
Rovo is designed for transparency
Our no b.s. commitment to open communication, accountability, and helping teams to use AI responsibly.
Rovo Search, Chats, Agents, and Skills
Select a Rovo feature below to get a transparent look at use cases and data use.
Cómo usan los Agentes Rovo la IA
Rovo Agents are powered by large language models developed by OpenAI and Google, as well as a combination of open-source large language models (including the Llama series) and other machine learning models. These models include OpenAI’s GPT series of models and Google’s Gemini series of models.
Rovo Agents use these models to provide functionalities to analyze and generate responses to prompts in natural language, and provide relevant responses from Atlassian and connected third-party products.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text, based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models and Google’s models. For more information on open-source language models, see information on the Llama series.
Acerca de los Agentes Rovo
With Rovo, we provide a number of out-of-the-box Agents, ready to use for a variety of tasks like, helping with decisionmaking, publishing knowledge documentation, and tidying up or organizing work items in Jira. Agents are specialized AI teammates that can assist human teams with moving work forward quickly and effectively.
You can:
- Create your own Agents using natural language and customize their instructions, knowledge, and actions
- Create your own Agents using code (Forge) to have more advanced integrations with your organization’s internal tools
- Install Agents via apps by developed by third-party Marketplace developers, customers, and Atlassian partners, which are typically more sophisticated Agents created using code (Forge).
We believe that Rovo Agents work best in scenarios where:
- A repetitive or formulaic task that would be easily automated through an Agent. Read more about automations
- Generating ideas or draft content to work from
- Performing fast searches for specific information
Limitaciones y consideraciones a la hora de utilizar Agentes Rovo
It’s important to remember that because of the way that the models used to power Rovo Agents work, these models can sometimes behave in ways that are inaccurate, incomplete or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We’ve found that Rovo Agents are less useful in scenarios where:
- You need current and accurate information about people, places and facts.
- You need information from your workspace that you don’t have access to within your workspace.
- You need very consistent and regular output that is not subject to any variation as answers can vary for a number of reasons including user permissions.
- Humans aren’t involved. Agents work best with a human to verify or guide the desired output.
For this reason, we encourage you to think about the situations when you use Rovo and review the quality of the responses you receive before sharing them with others.
You might also want to think about:
- Being as specific as possible in what you ask Rovo to do.
- Reading best practice guides on writing prompts for Agents.
- Finding efficient or tested ways to incorporate Agents in your team. You can see how others on your team are using Agents by viewing Agent profiles.
Tus datos y los Agentes Rovo
We understand you may have questions about how Rovo Agents use your data. This section supplements the information available on our Trust Center.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as a Confluence page, Jira work item or a document from a connected third-party product. If you use Agents in Rovo Chat via the Chrome browser extension, this includes the public webpage or Google Doc you have open.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, Rovo Agents apply the following measures.
Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third party LLM provider other than OpenAI or Google.
- Are not stored by any third-party LLM provider.
- Are not used to improve any large language models.
- Are retained by Atlassian for 30 days for safety and security purposes.
- OpenAI is a subprocessor on our List of Subprocessors. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows the permissions in your instance. For example, if you do not have access to a certain Confluence page, you will not be suggested content from that page in the response you receive. If you do not want your content to be available in responses to other users in your instance, please work with your org admin to ensure your permissions are set appropriately.
The details above apply to Agents that are provided by Atlassian. For more information about Agents provided by Atlassian, please see Rovo data, privacy and usage guidelines | Rovo | Atlassian Support. For information about Agents provided by third parties, please reference the Vendor Terms provided by those third parties for their apps.
How Rovo Chat uses AI
Rovo Chat is powered by large language models developed by OpenAI, Anthropic, and Google, as well as a combination of open-source large language models (including the Llama series) and other machine learning models. These models include OpenAI’s GPT series of models, Anthropic’s Claude series of models, and Google’s Gemini series of models.
Rovo Chat uses these models to analyze and generate responses to prompts using natural language, and provide relevant responses from Atlassian and connected third-party products.
Responses are generated by these large language models based on your inputs and are probabilistic in nature. That means that responses from large language models are generated by predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models, Anthropic’s models, and Google’s models. For more information on open-source language models, see information on the Llama series.
Use cases for Rovo Chat
Rovo Chat allows you to access the knowledge of your organization through a conversational interface. This means you can ask Rovo to write, read, review, or create content within your workflow. Rovo Chat understands the context of your work, using information from your Atlassian products, connected third-party products and uploaded files to answer questions or offer ideas and insights.
With Web Search, Rovo Chat can also access data sourced from the internet to help provide contextually relevant responses.
We believe that Rovo Chat works best in scenarios where:
- Your organization wants to provide users with an easy, conversational way to access company knowledge, based on each user’s access permissions.
- You prefer a conversational interface when performing tasks like creating pages or Jira work items, finding stakeholders, or finding information that is relevant to your work.
- You would like help finessing your writing.
- You would like help understanding long or complex knowledge pages.
- You would like help finding gaps in, or areas of improvement on, Jira work items or Confluence pages.
