Requirements Traceability Matrix for AI Software Delivery—From Issue to Pull Request

Your issue tracker and version control platform hold most of the requirements traceability matrix an auditor wants, in a form nobody in your org can query.

On a Tuesday, an auditor asks a simple question. Which requirement produced the code in release 4.2? Who verified that it worked? Instead of answering from Jira and GitHub, someone spends two days reconstructing the story from merged pull requests and old release notes. It’ll still be a best guess, delivered with total confidence.

You need a requirements traceability matrix. It connects each requirement to the work that implements it, then traces every delivered artifact back to the reason it exists. Its purpose is to make the auditor’s question boring.

In most engineering orgs, it does the opposite, because someone has to keep it current by hand. That manual work doesn’t scale, and nobody was ever promoted for updating it. Meanwhile, your development workflow is already creating the links the matrix needs, inside systems you already pay for, where nobody can query them.

What the Matrix Was Supposed to Do

Image generated with AI

The job of the matrix is to connect the reason for a change to the issue and pull request that carried it. It also ties that work to the test run and approval that checked it. It answers two questions. Given a requirement, what proves it shipped? Given a diff, what business reason justifies it?

That first question matters outside an audit. A team that can trace a requirement to the work behind it can understand unfamiliar code faster.

But the matrix usually gets treated as audit paperwork. When that happens, it never earns budget. It runs on whatever time is left over, and that time keeps failing to materialize.

The research says underfunding traceability costs more than it saves. A 2015 controlled experiment gave participants real maintenance tasks on unfamiliar projects. Some tasks included traceability links between requirements and code. Participants who could follow those links finished 24% faster on average and produced 50% more correct solutions.

Faster work frees capacity, and capacity determines what your team can take on. Traceability, it turns out, is a productivity feature wearing a compliance costume.

The Chain Your Workflow Leaves Behind

Set the hand-kept matrix aside and trace a change through your org. Each handoff leaves a record whether anybody wants it to or not:

• Requirements arrive from customer commitments or regulations.
• A Jira or Azure DevOps issue names the requirement.
• The implementation plan lives in the issue description or a linked design doc.
• Commits on a branch reference the issue by number.
• CI runs unit tests and acceptance tests against those commits.
• A reviewer approves or rejects the pull request, on the record, with a timestamp.
• The pull request merges and, closing keyword satisfied, closes the issue it named.

Requirement to issue to plan to code to tests to review to merge.

How much of that chain your platforms hold varies by team. One team writes fixes #4127 in every commit. The next puts PROJ-4127 in the pull request title. A third uses only a branch name like adam/auth-timeout-fix, because its tech lead is confident the branch name speaks for itself. It doesn’t.

Those references don’t join unless your teams use the same pattern. GitHub accepts several closing keywords, for example. Picking the one every team uses and enforcing it with a required status check is this quarter’s decision. It’s a boring decision, which is why it keeps not getting made.

Agent Volume Breaks the Hand-Kept Matrix

Now put coding agents into that workflow. The number of branches and pull requests rises, while human memory of why each change happened stays thin. The hand-kept matrix stops getting updated because somebody has to open it and add the link. That somebody has a release to ship, and the matrix loses that argument every single time.

Agent-authored change puts three questions on your desk. Who reviewed the work? Where is the agent’s implementation plan stored? What Git author or pull request author identifies the agent and the human who directed it? A shared bot account answers none of this, though it does look tidy in the commit log.

The honest objection is that this chain proves custody, not correctness. Commit text like fixes #4127 proves only that somebody typed a number. Tests and review still carry the correctness question. GitLab, for instance, removes existing approvals when new commits land on the source branch by default, so an agent push resets them.

Traceability Check: Pick one merged pull request. Working from Jira and GitHub alone, name the requirement it served. Then name the person who approved it. If you have to ask a developer, that record does not exist.

Your CI Retention Expires the Evidence on a Timer

Agent volume is not the only clock running on the evidence chain from issue to merge. If you ship a high-risk AI system into the EU market, Annex IV of the EU AI Act requires technical documentation carrying test logs and reports dated and signed by the responsible persons, and Article 11 requires it stay current.

Staying current under a rule like that assumes the evidence still exists. Test runs and approvals are worth only what your platforms keep. Most hosted CI platforms delete workflow logs and test reports after a default retention period nobody in your org chose.

GitHub is the worked example. It retains workflow artifacts and logs for 90 days by default, adjustable up to 400 days on private repositories. The JUnit report and workflow log for a March release can be gone by June unless somebody changed a setting that has never once come up in a planning meeting.

Months later, your biggest customer’s security team asks for the verification record of one release. The issue and pull request are still there. The test run was deleted 90 days after it passed, and the reviewer who approved the code has left. You’re attesting to work you can no longer show, which is a memorable position to negotiate a renewal from.

Traceability Is a Read Problem

Retention keeps the evidence alive; finding it is a different problem. Most traceability programs fail on the write side, which your workflow already handles. What your org lacks is a read path.

One query should answer: show me requirement REQ-128, the issue that implemented it and the pull request that merged it. Then show the tests that passed and the reviewer who approved it.

Software supply chains already treat machine-emitted records as proof. SLSA provenance makes the build platform record how an artifact was made. Human-written claims about a build stay claims; machine-emitted ones carry their own proof.

Your workflow leaves the same kind of records behind. Building the read path means a scheduled job that calls the Jira and GitHub APIs, then stores issue IDs and approval timestamps.

Raise your CI artifact retention window this week, before you scope that job. On a private repository, that one setting buys 310 more days of evidence, which is the best return on a mouse click you’ll get this quarter.

Explore Progress Forge Orchestration

Turn AI-assisted development into a repeatable engineering process by orchestrating the coding agents you already use through structured, configurable development workflows—from work item to pull request.

Your coding agents execute the work. You own the process. Forge orchestrates it.

Try Forge Now