The available quantitative evidence shows a clear expansion of AI across software-engineering tasks, but it does not yet provide a reliable breakdown of how developers divide their working hours.
In Jellyfish’s March 2026 survey, 53.1% of respondents reported using AI for code writing, 49.2% for code review, 42.8% for code explanation, 37.9% for refactoring, 35.8% for requirements analysis, 33.5% for documentation, 30.8% for test-case creation, 27.4% for debugging, and 24.2% for specification creation.
These figures measure whether developers use AI for particular activities. They do not show how much time developers spend on those activities during a typical workday.
That distinction is important. A robust 2026 time-and-motion dataset dividing an engineer’s week among manual coding, prompting, code reading, diff review, testing, architecture, debugging, multi-agent coordination, documentation, and final approval does not yet appear to exist. As a result, claims such as “developers now spend 70% of their day reviewing AI-generated code” go beyond what the current evidence can support.
There are, however, strong indications that direct implementation effort is declining among the most intensive AI adopters.
Jellyfish’s analysis of Claude Code sessions found that, among the longest 1% of sessions by duration, low-intervention sessions produced roughly 11 times more net code than heavily supervised sessions of similar length. Company-level telemetry also showed a sharp difference between leading adopters and the median organization: by April 2026, companies at the 90th percentile had approximately 35% of pull requests opened by agents, while the median remained close to 2%.
This suggests a highly uneven transition. For organizations at the frontier of AI adoption, supervising agent-generated implementation is already becoming an important part of engineering work. Across the broader industry, however, agent-authored pull requests are not yet the dominant workflow.
Review behavior shows a similar pattern.
In Jellyfish’s August telemetry, an “AI-only review” was defined as a merged pull request for which all recorded review activity came from AI, with no human review. At the median company, only 0.7% of merged pull requests met that definition. Among leading adopters, the share exceeded 3.5%.
Despite rapid improvements in code-generation capability, human review therefore remains the normal path to merging code in most observed organizations.
Sonar’s January 2026 publication points in the same direction, although its survey was conducted in October 2025. Among developers who used AI, 95% said they spent at least some effort reviewing, testing, or correcting AI-generated output, while 59% described that effort as moderate or substantial.
Reviewing AI-generated code was not necessarily easier than reviewing human-written code. Thirty-eight percent of respondents said it required more effort, while 27% said it required less. In addition, 96% did not fully trust AI-generated code to be functionally correct.
The same survey also highlights a gap between responsibility in principle and verification in practice. Forty-eight percent of respondents completely agreed that they always check AI-assisted code before committing it, while another 27% somewhat agreed.
These figures should not be interpreted as meaning that only 48% of developers review AI-generated code. They do, however, show that consistent and rigorous human verification cannot simply be assumed.
So, has software engineering moved from “hands on keyboard” to “eyes on code”?
Among highly AI-enabled teams, the shift is already substantial. Across the profession as a whole, the evidence supports a more cautious conclusion: manual implementation is becoming less dominant, while directing, reading, validating, testing, and taking responsibility for machine-generated implementation are becoming increasingly important.
What the evidence does not yet establish is that the average software engineer now spends more than half of the working day reviewing code rather than writing it.