State Of AI-Assisted Software Development

DORA · pdf · 2025-09-23 Reader 🖥️

practical · mixed — The primary DORA 2025 'State of AI-assisted Software Development' report: a ~39k-word mixed-method survey report (nearly 5,000 respondents, 78+ interviews) interleaved with signed guest essays from DORA/Google researchers and outside experts (Gene Kim, Eirini Kalliamvakou, Matt Beane). Practical throughout — every empirical finding is paired with prescriptive advice for technology leaders — but grounded in statistical analysis (cluster analysis, SEM/CFA) rather than opinion.

Structure

Key terms

• AI as amplifier AI as Amplifier
  AI's primary causal role in software development: it magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones, rather than having a uniform effect.
• Team archetype / team profile
  One of seven statistically-derived clusters of teams (e.g. 'Harmonious high-achiever', 'Legacy bottleneck', 'Constrained by process') defined jointly by throughput, instability, team performance, product performance, individual effectiveness, valuable work, friction, and burnout — not by delivery metrics alone.
• DORA AI Capabilities Model
  A set of seven organizational capabilities (clear AI stance, healthy data ecosystem, AI-accessible internal data, strong version control, working in small batches, user-centric focus, quality internal platform) empirically shown to interact with AI adoption — i.e., to amplify (or in one case, without them, reverse) AI's effect on outcomes.
• Value stream management (VSM)
  The practice of visualizing, analyzing, and improving the end-to-end flow of work from idea to customer; in this report specifically framed as the mechanism that keeps local AI-driven acceleration from being absorbed or wasted downstream.
• Systems view / organizations as systems
  The view (credited to Deming and the Theory of Constraints) that organizational performance emerges from the interaction of interdependent parts — teams, processes, policies, infrastructure — such that improving one part (e.g. code-generation speed) does not improve overall throughput unless the constraining downstream parts (testing, review, deployment, compliance) evolve too.
• Continuous AI
  A proposed AI-native collaboration pattern (credited to GitHub Next) in which an AI system persists as a living, continuously-updated participant in the team's workflow rather than a one-shot tool invocation — perceiving project events, adjusting direction with the team.

