Skip to content

Instantly share code, notes, and snippets.

@simonw

simonw/report.md Secret

Last active September 29, 2024 18:28
Show Gist options
  • Save simonw/862733da36e33b475717321a0aaecbc0 to your computer and use it in GitHub Desktop.
Save simonw/862733da36e33b475717321a0aaecbc0 to your computer and use it in GitHub Desktop.

Can Generative AI Improve Developer Productivity?

HERE'S WHAT THE REAL-LIFE DATA SUGGESTS

More engineering teams are experimenting with generative AI tools like Copilot to ostensibly improve productivity and developer experience. But what does the quantitative data say?

Using actual engineering data across a sample of nearly 800 developers in Uplevel's customer population, Uplevel Data Labs analyzed the difference in how teams with and without Copilot access performed according to objective metrics like cycle time, PR throughput, bug rate, and extended working hours ("Always On time").

The expectation is that Copilot helps developers write code faster and smarter, which should lead to lower cycle time, more PRs, and fewer bugs without increasing the risk of burnout. Here's what we found:

  • Copilot access provided no significant change in efficiency metrics.
  • Developers with Copilot access saw a significantly higher bug rate while their issue throughput remained consistent.
  • Copilot access was not effective in mitigating the risk of burnout.

Figure 1: Copilot's Impact on Efficiency Metrics

  • When comparing PR cycle time, throughput, and complexity along with PRs with tests, Copilot neither helped nor hurt developers in the sample, and also did not increase coding speed.
  • While some of these metrics were statistically significant, the actual change was inconsequential to engineering outcomes (e.g. cycle time decreased by 17 minutes)

Figure 2: Copilot's Impact on Bug Rate

  • Developers with Copilot access experienced a +41% increase in bug rate
  • This suggests that Copilot access may impact code quality. (The fact that PR throughput was unchanged further supports this possibility)

Figure 3: Copilot's Impact on Burnout Risk

  • Uplevel's "Sustained Always On" metric (extended working time outside of standard hours and a leading indicator of burnout), decreased for both groups.
  • But it decreased by 17% for those with Copilot access and by almost 28% for those without.

What Does This Data Mean?

Access to generative AI tools like Copilot has raised a number of important questions: Will AI help developers ship faster? Can it help them write better code and avoid burnout?

Not yet, for this population. But innovation moves fast, and GitHub reports that Copilot does improve developer satisfaction. Engineering leaders may benefit from adopting a conservative Copilot adoption strategy to prepare for further advancements in the tool:

  • Set specific goals. What specifically are the outcomes that you are wanting to achieve by including Copilot in your team's workflow?
  • Offer training to your teams. Onboarding can be a good way to lay out where Copilot should and shouldn't be used and what safeguards are in place as an organization.
  • Continue to experiment with generative AI. Seek out specific use cases in which Copilot can be helpful and the prompts that yield the best results. Share these findings across your organization so that success can be replicated.
  • Monitor the engineering effectiveness metrics that Copilot might impact. Start A/B testing on your own to gain objective, quantitative insight into whether AI is actually improving developer productivity and/or helping you reach your operational goals.

About the Study

  • Metrics:

    • Metrics were evaluated prior to implementation of Copilot from January 9 through April 9, 2023 versus after implementation from January 8 through April 7, 2024. This time period was selected to remove the effects of seasonality.
    • The results are based on t-tests for numerical metrics and z-tests for proportions to understand any impacts to each metric. Analysis is based on whether individuals had access to Copilot, not actual usage, because Copilot does not make that data available at the individual level. All results are observational, limited to the developers included, and not causal.
  • Data:

    • Data on Copilot access was provided to Uplevel Data Labs across several enterprise engineering customers, for a total of 351 developers in the TEST group (with Copilot access) and 434 in the CONTROL group (without Copilot access).
    • The developers in the CONTROL group were similar to those in the TEST group in terms of role, working days, and PR volume in each period.

Navigation for Engineering Leaders

Uplevel is the only holistic system of decision for enterprise engineering organizations. Applying advanced data science to tooling and collaboration data, Uplevel surfaces and interprets the hard-to-find signals that you need to focus your efforts, prioritize initiatives, and build an effective engineering culture.

Learn more at uplevelteam.com


@simonw
Copy link
Author

simonw commented Sep 29, 2024

I used https://gemini.google.com directly to generate this - my prompt was:

Extract full text from this image as markdown in a fenced code block, expand any figure into detailed text descriptions

IMG_1026

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment