Two reports published on October 20 ask the same question from different sides. AI coding tools help developers produce more. Is the business actually getting more?
Gergely Orosz of The Pragmatic Engineer spoke with engineers and engineering leaders at 27 companies about what they spend on AI and what changed. ActiveState, a company that sells tools for managing open source code, surveyed 307 engineering, security and platform leaders about the work AI leaves behind.
Together, they say something useful for anyone who pays for software to be built.
More pull requests is not the same as more delivered
A pull request is a proposed code change waiting for review. It is easy to count, so it often gets used to measure output.
Data that CodeRabbit, an AI code review company, shared with Orosz shows developers on its platform opening 113% more pull requests per week than a year ago.
Delivery is a different story. At one publicly traded SaaS company with around 600 engineers, pull request volume went up, but the pace of merged work barely changed. At another company, the number of pull requests stayed flat while each change grew to about three times as many lines of code.
Spending ranged from a $150 monthly cap per developer at a public company with 1,500 engineers to about $3,000 per developer at a two-person shop in Europe. Few companies could say what return they were getting.
The clearest gain came from a team that measured something else. A team lead at a multinational said his team now closes about three times as many user stories per month, and for the first time closes more than come in.
The rework is real, and it lands in other budgets
ActiveState’s numbers point to why output and delivery drift apart. 93% of respondents report higher engineering productivity from AI. But counting each answer at the low end of its range, engineers spend at least 139 hours a year, more than 17 working days, fixing or revisiting AI-assisted work.
Other costs land outside engineering:
- 76% had a software release delayed by unmanaged open source dependencies in the past year.
- Half say new dependencies, meaning outside code libraries, can reach production with no review first.
- 83% say developer productivity is outpacing their ability to govern it.
Keep the source in mind. The study was commissioned by a vendor that sells a fix, and it covers companies with 1,000 or more employees in North America and Western Europe. I read it as a signal, not an exact figure.
One engineer told Orosz that human reviewers still catch changes that were not needed at all, or ten times bigger than they should be.
What this means if you are paying for software
Writing code was rarely the only slow part of a business software project. Understanding the workflow, deciding what to build, testing and adoption still take judgment.
In the ERP and custom software projects I have worked on, the slow part was usually agreeing on how the process should work, not writing the code.
So when a vendor or your own team says AI makes them three times faster, it is fair to ask: three times more of what? Lines of code, pull requests, or features your staff actually use?
The gains can be real. The same SaaS engineer said triaging a customer issue, which sometimes took a whole day, now takes minutes plus up to an hour of human checking. Someone still owns the checking.
Practical takeaways
Whether you run a development team or hire one:
- Measure finished work the business can use, like features shipped or issues resolved, not code volume.
- Plan review time on purpose. Faster writing means more to check.
- Ask for smaller changes. A huge AI-generated change is harder to review well.
- Track rework on AI-assisted code as its own category, as LeadDev suggests, so it stops hiding in carryover.
- Make approving a new outside library part of code review.
I am still learning what good measurement looks like here. How do you judge whether AI is actually making your software work faster?
Sources
- The Pragmatic Engineer: What’s the ROI of software engineers’ spending on AI? (Oct 20, 2026)
- ActiveState press release: The AI Remediation Bill study (Oct 20, 2026)
LeadDev: Engineers lose 17+ days a year to AI rework (Oct 20, 2026)