As AI assistants start working with Jira data, I'm curious how administrators are handling access and accountability. Do you restrict Claude to specific projects, keep it read-only by default, or give it the same permissions as the person using it? Do permissions vary by user group, and are prompts and actions logged centrally so they can be reviewed later? I'd especially like to hear how others are balancing productivity with the risks of accidental changes, poor ticket hygiene, or prompt injection.
4 Answers
A sensible baseline is read-only access, limited to the projects and issues the current user can already see. Any write or administrative actions should require an explicit approval step, and every action should be tied to the individual user and recorded in the audit trail.
Giving the assistant unrestricted access just to move quickly is risky. Permissions should follow the user’s existing access, with project and operation limits where possible. Read-only access is a good starting point, and elevated actions should be approved and logged.
We authenticate Claude separately for each user through the official integration, so all activity is attributed to a real person. The assistant is treated as a tool rather than an autonomous employee: users are expected to review its output, maintain good ticket hygiene, and accept responsibility for changes made under their account. I’d also account for prompt-injection risks, especially when the assistant can read untrusted ticket content.
We use delegated access, so Claude operates with the same permissions as the authenticated user. Actions are treated the same as if that person had used Jira directly, meaning the user is responsible for anything done under their account. We capture prompts centrally and send the records to our security monitoring system for review.
That approach also makes it easier to start small. You can expand access later once you understand the failure modes and have appropriate controls in place.