Whereas a standard Rovo Chat session generates quick responses, Deep Research works best when you would like to create in-depth research reports on specific topics. Deep Research is designed to perform in-depth, multi-step research by breaking your prompt into smaller, discrete steps, searching your connected data sources, refining results based on each discovery, and finally compiling the information it finds into one report with citations.
Rovo Chat can also be accessed by users via a Chrome browser extension. Learn more about accessing Rovo Chat here and here.
Considerations when using Rovo Chat
It’s important to remember that because of the way that the models used to power Rovo Chat work, these models can sometimes behave in ways that are inaccurate, incomplete or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We’ve found that Rovo Chat is less useful in scenarios where:
- You need deterministic, absolute, and accurate information about people, places and facts.
- You need information that you don’t have access to within your workspace.
- You need to ask questions and receive responses in languages other than English.
- You need to use modalities other than text such as video.
For this reason, we encourage you to think about the situations when you use Rovo Chat and review the quality of the responses you receive before sharing them with others.
You might also want to think about:
- Being as specific as possible in what you ask Rovo Chat to do. For example, you can include in your request the specific people, pages, projects, times or other context relevant to your question, or add clarifications in your follow up conversations with Rovo Chat. Tips on writing prompts for Chat.
- Taking the time to refine your prompts and engage in a bit of back and forth with Rovo Chat to get better results.
- When using features like Deep Research, take extra care to examine reasoning steps, consulted sources, and generated results. Complex features process larger volumes of data and have a have a higher tendency to go down an incorrect reasoning path.
- When using web results in Rovo Chat, carefully review the responses to ensure the sources are accurate and appropriate.
- When using Rovo Chat to generate code, take extra care to verify the code before executing it.
Your data and Rovo Chat
We understand you may have questions about how Rovo Chat uses your data. This section supplements the information available on our Trust Center.
We process:
Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as a Confluence page, Jira work item, or a document from a connected third-party product. If you use Rovo Chat via the Chrome browser extension, this includes the public webpage or Google Doc you have open.
- Files you uploaded to Rovo Chat which are used as context for your prompts.
- Informaion retrieved from Web Search.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, Rovo Chat applies the following measures.
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third party LLM provider other than OpenAI, Google, or Anthropic on AWS Bedrock.
- Are not stored by any third-party LLM provider.
- Are not used to improve any large language models.
- Are retained within your Chat history for 30 days so that you can go back and reference previous conversations.
- Are retained by Atlassian for 30 days for safety and security purposes.
OpenAI, Google, and AWS Bedrock are each a subprocessor on our List of Subprocessors. They do not use your inputs and outputs for any purpose besides processing or acting on your request.
This feature follows the permissions in your instance. For example, if you do not have access to a certain Confluence page, you will not be suggested content from that page in the response you receive. If you do not want your content to be available in responses to other users in your instance, please work with your org admin to ensure your permissions are set appropriately.
Files are uploaded to Rovo Chat are not retained beyond the Chat history.
When you access Rovo Chat from the Chrome browser extension, Chat will only read the content on the web page you are viewing to complement the content already within your Atlassian products. No additional data from the web page will be stored.
Read more about Rovo
Discover more about using Rovo
How Rovo Search uses AI
Rovo Search uses AI to enhance search capabilities across Atlassian and third-party tools.
Rovo Search is powered by large language models developed by OpenAI and Google, as well as a combination of open-source large language models (including the Llama series) and other machine learning models. These models include OpenAI’s GPT series of models and Google’s Gemini series of models.
Rovo Search uses these models to provide functionalities like semantic search, relevance ranking, and natural language processing. This includes analyzing and generating responses to search queries in natural language, and providing relevant responses from Atlassian and connected third-party products.
Responses are generated by these large language models based on your inputs and are probabilistic in nature. That means that responses from large language models are generated by predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models and Google’s models. For more information on open-source language models, see information on the Llama series.
Use cases for Rovo Search
Rovo Search enables users to search across multiple tools and platforms, including Atlassian products and connected third-party products, providing contextual and relevant results to enhance team collaboration and productivity.
We believe that Rovo Search works best in scenarios where:
- You need to search across various Atlassian products (like Jira and Confluence) and connected third-party products (such as Google Drive and SharePoint) in a single interface, without switching contexts.
- Your teams require quick access to relevant information from different sources to save time searching across all of those sources separately.
- You want search results to return results based on what your users have permission to access.
Considerations when using Rovo Search
It’s important to remember that because of the way that the models used to power Rovo Search work, these models can sometimes behave in ways that are inaccurate, incomplete or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We’ve found that Rovo Search is less useful in scenarios where:
- Your team primarily uses a single product or platform and doesn’t require cross-platform search capabilities.
- Your organization is using niche or highly specialized third-party tools that are not supported by Rovo's connectors.
- You have small teams or projects with minimal data and documentation needs.
For this reason, we encourage you to think about the situations when you use Rovo and review the quality of the responses you receive before sharing them with others.
You might also want to think about:
- Before connecting a third-party product, you should check the permissions set up for that product. Rovo Search will respect permissions, but you may need to set appropriate permissions for your space or content.
