Written by: Mark Hull, Co-Founder and CEO, Exceeds AI | Last updated: July 15, 2026
What Engineering Leaders Can Expect From Claude Code
- Self-reported productivity gains from Claude Code usually fall between 30–60%, with some boilerplate tasks reaching up to 10x speedups, although these numbers come from surveys and metadata, not objective commit-level data.
- Controlled studies reveal a productivity paradox: developers often feel faster with AI tools, yet objective measurements show minimal or even negative changes in task completion times.
- Organizational delivery metrics improve far less than individual coding speed, typically only 5–15%, because review overhead, code churn, and downstream processes absorb most of the gains.
- Accurate measurement of AI impact requires commit-level attribution that tracks which lines were AI-generated and connects them to real outcomes over time, which most current tools cannot do.
- Exceeds AI provides a portable, auditable solution for commit-level AI attribution, so engineering leaders can prove real ROI — book a demo today.
How This Report Was Built and What Data It Uses
This report synthesizes data from Anthropic internal surveys, METR randomized controlled trials, DX longitudinal analysis, Faros AI telemetry, LinearB benchmark reports, Swarmia engineering organization analysis, BNY Mellon mixed-methods research, and GitClear longitudinal code churn studies. Each source uses a different measurement approach, and those differences strongly shape how you should interpret the results.
Surveys and metadata tools share a structural limitation: they cannot isolate AI’s contribution to a specific outcome. BNY Mellon researchers studying 2,989 developers found only a weak correlation (r=0.34) between satisfaction and perceived time savings, and they could not obtain reliable objective metrics across all engineers, so they had to rely on self-reported data. Knowing that a developer accepted 200 Copilot suggestions does not reveal whether that code was committed, merged, or improved delivery outcomes. Metadata tools that track PR cycle times, commit volumes, and review latency remain blind to which lines are AI-generated versus human-authored.
Commit and PR-level attribution closes this gap by tying every AI-touched line back to the specific tool, model, session, and interaction mode that produced it. Exceeds Ink is the only portable, auditable implementation of this approach on the market. It writes a structured Git Note at refs/notes/exceeds-ink alongside every commit, which is line-level, machine-readable, and stored in your own repository rather than a proprietary cloud. Heuristic and watermark-based detection typically reaches only 20–25% accuracy and cannot distinguish how an engineer worked with the AI tool.
Key Findings Summary
The productivity picture shifts depending on who measures it and how. Anthropic’s internal and user surveys show strong self-reported gains, while independent researchers and industry telemetry reveal a more mixed reality. With those differences in mind, here is what the major sources report:
- Anthropic internal survey (132 engineers, August 2025): self-reported productivity gains of 50%
- Anthropic survey of 81,000 Claude users (April 2026): mean self-reported productivity rating of 5.1 on a 1–7 scale, with 48% citing scope expansion as the primary benefit
- METR randomized controlled trial (experienced open-source developers): developers took 19% longer to complete tasks while believing they were 20% faster, a 39-percentage-point perception gap
- DX longitudinal analysis (400+ companies, Q1 2026): median PR throughput increased only 8% despite a 65% rise in AI usage
- Faros AI telemetry (10,000+ developers, 1,255 teams): 21% more tasks completed, 98% more PRs merged, and 91% longer PR review times
- LinearB 2026 Benchmarks (8.1 million PRs, 4,800 teams): AI-generated code waits 4.6x longer for review, with an acceptance rate of 32.7% versus 84.4% for human-written code
- GitClear longitudinal analysis: code churn rose from a 3.3% pre-AI baseline in 2021 to 7.1% in 2025
Compare these industry ranges to your own repos with commit-level data — book a demo.
Where Claude Code Helps Most and Least
Task-Type Ranges: Greenfield vs. Maintenance
Productivity gains from Claude Code vary sharply by task type. A solo developer case study reported roughly 10x speedup on boilerplate and scaffolding, with smaller gains on other tasks. Gains tend to be higher for boilerplate and CRUD work, unit tests, and documentation than for complex algorithms. Greenfield work, where context is thin and scaffolding dominates, consistently shows the largest gains. Maintenance work on mature, complex codebases shows the smallest returns and sometimes even negative returns.
