Written by: Mark Hull, Co-Founder and CEO, Exceeds AI | Last updated: July 9, 2026
Key Takeaways
- AI coding tools now generate or assist 42–54% of committed code, with adoption expected to reach 65% by 2027. Most leaders still lack visibility into which lines are AI-generated and how those lines perform over time.
- Productivity gains are real, with up to 66% more epics completed and 60% higher PR throughput. These benefits come with 1.7x more issues, 54% higher bug rates, and 45% of AI code containing security vulnerabilities.
- AI-generated code shows 3x more readability issues, 2x more error-handling gaps, and significantly higher churn rates. Teams rewrite AI code about 1.5x faster than comparable human-authored code.
- ROI calculations must include token costs of $200–$2,000+ per engineer monthly and account for 4–6 hours of rework for every 10 hours saved. Net productivity gain often lands near 40–60% without strong governance.
- Engineering leaders can unlock durable AI ROI by implementing commit-level attribution and longitudinal tracking. Start your free pilot with Exceeds AI today.
How This Data Was Measured
This report relies on diff-level analysis at the commit and PR level across multiple AI coding tools. The system establishes attribution through per-tool checkpoint materializers that resolve edit evidence against the actual working tree at commit finalization. It does not rely on post-hoc heuristics applied only to merged diffs.
Each line carries a tool, model, session, interaction mode, and timestamp. Lines that cannot be attributed with confidence are recorded as unknown rather than silently assigned to AI or human categories. Longitudinal outcome tracking follows AI-attributed code for 30 or more days, measuring incident rates, follow-on edits, test coverage, and rework patterns against a matched human-authored baseline.
Productivity Lift From AI-Assisted Coding
Faros AI’s 2026 AI Engineering Report, drawing on two years of telemetry from 22,000 developers across 4,000+ teams, found that moving from low to high AI adoption correlated with a 66% increase in epics completed per developer, a 33.7% increase in task throughput per developer, and a 16.2% increase in PR merge rate per developer.
GetDX’s Q4 2025 AI-Assisted Engineering Impact Report, covering 435 companies and 85,350 developers, shows a similar throughput pattern. Daily AI users merged 2.3 PRs per week at the median versus 1.4 PRs per week for non-users, a 60% throughput advantage based on merged PR data.
Larridin’s 2026 Developer Productivity Benchmarks report that cycle time for AI-assisted PRs is longer than for human-only PRs, even as elite teams achieve sub-8-hour AI-assisted PR cycle times. The industry-average AI velocity multiplier sits at 1.7x complexity-adjusted throughput, with elite teams reaching 1.8–2.0x. Throughput gains therefore come more from higher volume than from faster completion of each individual task.
METR’s 2025 randomized controlled trial found that 16 experienced open-source developers completed 246 real tasks 19% slower when AI tools were allowed, despite expecting a 24% speedup beforehand. The gap between perceived and measured productivity appears consistently across multiple independent studies.
Bug and Logic Error Rates in AI Code
Faros AI’s 2026 report found bugs per developer increased 54% as AI adoption grew, compared to a 9% increase reported in its 2025 study. The incidents-to-PR ratio increased 242.7%, and monthly incidents increased 57.9% as teams moved from low to high AI adoption.
A large-scale empirical study of 304,362 verified AI-authored commits across 6,275 GitHub repositories found that more than 15% of AI-authored commits from every coding assistant introduce at least one quality issue.
These productivity gains, however, come with a quality cost that adoption metrics alone do not capture. When teams track what happens to AI-generated code after merge, they see a consistent pattern of elevated defect rates across independent datasets.
Security and Maintainability Differences
Veracode’s 2025 GenAI Code Security Report, testing 80 coding tasks across 100+ LLMs, found that 45% of AI-generated code contains a security vulnerability, with Java showing a 72% failure rate. Veracode also found that AI tools fail to defend against cross-site scripting in 86% of relevant code samples and log injection appears in 88% of AI-generated outputs. These findings highlight concentration in well-known attack vectors rather than rare edge cases.
Apiiro’s September 2025 report monitoring 7,000+ developers across 62,000 repositories found a 10x increase in AI-assisted security findings over six months, with AI-assisted developers exposing cloud credentials at nearly 2x the rate of non-AI peers. This production data reinforces the lab results from Veracode.
On readability and maintainability, CodeRabbit found readability issues 3x more frequent and error-handling gaps nearly 2x more common in AI-co-authored code. The Google 2025 DORA Report found that a 90% increase in AI adoption correlates with a 91% increase in code review time and a 154% increase in pull request size. The review burden mentioned earlier, with 38% of developers reporting that AI code takes more effort to review, compounds when security-specific checks are required.