Propositions

• explicitAI functions primarily as an amplifier of the underlying organizational system: it magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones, rather than acting as a uniform force.conf: high
  extendsAI-Assisted Software Development: Team Profiles and Capabilities for Putting Research into Action · p1 — Same amplifier thesis; this is the primary report, the InfoQ piece a compressed secondary write-up of it.
• explicitA cluster analysis of software delivery performance, team well-being, and product outcomes across nearly 5,000 respondents reveals seven distinct team archetypes (from 'Harmonious high-achiever' to 'Legacy bottleneck' and 'Foundational challenges'), showing that delivery metrics alone cannot distinguish whether low performance is caused by technical debt, bureaucratic process, or burnout.conf: high
  extendsAI-Assisted Software Development: Team Profiles and Capabilities for Putting Research into Action · p2 — Adds full per-cluster percentages, performance/well-being/stability descriptions, and the explicit 'speed vs. stability is a myth' finding (prop 3) absent from the secondary summary.
• explicitThe long-assumed 'speed vs. stability' trade-off is empirically false at the high-performing end: the two best clusters ('Pragmatic performers' and 'Harmonious high-achievers', together ~40% of respondents) excel simultaneously at both throughput and stability, while a 'High impact, low cadence' cluster shows that speed without stability is a dangerous, unsustainable combination.conf: high
• explicitestablishedAI adoption now measurably improves software delivery throughput — a reversal from the 2024 report's finding — but still increases delivery instability, indicating teams are adapting their individual work for speed before their surrounding systems have evolved to safely absorb AI-accelerated change.conf: high
  extendsAI-Assisted Software Development: Team Profiles and Capabilities for Putting Research into Action · p5 — Frames the same instability finding explicitly as a reversal from the 2024 report on throughput specifically, which the secondary article doesn't state.
  supports ←How to measure AI developer productivity in 2025 | Nicole Forsgren · p5 — Both describe the same pattern -- AI genuinely speeds up coding/throughput, but new instability, bottlenecks, and tooling unreliability absorb much of the gain -- Forsgren (a DORA co-creator) giving the practitioner-level account of what the 2025 DORA report measures at the dataset level.
  supports ←Bugs in Large Language Models Generated Code: An Empirical Study · p3 — A concrete mechanism for why AI adoption increases delivery instability even as it improves throughput: a non-trivial share of AI-generated bugs (the generatively-caused categories) are not addressable by better prompting and will recur regardless of how skilled the prompting team becomes.
• explicitcorroboratedSeven organizational capabilities — a clear and communicated AI stance, a healthy data ecosystem, AI-accessible internal data, strong version control practices, working in small batches, a user-centric focus, and a quality internal platform — each measurably amplify AI's positive effect on individual, team, or organizational outcomes.conf: high
  extendsAI-Assisted Software Development: Team Profiles and Capabilities for Putting Research into Action · p3 — Adds the detailed measurement basis, certainty levels, and specific outcome pathway for each of the seven capabilities (the InfoQ piece only lists them).
  extends ←How to measure AI developer productivity in 2025 | Nicole Forsgren · p3 — Forsgren's point that DORA's four metrics can't be applied blindly in the AI era because feedback loops must now happen earlier and faster is the practitioner-voiced version of DORA 2025's own finding that AI ROI is an organizational-system property (data ecosystems, version control, quality platforms) rather than a tooling property measurable by the old metrics alone.
  supports ←The Pragmatic Engineer 2025 Survey: What's in your tech stack? · p3 — DORA names 'a clear and communicated AI stance' as one of seven organizational capabilities correlated with better outcomes; this survey's finding that larger companies converge on narrower, enterprise-standardized tool policy (and smaller companies on permissive 'use whatever' policy) is a concrete, tool-level instance of organizations differing precisely on how clear/enforced that AI stance is.
• explicitIn the absence of a user-centric focus, AI adoption has a measurable negative effect on team performance — meaning AI can actively make a team's performance worse, not merely fail to help it.conf: high
• explicitA quality internal platform amplifies AI's benefit to organizational performance partly by increasing, not decreasing, developer friction — because in the absence of settled best practices for AI-assisted development, the platform's main current function is gatekeeping inappropriate use (e.g. internal-only APIs with tighter security controls) rather than frictionlessly expanding access.conf: medium
• explicitAI adoption has not, so far, meaningfully changed developers' psychological relationship to their work — no measured shift in sense of meaning, need for cognition, or existential connection to colleagues, and no change in the perceived importance of most job skills except prompt engineering and (surprisingly) syntax memorization — with the exceptions that higher AI adoption is associated with greater authentic pride (via more time on valued work) and carries a small (21%) risk of reduced psychological ownership of one's code.conf: high
• explicitestablishedDefault AI usage patterns disrupt the traditional three-generation (junior/mid/senior) apprenticeship model, because experts can now self-serve answers from AI and therefore stop creating the hands-on, high-stakes opportunities junior engineers historically needed to develop tacit skill — a pattern Matt Beane has documented across more than 31 occupations and now sees emerging specifically in software engineering.conf: medium
  supportsWhen AI can do more than we can check · p13 — Beane's cross-occupation evidence of apprenticeship-pipeline erosion corroborates the premise behind that essay's proposed 'synthetic apprenticeship' fix — that the pipeline producing expert judgment is genuinely breaking.
  supportsResearch: Why Some Junior Employees Work Well with AI—and Others Don’t · p1 — A large-sample, cross-occupation finding (Beane) and a large-sample, single-company finding (KPMG study) converge on the same underlying phenomenon: junior/early-career engagement with AI is heterogeneous and a real risk category, not a uniform benefit.
  supports ←The Bet On Juniors Just Got Better · p6 — DORA documents the traditional junior/mid/senior apprenticeship model being disrupted by default AI usage patterns that remove hands-on learning opportunities; Beck's compounding-mentorship point is the positive case for why deliberately preserving and accelerating that apprenticeship chain (rather than letting it erode) pays off beyond any one developer.
  supports ←Redefining the Software Engineering Profession for AI · p2 — Both describe generative AI disrupting the traditional junior/mid/senior apprenticeship model by removing the hands-on opportunities EiC/junior engineers historically needed.
• explicitSustainable high performance requires jointly optimizing for productivity and skill development, not optimizing for productivity alone — in Beane's broader research, the greatest productivity gains were achieved only when organizations simultaneously insisted on skill development.conf: medium
• explicitcorroboratedOrganizations are systems, not sums of individuals: because downstream capacity (testing, review, integration, deployment, compliance) is a fixed constraint, individual-level AI productivity gains are absorbed rather than converted into organizational throughput unless the surrounding workflow itself is redesigned — the same dynamic DORA observed during the shift to cloud and the adoption of Agile/DevOps.conf: high
  supportsIs AI coding on average 17.3x faster or 1.3x faster? Both. — LinkedIn post by Fabrice Bernhard · p11 — DORA's Theory-of-Constraints argument (downstream capacity absorbs upstream AI gains) gives the empirical/theoretical mechanism behind Bernhard's claim that AI coding adoption targets the wrong point in the value stream.
  supports ←Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools · p3 — DORA's survey-based claim that downstream capacity (testing, review, integration, deployment) is a fixed constraint limiting local gains is independently confirmed here by a causally-identified, large-N estimate of exactly that constraint (elasticity of substitution of 0.25 between AI output and human downstream effort).
• explicitAI adoption should be managed as a deliberate organizational transformation effort — with executive-sponsored, cross-functional value-stream mapping and redefinition of roles (verification, orchestration, workflow design) — rather than treated as a tooling procurement decision.conf: high
• explicitValue stream management acts as a force multiplier for AI investment: by exposing where coordination costs, delays, and bottlenecks accumulate, it ensures local AI-driven productivity gains are converted into organizational-level performance gains instead of being lost to downstream chaos.conf: high
• implicitThe report's own statistical framing explicitly disclaims causal interpretation (its authors describe their regressions as 'comparisons,' not 'effects') — yet nearly every practical-advice section ('invest in your internal platform,' 'clarify and socialize your AI policies') is phrased as a causal prescription, leaving open whether building a capability causes AI's amplified benefit or whether already well-run organizations simply both have the capability and extract more from AI for independent reasons.conf: medium