- When searching, be aware that results differ for individual users, because Rovo respects permissions and Search includes results that only the current user has access to.
- Rovo Search is able to give back richer people and team results if you sync your org chart with Atlassian Home.
- Encouraging your team to not only keyword search, but ask questions (for example, “Who is Jane Smith?” or “What is the vacation policy?”) while using Rovo Search.
Your data and Rovo Search
We understand you may have questions about how Rovo Search uses your data. This section supplements the information available on our Trust Center.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as document content from your Atlassian products and connected third-party products, document permissions, and file metadata.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, Rovo Search applies the following measures.
- Your prompts and responses:
- Are not available to other customers.
- Are not sent to any third-party LLM provider other than OpenAI or Google.
- Are not stored by any third-party LLM provider.
- Are not used to improve any large language models.
- OpenAI is a subprocessor on our List of Subprocessors. They do not use your inputs and outputs for any purpose besides processing your request.
- Rovo Search follows the permissions in your products.
- Data is stored in Atlassian's cloud-based search index and entity storage system. Rovo Search synchronizes with access control list (ACL) data from your connected third-party products and Atlassian products to ensure users only see content they have permission to access.
- For example, if you do not have access to a private Confluence page or Google Drive document, you will not be suggested that content in the response you receive. If you do not want your content to be available in responses to other users in your Atlassian and third-party products, please work with your org admin to ensure your permissions are set appropriately.
- Atlassian may store your search queries to improve the search experience for your users.
- We ‘listen’ for content deletion, and update the Rovo index with changes. For example, once the Rovo index is updated, content that’s deleted from a connected third-party product will not appear in Rovo results. More about Rovo data usage.
Read more about Rovo
Rovo: Unlock organizational knowledge with GenAI | Atlassian
How Rovo Skills use AI
Rovo Skills are powered by large language models developed by OpenA and Google, as well as a combination of open-source large language models and other machine learning models. These models include OpenAI’s GPT series of models and Google’s Gemini series of models.
Rovo Skills uses these models to analyze and generate natural language, and to perform actions, across Atlassian and connected third-party products.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text, based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models and Google’s models.
About Rovo Skills
A Rovo Skill automates specific, repeatable workflows with precision, enabling Rovo to execute tasks using packaged instructions.
With Rovo, we provide a number of skills that support tasks such as managing work in Jira (e.g. helping a user create, update, or organize Jira work items), communicating progress and outcomes (e.g. drafting project status updates or summarizing a set of issues for a release), and supporting Atlassian administration (e.g. assisting an administrator with user access, permissions, or related settings).
Rovo Skills can be invoked explicitly via slash command in Rovo Chat (e.g., /create-jira-issue). Over time, Rovo will be able to detect your intent to invoke specific skills implicitly without the need for a slash command. This capability will be added incrementally for different skills.
A list of available Skills is viewable within the Rovo Chat interface.
We believe that Rovo Skills work best in scenarios where:
- You want to generate, refine, or standardize structured content from existing work, like drafting releases.
- You want to quickly audit new or existing work, such as using a skill to check if a bug report has all necessary fields.
- You want to quickly retrieve, combine, or analyze specific information from Atlassian products such as summarizing release notes or generating insights and charts in Jira.
- You want to analyze sentiment and intent across a high volume of un-categorized items (like support tickets), automatically applying labels and priority so your team can focus on what matters most.
In general, we suggest using Rovo Skills when you want Rovo to complete a specific task or take action (like creating or updating issues, pages, or messages), based on clear and consistent workflows. Use Rovo Chat and Search when you want to find information, explore ideas, or ask open-ended questions.
Limitations and considerations when using Rovo Skills
It’s important to remember that because of the way that the models used to power Rovo Skills work, the outputs they generate can sometimes be inaccurate, incomplete, or unreliable.
The natural language responses you see, or the actions that a skill helps initiate, may not accurately reflect the underlying content or configuration they are based on, and may include information that sounds reasonable but is false, incomplete, or out of date.
We’ve found that Rovo Skills are less useful in scenarios where:
- You need current and accurate information about people, places, or other real‑world facts outside your Atlassian products.
- You need to access information in your Atlassian products that you do not have permission to view. Rovo Skills respect existing product permissions and cannot surface or act on content that you cannot otherwise access.
- You require fully consistent, deterministic output with no variation between runs (for example, regulatory or contractual text that must not change). Natural language outputs may vary for a number of reasons, including model behavior and user context. In these scenarios, we recommend setting automation rules.
- You need to work with modalities other than text that are not currently supported by the relevant skill (for example, extracting reliable information from images or video).
We recommend against using Rovo Skills when no human is available to review or supervise the output or resulting changes. Rovo Skills are designed to assist humans, not replace human judgment, and you should review and validate outputs and changes before relying on them or sharing them with others
For these reasons, we encourage you to think about the situations when you use Rovo Skills and review the quality of the responses you receive before sharing them with others.
