Written by: Mark Hull, Co-Founder and CEO, Exceeds AI | Last updated: July 30, 2026
Key Takeaways for AI-Aware DORA Measurement
- DORA metrics still anchor software delivery performance, yet they cannot separate AI-generated outcomes from human-authored work.
- AI coding tools now generate over 50% of production code, while traditional DORA dashboards miss AI-driven technical debt, delayed incidents, and rework inflation.
- Metadata-only tools remain blind to line-level AI attribution, interaction-mode classification, and long-term outcome tracking that distinguish throughput gains from quality erosion.
- AI-impact observability platforms join DORA delivery metrics with commit-level AI provenance to create portable attestations that survive forks and support governance-ready reporting.
- Exceeds AI defines this category, using Exceeds Ink’s lightweight Git Notes attestation layer to deliver first insights within 60 minutes; start your free pilot today.
The Problem: AI Blind Spots That Break DORA
AI coding tools now sit in nearly every engineering workflow. Eighty-four percent of developers are using or planning to use AI tools, and 51% of professional developers use them daily. Earlier statistics show that AI now generates more than half of production code. The volume is real. The measurement infrastructure has not kept pace.
Attribution sits at the center of the problem. DORA metrics track outcomes at the deployment and pipeline level, yet they cannot reveal whether a fast lead time reflects a skilled engineer or an AI tool that produced 600 lines in 90 seconds. DORA cannot distinguish whether improved deployment frequency results from AI-generated code or whether rising change failure rates stem from harder-to-review AI code. That attribution gap creates direct, measurable consequences.
Rework inflation appears first. Seventy-four percent of organizations report that at least 25% of AI-generated code requires significant post-deployment rework due to poor context, incomplete data, or flawed system assumptions. This rework burden translates directly to token economics. Of every dollar spent on AI tokens, $0.44 goes to fixing bugs introduced by AI, $0.27 to rewriting or reworking AI-generated code, and $0.11 to review friction and merge overhead, leaving only $0.18 as shipped product value.
Incident rates then rise on a delay that standard DORA windows never capture. Organizations report spikes in production incidents tied directly to AI-generated code and have suffered major production failures caused by AI code. Frequent users of AI coding tools report that AI-generated code often appears in deployment issues. These incidents surface 30, 60, or 90 days after merge, well outside the window any standard DORA dashboard monitors.
Technical debt accumulation follows a structural pattern. A large-scale empirical study analyzing AI-authored commits from 6,299 GitHub repositories identified a large number of distinct issues introduced by AI-generated code, most of which were code smells with a substantial portion surviving in the latest repository version. A study of 8.1 million pull requests confirmed that technical debt increases 30% to 41% after AI tool adoption. GitClear’s 2026 longitudinal study of 211 million changed lines found two-week churn rose from 3.1% to 5.7%, refactoring fell from 25% to under 10%, and cloned code rose from 8.3% to 12.3% between 2020 and 2024. Together, these findings show AI-driven debt building quietly behind apparently healthy delivery metrics.
Metadata-only tools, which track PR cycle times, commit volumes, and review latency without reading code diffs, cannot detect any of these patterns. They see that PR #1523 merged in four hours. They cannot see that 623 of the 847 changed lines were AI-generated, that those lines required extra review iterations, or that the AI-touched module caused two incidents 45 days later. The measurement gap is not a feature gap. It is a category gap.
The Solution: AI-Impact Observability for DORA Teams
Closing this category gap requires a measurement layer that can see inside the code while preserving existing delivery metrics. The required augmentation layer combines DORA’s delivery metrics with code-level AI provenance. This category, AI-impact observability platforms, preserves every DORA signal while adding line-level AI attribution, interaction-mode classification, and longitudinal outcome tracking anchored to commit-time attestations. The result is a system that separates AI-driven throughput gains from AI-driven quality erosion and connects both to business outcomes.
Exceeds AI is the platform built for this category. Powered by Exceeds Ink, a lightweight on-machine provenance layer that writes a portable, line-level attestation alongside every commit as a Git Note, Exceeds delivers commit and PR-level fidelity across Cursor, Claude Code, Codex, GitHub Copilot, and Windsurf. Setup takes hours, not months. First insights are available within 60 minutes.
See line-level AI attribution in your own repos—start your free pilot.
Why Traditional DORA Measurement Breaks with AI
The 2024 DORA report found that a 25% increase in AI adoption was associated with a 1.5% drop in delivery throughput and a 7.2% drop in delivery stability. That pattern runs opposite to what adoption dashboards usually show. The divergence appears because AI inflates the volume metrics DORA measures while degrading the quality signals DORA does not track.
DORA’s Lead Time for Changes metric drops when AI generates code in seconds, yet a shorter lead time may simply mean a developer accepted an AI suggestion without thorough review rather than indicating a healthy CI/CD pipeline. DORA’s Change Failure Rate may temporarily appear stable while rising code churn shows that AI-generated code is being quietly rewritten or reverted within days of merge.