Arguments

Narratives

• At Adidas's gen-AI pilot (nearly 1,000 developers), teams with loosely-coupled architectures and fast feedback loops saw 20-30% productivity gains and 50% more 'Happy Time,' while teams tightly coupled to ERP systems with slow feedback loops saw little or no AI benefit at all.
  → Concretizes proposition 1 (AI as amplifier) and the implicit premise that architecture is a prerequisite, not a side effect, of AI payoff.
• At Booking.com (3,000+ developers), uneven uptake of coding assistants was traced not to tool quality but to missing training; once developers learned to give AI more explicit instructions and context, merge requests rose up to 30% and job satisfaction increased.
  → Foreshadows the Capabilities Model's central claim that building organizational/behavioral capability — not distributing the tool — is the actual lever on AI payoff.
• An interviewed developer describes self-censoring their AI experimentation because they didn't know whether their organization's unstated norms would approve of it, saying they didn't want to use AI 'in secret' but also weren't sure it was encouraged.
  → Concretizes proposition 5/argument 3 (clear AI stance) by showing the lived cost of ambiguity: both under-use from fear and the discomfort of operating in a policy vacuum.

Questions addressed

• ✅ Does AI actually enable better software delivery and organizational performance, the same question DORA asked of DevOps a decade ago?
  Conditionally yes: AI now improves throughput (a reversal from 2024) but still increases instability, and the net benefit depends on whether the organization has built the seven capabilities in the DORA AI Capabilities Model.
• ✅ What organizational conditions determine whether AI adoption helps or harms a team, rather than just how much AI individuals use?
  The DORA AI Capabilities Model: seven capabilities (AI-stance clarity, data health, AI-accessible data, version control discipline, small batches, user-centricity, platform quality) that interact with AI adoption to amplify its benefits — and, for user-centric focus specifically, whose absence turns AI adoption harmful.
• ✅ Does using AI assistance erode developers' sense of ownership, meaning, or motivation in their work?
  Mostly no measurable change across most sociocognitive constructs tested, with a modest increase in authentic pride and a small (21%) risk of reduced psychological ownership.
• ⭕ What happens to the junior-to-senior skill pipeline when experienced engineers can self-serve answers from AI instead of mentoring juniors through hands-on work?
  Named and partially explored only in Matt Beane's standalone guest essay, which reports early, cross-occupation evidence of apprenticeship disruption but offers no resolution specific to software engineering and proposes only a future research agenda (AI-assisted measurement of skill development).