The Productivity Paradox in Practice
ByteDance’s TRAE team generates more than 90% of its code with AI while achieving only a 60% improvement in per-capita throughput, so code volume increases tenfold while delivered output barely moves. Swarmia analysis found median batch size roughly doubled between Q1 2025 and Q1 2026. Individual developers code faster, yet organizational output grows only modestly when processes for scoping, delegation, review, and maintenance stay the same.
Sourced Comparison of Anthropic and Third-Party Productivity Figures
Anthropic’s internal numbers, independent trials, and industry telemetry often point in different directions. Reviewing them side by side helps leaders separate perception from measurable delivery outcomes and decide which figures should guide investment decisions.
See how your team compares to these benchmarks with commit-level analytics.
Does Claude Increase Productivity?
Claude’s impact depends on what you measure, who uses it, and which tasks they perform. Anthropic’s own data shows strong self-reported gains and significant PR volume increases. Third-party controlled studies present a more complicated picture. Industry benchmarks from 2026 show productivity multipliers for developers using AI coding tools, while organizational productivity gains remain more modest.
Anthropic’s user survey shows the 50% self-reported gains mentioned earlier, but that figure carries an important caveat: respondents were active Claude users willing to take a survey, which introduces selection bias. The METR controlled trial, by contrast, used experienced developers on their own codebases, conditions that favor AI tools, and still found a 19% slowdown. The 2025 Stack Overflow Developer Survey reported mixed productivity outcomes from AI. Claude Code increases productivity for specific task types and experience profiles. It does not reliably increase organizational delivery velocity without deliberate process redesign.
Turn these mixed signals into hard numbers for your org with Exceeds AI.
How to Measure AI Productivity Gains from Claude
Current measurement approaches fall into three categories, and each has documented limitations.
Self-reported surveys capture sentiment but not outcomes. Developers are poor at estimating counterfactuals, such as how long tasks would have taken without AI assistance, which makes self-reported time savings structurally unreliable.
Metadata tools that track PR cycle time, commit volume, and review latency cannot distinguish AI from human contributions. Most homegrown tools never attempt commit-level fusion of AI usage data because the data model is too complex to build ad hoc. Platforms like LinearB and Swarmia were built for the pre-AI era and remain blind to which lines are AI-generated.
Vendor telemetry such as GitHub Copilot Analytics and Cursor’s AI Code Tracking provides tool-specific usage stats but does not connect AI usage data to downstream outcomes such as cycle time, PR merge rates, or defect rates. Cursor’s tracking also requires developers to commit from the same machine where AI code was written, and automated code formatting can invalidate diff signatures.
Commit and PR-level attribution is the only method that closes these gaps. It requires capturing what actually happened on the engineer’s machine, including which tool, which model, which interaction mode, and how many tokens, then attesting that information alongside the commit in a portable, auditable format. The Git Notes attestation described earlier is the only implementation that does this without a long-lived daemon, without modifying commit messages, and without locking provenance data inside a proprietary cloud.
Connect your repo and see commit-level attribution running on your own codebase.
Claude Code Productivity Paradox
The productivity paradox describes a consistent pattern: individual-level speed gains from Claude Code do not translate into organizational delivery velocity. Three mechanisms explain this gap.
Review overhead. Teams with high AI adoption merged 98% more PRs but experienced a 91% increase in review time, which produced organizational productivity improvements of roughly 10%. AI-generated PRs are larger, less familiar to reviewers, and harder to spot-check for subtle errors. SonarSource’s 2026 survey of more than 1,100 developers found that 38% report reviewing AI-generated code takes more effort than reviewing a coworker’s code.
Code churn. Faros AI’s telemetry from 22,000 developers across 4,000 teams shows AI-heavy organizations experience 10x code churn compared to baseline. GitClear’s longitudinal analysis of 211 million lines of code found AI-assisted development caused 4x code duplication and refactoring to collapse from 25% to under 10% of changes.
