Recommended Free Tools
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
AI can speed up some coding tasks, but that does not guarantee a team will ship software faster or more reliably. The evidence also does not establish that AI-generated code is inherently harder to understand: measured complexity and documentation sometimes improved, while questions about review depth and delivery stability remain.
Why faster coding and faster delivery are different claims
“Building software faster” can refer to finishing a bounded coding task sooner. Delivering software also involves testing, review, integration, release, and keeping changes stable in production. A tool can help with the first activity without improving—and potentially while complicating—the whole delivery process.
That distinction helps explain why studies can point in different directions without directly contradicting one another. They measure different outcomes, use different methods, and involve different developers, codebases, tools, and work conditions. A task-time result is not a measurement of organization-wide throughput; a survey response about productivity is not a randomized estimate of time saved.
What the studies actually found
| Evidence | Setting and method | Finding | What it does—and does not—show |
|---|---|---|---|
| Google Research, 2024 | Randomized trial with 96 full-time Google software engineers working on one complex, enterprise-grade task. | Researchers estimated that AI features shortened task time by about 21%; the estimate had a large confidence interval. | This is evidence about one defined task and study setting, not a general productivity rate. The authors caution that results may not transfer to other tools or settings. |
| METR researchers, July 2025 | Randomized trial involving 16 experienced open-source developers and 246 tasks in mature repositories. Developers averaged five years of familiarity with the projects. Participants primarily used Cursor Pro and Claude 3.5/3.7 Sonnet, tools available in early 2025. | When AI was allowed on randomly assigned tasks, participants took 19% longer on average in this experimental setting. | The result applies to this small group, task set, and tool period. The authors say possible experimental artifacts cannot be entirely ruled out; it is not a finding that AI slows all coding work or current AI systems. |
| DORA, 2024 analysis | Organizational analysis modeling outcomes associated with a 25% increase in AI adoption. DORA reports estimates with 89% uncertainty intervals. | Estimated delivery throughput decreased 1.5% and delivery stability decreased 7.2%. At the same adoption increment, documentation quality increased 7.5%, code quality 3.4%, review speed 3.1%, and approval speed 1.3%; measured code complexity decreased 1.8%. | These are modeled associations, not universal causal effects. Faster review or approval does not necessarily mean review was more thorough. |
| Microsoft Research, August 2025 | Mixed-methods study using survey responses from more than 500 developers, alongside interviews and observational work. | Developers widely perceived productivity and satisfaction benefits, particularly for routine work. Outcomes varied with task complexity, individual usage, and team adoption; evidence of collaboration impact was less pronounced. | This describes developer experience and perception, not a universal causal estimate of time saved. |
The Google trial’s estimated speedup and the METR trial’s slowdown should not be averaged or treated as replications. Their participants, tasks, repositories, tools, and work settings differed. Together, they show why a single headline percentage cannot answer whether AI makes software development faster in general.
#1 Best Overall
What the delivery research says about organizational effects
DORA’s 2024 report presents a mixed picture: AI adoption was associated with higher individual productivity, flow, and job satisfaction, but also with lower software delivery stability and throughput. Its estimates above suggest that documentation, code quality, measured complexity, and review or approval speed can move in a favorable direction even as delivery outcomes move in an unfavorable one. The report explicitly cautions that faster review and approval do not prove that reviews became more thorough.
DORA hypothesizes that increased code generation could lead to larger change batches, which earlier DORA findings associate with slower delivery and instability. That is a proposed explanation, not proof that AI caused larger changes or that this mechanism explains the modeled results. DORA’s practical emphasis remains on delivery fundamentals such as small batch sizes and robust testing, alongside user-centricity, stable priorities, experimentation, and continuous learning. See the 2024 report overview and its analysis document.
The 2025 DORA report broadens the organizational framing. Drawing on more than 100 hours of qualitative data and nearly 5,000 technology-professional survey responses from around the world, it describes AI as an “amplifier” of existing organizational strengths and weaknesses. The public overview does not give complete sampling details or precise fieldwork dates, so those figures alone do not establish that the respondents represent every technology organization. The framing is nevertheless useful: adding an assistant to a healthy delivery system is not the same intervention as adding one to a system with weak testing, oversized changes, or unclear priorities. See DORA’s 2025 overview and Google Research’s report record.
Does AI make code harder to understand?
The available evidence does not establish a general rule that AI-generated code is harder to understand. In DORA’s 2024 analysis, the modeled association for measured code complexity was favorable, and documentation quality also increased. Those measures are relevant, but neither proves that developers can correctly reason about every AI-assisted change or that a project’s architecture remains easy to navigate.
Review speed is also an incomplete proxy for comprehension. A quicker review could reflect a smaller or clearer change, better tooling, or less scrutiny; the metric alone cannot distinguish among those possibilities. DORA specifically warns that faster reviews do not necessarily mean more thorough reviews. Delivery stability matters too: if changes are harder to verify or integrate, that concern may appear in testing or release outcomes rather than in a complexity score.
So the title’s “harder to understand” claim is a question worth testing in a particular codebase, not a settled research conclusion. Teams should distinguish code-level measures from whether maintainers can explain a change, assess its behavior, and safely modify it later.
Rank #4
How teams can tell whether AI is helping their delivery
Do not infer an organization-wide gain from faster completion of a few coding tasks, and do not assume that more generated code equals more delivered value. Establish a baseline and observe both the work people feel and the outcomes the delivery system produces.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- Track task work separately from delivery. Record time or effort for comparable tasks, then examine throughput and stability across releases. Keep the measures distinct rather than collapsing them into one “productivity” score.
- Keep changes small and test them robustly. DORA emphasizes small batch sizes and robust testing. These practices make it easier to isolate regressions and understand what a change does.
- Evaluate review quality, not only review speed. Check whether reviewers can explain the change, identify relevant risks, and verify tests and behavior. A shorter approval time by itself cannot answer those questions.
- Segment results by task and team conditions. Separate routine work from complex changes, and account for tool use, project familiarity, and team adoption. The Microsoft study found that outcomes varied with complexity and usage patterns.
- Reassess as tools change. The randomized studies here tested specific tools available in 2024 and early 2025. They are not a live comparison of later model versions, and their results should not be transferred automatically to a different tool or workflow.
DORA’s 2025 report calls AI an amplifier of an organization’s existing strengths and weaknesses. In practice, that means teams should judge an assistant within the delivery system they actually have—and improve the system as well as the tool when the results are poor.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
