DORA in the AI era: the metrics held, the practices got louder

Overview

The four keys survived AI. Everything around them shifted. Between September 2023 and September 2026 DORA renamed one metric (time to restore became failed deployment recovery time, redefined in 2023 to isolate failures caused by a production change), reclassified another (recovery time moved from stability to throughput in 2024), added a fifth (rework rate, unplanned deployments fixing user-facing bugs), and in 2025 rebranded the whole programme as the State of AI-assisted Software Development, retiring low-to-elite performance tiers in favour of seven team archetypes. Sources: DORA / Google Cloud (2026) (↗); Continuous Delivery Foundation (2025) (↗); redmonk.com (2024) (↗); RedMonk (2025) (↗)

The defining shift of the past eighteen months is the collapse of the simple speed-up narrative. DORA 2024 measured a 25% increase in AI adoption as correlating with a 7.2% reduction in delivery stability and a 1.5% reduction in throughput, even as individual productivity and flow rose. METR's randomised trial then found experienced developers 19% slower with AI while believing themselves 20% faster. By 2025 DORA had replaced trade-off language with an amplifier thesis: AI magnifies the strengths of well-run organisations and the dysfunctions of struggling ones. Sources: dora.dev (n.d.) (↗); metr.org (2025) (↗); Google Cloud Blog (2025) (↗)

For a leader deciding what to change on Monday, the practical message is conservative. The capabilities that predicted the four keys before AI (small batches, version control discipline, continuous testing, platform quality) are the same ones DORA's 2025 AI Capabilities Model found determine whether AI helps or harms delivery. AI raises the price of skipping them, because it makes it trivially easy to enlarge batch size, duplicate code, and outrun review capacity. Sources: Google Research (2025) (↗); Substack (practitioner) (2024) (↗)

Timeline

H2 2023

• DORA redefines recovery metric as failed deployment recovery time, isolating change-caused failures

H1 2024

• GitHub and Accenture publish enterprise Copilot ROI figures, setting the vendor benchmark for gains

H2 2024

• DORA 2024 measures AI adoption reducing throughput and stability
• Rework rate added as fifth metric, recovery time reclassified as throughput

H1 2025

• GitClear reports copy-pasted code exceeding refactored code for the first time
• Cursor maker Anysphere valued at 9.9 billion dollars, capital fully committed to the acceleration thesis

H2 2025

• METR RCT finds 19% slowdown against a forecast 24% speedup
• DORA rebrands as State of AI-assisted Software Development, ships the seven-capability AI model and seven team archetypes
• Gartner publishes first Magic Quadrant for AI code assistants

H1 2026

• METR abandons its follow-up study after developers refuse the no-AI arm
• GitClear extends telemetry, block duplication up 81% on the 2023 baseline
• Vibe coding gives way to disciplined context engineering per Thoughtworks

Key Findings

Throughput and stability are now framed as companions, not a trade-off, and the metric groupings changed to enforce that. DORA 2024 formally paired deployment frequency and lead time as throughput and change failure rate and recovery time as stability, then moved recovery time into throughput and treated change failure rate as a proxy for rework, adding deployment rework rate to measure it directly. RedMonk's Rachel Stephens flagged this as the report's most consequential structural move. The decade-long finding behind it is that high performers achieve both at once, so any local improvement that trades one for the other is a mis-optimisation. Sources: redmonk.com (2024) (↗); Continuous Delivery Foundation (2025) (↗); DORA / Google Cloud (2026) (↗)

The strongest levers per metric are boring and repeated across report years. Lead time responds to batch size and trunk-based version control practice: DORA's explanation for the 2024 AI penalty was precisely that AI enlarges changesets, and "DORA's data has consistently shown that larger changesets introduce risk". Change failure rate and rework respond to continuous testing and review discipline. Recovery time responds to version control hygiene and rollback capability, which is why "strong version control practices" appears verbatim in the 2025 AI Capabilities Model. The Monday action is mechanical: cap PR size, keep AI-generated work on short-lived branches, and refuse to let scaffolding volume masquerade as throughput. Sources: thenewstack.io (2024) (↗); Google Research (2025) (↗); aviator.co (2026) (↗)

DORA 2024's headline was a measured contradiction, and it replicated only partially. With roughly 39,000 respondents and about 3,000 answering AI questions, the 2024 Bayesian model tied a 25% AI adoption increase to a 7.2% stability drop, a 1.5% throughput drop, and a 2.1% individual productivity gain. The 2025 wave (near 5,000 respondents, over 100 hours of interviews) flipped the throughput sign positive while instability persisted; DORA's own language was that AI "not only fails to fix instability, it is currently associated with increasing instability". The stability finding has now held across two survey waves. The throughput finding has not, so treat it as unsettled. Sources: dora.dev (n.d.) (↗); RedMonk (2025) (↗); Research-Driven Engineering Leadership (Substack) (2025) (↗)

