Written by: Mark Hull, Co-Founder and CEO, Exceeds AI
Key Takeaways for AI-Heavy Engineering Orgs
- Most leaders lack clear visibility into whether AI coding investments deliver measurable ROI across speed, quality, and risk.
- Traditional DORA metrics and metadata-only platforms cannot distinguish AI-generated code from human-written code or tie outcomes to specific tools and interaction modes.
- The five-layer metrics model extends DORA with AI adoption tracking, commit-level AI vs. human outcome analytics, rework pattern detection, and 30-plus-day longitudinal incident monitoring.
- Exceeds Ink provides a portable, line-level provenance layer that records tool, model, session, and interaction mode for every commit across Cursor, Claude Code, Codex, GitHub Copilot, and Windsurf.
- Exceeds AI delivers all five layers with repo access; connect your repo today to start a free pilot and prove AI ROI in weeks.
Why DORA Alone Is Insufficient for AI Teams
Classic DORA metrics, including deployment frequency, lead time for changes, change failure rate, and mean time to restore, measure throughput and stability in human-authored delivery pipelines. They remain useful baselines for any engineering organization. They no longer cover what leaders need to know about AI-assisted development.
The core gap is attribution. DORA metrics capture that a PR merged in four hours, but they cannot tell you whether 600 of the 800 lines in that PR were generated by Cursor in agent mode. That distinction matters because AI-generated code may require extra review iterations or introduce defects that surface 45 days later, outcomes that merge-time metrics cannot predict. Without that attribution, leaders cannot connect AI adoption to delivery outcomes.
The span-of-control problem compounds this attribution gap. Microsoft’s 2008 ICSE study found organizational-complexity metrics, including team size and management span, to be among the strongest predictors of defect-proneness. As manager-to-IC ratios stretch from the typical 1:5 toward 1:8 or higher, bandwidth for code review and mentorship shrinks. AI tools accelerate output into that gap, and without code-level observability, the additional risk stays invisible.
Layers 2 through 5 of the five-layer model close this gap by connecting AI adoption rates, commit-level outcomes, rework patterns, and longitudinal incident data to the delivery signals DORA already tracks. DORA becomes the floor, not the ceiling. The next section breaks down each of those five layers and explains how they work together.
Start tracking AI attribution in your delivery pipeline
The Five-Layer Metrics Model for AI Delivery
Software delivery performance analytics for AI-assisted development teams relies on five distinct measurement layers. Each layer depends on the one below it. Only Layers 3 through 5 require repo access, and only repo access makes those layers possible.
Layer 1 – Classic DORA Metrics. Deployment frequency, lead time, change failure rate, and MTTR provide necessary context for throughput and stability. They remain essential but are insufficient on their own for AI-heavy teams.
Layer 2 – AI Adoption Rates Across Tools. This layer shows which teams use which tools, and at what frequency, across Cursor, Claude Code, Codex, GitHub Copilot, and Windsurf. It provides adoption context, but adoption rates without outcome data are vanity metrics, which is why the next layer focuses on results.
Layer 3 – AI vs. Non-AI Outcome Analytics at Commit and PR Level. This layer compares cycle time, review iterations, and defect density between AI-touched and human-authored code at the line level. Leaders see where AI improves throughput without harming quality and where it quietly increases risk. This is where ROI proof begins.
Layer 4 – Review and Rework Patterns. This layer reveals whether AI-generated lines require more review iterations and whether specific tools or interaction modes, such as agent, edit, or plan, correlate with higher rework. Many agent-assisted pull requests require human revisions to align with project-specific standards. Layer 4 identifies which patterns drive that need up or down.
Layer 5 – 30-Plus-Day Longitudinal Incident and Maintainability Tracking. This layer answers whether AI-touched code that passes review later causes incidents at 30, 60, or 90 days and whether it accumulates follow-on edits. It anchors those outcomes to Exceeds Ink’s per-commit attestation, so leaders can see which tools and interaction modes create durable code and which quietly accumulate debt.
