Rovo in Jira Service Management works when each agent has one job.
Atlas Bench, an Atlassian Platinum Solution Partner, helps service teams design Rovo agent workflows in Jira Service Management: we pick each agent's job, write the decision into its instructions, connect its knowledge, and test it in Rovo Studio. Our Rovo consulting puts approval and change control in place before any agent changes a work item, and our managed support keeps the agents tuned after launch.
Atlas Bench, an Atlassian Platinum Solution Partner, designs, builds, and supports Rovo agent workflows for service teams in Jira Service Management. Our Rovo consulting picks the job each agent does on your service desk, writes the decision into its instructions, and tests it before it touches a live request. After launch, we run and tune the agents as part of managed support.
The problem
Most service desks meet Rovo as a general helper: one agent, broad knowledge, and no written rule for when it answers and when it hands off. It handles some requests well and guesses at the rest, and nobody can say which decisions it may make on its own, or what it will cost once usage billing starts.
The fix is narrower than it sounds. Give each agent one job, write the decision it makes into its instructions, connect only the knowledge that job needs, and test it against your own requests before launch. Then decide who approves anything it changes, because Atlassian notes that guardrail instructions are suggestions, not guarantees.
What the work is
Jobs for the agent, jobs for people
We review your queue and pick the jobs an agent should own, such as triage and categorization, answers from your knowledge base, a summary of the work item for the next agent, or a draft reply or knowledge article update for a person to review. Approvals, exceptions, and anything that needs judgment stay with people.
Where each agent runs
Customers reach a Rovo agent or the virtual service agent in the help center or portal. The virtual service agent needs a Service Collection Premium or Enterprise plan, and a Rovo agent in a portal currently serves only customers with Atlassian accounts, so the choice depends on who your customers are. Your service desk agents get Service Request Helper and Resolve with Rovo in the work item view, and a Jira automation rule can call an agent with the Use agent action.
Instructions, knowledge, and tools
We write each agent's instructions so the decision is in the prompt: when to answer, when to hand off to a person, and what the customer is told. We connect only the knowledge sources the job needs, such as the Confluence spaces behind your knowledge base, and keep tools to the few the job uses, in line with Atlassian's advice of no more than five.
Identity and who can build
By default, anyone with access to Studio can create agents, so we agree with you who should and set it. Each agent runs on the user's account, with that person's permissions, or on its own agent's account, which reaches only the spaces and content it has been granted. The agent's account needs your organization admin to set default access for agent accounts first; until then, agents use the user's account. We choose per agent and grant only what its job needs.
Testing and approval
We test each agent with evaluations in Rovo Studio: a CSV of up to 100 prompts drawn from your own requests, scored for response accuracy or resolution rate, or reviewed by hand. For agents called from automation, the setting that lets an agent create, edit, or delete content stays off until the owner approves those actions through your change control.
Rovo credits and usage
Each agent invocation is a billable event in Rovo credits. A Use agent step draws Rovo credits instead of an automation step, and the rule's other steps still count toward your automation allowance. We estimate credits per job before launch and write narrow triggers, so an agent runs only on the requests it is meant for.
How it runs
-
Design
2 to 3 weeksQueue review, job selection, where each agent runs, the account each agent uses, and a credit estimate, approved by the service owner before anything is built.
-
Build and test
3 to 5 weeksInstructions written, knowledge and tools connected, evaluations run against your own requests, and any write actions approved through change control before launch.
-
Run and tune
Ongoing, under managed supportWe rerun evaluations after each change, adjust instructions and knowledge as your services change, track credit use against your allowance, and retire agents that no longer earn their place.
What you get
- A job map for your service desk: what each agent owns and what stays with people
- Instructions for each agent with the answer or hand-off decision written in, plus its knowledge sources and tools
- An identity decision per agent (user's account or agent's account), the spaces it can reach, and your Studio creation setting
- An evaluation set of up to 100 prompts per agent, with results recorded before launch
- An approval path in your change control for any agent that changes a work item
- A Rovo credit estimate per agent, and a managed support plan for tuning after launch