Organizational velocity gaps. CircleCI’s 2026 report shows a widening gap in throughput under AI-driven code growth, but does not prove that only 1 in 20 teams convert coding speed gains into delivery speed gains. Coding accounts for only 14–35% of a developer’s day, so AI accelerates that slice without affecting planning, design, QA, or operations. DX identifies five structural reasons AI productivity gains fall short of vendor claims: coding represents only about 14% of a developer’s day, AI creates new review and integration bottlenecks, social friction between pro- and anti-AI engineers, skill gaps, and lack of institutional context in tools.
Exceeds AI shows whether your AI rollout is compounding these costs or actually improving delivery.
What These Findings Mean for Engineering Leaders
Board-ready AI ROI reporting requires three capabilities that self-reported surveys and metadata tools cannot provide: attribution that identifies which lines are AI-generated, causation that shows whether AI caused the outcome, and longitudinal tracking that reveals whether AI-touched code holds up 30, 60, and 90 days later.
Many engineering leaders say their current metrics do not fully capture AI impact. The gap comes from data infrastructure, not analytical skill. Without commit-level provenance, leaders end up reporting sentiment instead of evidence.
Exceeds AI connects AI adoption directly to productivity and quality outcomes at the commit and PR level, across every AI tool your team uses. The platform’s AI vs. Non-AI Outcome Analytics tracks immediate outcomes such as cycle time and review iterations, along with longitudinal outcomes such as incident rates 30 days later, follow-on edits, and test coverage. This approach is the only reliable way to detect AI technical debt before it surfaces in production.
Get board-ready AI ROI evidence from your own repos — book a demo.
Variations and Exceptions Across Teams
Productivity outcomes from Claude Code vary significantly by experience level, task type, and organizational maturity.
Experience level. Productivity gains shift with experience. Senior developers with 10+ years of experience report 81% productivity gains in some research, while junior developers show no measurable output improvement in the same study, and these findings conflict across datasets, which reflects genuine variation in how experience interacts with AI tool design. The METR study found the largest slowdowns among developers most familiar with the codebase.
Task type. Greenfield, boilerplate-heavy, and test-generation tasks consistently show the largest gains. Complex maintenance on mature codebases shows the smallest gains. In boilerplate-heavy codebases, teams using AI code generation tools achieved a 2.1x throughput increase, while teams adopting AI tools superficially without workflow adaptation saw almost no measurable productivity benefit.
Organizational maturity. A synthesis of 67 sources identifies three variables that largely determine whether AI coding tools help or hurt: task abstraction level, codebase maturity, and developer experience level. Productivity gains from AI coding tools follow a non-linear adoption curve, with 10–15% improvement in month 1, a dip in months 2–3 due to review overhead, and compound benefits appearing in months 4–6 once prompting and review practices adapt.
Run a free pilot to see where your team sits on this curve.
Practical Takeaways for Measuring Claude Code ROI
Engineering leaders who need defensible AI ROI data can follow a clear measurement sequence.
- Establish baselines before expanding adoption. Cycle time, deployment frequency, change failure rate, and MTTR must exist as baselines before introducing AI coding tools; otherwise improvements cannot be attributed to AI.
- Move beyond acceptance rates. Once you have baselines, stop treating suggestion acceptance as your primary metric. GitHub Copilot’s acceptance rate is a vanity metric, because a developer accepting 90% of suggestions may be building the wrong thing faster. Track commit-level attribution, not suggestion counts, and compare those baselines for AI and non-AI work.
- Track review overhead explicitly. AI Review Tax, which covers whether AI-assisted PRs take longer to review, require more review rounds, or generate more comments, reveals whether AI shifts work to reviewers and reduces net productivity gains.
- Monitor longitudinal quality. Heavy AI adoption correlates with knowledge erosion signals such as increased refactor commits after AI-assisted features, higher revert rates, and more fix commits within 48 hours of AI-heavy PRs. Tracking these signals shows whether AI-generated code ages well.
- Require portable, auditable provenance. The Git Notes attestation described earlier is readable by any Git client, portable across forks and mirrors, and auditable by anyone with repo access, which makes it the only provenance approach that survives outside a vendor’s platform.
Get a measurement plan and live dashboards tailored to your stack — book a demo.
