How Much Faster Can AI Make Us in Software Development? A Concise Overview of Current Studies and Surveys
Martin Zoeller
In the context of my workshop, I’m frequently asked by managers and engineering leads: “How much faster do we actually get when we use AI agents in software development?”
A blanket answer is difficult here, since your team’s speed depends on many factors — not least how much the act of writing code is actually a bottleneck in your workflows. That’s why in my workshop, I don’t prescribe a cookie-cutter pattern for how your engineers should use AI to my liking. Instead, I empower you as a team to design your own workflows and make informed decisions about how to use AI agents.
The answer above has a problem: it doesn’t contain numbers, and many of the people I talk to want numbers. So here’s an overview of the latest studies and surveys on AI adoption in software development and its impact.
So: What Do the Numbers Look Like?
In short: They give reason to hope that using AI agents in software development pays off — and massively so, when done right. However, we also have to acknowledge the nearly endless degrees of freedom and unknowns underlying the data in these studies.
Here’s a quick overview of the most methodologically rigorous and up-to-date studies.
1. Faros AI: AI Acceleration Whiplash
Who?
Faros AI is an observability platform for engineering workflows — a platform for measuring and analyzing engineering teams. Their customers are typically large tech companies and engineering leaders who want to make data-driven decisions.
When?
April 2026
Data
Two years of automated telemetry from commits, PRs, reviews, incidents, and deployments. 22,000 developers across 4,000 teams. Important: all teams use Faros AI, meaning there’s an inherent bias toward valuing performance and its measurement.
Numbers
Here’s how AI adoption affects engineers whose performance is measured with Faros AI:
| Metric | Change |
|---|---|
| Completed epics per developer | +66.2% |
| Bugs per developer | +54% |
| Incidents per PR | +242.7% |
| Code churn | +861% |
| PR size | +51% |
| Bugs per PR | +28% |
| Median review time | 5× higher |
| Code with no review at all | +31% |
Takeaway
The numbers sound dramatic: AI adoption increases the volume of code, but also the volume of bugs and incidents. However, there are outliers. High-performing teams combine AI with a rigorous code review culture and use senior engineers as a quality gate. Faros AI observes that senior engineers achieve 2–3× more output through AI compared to juniors (1.2–1.5×) — experience is a decisive factor in how well you can contextualize and review AI-generated code.
Source
faros.ai/blog/ai-acceleration-whiplash-takeaways
2. METR Experiments
Who?
METR (Model Evaluation & Threat Research) is an independent nonprofit research organization specializing in the scientific evaluation of AI capabilities and risks.
When?
July 2025, follow-up February 2026
Data
Controlled experiments in which experienced open-source developers were randomly assigned to control groups with or without AI tools.
Numbers
Early 2025: AI slows things down by +19% (confidence interval +2% to +39%).
Late 2025/Early 2026: AI speeds things up by -18% (confidence interval -38% to +9%).
Takeaway
The result is not statistically significant, but it does suggest improvement compared to the previous year.
Sources
- July 2025: arxiv.org/abs/2507.09089
- February 2026: metr.org/blog/2026-02-24-uplift-update
3. Google DORA Survey
Who?
DevOps Research and Assessment: A long-running research program founded in 2014 and acquired by Google in 2018. DORA developed the five standard metrics for measuring software delivery performance:
- Deployment Frequency
- Lead Time for Changes
- Change Failure Rate
- Mean Time to Recovery
- Reliability
When?
September 2025
Data
Self-assessments from approximately 5,000 respondents worldwide.
Numbers
- 90% of surveyed engineers now use AI.
- More than 4 out of 5 report increased productivity.
- Median: 2 hours of work with AI per day.
- But: meetings, interruptions, and review delays cost more time than AI saves.
Takeaway
The numbers are hard to evaluate since they are based on self-assessments. Likewise, it’s unclear how much time per day is actually spent in code at the median, and what proportion of that time is spent with AI.