METR's RCT is the field's only controlled evidence, and it broke the self-report link. Sixteen experienced open-source developers, 246 real tasks in repositories they had maintained for an average of five years, mostly Cursor Pro with Claude 3.5/3.7 Sonnet: measured completion time rose 19% (95% CI roughly -40% to -2%) against a forecast 24% speedup and a post-hoc estimate of 20% faster. Independent readings from LessWrong and others argue the result may not generalise beyond experts in familiar code, and METR itself now labels it historical, stating in February 2026 that it believes developers are more sped up by early-2026 AI than the study measured. The durable lesson is narrower and more useful: developer self-report is an unreliable instrument for AI productivity, so measure delivery outcomes instead. Sources: arXiv / METR (2025) (↗); LessWrong (2025) (↗); METR (Substack) (2026) (↗); Reuters (via Investing.com) (2025) (↗)

GitClear's telemetry corroborates the instability side, not the productivity side. Across 211 million changed lines (2020 to 2024), copy-pasted code exceeded moved (refactored) code for the first time in 2024, duplicated blocks rose roughly eightfold, and two-week churn climbed from about 3.1% in 2020 to 5.7% in 2024, with moved code falling from about 25% of changed lines in 2021 to under 10%. The 2026 follow-up found block duplication up 81% on the 2023 baseline, and, tellingly, that heavy AI users' 4 to 10x output advantage over non-users mostly predates their AI adoption, with a more modest 25% velocity gain against their own past output. This is vendor telemetry from one company's customer base, but it is the only commit-level dataset at this scale and it points the same direction as DORA's stability finding. Sources: DevClass (2025) (↗); particula.tech (2026) (↗); DevOps.com (citing Wall Street Journal) (2025) (↗)

The 2025 AI Capabilities Model is mostly the old capability list with an AI prefix, and that is its point. Built from 78 interviews and validated against the near-5,000 sample, its seven capabilities are a clear and communicated AI stance, healthy data ecosystems, AI-accessible internal data, strong version control practices, working in small batches, user-centric focus, and quality internal platforms, selected from fifteen candidates for showing a statistically significant interaction with outcomes. Three of the seven (small batches, version control, platforms) are pre-existing DORA capabilities. The genuinely new ones concern making internal context available to AI tools and telling people what they are allowed to do with them, both cheap to implement and, per DORA, both currently rare. Sources: Google Research (2025) (↗); services.google.com (n.d.) (↗); Google Cloud Blog (2025) (↗)

Platform engineering shows the same bifurcation as AI itself. DORA 2024 found internal platforms lifted individual productivity by roughly 8% and organisational performance by roughly 6% while reducing throughput by about 8% and stability by about 14%. The pattern (individual gains, delivery losses) mirrors the AI finding, suggesting a general failure mode: tools that make individuals feel faster while delivery discipline erodes. A platform is a lever only when paired with the batch-size and testing practices it is supposed to encode. Sources: thenewstack.io (2024) (↗); redmonk.com (2024) (↗)

SPACE, DevEx and DX Core 4 now bracket DORA rather than compete with it. The DevEx framework (Noda, Storey, Forsgren, Greiler, ACM Queue 2023) explicitly measures the experience "before and after" the commit-to-production window DORA covers, via feedback loops, cognitive load and flow. DX Core 4 (2024) consolidates the lot. GetDX's practitioner argument is that the four keys under-measure AI-assisted teams because deployment frequency inflates with AI scaffolding while review overhead for AI code often exceeds that for human code, so review load and cognitive load belong on the dashboard next to the five keys. Sources: The Pragmatic Engineer (Gergely Orosz) (2024) (↗); GetDX Newsletter (Substack) (2024) (↗); Unblocked (company blog) (2026) (↗); Jens Oliver Meiert (personal blog) (2025) (↗)

A good team in 2026 looks like DORA's "harmonious high-achievers" archetype, and its numbers are behavioural, not threshold-based. Ninety per cent of respondents use AI daily at a median two hours, yet 30% report little or no trust in AI output, and the good teams treat that distrust as process: small batches enforced on AI-generated changes, rework rate tracked as a first-class stability signal, internal data wired into tools, and an explicit organisational AI stance. Thoughtworks' Rachel Laycock reports vibe coding "practically disappeared" by late 2025 in favour of context engineering, which is the same conclusion Simon Willison reached from practice: AI multiplies existing discipline in testing, planning, review and documentation. Elite multipliers such as 127x lead time still circulate from the 2024 report, but the tier structure behind them no longer exists, so quote them with a year attached or not at all. Sources: RedMonk (2025) (↗); cloud.google.com (n.d.) (↗); Thoughtworks (2026) (↗); Simon Willison's Weblog (2025) (↗)

