How to make Rovo useful in JSM: put the decision in the agent's prompt

To make Rovo useful in Jira Service Management, match your most common ticket keywords to knowledge base articles, then let the agent's prompt decide whether to answer or hand off to a person. That was our lesson: an automation rule that parsed the agent's output kept failing, and moving the decision into the prompt gave us the behavior we wanted in testing.

This post covers how we built it and what broke. It also covers what has changed since: the names, the credit costs, and what to settle before an agent answers real requests.

Start with your tickets, not the agent

We began by looking at the tickets we already had.

1. Analyze the tickets. We used Rovo to pull the 10 most frequent words in incoming tickets.
2. Pick the keywords. The top 5 were Data, Privacy, AI, Training, and Model. Yours will differ, and they should come from your own queue.
3. Link the articles. We created five Confluence articles, each with a header that exactly matched one keyword, so a ticket about privacy pointed to the Privacy article.
4. Find the gaps. We ran a gap analysis, wrote articles for missing topics, and labeled them.

Build one agent with one job

We created a Rovo agent called Ticket Assistant and kept its setup small:

• Instructions: analyze incoming tickets, identify keywords, provide the matching article, and stay polite. If a ticket mentions Data, provide the Data article.
• Knowledge: one Confluence space of knowledge base articles.
• Actions: add a comment to a work item, and assign a work item.

Why parsing the agent's output failed

Our first design kept the decision in the automation rule. The rule ran when a new ticket reached Waiting for support and sent the agent this prompt: find the article most likely to help, and rate your confidence from 1 to 10. Above 8, return STATUS:ready_to_go with the article URL. Otherwise, return STATUS:human_action_needed.

The rule then branched. On ready_to_go, it added a public comment with the article link. On human_action_needed, it commented that a person would take over and assigned the ticket to someone on the team.

To read the status, we used a smart value that split the agent's response and took the first word. It worked on the first run. Later runs failed with this error in the audit log: Failed to get value for agentResponse.split(' ')[0].

That error means the value was empty or null when the rule read it. It made little sense to us, because another rule showed the agent did return a string. We checked the conditions, actions, and syntax and never found the cause, so we changed the design.

What worked: move the decision into the prompt

The second rule used the same trigger, and we added these instructions to the same prompt:

• If you return ready_to_go, attach the document and answer the customer.
• If you return human_action_needed, tell the customer a person needs to take this.
• Be kind, professional, and knowledgeable, and do not reveal your return value or your internal thought process.

The rule no longer had to interpret anything. Two test tickets showed the difference:

• "I would like to know more about privacy inside my company." The agent replied with a summary and a link to the Privacy article.
• "Can I go on holidays on Christmas?" No article matched, so the agent said a person would follow up, and the ticket was assigned to one.

Test this in your own site first

Atlassian's pages differ on whether this design is available. Automate Rovo agents says an agent in an automation flow can act directly, unless an admin or user has prevented it. Add tools to Rovo agents says an agent in a flow cannot use its own tools and can only return text through the agentResponse smart value. So run one test ticket and check the audit log before you commit to either design.

A prompt is also not a control. Atlassian notes that natural language instructions can be interpreted differently from one conversation to another. Where a tool supports it, set fixed values in the tool's configuration. Raise a request, for example, lets you fix the service desk and request type.

What an agent run costs now

Agent runs spend Rovo credits. Atlassian's Rovo credits page sets the rates by feature:

• Basic Rovo agents (no model selection): 10 credits per billable event.
• Advanced Rovo agents (with model selection): variable credits.
• Self-service Agent: variable credits. More reasoning, search, and tools generally means more credits.
• Request Resolver: Atlassian's pages differ. The credits page lists it as variable, while the Request Resolver page gives a set number of credits per resolution plan. Check both before you budget.

Each invocation of an agent is a billable event, and the page makes no exception for a run that ends in a handoff. A rule that fires on every new ticket spends credits on every new ticket, so write the trigger and conditions narrowly. Atlassian's Rovo licensing FAQ adds that one rule can draw on two meters in a run: the agent step uses Rovo credits, and triggers, conditions, and standard actions use Automation steps.

The virtual service agent is documented separately. Its usage page counts assisted conversations, and the Rovo credits page does not list it. We could not confirm from Atlassian's pages whether its AI answers also draw Rovo credits. Run a test, then export the Rovo credits usage data, which shows feature, agent, and credits consumed.

Usage billing takes effect December 3, 2026 for most meters, and timing can vary by capability and billing cadence. Service Collection includes 250 credits per user per month on Standard, 700 on Premium, and 1,500 on Enterprise. Allowances pool across your organization, reset monthly, and do not roll over. Before December 3, check the extra usage setting on each meter. Our guide to usage billing walks through what to set.

Want to check the cost first? Run the Rovo ROI calculator, and an architect can review your numbers.

Current names for Studio, service agents, and help centers

We also created a customer-facing agent in Studio and activated it for customers in the help center, so they could get one-on-one help when submitting requests. Earlier versions of this post used older names for those pieces. Here is what Atlassian's pages say now.

Check which help center you have, because it changes the setup. Every help center is labeled Modern or Legacy. A Rovo agent can go on either, but the virtual service agent is only available in Legacy help centers.

What to decide before an agent answers requests

Settle these four things first. The full checklist is in how to approve and audit Rovo agents.

Who owns it

Name one owner and at least one manager. Managers can edit and delete an agent and add editors, and a manager can take ownership of an agent that has no owner. Creating agents is open to everyone by default until a Studio admin limits it. Our change control guide explains why an unowned agent is the failure to watch for: it keeps its permissions, and nobody wants to turn it off.

What knowledge it can read

Every agent has access to all organizational knowledge unless you configure it differently. Choose Custom knowledge, then narrow it to specific spaces or one branch of a Confluence content tree. Atlassian's pages differ on what the person asking can get back. The knowledge page says agents don't return information that person has no clearance to view. Its best practices page says an agent on its own account can share what it has access to with anyone who can use the agent. So check which account the agent uses before you rely on either. Our approval guide covers agents that run on their own account.

When it hands off to a person

Write the triggers down, then give the agent a way to act on them. For requests, Atlassian recommends its Jira Service Management request tool over the general create work item tool in a portal. Its pages call this tool both Raise a request and JSM Create Request. Also decide who receives the ticket and what the customer is told, because Automate Rovo agents says an agent can be configured to run without confirming each action.

How you check its answers

• Before launch: Evaluation takes a CSV of up to 100 prompts, with an optional column of expected responses. It offers Response accuracy, Resolution rate, and Manual testing.
• After launch: live conversation review lets agent creators and managers read transcripts and user feedback. End users see a privacy disclaimer. Transcripts expire, and that page gives both 30 and 28 days, so review early and confirm the period with Atlassian.
• On a schedule: we recommend a quarterly review for any agent with write tools, and a way to stop it that does not depend on the agent.

If you want this built with your team, our agent guardrails work sets the owner, scope, and off switch for each agent, and our service management work designs the request catalog and change control around it.

Next step

Join our free live webinar, The complete Rovo briefing for Atlassian teams, on Thursday, November 19, 2026, from 1:00 PM to 2:00 PM EST. It covers agents, usage-based pricing, and admin controls.