Source
dora.dev/research/2025/dora-report
4. GitLab Global DevSecOps Survey 2026
Who?
An annual survey conducted by GitLab.
When?
November 2025
Data
Self-assessments from approximately 3,266 respondents worldwide.
Numbers
- 98% report increased efficiency.
- Teams lose 7 hours per week due to inefficient AI processes.
- 49% of respondents use more than five different AI tools.
What are inefficient AI processes?
Three specific causes:
- Lack of cross-team communication: Generated code doesn’t match the expectations of other teams.
- Insufficient knowledge sharing: Only the engineer who wrote the prompt knows what’s behind it in terms of content and logic. Onboarding and knowledge transfer fall short.
- Tool sprawl: Half of respondents use more than five AI tools, leading to inconsistent output quality, duplicated work, and costs from context switching.
Takeaway
GitLab’s survey confirms the split picture: perceived efficiency gains collide with significant losses from inefficient workflows and poor tool choices.
Source
about.gitlab.com/press/releases/2025-11-10-gitlab-survey-reveals-the-ai-paradox
5. Capgemini Research Institute Survey
Who?
Capgemini SE is a publicly traded French IT consulting and technology company headquartered in Paris. It is the largest IT services provider of European origin, with over 340,000 employees in more than 50 countries.
When?
April 2024
Data
Self-reports from
- 1,098 senior executives (director level and above),
- 1,092 software professionals (developers, testers, architects, PMs),
- All from companies with >$1B in annual revenue,
- Plus 20 qualitative deep-dive interviews with industry leaders.
Only companies with active pilots were included, so the sample is not a representative cross-section of the industry.
Numbers
This study is frequently cited because it is methodologically solid and promises real gains:
| Task | Average Time Savings | Maximum Time Savings |
|---|---|---|
| Coding Assistance (writing/completing code) | 9% | 34% |
| Creating/updating documentation | 10% | 35% |
| Debugging & Testing | 5% | 20% |
| Project management (tasks, tickets) | 1% | 20% |
Takeaway
One notable success story is a 30% increase in test coverage through GitHub Copilot at an Australian company. Low average values suggest uneven and incomplete AI adoption and a lack of expertise.
Source
capgemini.com/…/Final-Web-Version-Report-Gen-AI-in-Software-Engineering.pdf
What Do All These Numbers Tell Us?
You may have noticed that the studies paint a split picture: where hard telemetry data is available (Faros AI, METR), the results are significantly more sobering than where developers self-assess their own productivity (DORA, GitLab, Capgemini).
When I talk to engineers, I hear this over and over: working with AI agents feels extremely fast — but as soon as I ask for measurements, things go quiet. Product teams need to address exactly this gap: “What works for us” has to be the central question.
On top of that, the tool landscape is moving so fast that each of these studies already reflects an older generation of tools. So the numbers are less of a verdict and more of an outlook: they show the enormous opportunities, but also the dangers of teams deploying AI without a methodology.
Overall Conclusion
Those who deployed AI in software development in 2025 were already able to achieve >50% performance gains for senior engineers with the right processes. With stronger models in 2026, we can hope for a significant improvement, and early data points suggest a potential efficiency gain of 1.2× to 3× (Faros AI).
But: AI — and AI agents in particular — are a catalyst, both for the strongest engineers on the team and for the biggest inefficiencies and shortcomings in workflows, communication, and team dynamics. Anyone who throws AI at their problem and hopes for the best should expect serious issues, such as a massive increase in bugs per engineer, more incidents, longer review times, and dangerous blind spots in the product.
Valuable AI workflows emerge in teams that are willing to experiment and learn, and that can maintain an open dialogue about what works and what hurts.
So: Will We Actually Get Faster?
In short: yes, probably.
Will you get significantly faster without setting major fires in the process? Quite possible. With the right methodology, open communication, and a high degree of self-discipline, you dramatically increase your chances.