You might also want to think about:
- Verifying the Source of Truth: Before running a skill that summarizes or acts on data (like /summarize-release), ensure the underlying Jira issues or Confluence pages are up to date. The Skill's output is only as accurate as the data it’s pulling from.
- Reviewing before Sending: For Skills that draft communications (like Slack messages or emails), treat the output as a first draft. Always review for tone and accuracy before hitting send or publishing.
- Being specific with your Slash Commands: When invoking a skill explicitly (e.g., using /), include as much context as possible in the initial prompt. This helps the Skill map your request to the correct parameters (like the right Jira Project or Confluence Document) without needing multiple rounds of clarification.
Your Data and Rovo Skills
We understand you may have questions about how Rovo Skills use your data. This section supplements the information available on our Trust Center.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as a Confluence page, Jira issue or a document from a connected third-party product.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, Rovo Skills apply the following measures.
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third party LLM provider other than OpenAI or Google
- Are not stored by any third-party LLM provider.
- Are not used to improve any large language models.
Note: Your skill invocations and responses in Rovo Chat will be visible in conversation history.
- OpenAI and Google are each a subprocessor on our List of Subprocessors. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows the access and action permissions in your instance. For example, if you do not have access to a certain Confluence page, you will not be suggested content from that page in the response you receive. If you do not have the ability to take an action, the Skill will not be able to execute that action. If you do not want your content to be available in responses to other users in your instance, please work with your org admin to ensure your permissions are set appropriately.
- Where you invoke your skill via Rovo Chat, you can review those interactions in your Rovo Chat conversation history.
The details above apply to Skills that are provided by Atlassian. For more information about Skills provided by Atlassian, please see Rovo data, privacy, and usage guidelines | Rovo | Atlassian Support.
Read more about Rovo at Atlassian
Accelerate work with AI
Select a Rovo feature below to get a transparent look at use cases and data use.
How Rovo Customer Service uses large language models
Rovo Customer Service is powered by large language models developed by OpenAI, Google, and Anthropic, as well as a combination of other machine learning models. These large language models include OpenAI's GPT series of models, Google's Gemini series of models, and Anthropic's Claude series of models.
Rovo Customer Service uses these models to analyze and generate natural language within our products.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI's models, Google's models and Anthropic's models. For more information on open-source language models, see information on the Llama series and the Phi series.
Use cases for Rovo Customer Service
Rovo Customer Service is the AI agent capability in Customer Service Management (CSM). It allows organizations to configure their own AI agent that can answer customer queries and requests in seconds, and retrieve answers or automate actions to resolve requests. Organizations can continually monitor and enhance agent effectiveness. Admins can also use Rovo Customer Service to analyze past support interactions, identify gaps in their agent's knowledge, and generate suggested knowledge content for admin review using the CSM Optimize feature.
We believe that Rovo Customer Service works best in these scenarios:
• Your organization has a well-structured knowledge base.
• Your organization has a meaningful portion of customer queries that can be resolved through knowledge articles or simple actions.
• When using CSM Optimize, your organization has past support interactions that are relevant and appropriate to use for improving the AI agent's knowledge. Having at least 500 past support interactions is required to use the automated optimization.
• Your organization is willing to continuously evaluate and use available configuration features in the app to refine the AI agent's performance.
Considerations when using Rovo Customer Service
It's important to remember that because of the way that the models used to power Rovo Customer Service work, these models can sometimes behave in ways that are inaccurate, incomplete or unreliable.
For example, the responses that you and your end users receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We've found that Rovo Customer Service is less useful in scenarios where:
• You need current and accurate and up to date information about people, places and facts.
• You need Rovo Customer Service to have access to information that isn't readily available to you (for example, connected knowledge base) to properly answer your request.
• Your connected knowledge base for Rovo Customer Service is outdated or contains wrong information, or does not contain enough information.
• Your instructions or knowledge base provided to Rovo Customer Service contains conflicting information.
• You should review any suggested knowledge content created by CSM Customize before publishing it internally, or making it available to Rovo Customer Service.
Rovo Customer Service should not be used in scenarios that are outside of customer service, or for any purpose that would violate Atlassian's Acceptable Use Policy.
For this reason, we encourage you to think about the situations when you use Rovo Customer Service and review the quality of the responses you receive before enabling Rovo Customer Service for your end users.
You might also want to think about:
• Being as specific as possible in what you ask Rovo Customer Service to do.
• Ensure the knowledge sources you provide to Rovo Customer Service are reviewed and suitable for end users. We strongly recommend removing internal-only content, confidential information and simplifying language to deliver a clear, accurate customer experience.
• Regularly review end-user conversations with Rovo Customer Service, and make improvements where needed - whether by updating the knowledge base, adjusting actions, or adding guidance.
Your data and Rovo Customer Service
We understand you may have questions about how Rovo Customer Service uses your data. This section supplements the information available on our Trust Center.
We process:
• Your prompts (inputs) and responses (outputs).
• Context from your instance relevant to your prompt, such as the current conversation, connected knowledge base, available actions, and your admin-configured guidance - to deliver relevant responses.
