AI coding productivity surges as acceptance rates mask rising code churn
New analytics show AI coding productivity increases accepted code but raises code churn and costs, prompting firms to rethink how they measure developer output.
Software teams report clear early gains in AI coding productivity, with engineers producing larger volumes of accepted code using tools such as Claude Code, Cursor and Codex. However, new analytics from engineering-intelligence vendors indicate that much of that output requires revision in the weeks that follow, reducing the net value of generated code. As companies race to adopt coding agents, managers are confronting a mismatch between input metrics like token budgets and the lasting quality of software delivered.
High acceptance rates and the revision problem
Engineering analytics firms are finding that initial acceptance metrics overstate lasting gains in output. Waydev, which monitors thousands of engineers across multiple customers, says AI-generated code is being approved at rates between roughly 80% and 90% on initial review.
Those headline acceptance numbers obscure a second-stage effect: a substantial share of accepted AI output is later revised, reviewed again or deleted. Waydev estimates that follow-on churn reduces the effective acceptance rate by about 10% to 30%, meaning a meaningful slice of what passes code review does not persist in production-quality repositories.
Industry reports quantify the churn
Independent analyses reinforce the pattern of rising churn alongside increased output. A January study from GitClear reported that regular AI users produced higher raw productivity but also averaged roughly 9.4 times more code churn than their non-AI peers. That level of churn, the report argued, more than offset many of the efficiency gains the tools initially delivered.
Faros AI examined two years of customer telemetry and found even larger relative changes under intense AI adoption: code churn — measured as lines deleted versus lines added — rose dramatically, signaling that teams were expending effort to rewrite or remove substantial portions of generated code. These findings point to a consistent industry signal: volume is up, but permanence and quality are not keeping pace.
Token budgets drive volume but not proportional value
Companies and developers have taken to tracking token budgets — the allocation of AI processing capacity — as a measure of adoption and usage. But token consumption is an input metric: high budgets often correlate with more generated code rather than better outcomes. Data collected by Jellyfish on thousands of engineers in the first quarter of 2026 shows that engineers with the largest token allocations produced the most pull requests, yet the productivity gain did not scale linearly with cost.
Jellyfish found roughly twice the throughput at roughly ten times the token expense for the highest-consuming engineers, a disparity that highlights diminishing returns. For teams seeking efficiency, these results suggest token-maximizing strategies risk creating cost inflation without equivalent improvements in long-term code stability or reduced maintenance burden.
Seniority gap and accumulating technical debt
Developer experiences with coding agents differ markedly by experience level, and that variation is shaping team outcomes. Senior engineers tend to scrutinize and adapt AI suggestions more aggressively, while junior engineers accept a higher share of generated code and subsequently spend more time rewriting or fixing it. That dynamic concentrates revision work and technical debt in particular parts of organizations.
Managers report that code review backlogs and follow-on maintenance are increasing even as engineers enjoy the immediate productivity lift from AI assistance. The result is a layering of short-term throughput that can mask growing architectural inconsistencies and future remediation costs if not proactively managed.
Toolmakers and platforms respond to new needs
The analytics vendors themselves are adapting their product road maps to capture agent-driven workflows and metadata. Waydev recently refactored its platform to ingest AI-agent metadata and surface measures that link token use, code quality and downstream revisions. Other firms, including GitClear, Faros AI and Jellyfish, are publishing reports and building dashboards designed to quantify the trade-offs between speed and rework.
Market consolidation is also underway: incumbent software companies are investing to offer customers clearer return-on-investment signals for coding agents. Atlassian’s acquisition of an engineering intelligence startup last year underscored buyer demand for tools that make AI-driven development economically visible and operationally tractable.
What managers should measure next
The emerging consensus among analysts and practitioners is that teams should shift focus from inputs like token consumption to output durability and maintenance cost. Practical measures include long-window acceptance rates that track whether AI-generated code remains untouched after several weeks, churn-adjusted throughput, and metrics tying generated code to bug rates or post-release fixes.
Organizations with disciplined instrumentation and feedback loops can still capture genuine productivity benefits from AI coding tools while preventing unchecked technical debt. Establishing guardrails for junior engineers, investing in pair-review practices and tying token budgets to measurable quality outcomes can help teams convert volume into lasting value.
Adoption of coding agents is accelerating, but the gains are nuanced: more code does not automatically mean more efficient software delivery. As firms deploy analytics and adjust managerial incentives, the debate is shifting from how much AI we use to how well AI improves durable outcomes.