Evidence & Data

The load-bearing numbers, by source type. DORA survey (organisational, correlational): 2024, 25% more AI adoption maps to -7.2% stability, -1.5% throughput, +2.1% individual productivity, on ~39,000 respondents; 2025, throughput turns positive, instability persists, 90% daily AI use, 30% low trust, on ~5,000 respondents. Controlled trial (individual, causal): METR, 16 developers, 246 tasks, +19% completion time against forecast -24%. Commit telemetry (codebase, observational): GitClear, churn 3.1% to 5.7% (2020 to 2024), duplication up roughly eightfold in 2024, up 81% on 2023 by 2026, cloned code linked to 15 to 50% more defects. Sources: dora.dev (n.d.) (↗); Google Blog (2025) (↗); metr.org (2025) (↗); particula.tech (2026) (↗)

These are different units of analysis and should never be merged. Cui and colleagues' pooled field experiments across Microsoft, Accenture and a Fortune 100 firm (4,867 developers) found a 26.08% increase in completed tasks, with larger gains for juniors, while Peng's original Copilot RCT found 55.8% on a single scripted task. Both measure task output, not delivery. Bain puts realised team-level gains at 10 to 15%, noting coding is only 25 to 35% of the lifecycle; Gartner forecasts 30% "through 2028" as projection, not measurement; McKinsey's Jellyfish-partnered claim of over 110% gains at 80 to 100% adoption is vendor-sourced and unreplicated. Sources: cmustatistics.github.io (2025) (↗); Bain & Company (2025) (↗); Gartner (2025) (↗); McKinsey & Company (2025) (↗)

The market priced the bull case regardless: Anysphere at 9.9 billion dollars in June 2025 past 500 million ARR, past 2 billion ARR by early 2026, with talks at a 50 billion valuation. Sources: TechCrunch (citing Bloomberg) (2025) (↗); TechCrunch (citing Bloomberg) (2026) (↗)

Signals & Tensions

Measured versus asserted. DORA's Bayesian survey model, METR's RCT and GitClear's telemetry form the measured tier. Vendor telemetry (Faros, Jellyfish), Copilot ROI studies and Magic Quadrant placements form the asserted tier, and consistently report larger, more favourable numbers. The gap between the two tiers runs 30 to 50 percentage points. Sources: GitHub Blog (2024) (↗); faros.ai (2026) (↗); The Register (2025) (↗)

Instrumentation drift. DORA's canonical definitions come from self-report, while tool vendors instrument the same names through CI/CD data with materially different choices, such as whether rollbacks count as deployments or lead time starts at ticket rather than commit. Both errors flatter the numbers, and the survey-versus-telemetry gap is largely unquantified. Sources: getdx.com (2026) (↗); dynatrace.com (2025) (↗)

The METR refusal problem is itself a finding. The follow-up died because developers would not accept randomisation into a no-AI arm, which biases future estimates downward and means the cleanest study design in the field may be unrepeatable. Sources: METR (Substack) (2026) (↗); METR (2026) (↗)

Capability trend versus delivery trend. METR's time-horizon evaluations show autonomous coding capability climbing steadily (GPT-5 at roughly 2h15m at 50% reliability against o3's 1h30m) while team-level stability telemetry worsens with adoption. Both can be true; the bottleneck has moved to integration, review and rework. Sources: METR (2025) (↗); metr.org (2025) (↗)

Underreported: skill formation. A 2026 randomised study found AI use impaired conceptual understanding and debugging without average efficiency gains, a cost no delivery metric currently captures. Sources: ainewstoday.substack.com (n.d.) (↗)

Open Questions

Whether the 2025 throughput reversal is a real learning effect or an artefact of a changed sample and instrument, since DORA's waves are not a panel and question wording shifted with the rebrand. Sources: DORA (2025) (↗)

Whether rework rate and the AI-era proposals (AI-attributable rework, acceptance rate, trust, review load) survive a second survey wave, or join the churn of retired constructs like the elite tier. Sources: RedMonk (2025) (↗); GetDX Newsletter (Substack) (2025) (↗)

Whether GitClear's duplication and churn trends replicate outside its own customer base; the first independent RCTs on AI-assisted maintainability are only now appearing on arXiv. Sources: DevClass (2025) (↗)

How much of the individual-versus-organisational gap is review bottleneck, measurable as review load, versus genuine quality decay, measurable as defects, since the interventions differ. Sources: Unblocked (company blog) (2026) (↗)

Whether a stable elite profile for AI-assisted teams exists at all, given the archetypes are one wave old and the cluster analysis that once defined elite has already failed once, in 2022. Sources: cloud.google.com (2020) (↗); Octopus Deploy blog (2024) (↗)

Whether any credible RCT of AI-era delivery is still possible now that developers refuse control conditions, or whether the field is permanently limited to observational telemetry and survey correlation. Sources: METR (Substack) (2026) (↗)

The Monday takeaway holds either way: the five keys still measure the right things, AI has not changed which practices move them, and the teams winning in 2026 are the ones that got stricter about small batches and review exactly when their tools made it easiest to abandon both.