• Past support interactions that admins choose to use for CSM Optimize, and suggested knowledge content generated from those interactions.
• Data about how you interact with our features, such as clickstream data and the people you work with.
• Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, Rovo Customer Service applies the following measures:
Your prompts (inputs) and responses (outputs):
• Are not available to other customers.
• Are not sent to any third party LLM provider other than OpenAI, Google, or Anthropic on AWS Bedrock.
• Are not stored by any LLM provider.
• Are not used to improve any large language models.
OpenAI, Google, and Anthropic on AWS Bedrock are subprocessors on our List of Subprocessors. They do not use your inputs and outputs for any purpose besides processing your request.
This feature follows the permissions in your instance. Rovo Customer Service only has access to knowledge and support-interaction data that you manually connect, select, upload, or otherwise make available to Rovo Customer Service.
Admins can choose whether to use CSM Optimize features and can review, edit, accept, reject, or delete suggested knowledge content before it is published or used with end users.
Your conversation history will be retained in your Rovo Customer Service instance and available to admins, while you maintain an active license, unless you manually delete it.
Admins have access to an evaluation capability, which enables the admin to upload a predefined data set, and use that data set to evaluate the performance of the AI agent. You can delete evaluations and conversations that are generated with the data set. When an evaluation is deleted, all metrics for the evaluation, will also be deleted.
How AI ChatOps for incident management uses AI
The AI ChatOps for incident management features are powered by large language models developed by OpenAI, as well as a combination of open-source large language models (including the Llama series) and other machine learning models. These large language models include OpenAI’s GPT series of models.
Rovo uses these models to analyze and generate natural language within our products.
These large language models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models or about this approach in OpenAI’s research papers. For more information on open-source language models, see information on the Llama series.
Use cases for AI ChatOps for incident management
AI ChatOps for incident management help your users to expedite the incident resolution process by offering a summary of the relevant incident and all the conversations about it so far to new users when they’re added to the Slack channel which is linked to the incident work item in Jira Service Management. Additionally, AI ChatOps for incident management can log the conversations that occur in Slack to Jira Service Management as a timeline for future reference.
We believe that AI ChatOps for incident management work best in scenarios where:
- Your organization uses Slack as a primary collaboration tool for triaging and resolving incidents.
-
Your organization follows a protocol of documenting incident-related conversations within the incident work item in Jira Service Management.
Considerations when using AI ChatOps for incident management
Remember that because of the way that the models used to power AI ChatOps for incident management work, these models can sometimes behave in ways that are inaccurate, incomplete, or unreliable. For example, the responses that you receive might not accurately reflect the content that they’re based on or include content that sounds reasonable but is false or incomplete.
We’ve found that AI ChatOps for incident management are less useful in scenarios where:
-
You need current and accurate information about people, places and facts.
-
You don’t use Slack as your organization’s primary collaboration tool for triaging and resolving incidents.
-
Your team lacks access to the incident work item in Jira Service Management or the Slack channel linked to the incident work item that’s being triaged.
For this reason, we encourage you to consider situations where you use AI ChatOps for incident management and review the quality of the responses you receive before sharing them with others.
You might also want to think about reviewing permissions to ensure that users have appropriate levels of access to relevant incident work items and Slack channels.
-
Your data and AI ChatOps for incident management
We understand you may have questions about how AI ChatOps for incident management uses your data. This section supplements the information available on our FAQ page.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as incident data and messages in your connected Slack channels.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Any feedback you provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, AI ChatOps for incident management apply the following measures:
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third-party LLM provider other than OpenAI.
- Are not stored by OpenAI.
- Are not used to improve OpenAI or any other third party LLM provider.
- OpenAI is a subprocessor on our Subprocessors page. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows your product’s permissions. For example, if you don’t have access to an incident work item, you won’t find enough of its context when you generate AI ChatOps for incident management. If you don’t want your content to be available in response to other users of your product, work with your organization admin to ensure the team’s access permissions are configured appropriately.
How AI drafts uses AI
AI drafts is powered by large language models developed by OpenAI and Anthropic, as well as a combination of open-source transformer-based language models and other machine learning models. These large language models include OpenAI’s GPT series of models and Anthropic’s Claude series of models.
Rovo uses these models to analyze and generate natural language within our products. The open-source encoder models convert your textual inputs into numerical forms (embeddings) which are used for identifying and forming topics from your inputs.
These large language models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models and Anthropic’s models. For more information on open-source language models, see information on embedding models.
Use cases for AI drafts
AI drafts suggests pre-generated drafts of knowledge articles for your admins and agents to consider. This feature generates drafts for the most common work items in a Jira Service Management project, using the details and comments within work items in that project.
This allows your service teams to increase the coverage of your knowledge base articles more quickly and easily, which will, in turn, benefit the performance of other features in Jira Service Management. For example, knowledge articles created using AI drafts can then be used by Virtual Service Agent’s AI answers feature to resolve requests from your help seekers more easily and quickly.
Read more about AI answers in virtual service agent.
We believe that AI drafts work best in scenarios where:
- Your project receives more than 40 requests in any month. Projects with fewer than 40 requests may not include enough information to generate AI drafts.
