a reliability concern as well as a code-style difference.
Security findings were amplified in CodeRabbit's AI group, reaching 2.74 times the human-only rate for the strongest individual pattern. Improper password handling and insecure object references featured prominently; no vulnerability class appeared exclusively in AI code.
Excessive input/output operations were about eight times more common in the AI-co-authored PR sample. The absolute number of performance findings was small, so CodeRabbit presents this as a strong skew in a comparatively rare category rather than a universal AI-code trait.
Formatting findings occurred 2.66 times as often in CodeRabbit's AI-co-authored PRs, including spacing, indentation and structural inconsistencies. The gap persisted in repositories that already used formatters and linters.
AI-co-authored changes contained nearly twice as many naming inconsistencies in CodeRabbit's study. Generic identifiers and terminology that drifted from repository conventions added review and maintenance work even when the code remained functional.
Using high-frequency payroll data from the largest US payroll processor, Stanford researchers found early-career workers (ages 22-25) in the most AI-exposed occupations saw a 16% relative decline in employment since generative AI's widespread adoption, while more experienced workers in the same occupations, and younger workers in less-exposed occupations, saw stable or growing employment. A follow-up note ruled out interest rates as an alternative explanation and found the AI-linked decline becomes statistically distinguishable starting in 2024.
GitHub's 2025 Octoverse report, drawing on activity across its 180 million+ developer accounts, found nearly 80% of newly joined developers used GitHub Copilot within their first week on the platform. The same report counted 4.3 million AI-related repositories on GitHub, nearly double the count from two years earlier, and 1.1 million public repos now importing an LLM SDK, up 178% year-over-year.
JetBrains' State of Developer Ecosystem 2025 survey of 24,534 developers across 194 countries found 85% now use AI tools regularly for coding and development, but only 62% rely on at least one dedicated AI coding assistant, agent, or code editor — the rest lean on general-purpose chat tools instead. 15% of developers hadn't adopted AI tools in their daily work at all.
Google's 2025 DORA report, based on survey responses from nearly 5,000 technology professionals, found AI adoption reaching 90%, with more than 80% reporting a productivity increase. For the first time, the report found a positive relationship between AI adoption and both software delivery throughput and product performance — but AI adoption still correlates with reduced delivery stability, and 30% of respondents report little or no trust in AI-generated code.
Anthropic's Economic Index, which studies real, privacy-preserved usage patterns rather than survey responses, found computer-and-mathematical tasks accounted for 36% of Claude.ai conversations in August 2025 — more than any other occupational category. On the API, used mostly by businesses building their own products, that share climbs to 44% of traffic, and software development is the most common use of Claude across every country studied.
Apiiro compared several thousand developers across tens of thousands of repositories at Fortune 50 enterprises. AI-assisted developers produced three to four times as many commits, while their changes were consolidated into fewer and larger pull requests.
Apiiro's first-party repository analysis recorded ten times as many security findings as AI coding use accelerated. The study covered open-source dependencies, insecure code, exposed secrets and cloud misconfiguration rather than a single vulnerability scanner category.
Pull-request volume fell by nearly a third in Apiiro's enterprise dataset even as AI-assisted developers created three to four times more commits. The combination meant more change was packed into each review unit, increasing scope and the number of services touched per PR.
By June 2025, Apiiro was detecting more than 10,000 new security findings a month in AI-generated code across its study population. That monthly count was ten times the December 2024 level and came from customer repository data rather than a survey.
Apiiro found a 76% drop in trivial syntax errors associated with AI-written code. The apparent quality gain did not carry through to deeper design risks: the same data showed sharp increases in privilege-escalation paths and architectural flaws.
Logic bugs fell by more than 60% in Apiiro's analysis of AI-written code. Apiiro distinguishes these surface-level improvements from the systemic security problems that increased in the same repositories, so the result should not be read as an overall defect-rate reduction.
Privilege-escalation paths rose 322% in Apiiro's enterprise repository analysis. These are contextual, architectural risks involving permissions and trust boundaries, which can survive syntax checking and narrow unit tests.
Apiiro recorded a 153% increase in architectural design flaws as AI coding expanded in the organisations studied. The result points to problems spanning services, authentication flows and system boundaries rather than local syntax or formatting defects.
Stack Overflow's 2025 Developer Survey (49,000+ respondents) found 84% of developers now use or plan to use AI tools, up from 76% in 2024. Trust moved the opposite direction: only 33% say they trust AI output accuracy (just 3.1% "highly"), while 46% actively distrust it — up sharply from 31% distrust the year before. Experienced developers were the most skeptical of all.
In Stack Overflow's 2025 Developer Survey, 66% of respondents who described problems with AI tools selected plausible-looking answers that fall short of being correct. The question drew 31,476 responses and allowed people to select more than one frustration.