Architecture and Long-Term Outcomes
GitClear’s analysis of 211M+ changed lines shows code churn within two weeks rising from a 3.3% pre-AI baseline in 2021 to 5.7% in 2024 and 7.1% in 2025. The AI vs Human Turnover Ratio, which measures how much faster AI-generated code is rewritten relative to human-authored code, signals insufficient review discipline when it exceeds 1.5x.
Engineering leaders report rework rates increase by 30–60% within six months of heavy AI adoption, with Stage 5 “Velocity Collapse” occurring within 6–12 months, where refactoring becomes a prerequisite for new work. An Anthropic randomized controlled trial (arXiv:2601.20245, 2026, 52 engineers) found that developers who used AI scored 17% lower on tests of understanding their own code, with the largest drop in debugging performance.
Adoption Patterns by Team Size, Tool Mix, and Project Type
Daniotti et al. (Science, 2026), analyzing more than 30 million GitHub commits by 160,097 developers, found that by the end of 2024 an estimated 29% of Python functions in the US were produced with substantial AI support, while France and Germany reached 23–24% and India showed considerable catch-up. Exploration gains from AI, such as experimenting with new libraries, accrued exclusively to senior-level developers. Early-career developers used AI tools more frequently but showed no statistically significant productivity or exploration benefits.
Industry-average AI-assisted code share is 15–25% of lines written, with top-quartile teams at 40–60% and elite teams above 75%. Block reports that approximately 95% of their engineers regularly use AI. In the Faros AI dataset, 80% of teams exceed the 50% weekly active user threshold for AI tools and the acceptance rate of AI-generated code has risen from 20% to 60%.
Faros AI’s 2026 report found pull requests merged without any review up 31.3% under high AI adoption, and median time to first PR review up 156.6%. These patterns are more pronounced in teams with stretched manager-to-IC ratios, where bandwidth for mentorship and code inspection is already limited. If your organization is seeing similar review bottlenecks as AI adoption scales, commit-level attribution can highlight which teams need intervention before quality degrades further, so you can start a targeted free pilot with Exceeds AI.
Interpretation: Governance, Technical Debt, and ROI
Healthy ROI on AI coding tools in 2026 is 2.5–3.5x at the industry average and 4–6x for top-quartile organizations, when the cost denominator includes actual token and usage-based costs rather than just seat licenses. Total AI cost per developer per month averages $200–$600, with agentic tools like Claude Code adding $200–$2,000+ per engineer monthly in token spend beyond seat licenses.
The ROI calculation breaks down when technical debt is excluded. For every 10 hours saved by AI code generation, 4–6 hours are spent on rework, debugging, and incident response, yielding only 40–60% net productivity gain. Lightrun’s 2026 State of AI-Powered Engineering report found that 43% of AI-generated code requires manual debugging in production, and 88% of organizations need 2–3 redeploys to fix an AI-generated change.
Governance decisions require commit-level attribution. Token spend, tool mix, interaction mode, and longitudinal outcome data must connect in a single view before leaders can answer whether a given AI investment is generating durable value or accumulating hidden cost. Metadata-only platforms cannot provide that connection, so leaders need to see what commit-level visibility reveals in their own repositories through a free Exceeds AI pilot.
Where AI Code Performs Well
AI code performs comparably or better under specific conditions. In an architect-led AI-assisted modernization engagement for a Kansas City insurance platform, teams compressed an estimated 18–24 month effort into roughly five months, cutting development costs by more than 75%. TELUS teams created over 13,000 custom AI solutions while shipping engineering code 30% faster and saved over 500,000 hours with an average of 40 minutes saved per AI interaction.
GetDX Q4 2025 data shows that increasing GenAI enablement by 25% correlates with roughly 8% higher code maintainability, 10.6% higher change confidence, and 18.2% less time loss. The common thread in positive outcomes is disciplined engineering practice: architecture-first delivery, automated testing, mature CI/CD pipelines, and active code review. DORA research confirms that AI primarily amplifies an organization’s existing delivery capabilities, with teams that have strong baselines seeing throughput improvements and teams without those capabilities often experiencing increased instability despite faster coding.
Practical Takeaways: Evaluation Criteria for Leaders
Engineering leaders evaluating AI coding tools and their impact need criteria that go beyond acceptance rates and seat utilization. The following framework applies regardless of which tools a team uses.
- Establish commit-level attribution before expanding adoption. Aggregate adoption stats cannot distinguish which lines are AI-generated, which tool produced them, or which interaction mode was used. Attribution at the line level is the prerequisite for every other measurement.
- Track the AI vs Human Turnover Ratio. A ratio above 1.5x indicates AI-generated code is being rewritten at a significantly higher rate than human-written code, signaling insufficient review discipline or poor prompt practices.
- Measure longitudinal outcomes, not just cycle time. The incidents-to-PR ratio increased 242.7% under high AI adoption in the Faros AI dataset. Incident rates 30+ days after merge provide a more reliable quality signal than initial review pass rates.