- The content in work items such as summary, description, and other fields is primarily in English. The quality of AI drafts may vary for other languages.
Considerations when using AI drafts
It’s important to remember that because of the way that the models used to power AI drafts work, these models can sometimes behave in ways that are inaccurate, incomplete, or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We’ve found that AI drafts are less useful in scenarios where:
- Your project receives less than 40 requests in a month.
- Your project creates requests automatically from alerts as these might not require a knowledge base article, but this feature will still suggest AI drafts to you in relation to those requests.
- You need drafts for work items where the content is in languages other than English.
For this reason, we encourage you to think about the situations when you use Rovo and review the quality of the responses you receive before sharing them with others.
You might also want to think about:
- Putting in place measures ensure that your team members thoroughly review AI drafts before proceeding to create knowledge articles based on them
- Asking your project admins to identify experts within their teams who will be responsible for reviewing and editing AI drafts in order to maintain high-quality knowledge base articles
Your data and AI drafts
We understand you may have questions about how AI drafts use your data. This section supplements the information available on our Trust Center.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to the feature, such as data in a work item like field values, summary, and description to generate drafts using Rovo.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, AI drafts apply the following measures.
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers
- Are not sent to any third party LLM provider other than OpenAI or Anthropic on AWS Bedrock.
- Are not stored by any LLM provider.
- Are not used to improve any third-party LLMs.
- OpenAI and AWS Bedrock are each a subprocessor on our List of Subprocessors. They do not use your inputs and outputs for any purpose besides processing your request.
- We only process work items that are accessible to everyone in a service project on your site. Any issues with work item permissions applied are not considered to generate drafts.
- We refresh AI drafts every 7 days to ensure that content is up-to-date.
Read more about Rovo
How AI Summaries in Company Hub Cards uses AI
AI Summaries in Company Hub Cards is powered by large language models developed by OpenAI. These large language models include OpenAI’s GPT series of models.
Rovo uses these models to analyze and generate natural language within our products.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models.
Use cases for AI Summaries in Company Hub Cards
AI summaries in Company Hub Cards helps you to quickly populate and publish your Company Hub, by suggesting AI-generated descriptions for cards. This feature uses AI to generate a summary of the linked Confluence page or blog post for your Company Hub card.
We believe that AI Summaries in Company Hub Cards works best in scenarios where:
- You like to use Company Hub cards in Confluence to link to text-heavy pieces of content.
- You are looking for a concise summary of text-heavy Confluence content to be displayed in Confluence card descriptions.
Considerations when using AI Summaries in Company Hub Cards
It’s important to remember that because of the way that the models used to power AI Summaries in Company Hub Cards work, these models can sometimes behave in ways that are inaccurate, incomplete or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We’ve found that AI Summaries in Company Hub Cards is less useful in scenarios where:
- You need current and accurate information about people, places, and facts.
- You need a summary of a very short Confluence page where there is not enough content to generate a useful summary.
- You need a summary of a Confluence page where most of the content is in tables or expands.
- You need a summary of a Confluence page with most of the content in macros.
For this reason, we encourage you to think about the situations when you use Rovo and review the quality of the responses you receive before sharing them with others.
Your data and AI Summaries in Company Hub Cards
We understand you may have questions about how AI Summaries in Company Hub Cards uses your data. This section supplements the information available on our Trust Center.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as content from the Confluence page you want to summarize.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, AI Summaries in Company Hub Cards applies the following measures.
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third-party LLM provider other than OpenAI.
- Are not stored by any LLM provider.
- Are not used to improve large language models.
- Open AI is a subprocessor on our List of Subprocessors. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows the permissions in your instance. For example, if you do not have access to a Confluence page, you will not be able to use this feature to retrieve a summary using Rovo. If you do not want your content to be available to other users in your instance, work with your org admin to ensure your permissions are set appropriately.
Read more about Rovo
How AI related resources uses AI
AI related resources is powered by large language models developed by OpenAI, as well as a combination of open-source large language models (including the Llama series) and other machine learning models. These large language models include OpenAI’s GPT series of models.
Rovo uses these models to analyze and generate natural language within our products, and to provide relevant responses from Atlassian and connected third-party products.
These large language models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models or about this approach in OpenAI's research papers. For more information on open-source language models, see information on the Llama series.
Use cases for AI related resources
Rovo enables your users to speed up the process of resolving incidents by suggesting a list of resources that they can refer to, across your linked knowledge base spaces and articles, Jira work items, and any third-party products you have integrated through Rovo. Read more about Rovo and third party tools.
We believe that AI related resources work best in scenarios where:
-
Your organization has a lot of documentation in your linked knowledge base spaces and connected third-party tools that are relevant to and helpful for resolving incidents that occur in your organization.
-
Your team needs quick access to resources that are likely relevant to the incident when they are triaging the incident.
-
Considerations when using AI related resources
Remember that because of the way that the models used to power AI related resources work, these models can sometimes behave in ways that are inaccurate, incomplete, or unreliable. For example, the responses that you receive might not accurately reflect the content that they are based on or include content that sounds reasonable but is false or incomplete.
