Key takeaways:
- AI adoption is up 65%, but median PR throughput rose just 8%.
- Coding is only ~16% of engineer time, so speeding it up barely moves overall throughput.
- To close the gap, target AI at non-coding bottlenecks, and track utilization, impact, and cost together.
Most engineering leaders I talk to are quietly worried they’re falling behind.
AI models and coding tools are launching constantly, often with claims of 3x or 10x productivity improvements. When their teams see more modest results, they assume something is wrong.
The concern is understandable. But in most cases, it’s the benchmark that’s wrong, not the results.
Our team at DX analyzed engineering velocity from November 2024 to February 2026 across a sample from 400+ companies where AI adoption rose sharply. We found a 7.76% median increase in PR throughput – the best, though imperfect, proxy metric we currently have for directionally measuring velocity.
That’s a real gain. It’s also well below what most leaders expect, and well below what vendor marketing typically promises.
If your numbers are in that range, you are not behind. You are in line with the industry.
What the data actually shows
Our sample covered technology, financial services, retail, and healthcare companies ranging from approximately 150 to 10,000 engineers. During the study period, AI tool usage increased by an average of 65%. Median PR throughput increased by 7.76%.
The mean gain was 13.1%, higher than the median due to a small number of high-performing companies. Even at the 90th percentile, gains approached 44% — still far below the order-of-magnitude improvements often cited in executive circles.
For most organizations, today’s AI coding tools are delivering a 5–15% throughput gain. To put that in concrete terms: an organization with 500 engineers seeing a 10% improvement means potentially shipping 10% more value per engineer, without the headcount cost. That is a meaningful return. The mistake is measuring it against an expectation of 2–3x productivity gains.
Why the gap exists
If AI adoption increased 65%, why has throughput moved less than 8% at the median? We interviewed developers across our sample to understand what’s holding them back.
Coding isn’t the main bottleneck.
AI accelerates code generation, but coding represents approximately 16% of how engineers spend their time. Even cutting that time in half doesn’t meaningfully move overall throughput. As one developer put it: “A four-day task might take three. But that doesn’t mean I’m shipping 3x more PRs.”
New bottlenecks emerge
AI has accelerated code generation, but code review and integration remain largely unassisted. Several developers described time saved writing code being consumed by the extra scrutiny AI-generated output requires, resulting in a net-zero productivity gain.
Skill and tooling gaps compound each other.
Effective use of AI tools and agents is its own discipline, and most developers are still early in developing it. Immature tools steepen the learning curve; developers early on the curve extract less value.
In our research, we have found that every business firmographic and developer demographic undergoes a measurable J-curve in speed and quality when first learning about these tools. With additional practice and education, this curve corrects, and advanced users consistently outperform their peers.
AI tools lack institutional context.
AI performs well on self-contained, well-documented problems. Most real engineering work is neither. The context typically lives in people’s heads or is distributed across heterogeneous systems of record.
More like this
Three steps to unlock more from AI investment
The 5–15% gains we’re seeing today are not the ceiling. Closing the gap between current results and the potential of AI will require looking beyond code generation.
Based on our research, here’s where I’d focus.
- AI tools are amplifiers. They take what developers are already doing and scale it, for better or worse. A well-documented, well-structured codebase with low friction makes it easier for developers to use AI effectively. Poor documentation and high friction make it harder. Assess your codebases and workflows at the service and team level before scaling AI use.
- Since coding represents only ~14% of a developer’s time, the highest-leverage opportunities are likely elsewhere: in planning, code review, documentation, and operations. Use developer experience surveys and system telemetry to identify where friction is highest, then direct AI investment toward those specific pain points.
- Increased velocity from AI comes with risks that, left unmonitored, can erode the gains. Track not just whether throughput is improving, but whether that improvement is coming at a cost. The DX AI Measurement Framework covers three dimensions – utilization, impact, and cost – giving leaders validated signals across the full AI adoption lifecycle. This multi-dimensional approach matters, because improvements in one dimension can come at the expense of another.
Where to focus next
The organizations that capture the next wave of AI-driven productivity gains will be the ones that invest in the full breadth of their development lifecycle and in the human and contextual conditions that allow AI to be effective.
That starts with three steps: assess your foundational readiness by service and team; identify where the highest-leverage AI opportunities exist; and measure progress using validated signals across utilization, impact, and cost.
The gains are real. The ceiling is higher than where most organizations are today. Getting there requires a more complete view of where AI coding tools are and aren’t working.
Promoted Partner Content