- Include token spend in the ROI denominator. Agentic tools can add $200–$2,000+ per engineer monthly in token spend beyond seat licenses. ROI calculations that omit usage-based costs overstate returns.
- Segment findings by team, tool, and interaction mode. Org-wide averages mask the patterns that matter. A team using agent mode without a plan phase produces a different risk profile than one using structured ask-then-edit workflows.
- Apply higher test coverage thresholds to AI-generated code. ContextQA’s 2026 review recommends coverage thresholds of 85–90% for AI-generated code versus 70–80% for human-written code.
These criteria require commit-level data that most analytics platforms do not provide, so leaders should connect their repositories to Exceeds AI and receive their first attribution report in under 10 minutes.
Frequently Asked Questions
How does commit-level AI attribution differ from the analytics built into individual AI coding tools?
Tool-native analytics such as GitHub Copilot’s built-in dashboards report usage statistics like acceptance rates, lines suggested, and active users, but they cannot connect those signals to downstream outcomes. They do not show whether AI-touched code had higher incident rates 30 days later, required more follow-on edits, or introduced security vulnerabilities that passed initial review.
They are also blind to every other AI tool in the stack. If engineers use Cursor, Claude Code, or Windsurf alongside Copilot, those contributions remain invisible to Copilot’s own analytics. Commit-level attribution resolves edit evidence against the actual working tree at the moment of commit, assigns each line to a specific tool, model, session, and interaction mode, and then tracks those lines longitudinally. That approach connects AI usage to business outcomes rather than stopping at adoption counts.
How current is the data underlying these findings?
The findings in this report draw on sources published between late 2024 and mid-2026. These sources include Faros AI’s 2026 AI Engineering Report, which covers 22,000 developers and 4,000+ teams over two years of telemetry, and DX’s Q4 2025 AI-Assisted Engineering Impact Report, which covers 85,350 developers across 435 companies.
The dataset also includes Daniotti et al., published in Science in 2026 with more than 30 million GitHub commits, Sonar’s January 2026 State of Code Developer Survey with 1,100+ developers, and GitClear’s ongoing analysis of 211 million+ changed lines. The AI coding landscape is evolving rapidly, so any specific statistic should be treated as a point-in-time benchmark rather than a permanent baseline.
What are the main limitations of studies comparing AI and human code quality?
Several methodological constraints affect interpretation. Most large-scale studies rely on heuristic or watermark-based AI detection, which tops out at roughly 20–25% accuracy and cannot distinguish interaction modes or multi-turn sessions. Controlled experiments often use simplified tasks or short time horizons that do not capture the architectural and maintenance effects that emerge over months.
Selection effects also matter. Teams that adopt AI tools heavily may differ systematically from low-adoption teams in ways that confound quality comparisons. Aggregate metrics such as bug rates and churn rates obscure wide variation by team, tool, project type, and seniority level. The Daniotti et al. Science 2026 study, for example, found that exploration gains from AI accrued exclusively to senior developers, with no statistically significant benefit for early-career developers, which would be invisible in an org-wide average.
What governance questions should engineering leaders be able to answer before scaling AI adoption?
Leaders scaling AI adoption need answers to at least five questions before expanding tool access or increasing token budgets. First, they need to know which lines in production are AI-generated, by which tool, and in which interaction mode, and whether that information is auditable. Second, they must understand the 30-day incident rate for AI-attributed code compared to a matched human-authored baseline.
Third, they should know the total cost per engineer per month including token spend, not just seat licenses, and the resulting ROI at current utilization and rework rates. Fourth, they need to identify which teams or individuals show AI-assisted code with turnover ratios above 1.5x, indicating the code is being rewritten faster than human-authored equivalents. Fifth, they must be able to trace any production incident caused by AI-assisted code back to the exact session, prompt, developer, and tool. Organizations that cannot answer these questions with evidence rather than estimates are governing AI adoption on assumption.
Summary
The 2026 evidence on AI versus human code contributions shows a consistent pattern. AI coding tools deliver measurable productivity gains in throughput, cycle time, and onboarding speed while simultaneously introducing higher rates of logic errors, security vulnerabilities, code churn, and long-term technical debt when teams deploy them without commit-level visibility and governance. Faros AI’s 2026 data shows a 66% increase in epics completed per developer alongside an 861% increase in code churn and the incident-rate spike documented earlier under high AI adoption. Daniotti et al. estimate the productivity gains from AI coding tools translate into $23–38 billion in yearly value at the US economy level, but that value is not uniformly distributed and remains invisible without the measurement infrastructure to attribute it.
The organizations that will capture durable ROI from AI coding tools are those that move beyond adoption metrics to commit-level attribution, longitudinal outcome tracking, and tool-by-tool comparison. The data to make those decisions already exists in every repository. The remaining question is whether leaders have the infrastructure required to read it.