R&D reporting is slow because the evidence is scattered across commits, pull requests, issues and people's memories. It is usually pieced together months after the work, when details have been forgotten.
AI agents are good at the gathering, cross-referencing and first-draft parts of this job. They are not good at deciding what is true, so a human must stay accountable for the claim.
Point an agent at your repositories and have it collect, for each core activity:
Ask it to keep every PR, commit and issue link, so each statement in the write-up can be checked.
Reasoning often never reaches a commit message or PR. Point the agent at your past AI chats and coding-agent sessions to recover:
Many coding agents keep session transcripts on disk, so they can be searched.
Chat history is a lead, not proof. Confirm each decision against a PR, issue or architecture decision record, and cite that instead.
Check transcripts for secrets and customer data before sharing them or quoting them in a claim.
Figure: Use chat history to find the evidence, not as the evidence
Have the agent draft each core activity in the structure the claim needs:
Start from the core activities your team defined. See Do you define your core activities for R&D? and Do you know the three parts of R&D work?.
Use a separate agent, with fresh context, to check the draft against the R&D criteria and verify that every citation supports the sentence it is attached to. An agent that wrote the draft will not reliably find its own mistakes.
AI must not make the claim stronger than the record supports.
R&D work done with AI tools should say so. Use a consistent marker, such as a commit trailer or a footer in the PR description, so the number of AI-assisted pull requests can be counted accurately.
Check your counts against the full source. If your tooling truncates long PR bodies, it can cut off the footer and understate the figure.