Engineering leaders report that their current metrics often fail to reflect AI’s impact. Tech debt, validation time, and developer burnout rarely appear in those dashboards. Few leaders believe existing frameworks can fully address the gaps. Two teams can show identical DORA scores while one accumulates AI technical debt at a rate that will stall delivery within 18 months.
What Metadata-Only Engineering Tools Cannot See
Platforms that operate on metadata alone, such as PR cycle time, commit volume, and review latency, remain structurally blind to three critical signals.
- Line-level attribution. Without reading code diffs, no tool can determine which specific lines are AI-generated versus human-authored. Heuristic and watermark-based detection tops out around 20–25% accuracy. The only authoritative method uses client-level capture at the moment the work happens, observing what the engineer typed into the AI tool, how long they iterated, and which model produced the output.
- Interaction-mode classification. Whether an engineer used plan mode, agent mode, or headless mode with Claude Code or Cursor never appears in metadata dashboards. That signal forms the foundation for actionable coaching. Agent-mode commits without a plan phase represent a coachable pattern. Headless batch transforms carry different risk profiles than interactive edits.
- Longitudinal outcome tracking. AI-generated code often feels easier to understand yet harder to trust in production. Tracking whether AI-touched code causes incidents 30, 60, or 90 days after merge requires per-commit attestation that persists in the repository, not a metadata snapshot that expires with the sprint.
Core Technical Capabilities for AI-Impact Augmentation
An effective AI-impact observability platform relies on three technical capabilities working together.
Line-level AI attribution means every line in every commit carries a record of the tool, model, session, and turn that produced it. This record is written as a structured Git Note at refs/notes/exceeds-ink, machine-readable JSON that lives in the repository, travels across forks and mirrors, and remains auditable by anyone with repo access. Lines that cannot be confidently attributed are recorded as unknown_lines, not silently rolled into “human” or “AI.” This conservative approach makes the attestation trustworthy for governance, legal review, and patent examination.
Interaction-mode classification captures whether the engineer worked in plan, ask, agent, edit, or headless mode. This signal enables prescriptive coaching instead of static dashboards. This matters because knowing that AI code has high rework rates is useless without understanding why, and the why often sits in how the engineer used the tool. When Exceeds identifies that a team’s high-rework commits cluster in agent mode without a preceding plan phase, that pattern can be distributed as a versioned coaching skill directly into engineers’ Claude Code or Cursor agents through ink-prompting-coach.
Longitudinal outcome tracking monitors AI-attested code over 30-plus days for incident rates, rework patterns, and maintainability issues. Because every line carries a commit-time attestation, Exceeds can answer specific questions. Did the AI-touched code in PR #1523 cause incidents? Did it require follow-on edits? Did it have lower test coverage than the human-authored lines in the same PR? This mechanism surfaces hidden AI technical debt before it becomes a production crisis.
Track AI outcomes over time in your codebase—start your free pilot.
Practical Security and Deployment Details
Repo access enables code-level AI analysis and raises the primary security question for engineering leaders. Exceeds Ink is designed to pass enterprise security review by minimizing its attack surface and operational footprint. It is a single approximately 10 MB Rust binary with no Node or npm runtime dependency, which removes the supply-chain risks of large package ecosystems. It runs as short-lived hook processes invoked by standard Git hooks, with no long-lived daemon that could be exploited, no PATH-shimmed git binary that could intercept commands, and no global git config mutation that could affect other repositories. Per-repo opt-in ensures no engineer’s machine is affected without explicit enrollment.
Remote ingest uses HMAC-SHA256 signing with revocable per-machine tokens. Prompt content passes through an LLM-based redaction layer before persistence, which catches context-sensitive secrets that entropy-only redactors miss. An aggregate-only mode, enabled by a single environment variable, keeps transcripts off the wire entirely. Deployment options include local-only, where nothing leaves the machine, a self-hosted collector with bring-your-own Postgres, or Exceeds-hosted. Each option uses the same binary, the same Git Notes, and the same dashboards.
Regulated environments can use an in-SCM analysis option where analysis runs within the customer’s own infrastructure with no external data transfer. The platform has passed formal enterprise security reviews, including a Fortune 500 retailer’s two-month evaluation process.
How to Evaluate AI-Impact Observability Platforms
Engineering leaders at 50–1,000-engineer organizations can use the following dimensions to evaluate AI-impact observability platforms.
- Attribution method. Does the platform use client-level capture or heuristics? Heuristics remain fundamentally unreliable for attribution. Only client-level capture, which observes the AI tool on the engineer’s machine at commit time, produces authoritative attestation.
- Provenance portability. Does the attestation live in your repository or in a proprietary cloud? Git Notes attestations remain portable across forks and mirrors and survive outside any vendor’s platform. Proprietary cloud-only metadata creates vendor lock-in by design.
- Multi-tool coverage. Teams in 2026 use Cursor, Claude Code, Codex, GitHub Copilot, Windsurf, and others concurrently. A platform with adapters for only one vendor’s telemetry produces an incomplete picture.
