Skip to content

Knowledge base

The DORA report on AI-assisted development

The DORA 2025 report finds that AI amplifies the engineering system it lands in: strong review, pipelines and small batches gain speed, weak ones gain instability. Telemetry from Faros AI disputes that a strong foundation protects a team from downstream deterioration.

Written by

Updated:

Evidence 7

Our position. We read DORA’s finding the way we work: AI multiplies what a team already has, so fix the foundation first, then multiply it (AI is a multiplier (opens in a new tab)). Buying licenses is not an AI strategy. Building the engineering environment that amplifies AI’s benefits is.

A DX newsletter article on the DORA 2025 report supports that position and adds the conditions under which the foundation pays off (What the 2025 DORA report means for your AI strategy (opens in a new tab)). AI has to be aimed at system-level bottlenecks, measured with robust frameworks and backed by clear change management.

Does AI fix a weak engineering system?

The piece’s core thesis is that AI magnifies whatever practices already exist: teams with dysfunction see more instability and friction, while teams with solid practices gain speed and resilience. If code review, deployment pipelines, or prioritization are weak, AI simply accelerates the negative effects; if source control, observability, and small-batch shipping are strong, AI becomes an accelerant for positive outcomes.

Where should an AI rollout aim?

DORA’s 2025 report is described as a warning against “spray and pray” AI rollouts that give every developer a tool license but ignore system bottlenecks. The article recommends value stream management and techniques like problem-based learning to direct AI at real SDLC constraints rather than just speeding up individual tasks whose gains are lost elsewhere in the system.

What should you measure once nearly everyone uses AI?

The article stresses that high AI adoption (around 90%) makes “are people using AI?” the wrong primary question. Instead, it promotes a broader measurement strategy (e.g., AI-specific metrics plus existing DORA and DX indicators, combining survey and system data) to answer whether AI is improving outcomes for individuals, teams, products, and the organisation.

Why do internal platforms matter for AI?

Internal developer platforms are framed as key to scaling AI because they provide golden paths, self-service capabilities, and a distribution channel for AI across the SDLC. Investments in developer experience are portrayed as having outsized returns in an AI-enabled environment, though platforms can slightly increase instability if not managed well.

Which levers make AI adoption pay off?

Clear, explicit AI policies and expectations are presented as one of the strongest levers for positive AI impact, reducing friction and risk while enabling safe experimentation. The article also notes DORA’s shift away from elite/high/medium/low performance clusters toward seven data-derived team types that are better linked to concrete capability improvements, reinforcing the theme that measurement exists to drive continuous improvement, not just benchmarking.

Does telemetry confirm the amplifier thesis?

The amplifier thesis has a serious challenger. Faros AI’s 2026 report, built on two years of telemetry from 22,000 developers and 4,000+ teams, finds that strong pre-AI engineering practice offers no protection: organizations with mature DevOps, high DORA scores, and disciplined delivery show the same downstream deterioration as everyone else (Faros AI – The AI Engineering Report 2026: The Acceleration Whiplash (opens in a new tab), blog post of 12 April 2026; Faros sells the telemetry platform). Their explanation for the divergence is methodological, and it is worth taking seriously: DORA’s finding rests on surveys, and surveys capture how developers feel. At the individual level they are faster, so the perception is accurate and the downstream cost is invisible to them. Review queues backing up, incidents accumulating, and PRs merging unreviewed show up in telemetry months before they show up in sentiment. Perception lags; telemetry does not.

The two findings are not fully reconcilable, and we do not pretend otherwise. Treat “AI amplifies your existing system” as the working model for where to invest (platforms, small batches, change management all still pay), and treat the Faros numbers as the warning that a strong foundation is not a free pass. Both agree on the same intervention: fix the code at the point of authoring, not at the point of review. The numbers behind the whiplash are on Productivity evidence and quality.

Resources

Articles