We’ve found that AI related resources is less useful in scenarios where:
-
You need current and accurate information about people, places, and facts.
-
You need AI related resources to access information not readily available to you, like restricted Confluence documents (for example, access-restricted Confluence pages) to properly respond to your request.
-
You have minimal documentation available in your setup that Rovo can use to suggest related resources.
For this reason, we encourage you to consider situations where you use Rovo and review the quality of the responses you receive before sharing them with others.
You might also want to think about the following:
-
Reviewing permissions to ensure that users have appropriate levels of access to your relevant documentation and resources from third-party tools (like Slack channels).
-
Reviewing and confirming that your organization’s documentation is complete, up-to-date, and accurate.
-
Your data and AI related resources
We understand you may have questions about how AI related resources uses your data. This section supplements the information available on our FAQ page.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as incident data, alerts data, documentation in your Atlassian products, and connected third-party products such as Google Docs.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Any feedback you provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, AI related resources applies the following measures:
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third-party LLM provider other than OpenAI.
- Are not stored by OpenAI or any other third-party LLM provider.
- Are not used to improve OpenAI or any other third party LLM provider.
- OpenAI is a subprocessor on our Subprocessors page. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows your product’s permissions. For example, if you don’t have access to a certain Confluence page, you won’t find that page in the list of related resources. If you don’t want your content to be available in response to other users of your product, work with your organization admin to ensure the team’s access permissions are configured appropriately.
How AI suggestions in Jira Service Management use AI
AI suggestions in Jira Service Management is powered by large language models developed by OpenAI, and other machine learning models. These large language models include OpenAI’s GPT series of models.
Rovo uses these models to analyze and generate natural language within our products.
These large language models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models.
Use cases for AI suggestions in Jira Service Management
With AI suggestions in Jira Service Management, your team can quickly get up to speed by gathering important context about your service requests and incidents at a glance. Rovo helps your team to:
- For service requests, understand a brief summary of the work item, details of the work item reporter, and a list of suggested steps that could help agents resolve the work item. Rovo also suggests ways that your team can update service requests based on this context, such as changing their priority or adding an assignee.
- For incidents, understand a brief summary of the work item and details of the probable root cause. Rovo also suggests ways that your team can update incidents based on this context, such as identifying an incident as a major incident, adding responders or affected services, changing their severity or priority, or investigating the problem by creating a problem work item.
AI suggestions in Jira Service Management can also recommend that agents escalate a request or incident when the applicable SLA is about to be breached. In the case of service requests, this feature may also suggest that agents escalate that request where the models used to power these suggestions identify, based on the text of the reporter’s comments, a sense of urgency or anger with that request.
We believe that AI suggestions in Jira Service Management work best in scenarios where:
- Your projects receive a lot of similar requests or incidents.
- Your organization already records service requests and incidents in Jira Service Management with complete, accurate, and up-to-date information.
- Your team members' user profiles have been populated with information about their roles, departments, and locations.
- Your customers and team members keep a written record of all the conversations by adding comments on work items.
- Your projects include a number of each kind of service request or incident commonly encountered by your teams.
Considerations when using AI suggestions in Jira Service Management
It’s important to remember that because of the way that the models used to power AI suggestions in Jira Service Management work, these models can sometimes behave in ways that are inaccurate, incomplete or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on or include content that sounds reasonable but is false or incomplete.
We’ve found that AI suggestions in Jira Service Management are less useful in scenarios where:
- You need current and accurate information about people, places, and facts.
- You need suggestions for a service request or incident in circumstances where your project does not include any similar requests or incidents to learn from.
- You need information that you don’t have access to within your project.
- You need suggestions for assignees for whom information about their roles, departments, and locations has not been added to their profile.
For this reason, we encourage you to think about the situations when you use Rovo and review the quality of the responses you receive before sharing them with others.
You might also want to think about:
- Asking team members to make sure that their profile details (and past incidents/service requests) are fully populated before using AI suggestions in Jira Service Management.
Your data and AI suggestions in Jira Service Management
We understand you may have questions about how AI suggestions in Jira Service Management uses your data. This section supplements the information available on our Trust Center.
We process:
- The prompts (inputs) and responses (outputs)
- Context from your instance relevant to the feature, such as similar requests or incidents, list of assignees or teammates, assets, and data in a work item like field values, comments, etc.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, AI suggestions apply the following measures.
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third-party LLM provider other than OpenAI.
- Are not stored by OpenAI.
- Are not used to improve OpenAI.
- OpenAI is a subprocessor on our List of Subprocessors. They don't use your inputs and outputs for any purpose besides processing your request.
- This feature follows the permissions in your instance. For example, if you don't have access to a certain service request or incident, Rovo will not consider those requests or incidents while generating suggestions. If you don't want your content to be available to other users in your project or site, work with your org admin to ensure your permissions are set appropriately.