Frequently Asked Questions
What is a realistic Claude Code productivity improvement percentage for my engineering team?
The realistic range is 20–50% on self-reported measures for typical engineering tasks, with task-type variation that matters more than the headline number. Boilerplate, scaffolding, and test generation consistently show the largest gains, sometimes approaching 2x or higher. Complex maintenance on mature codebases shows the smallest gains, and controlled studies have found that experienced developers on familiar repositories can be slower with AI tools. The 30–60% figures commonly cited in vendor materials reflect self-reported data from active users, which carries selection bias. Organizational delivery metrics such as deployment frequency, change failure rate, and main branch throughput show much smaller improvements, typically in the 5–15% range, because coding is only a fraction of a developer’s day and downstream processes such as review, QA, and operations remain unchanged. The most defensible answer for your specific team requires commit-level attribution data, not survey averages.
Does Claude Code improve code quality, or does it introduce more bugs?
Evidence on quality is mixed and depends heavily on review discipline. Anthropic’s internal data shows Claude-written code reached rough parity with human-written code by May 2026 after being rated “somewhat worse” in late 2025. Third-party analysis tells a more cautious story: a study of 302,532 AI-authored commits found more than 15% of commits from every studied AI assistant introduced at least one issue, with 22.7% of AI-introduced issues still surviving at repository head. Faros AI telemetry across 1,255 teams found a 9% increase in bugs per developer alongside a 154% increase in average PR size. Teams that adapted code review practices specifically for AI-generated code, including complexity limits, test coverage requirements, and AI-specific linting, recovered most of the review overhead and saw stable or improved defect rates. Teams that did not adapt saw quality metrics deteriorate. Claude Code’s quality impact therefore depends on the review and governance infrastructure around it, not only on the tool itself.
How do I prove Claude Code ROI to my board without relying on self-reported data?
Board-ready AI ROI proof requires three data types that self-reported surveys cannot provide: attribution that identifies which specific commits and lines are AI-generated, outcome correlation that shows whether AI-touched code performed differently on cycle time, defect rates, and incident rates, and longitudinal tracking that reveals whether AI-generated code holds up 30, 60, and 90 days after merge. The measurement infrastructure to produce this data requires commit-level provenance that captures which AI tool wrote which lines, in which interaction mode, and at what token cost, then correlates that provenance with downstream outcomes in your actual codebase. Metadata tools such as LinearB, Swarmia, and Jellyfish cannot do this because they do not analyze code diffs. Vendor telemetry such as GitHub Copilot Analytics and Cursor dashboards cannot do this because it does not connect usage to outcomes. Exceeds AI’s platform, powered by Exceeds Ink’s line-level Git Notes attestation, is built specifically to produce this evidence, connecting AI adoption to productivity and quality outcomes at the commit and PR level, across every AI tool your team uses, in a format that survives outside any vendor’s platform.
Get the AI ROI evidence your board is asking for — book a demo.
Conclusion: Turning Claude Code Data Into Actionable ROI
The Claude Code productivity improvement percentage spans a wide, source-dependent range, around 50% on self-reported measures from Anthropic’s August 2025 survey and negative in METR’s controlled trial of experienced developers on mature codebases. Third-party telemetry consistently shows individual-level gains that do not fully translate into organizational delivery velocity, because review overhead, code churn, and downstream process bottlenecks absorb most of the speed increase.
Every figure in this report carries a measurement limitation. Self-reported data reflects perception, not outcomes. Metadata tools reflect volume, not attribution. Vendor telemetry reflects usage, not causation. Commit and PR-level attribution is the only method that produces defensible, board-ready AI ROI proof, because it ties every AI-touched line back to the tool, model, session, and interaction mode that produced it, then tracks what happens to that code over time.
Exceeds AI delivers that proof. Exceeds Ink writes a portable, auditable, line-level attestation alongside every commit, stored in your own repository, readable by any Git client, and ready for longitudinal outcome tracking that no metadata tool can match. Setup takes hours, first insights arrive within minutes, and board-ready ROI reports are available within weeks.
See your real Claude Code impact with commit-level analytics — book a demo.