A year ago, developers told Sonar that AI generated or heavily assisted 42% of the code they committed. They expected that share to reach 55% within a year and 65% within two. This year’s survey of 1,026 developers and engineering managers puts it at 65% already (Sonar, Nov 2026).
Review did not speed up to match. In the same survey, 76% agree with the statement “we generate AI code faster than we can verify it.” The share who completely agree that they always check AI code before committing fell from 48% to 39% (Sonar).
This piece reads five primary sources published between June and November 2026. Each one uses its own sample and its own definitions, so every number stays with its survey. The last section is a checklist for this week and this month.
What 65% of Committed Code Means
Sonar’s 65% is a self-reported share of committed code that AI generates or heavily assists. It is not the share written with no human involved. The same survey says 76% of respondents’ teams or companies now use AI coding tools for business-critical or mission-critical services, up from 58% (Sonar).
JetBrains asked a narrower question. Between May and July 2026, more than 15,000 professional developers split last month’s work code into three buckets. On average, about 47% was fully generated by agents, about 38% was written with some AI assistance, and about 27% was written by hand (JetBrains, Aug 2026). JetBrains averaged the midpoints of answer ranges, so the three can add up to more than 100%.
Do not place Sonar’s 65% next to JetBrains’ 47% as if one corrected the other. One counts generated plus heavily assisted code. The other counts only code an agent wrote end to end.
The averages also hide a wide spread. More than half of developers in the JetBrains data write less than 20% of their code by hand, and one in five writes none without AI. Yet the group that lets agents produce over 80% of its code is only around 22%. JetBrains sorted respondents into three profiles (JetBrains):
- Agentic coders, about 31% of developers, report 84% agent-generated code on average.
- AI-assisted coders, about 47%, report 40% agent-generated code and 60% AI-assisted.
- Manual coders, about 23%, still write 75% by hand.
Region and tool move the figure too. In China, Japan and South Korea, 32%–35% of developers generate over 80% of their code with agents, against around 16% in Europe and the UK. By most-used tool, that share is 42% for Codex, 32% for Claude Code and 28% for Cursor (JetBrains).
Review Time Moved, Confidence Did Not
The work shifted from writing to reading. According to Sonar, 53% of respondents now spend more time reviewing code than creating it. Only 18% say that was true before they adopted AI tools (Sonar). The 76% who say generation outruns verification answered after that shift, not before it.
Trust rose a little and stayed shallow. Agreement that AI-generated code is functionally correct went from 29% to 45%. Complete agreement went from 4% to 6%, which Sonar reads as 94% who still do not fully trust it (Sonar). Put that beside the drop from 48% to 39% in always checking before a commit. Code that few people fully trust is being checked less consistently.
Experience changes the bar. In JetBrains’ State of Developer Ecosystem 2026, based on 28,945 developers in 198 countries and regions, 60% of developers with one to two years of experience say AI-generated solutions are better than their own. Among those with 16 or more years, 47% say so (JetBrains, Nov 2026). JetBrains adds that senior developers go beyond prompts and give agents guidelines, specs and ways to verify their work.
A third source asks about trust directly. A DORA essay from June cites DORA’s State of AI-assisted Software Development report: about 30% of developers trust AI outputs “a little” or “not at all” (DORA, June 2026). That is a different question from Sonar’s, asked of a different sample. Read it as a second angle, not a matching figure.
Security Worries Outgrew the Controls
The top concern in Sonar’s survey is no longer the code AI writes. It is what attackers do with AI. Sonar found 61% are very or extremely concerned about attacks that use AI to find and exploit vulnerabilities faster and at greater scale in all code, existing code included (Sonar).
Four of the top seven concerns about AI-generated code are security concerns. At the same two concern levels, 44% worry about new or subtle vulnerabilities and 42% about severe ones (Sonar). Teams are raising autonomy anyway. Sonar reports that 5% are building or have built software factories where agents work on tasks autonomously, and 34% expect to be creating or using one within a year.
Most checks still happen late. Among respondents whose teams review any AI code, 81% say it is verified at code review or pull request time, 51% in the continuous integration and delivery (CI/CD) pipeline, and 43% in real time as the agent writes (Sonar).
Timing lines up with results. Teams that run automated review inline are more likely than teams without it to report a positive AI impact:
| Positive impact reported on (Sonar) | Inline review | No inline review |
|---|---|---|
| Code security | 61% | 46% |
| Vulnerability rates | 49% | 38% |
| Code quality | 75% | 63% |
| Time-to-market | 89% | 81% |
Sonar calls these “associations rather than proof of cause.” A team disciplined enough to wire checks into the editor may be better at other things too.
The 64-Point Gap Between Feeling Faster and Proving It
GitKraken surveyed 554 developers and engineering leaders from its own customer and user community. In that group, 96.4% of teams have adopted AI coding tools and 84% of developers feel more productive, 43% of them “much more.” Fewer than one in 20 feel slower (GitKraken, July 2026).
Then comes the number the report is built around. While 84% feel more productive, only 20% can measure it, a 64-point gap between belief and proof (GitKraken).
The other side of that gap has two parts. According to GitKraken, 39% of organizations have no way to measure AI’s impact at all, and another 33% lean on developers’ self-reports (GitKraken). The report adds those together when it says 72% of organizations cannot put a number on their AI return.
Company size changes the picture. The share with no way to measure falls from 47% at small organizations to 19% at enterprises. Use of DORA or engineering metrics rises from 6% to 21% across the same range (GitKraken).