Stack Overflow found that 45% of developers who reported AI-tool frustrations selected time-consuming debugging, second only to nearly-correct output. The result describes experienced frustration rather than a measured change in debugging duration.
Among 25,349 Stack Overflow survey respondents asked where AI fits into their development workflow, 76% ruled out deployment and monitoring. It was the workflow area with the strongest resistance to AI, reflecting caution around high-responsibility operational work.
Project planning was another boundary in Stack Overflow's 2025 workflow data: 69% of 25,349 respondents said they did not plan to use AI for it. The figure separates interest in coding assistance from willingness to delegate broader coordination decisions.
Stack Overflow found agents were not yet mainstream in 2025: 52% of respondents either did not use agents or stayed with simpler AI tools, while 38% explicitly had no plan to adopt agents. The agent question received 31,877 responses.
Of 12,301 Stack Overflow respondents who described their agent use at work, 84% of software developers selected software development as a purpose. This is a conditional result about existing agent users, not an 84% adoption rate across all developers.
Around 70% of the 12,823 Stack Overflow respondents answering the agent-impact question agreed that agents saved time on particular development tasks. The survey records users' assessments, so it does not establish the size of the time saving.
Only 17% of Stack Overflow respondents assessing agents' impact agreed they had improved team collaboration. It was the lowest-rated impact by a wide margin, contrasting with the much stronger reported gains in individual efficiency.
Stack Overflow's agent-challenges question drew 28,930 responses. Of those respondents, 81% agreed they had security and data-privacy concerns, showing that apprehension extends beyond whether an agent produces correct code.
Accuracy was the most widespread agent concern in Stack Overflow's 2025 survey, selected by 87% of the 28,930 people answering the question. The result includes non-users as well as users, making it a measure of broad developer concern rather than failure observed in production.
In a randomised controlled trial, 16 experienced open-source developers completed 246 real tasks in their own mature repositories, half with AI tools allowed and half without. Measured completion time was 19% slower with AI — yet the same developers estimated afterward that AI had sped them up by 20%.
DORA's April 2025 industry study found that 89% of organisations prioritised AI integration in applications, while 76% of technologists already used AI in some part of their daily work. The report combined developer interviews with quantitative organisational data.
DORA modelled a 25% increase in AI adoption against a 2.1% increase in self-reported individual productivity. The result is an observational association from workplace data, not a claim that every team receives the same causal gain.
DORA found that giving developers dedicated work hours to explore AI was associated with a 131% increase in team adoption. The number describes the relationship observed across organisations and does not mean experimentation time alone guarantees that increase.
Organisations with a clear policy for acceptable AI use showed 451% higher adoption in DORA's analysis than organisations without one. The finding supports policy clarity as an adoption condition, while its observational design does not isolate policy as the only cause.
In DORA's 2025 study, a 25% increase in AI adoption was associated with a 1.5% reduction in system-level delivery throughput, even while individuals reported productivity gains. The finding predates DORA's later 2025 report, which observed a positive throughput relationship as organisational practices changed.
DORA associated a 25% increase in AI adoption with a 7.2% reduction in delivery stability. The result helps explain why local coding speed can coexist with more difficult system outcomes when testing, review and feedback capacity do not scale with code output.
HackerRank's 2025 report combined a global survey of 13,732 people across 102 countries with platform data from more than three million annual assessments. Ninety-seven percent of developers said they used at least one AI assistant.
HackerRank found that 61% of developers used at least two AI tools at work, moving between general chat systems and development-specific assistants. The result shows that adoption often means a stack of tools rather than loyalty to one product.
ChatGPT appeared in 89% of developers' work toolkits in HackerRank's survey, 13 percentage points more than the year before. Most users of specialised tools also used it, including 80% of Cursor users and 84% of GitHub Copilot users.
Developers in HackerRank's survey estimated that AI generated 29% of their code on average. This is a self-reported share across respondents and roles, not a repository-level measurement of committed or production code.
Fifteen percent of HackerRank's developer respondents said AI produced a majority of their code. The share rose to 37% among developers who said they almost always used AI tools, showing how averages conceal a much more intensive adopter group.
Eighty-five percent of developers in HackerRank's survey reported faster project completion with AI. The report also notes a selection effect: the heaviest adopters may be unusually inclined to experiment and refine their workflows, so the survey does not attribute the entire difference to the tools.
HackerRank's heaviest AI users reported that 48% of their code was generated by AI, compared with 29% across developers overall. Within this heavy-use group, 37% said AI wrote the majority of their code.
HackerRank found that 67% of developers felt more pressure to deliver quickly because of AI. Eighty-four percent of engineering leaders said they had raised productivity expectations, a 17-point difference between management action and developer experience.
Nothing matches that search or tag. Try a broader term, or clear the tag filter.