- Longitudinal tracking. Can the platform monitor AI-attested code for incidents and rework 30, 60, and 90 days after merge? Snapshot-only tools miss the delayed quality signal entirely.
- Operational footprint. Does the capture layer require a long-lived daemon, a PATH-shimmed git binary, or a global git config mutation? Each of these creates operational drag, CISO friction, and fleet management overhead that a hook-direct architecture avoids.
- Behavior-change capability. Does the platform distribute coaching into the developer’s own AI agent, or does it stop at a dashboard? Measurement without actionability leaves managers staring at patterns they cannot act on.
- Pricing model. Per-contributor pricing penalizes team growth. Outcome-aligned pricing tied to manager seats aligns vendor incentives with engineering outcomes.
Frequently Asked Questions
Do DORA metrics still matter in 2026?
DORA metrics still provide the foundational framework for measuring software delivery performance. Deployment Frequency, Lead Time for Changes, Change Failure Rate, Failed Deployment Recovery Time, and Deployment Rework Rate give essential baselines for delivery speed and stability. The problem does not sit with DORA itself. DORA was designed before AI coding tools existed and cannot attribute outcomes to AI versus human code. Engineering leaders now need DORA plus an AI provenance layer, not a replacement for DORA.
Why can’t GitHub Copilot Analytics or similar vendor dashboards fill this gap?
Vendor-provided analytics such as GitHub Copilot Analytics report usage statistics like acceptance rates, lines suggested, and active users, yet they cannot prove business outcomes. They do not show whether Copilot-touched code has higher incident rates, more follow-on edits, or lower test coverage than human-authored code. They also remain blind to every other AI tool. If your team uses Cursor, Claude Code, or Windsurf alongside Copilot, those contributions never appear in Copilot Analytics. An AI-impact observability platform must stay tool-agnostic and outcome-connected to answer the ROI question.
How does Exceeds AI differ from metadata-only platforms like Jellyfish, LinearB, or Swarmia?
Jellyfish, LinearB, and Swarmia were built for the pre-AI era. They track metadata such as PR cycle times, commit volumes, and review latency, yet they cannot read code diffs to distinguish AI-generated from human-authored lines. That limitation prevents them from attributing outcomes to AI usage, detecting AI technical debt accumulation, or proving whether AI investments improve or degrade quality. Exceeds AI analyzes actual code diffs at the commit and PR level, powered by Exceeds Ink’s line-level attestation, to connect AI adoption directly to productivity and quality outcomes. Exceeds is designed to sit alongside these tools as the AI intelligence layer, not replace them.
What makes Exceeds Ink’s attestation more trustworthy than heuristic detection?
Heuristic and watermark-based detection, which looks at code patterns, commit timing, or vendor-inserted markers, tops out around 20–25% accuracy by Exceeds’ own assessment. It cannot reveal how the engineer worked, such as whether they had a thoughtful back-and-forth with the AI or simply said “build this for me.” Exceeds Ink uses client-level capture. It observes what actually happens on the engineer’s machine at the moment the work occurs, records the tool, model, session, turn, and interaction mode, and writes a structured attestation as a Git Note at commit finalization. That attestation remains conservative, recording lines that cannot be confidently attributed as unknown rather than silently assigning them to AI or human, which makes it suitable for governance, legal review, and board reporting.
How quickly can an engineering team get value from Exceeds AI?
Setup requires GitHub, GitLab, or Azure DevOps authorization plus a per-machine Exceeds Ink install. As mentioned earlier, first insights appear within 60 minutes. Complete historical analysis finishes within four hours, and real-time updates appear within five minutes of new commits. The Pro plan starts at $49 per manager per month with no per-contributor data tax, so pricing ties to manager seats and the insights used, not the number of engineers analyzed. A free seven-day pilot is available with no enterprise sales gate required.
Conclusion: Make DORA Honest About AI
DORA metrics remain valuable yet incomplete. AI coding tools have created a structural measurement gap where throughput metrics improve while quality signals degrade, and the degradation surfaces 30 to 90 days after merge, long after any standard DORA dashboard has moved on. The single biggest challenge with AI coding tools, named by 57% of tech leaders, is AI integration. Metadata-only platforms cannot close that gap because they cannot read the code.
Engineering teams now need an AI-impact observability platform that combines DORA’s delivery signals with commit-level AI provenance. Exceeds AI delivers portable, Git-Notes-based, per-tool attestations at commit time across Cursor, Claude Code, Codex, GitHub Copilot, and Windsurf, while keeping the operational footprint minimal. Exceeds Ink’s attestations live in your repository, travel with your code, and remain auditable by anyone with repo access. The platform connects those attestations to longitudinal outcome tracking, interaction-mode coaching, and board-ready ROI reporting. Setup takes hours. Continuing to report DORA numbers without AI-aware provenance means flying blind on AI code quality.
Stop flying blind on AI code quality—start your free pilot today.