Gaps

• incompleteThe DORA AI Capabilities Model's seven capabilities — the report's headline, survey-validated deliverable — include no apprenticeship or skill-transmission capability, even though Matt Beane's own essay in the same report argues default AI use measurably erodes the junior-to-senior pipeline across many occupations and (early signs suggest) in software engineering too. The model and the warning sit side by side without being reconciled.
  Read next: Track whether DORA's next annual survey operationalizes a skill-development or apprenticeship-health measure as an eighth capability; read Matt Beane's ongoing SkillBench research directly for software-specific findings.
• illogicalThe report's statistical methodology explicitly disclaims causal interpretation (its own footnotes describe the regressions as comparisons, not effects), yet its practical-advice sections phrase every one of the seven capabilities as something leaders should causally 'invest in' or 'build' to amplify AI's benefit. This leaves open whether building a capability causes the amplification, or whether already well-run organizations simply both have the capability and separately extract more value from AI.
  Read next: Look for a longitudinal/panel-data follow-up DORA report, or independent experimental/quasi-experimental evidence isolating any one of the seven capabilities from general organizational maturity.
• incompleteQuality internal platforms are reported to increase AI-era developer friction even as they improve organizational performance, but the report doesn't distinguish how much of that friction is the deliberate, 'right kind' of guardrail (e.g. security gatekeeping) versus avoidable platform-usability debt — a distinction a platform team would need in order to act on the finding.
  Read next: A practitioner-level study decomposing AI-era platform friction into 'intended guardrail' vs. 'usability debt' categories.
• incompleteThe report's two major empirical halves — the seven team archetypes (cluster analysis) and the sociocognitive impact findings (pride, ownership, skill reprioritization) — are never cross-tabulated against each other, so it's unknown whether, say, the 'Legacy bottleneck' cluster shows different ownership or pride effects from AI than the 'Harmonious high-achiever' cluster.
  Read next: A follow-up analysis joining cluster membership with the sociocognitive constructs in DORA's underlying dataset.

Closes gaps elsewhere

• KI, utvikling og følelser · g2 — Partially closes it: Beane's research gives an operational, institutional mechanism for the transmission problem (pair programming as tacit-knowledge transfer; the three-generation model; using AI itself to measure skill development from developer-AI interaction logs) — but doesn't engage the craft-history analogy (Sennett, Crawford) the gap also calls for, so the empirical/institutional half closes while the historical-analogy half remains open.
• It Ain't Broke: Why Software Fundamentals Matter More Than Ever · g5 — Beane's cross-occupation research (31+ occupations) directly answers whether AI-mediated 'gray box delegation' creates a competency gap: yes, by default, because experts self-serve with AI and stop creating the hands-on opportunities novices need — with early signs of the same pattern specifically in software engineering.
• The Future Of Software Engineering · g4 — Beane's essay supplies the apprenticeship-model literature and specificity the gap asked for: the at-risk 'fundamentals' are the tacit, hands-on skills transmitted via the three-generation (junior/mid/senior) model, and effective retraining means jointly measuring and incentivizing skill development alongside productivity rather than offering generic AI-tool training.

Related documents

• Is AI coding on average 17.3x faster or 1.3x faster? Both. — LinkedIn post by Fabrice Bernhard
• When AI can do more than we can check
• 2024_final_dora_report
  AI as Amplifier Burnout & Recovery DORA AI Capabilities Model (Team Profiles & Capabilities) DORA & Engineering Metrics
• AI-Assisted Software Development: Team Profiles and Capabilities for Putting Research into Action
  AI as Amplifier Burnout & Recovery DORA AI Capabilities Model (Team Profiles & Capabilities) DORA & Engineering Metrics
• Engineering Leadership Report 2026
  Burnout & Recovery DORA & Engineering Metrics Enterprise AI Adoption & Readiness Systemic Apprenticeship & Skill Transmission
• 🧠 Community Wisdom: How AI is changing product operating models, tracking work stress with Whoop, whether you nee…
  AI as Amplifier Burnout & Recovery Enterprise AI Adoption & Readiness
• Finding Comfort in the Uncertainty
  AI as Amplifier Enterprise AI Adoption & Readiness Platform Engineering
• How AI will change software engineering
  DORA & Engineering Metrics Enterprise AI Adoption & Readiness Systemic Apprenticeship & Skill Transmission
• The "funhouse mirror": How AI reflects the hidden truths of your software pipeline
  AI as Amplifier DORA & Engineering Metrics
• Can an AI-Powered Scribe Curb Physician Burnout?
  Burnout & Recovery Systemic Apprenticeship & Skill Transmission