See the full five-layer model on your own repos
How to Track AI Technical Debt Over Time
AI technical debt differs from traditional technical debt because it often stays invisible at the moment it is created. Developers report that generating code that appears correct but is unreliable is a common negative effect of AI, creating hidden defects and false confidence in code quality. Code passes review, ships to production, and fails later.
A study of 806 open-source repositories that adopted Cursor AI found a 41% increase in code complexity and a 30% increase in static-analysis warnings, with only transient velocity gains while the debt persists. The same research identifies a self-reinforcing feedback loop where accumulated debt later reduces development velocity.
Layer 5 longitudinal tracking breaks this cycle. Exceeds Ink writes a structured Git Note alongside every commit that records which tool, which model, which interaction mode, and which lines were involved. That attestation becomes the anchor for long-term analysis.
Thirty days after merge, Exceeds AI correlates incident rates, follow-on edit frequency, and test coverage changes back to the specific commits and tools that produced the code. Metadata-only tools cannot perform this analysis. They see merge status but not what happens to the code afterward, and they cannot trace an incident back to a Cursor agent-mode session from six weeks prior. Exceeds Ink’s per-commit attestation makes that connection possible.
See which AI-generated code is accumulating debt in your repos
What Multi-Tool AI Coding Analytics Actually Requires
Engineering teams in 2026 rarely rely on a single AI coding tool. They use Cursor for feature development, Claude Code for large-scale refactoring, Codex for batch transforms and headless workflows, GitHub Copilot for inline autocomplete, and Windsurf for specialized tasks. Leaders lack aggregate visibility into the combined impact of this toolchain when they depend on any single vendor’s telemetry.
Heuristic and watermark-based detection, the approach most platforms rely on, tops out around 20–25% accuracy. It cannot identify which tool wrote a given block, what interaction mode the engineer used, or how many tokens the session consumed. It guesses and leaves leaders with noisy, incomplete data.
Exceeds Ink uses per-tool checkpoint materializers for Claude Code, Cursor, and Codex that resolve edit evidence against the actual working tree at commit finalization. Every line carries its tool, model, session, interaction mode, and timestamp. The attestation is written as a Git Note at refs/notes/exceeds-ink, which stays portable, machine-readable, and stored in your own repo rather than in a proprietary cloud.
Jellyfish, LinearB, and Swarmia were built before multi-tool AI adoption existed. They track metadata such as PR cycle times, commit volumes, and review latency, and remain fundamentally blind to which lines are AI-generated and by which tool. DX adds survey data and an AI Code Insights module captured by a closed-source daemon, but attribution lives in DX Data Cloud rather than your repo, and the weakest capture tier falls back to filesystem-change heuristics. None of these platforms deliver authoritative, portable attribution across the full AI toolchain.
Get unified visibility across your entire AI toolchain
How to Turn Metrics into Coaching Actions
Metrics only matter when they drive behavior change. Exceeds AI converts Layer 3 through 5 data into two concrete outputs: Coaching Surfaces for managers and ink-prompting-coach delivered inside the developer’s own Claude Code or Cursor agent.
Coaching Surfaces identify which engineers use AI effectively, combining stable quality with productivity gains, and which struggle with high rework rates or disruptive context switching. This identification is possible because Ink’s interaction-mode classification, including plan, ask, agent, edit, and headless, provides the underlying signal. For example, a manager can see that spiky, agent-mode-heavy commits without a plan phase correlate with elevated rework, then act on that specific pattern rather than guessing at the root cause.
ink-prompting-coach installs as a SKILL.md and slash command directly into the developer’s own agent. Coaching appears where the work happens, inside the tools engineers already use. When a pattern that works on one team is identified, Skill Transfer distributes it as a versioned skill across the organization, with rollback available if it does not land.
