High-performing organizations are getting faster, while lower-performing organizations are getting slower. The 2025 DORA research makes this pattern explicit. AI is not a neutral accelerator, it’s an amplifier. It magnifies the strengths of mature engineering systems and intensifies the weaknesses of immature ones.
The Same Tools, Radically Different Results
What makes this divergence easy to miss is that the inputs look identical. The same code generation tools, the same underlying models, the same initial reduction in lead time. It’s tempting to attribute the difference in outcomes to tool usage or adoption speed, but that explanation doesn’t hold. What’s actually happening is a system-level stress test. Organizations with resilient delivery systems absorb the increased flow and accelerate. Organizations with fragile ones surface constraints that were previously tolerable but never resolved.
Every engineering organization is a system governed by constraints. Throughput is determined not by how fast any individual activity can go, but by the slowest, most fragile step in the flow.
In high-performing organizations, those constraints have already been mitigated:
- Code changes are small and reversible
- Validation is automated and reliable
- Review and integration are disciplined and fast
- Feedback from production and users arrives quickly
When AI accelerates code generation in these environments, the system absorbs the increase. Throughput improves. Learning accelerates.
In lower-performing organizations, the constraint typically lives downstream from code generation:
- Review capacity is limited
- Testing focuses on correctness, not value
- Feedback arrives late, if at all
- Quality ownership is implicit and diffuse
When AI accelerates code generation here, the constraints tighten. Queues grow. Teams have to choose between increased lead time or degraded platform stability.
The same optimization produces opposite outcomes.
AI Doesn’t Change How the Broader System Works
This is where many AI adoption efforts quietly go wrong. Today’s AI tooling makes the fast parts of the system faster. It does not make the slow parts more resilient. For most teams, code production was never the bottleneck. Review, validation, and learning are.
This is classic systems behavior. Optimizing a non-bottleneck does not increase throughput. It increases work-in-progress. In a factory, inventory piles up on the floor. In software, it piles up as pull requests, review queues, and partially validated ideas.
Why Some Teams Get Leverage and Others Get Noise
The most important insight in the 2025 DORA AI Capabilities Model is not about tools. It’s about foundational capabilities.
Teams that see positive outcomes from AI consistently operate with:
- Small batch sizes keep changes easy to review
- Internal platforms provide quality guardrails
- Feedback loops tie outputs to outcomes
- Etc
Teams that lack these capabilities experience something else entirely: more code, more instability, more friction, and more burnout. AI adoption combined with mature foundational capabilities produces outsized gains.
The Question Leaders Should Be Asking
The wrong question is:
“How do we roll out AI code generation?”
The right question is:
“Where is our system’s bottleneck today, and what happens when AI pushes harder on it?”
For many organizations, the answer is uncomfortable:
- Review queues are already long
- Test automation is incomplete, slow, and/or brittle
- Feedback loops are indirect
- Quality relies on individual heroics
- Deployments require a lot of manual QA
AI doesn’t fix these problems; it makes them worse.
What the Research Suggests You Do
The 2025 DORA AI Capabilities Model is clear: organizations that benefit most from AI do not start with tools. They start by strengthening the parts of their delivery system that AI will inevitably pressurize.
Before accelerating code generation, leaders need a clear view of where the system is constrained and which foundational capabilities govern flow, judgment, and feedback at that constraint. Skipping this step is where many AI rollouts go wrong.
At Pragmint, this insight has been translated into a practical, research-backed approach called the S.T.E.P. framework, supported by an open-source DORA Capabilities Maturity Assessment and a growing library of documented engineering practices.
The process begins with Survey. Pragmint’s open-source DORA Capabilities Maturity Survey helps teams assess adoption across 29 DORA capabilities. The objective is not scoring for its own sake, but identifying gaps that are likely acting as system-level bottlenecks.
From there, teams Target. Not every weakness deserves attention at once. Focus shifts to the capabilities that are currently constraining throughput, learning, or stability, prioritizing improvements where the system will experience the greatest relief.
Teams then Experiment. Rather than prescribing universal solutions, Pragmint maintains an open repository of engineering practices that have proven effective across many contexts. Each practice is explicitly mapped to one or more DORA capabilities, creating clarity around what to try, why it matters, and which system behavior it is meant to improve. Teams select a small number of high-impact experiments and integrate them into day-to-day work.
Finally, teams Polish or Pitch. Reflection focuses on evidence, not anecdotes. Metrics and signals attached to each practice help determine whether an experiment improved flow, quality, or learning. Successful practices are refined and reinforced. Unsuccessful ones are dropped. The cycle then repeats.
This approach strengthens delivery systems without pausing delivery itself.
Strengthening the capabilities that constrain flow doesn’t make teams “faster” in isolation. Increased system capacity allows AI-driven speed to become leverage rather than load.
The difference isn’t the tool or model you choose. It’s the system you ask it to scale.