Meanwhile the work is getting more autonomous. For 28% of developers, the primary way of working with AI is now autonomous, up from 7.6% in September 2025. A third (34%) keep agents running the entire workday, 47% in enterprises against 32% in small shops, and 2% run them around the clock (GitKraken).
The teams furthest along measure more. GitKraken says agent-native teams, those running parallel agents regularly, are more than three times as likely to track specific productivity metrics (34% vs. 10%). They are also less likely to work somewhere that measures nothing (28% vs. 54%) (GitKraken).
Why Token Counts Fail as an ROI Metric
The most common metric is not the one managers trust. In JetBrains’ survey, 51% of tech managers say their organizations track token usage, but only 30% count it among the most insightful metrics. Net time gain runs the other way: 39% find it insightful and only 28% track it (JetBrains).
Satisfaction with return on investment (ROI) does not depend on measuring it. More than half of tech managers at small organizations are satisfied with the return on AI spending, against 34% at large ones. And 40% of managers at organizations that do not measure ROI say they are satisfied anyway (JetBrains).
DORA’s Evan Conaway explains why token counts mislead. Some companies now run internal leaderboards that rank people by tokens consumed, a practice known as “tokenmaxxing.” A raw token count measures input only, so it is simple to game. DORA cites reports of developers running agents on nonsense projects to stay above the company average, and calls the result “productivity theater” (DORA).
The cost side is no better. Citing analyst Adnan Masood, DORA writes that teams with the highest budgets often get only a marginal increase in delivery throughput at a 10x increase in cost (DORA).
Sonar’s respondents describe the same shortfall from the inside. According to Sonar, 43% say AI coding has not yet delivered the return they expect. While 47% expect an ROI of 50% or more, 27% report one that high, and 22% are unsure what return they have seen (Sonar).
Two things absorb the gains. Respondents estimate that about 27% of their work is toil, up from 24%, and 54% say toil takes more than a quarter of their week, up from 38%. Concern about the cost of advanced AI coding tools and agents rose from 30% to 45% (Sonar).
Where Governance Changes the Numbers
More teams have rules for AI code than a year ago. In Sonar’s survey, 55% say their organization has guidelines, policies or quality gates for AI-generated code that differ from those for human-written code, up from 41%. Only 23% describe well-defined, distinct guidelines or automated checks, up from 14%. The most common approach is still informal rules enforced by hand, at 32% (Sonar).
The well-defined group reports better outcomes. Those teams are 31% more likely to say AI improved defect rates (66% vs. 51%) and 32% more likely to say it improved rework or patch costs (70% vs. 53%). They are also the only group where AI reduced compliance risk more often than it raised it, 23% against 16% (Sonar).
Return follows the same line. Sonar found that 32% of respondents with governed agents report an observed ROI of 50% or more. Among those with approved agents but informal, manual governance, 19% do (Sonar).
DORA’s essay names three companies that changed what they count. Salesforce introduced Agentic Work Units (AWUs), which measure completed work and ignore the compute used to get there. Shopify dropped its AI leaderboard for a usage dashboard and added “circuit breakers” that catch runaway agents and flag abnormal daily spend spikes (DORA).
Starburst, a data platform, is the third. It ignores token maximization and measures AI against standard DORA metrics, developer velocity and incident resolution times. DORA suggests pairing those core metrics with cost per accepted change and code rework rates. Its own position: “company-wide aggregates are meaningless, and individual scores are harmful,” so look at token usage by team, application or use case (DORA).
What These Surveys Do Not Cover
Every figure above is self-reported. None of the five sources counted AI-written lines in a repository or timed a review.
The publishers sell tools for the problems they describe. Sonar sells code verification, and its post ends with SonarQube-user comparisons that are relative likelihoods, such as 81% more likely to report the top level of positive impact on code security. Sonar notes those do not adjust for company size, region or role (Sonar). Do not mix them with the absolute shares in the inline review table.
GitKraken sells measurement software and says its sample “skews toward developers already engaged with modern Git and engineering tooling” (GitKraken). The DORA piece is commentary, not a survey. Its Salesforce, Shopify and Starburst cases come from press reports that DORA cites.
Primary Sources Behind This Analysis
Each source answers a different question with a different sample:
- Sonar, AI hit 65% of committed code a year early (November 12, 2026): code share, review time, trust, security concerns, governance and ROI. 1,026 developers and engineering managers.
- JetBrains, The State of Developer Ecosystem 2026 (November 10, 2026): experience gap and ROI metrics. 28,945 developers, surveyed May to July 2026.
- JetBrains, How Much Code Do Developers Really Let Agents Write? (August 26, 2026): agent-generated, AI-assisted and manual code shares. More than 15,000 professional developers from the same survey.
- GitKraken, The State of AI In Engineering (July 2026): felt productivity against measurement. 554 developers and engineering leaders.
- DORA, Finding balance in the era of tokenmaxxing (June 2, 2026): token leaderboards and outcome metrics. Commentary with cited cases.
Action Steps for Engineering Teams
This week
- Turn on inline static and security checks for AI-generated code in the editor or agent loop, and run the same checks on every pull request and in CI.
- Record a baseline for three numbers before changing anything else: review time per pull request, code rework rate and cost per accepted change. If you run a token leaderboard, replace it with a per-team usage view.
This month
- Write explicit AI guardrails for each codebase (a context document, permission boundaries and a rule that no model reviews its own work), then connect them to a CI gate.
- Track quarterly ROI with DORA throughput and stability metrics plus a completed-work unit in the style of Salesforce’s AWUs, reported by team and application instead of by person.