The result is a behavior-change pipeline, not just a measurement pipeline. Turn AI metrics into targeted coaching for your teams
Implementation Checklist for Exceeds AI
- GitHub or GitLab OAuth authorization — 5 minutes
- Per-repo opt-in and scoping — 15 minutes
- Exceeds Ink install on developer machines — same day
- First insights visible in the platform — within 60 minutes
- Complete historical analysis — within 4 hours
- Real-time updates after new commits — within 5 minutes
- Board-ready ROI report — within weeks
No global git config mutation. No long-lived daemon. No PATH-shimmed git binary. Per-repo opt-in only. Set up Exceeds AI and see insights in under an hour
Conclusion: Turning AI Coding into Measurable Performance
As established earlier, traditional DORA metrics cannot prove AI ROI or identify hidden technical debt on their own. The five-layer model described above is the only approach that connects AI adoption to delivery outcomes across the full multi-tool reality of 2026.
Exceeds AI is the only platform that delivers all five layers, from classic DORA baselines and AI adoption rates across every tool to commit and PR-level AI vs. human outcome analytics, rework pattern detection, and 30-plus-day longitudinal incident tracking. Setup is fast, insights are immediate, and ROI proof follows quickly.
Prove the impact of AI on your software delivery performance
Frequently Asked Questions
What is the difference between software delivery performance analytics and traditional DORA metrics for AI teams?
Traditional DORA metrics measure delivery throughput and stability but cannot attribute outcomes to AI usage. Software delivery performance analytics for AI-assisted teams extends DORA with four additional layers that connect AI adoption to delivery outcomes. These layers cover AI adoption rates across tools, commit and PR-level AI vs. human comparisons, review and rework pattern analysis, and longitudinal incident tracking tied to specific AI-generated commits. This extended model catches AI technical debt before it becomes a crisis.
How does Exceeds AI measure AI technical debt at the commit level?
Exceeds Ink writes a structured attestation as a Git Note alongside every commit. That note records which AI tool produced which lines, in which interaction mode, during which session. After merge, Exceeds AI monitors those attested lines over time, tracking incident rates, follow-on edit frequency, test coverage changes, and rework patterns. When AI-generated code that passed initial review begins accumulating follow-on edits or correlates with production incidents, the platform surfaces that signal and traces it back to the specific tool, session, and interaction mode that produced it. This longitudinal tracking requires repo access and a reliable per-commit provenance layer. Metadata-only tools see merge status but cannot see what happens to the code afterward.
Why cannot existing tools like Jellyfish, LinearB, or Swarmia track AI vs. human code outcomes?
Jellyfish, LinearB, and Swarmia were built before multi-tool AI adoption existed. Their data models rely on metadata such as PR cycle times, commit volumes, review latency, and deployment events. Metadata describes what happened at the process level but does not record which lines were written by a human and which were generated by Cursor in agent mode. Without that line-level distinction, these platforms cannot compare AI-touched PR outcomes to human-authored PR outcomes, cannot identify which AI tools drive better results, and cannot track whether AI-generated code accumulates debt over time. This represents a category gap rather than a feature gap. Closing it requires repo access and a provenance layer that captures AI authorship at the moment of creation, which Exceeds Ink provides.
How does Exceeds AI handle teams using multiple AI coding tools simultaneously?
Exceeds Ink uses per-tool checkpoint materializers for Claude Code, Cursor, and Codex that resolve edit evidence against the actual working tree at commit finalization. For GitHub Copilot and Windsurf, dedicated adapters capture authorship signals. Lighter-weight detection covers up to approximately 50 additional AI tools. Every attested line carries its tool, model, session, interaction mode, and timestamp. Leaders gain aggregate AI impact across the entire toolchain, tool-by-tool outcome comparisons, and team-by-team adoption patterns from a single platform. The attestation is stored as a portable Git Note in your own repo, not in a proprietary cloud, so it remains auditable by anyone with repo access and survives outside the Exceeds platform.
How quickly can an engineering team get actionable insights from Exceeds AI?
GitHub or GitLab authorization takes approximately five minutes. Per-repo opt-in and scoping takes fifteen minutes. Exceeds Ink installs on developer machines the same day. First insights are visible in the platform within 60 minutes of setup. Complete historical analysis, covering the full commit history of selected repositories, completes within four hours. Real-time updates appear within five minutes of new commits. Board-ready ROI reports are typically available within weeks. This timeline contrasts with metadata-only platforms that often require months of onboarding before delivering meaningful signal, and with Jellyfish, which averages approximately nine months before showing ROI.