Read more about Rovo
How Rovo Dev uses large language models
Rovo Dev is powered by large language models developed by OpenAI and Anthropic on AWS Bedrock and Anthropic on Google Vertex AI, as well as a combination of other machine learning models. These large language models include OpenAI's GPT series of models and Anthropic's Claude series of models.
Rovo Dev uses these models to analyze and generate natural language and code within our products.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI's models and Anthropic's models.
Use cases for Rovo Dev
Rovo Dev is a context-aware AI agent for professional software engineers. It understands natural language and code, and integrates with Atlassian tools like Jira, Confluence, and Bitbucket as well as third-party developer platforms such as GitHub. Powered by Atlassian’s Teamwork Graph, it draws on organizational context, business knowledge, and codebase history to accelerate the software development lifecycle. Rovo Dev assists with planning, code generation, reviews, and automating repetitive work at scale.
The Rovo Dev product suite includes:
- Rovo Dev in Bitbucket and GitHub: Rovo Dev integrates with both Bitbucket and GitHub to check pull requests against both technical requirements and business objectives. It helps you to spot potential issues before they reach production and checks that code changes align with Jira acceptance criteria, reducing review cycles and preventing rework.
-
Rovo Dev CLI: Rovo Dev CLI brings Rovo Dev directly into your command line interface, turning your terminal into an AI development agent that is seamlessly integrated with your Atlassian workflow.
We believe that Rovo Dev works best in scenarios where:
- You need assistance with coding tasks such as writing new code, debugging existing code, or understanding complex codebases.
- You want automated code review feedback on pull requests to catch potential issues and improve code quality.
- You need help with common development tasks like refactoring, testing, or documentation
- You want to resolve Jira work items more efficiently by automating code changes in linked repositories.
- Your development team wants to maintain consistent coding standards and best practices.
Considerations when using Rovo Dev
It's important to remember that because of the way that the models used to power Rovo Dev work, these models can sometimes behave in ways that are inaccurate, incomplete, or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We've found that Rovo Dev is less useful in scenarios where:
- You need current and accurate information about people, places and facts that are not contained within your codebase, development context, or reachable Atlassian content such as Confluence pages or Jira issues.
- You need Rovo Dev to have access to information that isn't readily available in your repository, development environment, or linked Atlassian tools.
- You need perfect accuracy for safety-critical or security-sensitive code without human review because Rovo Dev should not be used in these scenarios.
- You're working with proprietary or highly confidential code patterns that weren't part of the model's training data. You might not find Rovo Dev helpful when working with very new technologies that weren't included in the model's training data, because Rovo Dev may not have sufficient context on those technologies to provide useful outputs.
- Rovo Dev should not be used as the sole method for security reviews, compliance checks, or in scenarios where code correctness is critical for safety or regulatory requirements without appropriate human oversight.
-
For this reason, we encourage you to think about the situations when you use Rovo Dev and review the quality of the responses you receive before implementing suggested changes or sharing them with others.
You might also want to think about:
- Being as specific as possible in what you ask Rovo Dev to do, including providing clear context about your codebase and requirements.
- Reviewing and testing any code changes suggested by Rovo Dev before committing them to your repository.
- Ensuring your development environment has appropriate permissions and access controls configured.
- Keeping your repository documentation and code comments up to date to provide better context for AI assistance.
- Using Rovo Dev as a collaborative tool alongside human code review rather than a replacement for it.
Your data and Rovo Dev
We understand you may have questions about how Rovo Dev uses your data. This section supplements the information available on our Rovo Dev Page.
We process:
- Your prompts (inputs) and responses (outputs) when interacting with Rovo Dev
- Context from your development environment relevant to your prompt, such as code files and pull requests in Bitbucket, Jira work items, content from connected third-party apps such as Github or connected MCP servers, or local files on your device (if using RovoDev CLI)
- Data about how you interact with our features, such as command usage patterns and the development tools you work with
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback
When it comes to your data, Rovo Dev applies the following measures:
Your prompts (inputs) and responses (outputs):
- Are not available to other customers
- Are not sent to any third party LLM provider other than OpenAI, Google Vertex AI, and Anthropic on AWS Bedrock
- Are not stored by OpenAI, Google Vertex AI or Anthropic on AWS Bedrock
- Are not used to improve OpenAI or Anthropic models
Third-party processing:
- OpenAI, Google Vertex AI, and Amazon Web Services (AWS Bedrock) are subprocessors on our List of Subprocessors. They do not use your inputs and outputs for any purpose besides processing your request.
Permissions and access:
- Rovo Dev follows the permissions in your development environment and connected Atlassian tools. For example, if you do not have access to certain repositories or Jira projects, Rovo Dev will not access content from those sources when generating responses.
- If you do not want your code or content to be available in responses to other users in your organization, please work with your system administrators to ensure your permissions and access controls are set appropriately.
- In Rovo Dev CLI, users are able to add third-party Model Context Protocol (MCP) servers to provide Rovo Dev with access to third party content. Integration with any third-party MCP server requires approval from your user and you must ensure that you comply with any third-party terms of use. Users can enter prompts to send data outside of your Atlassian site, and data sent outside your site will not be subject to Atlassian terms.
Read more about Rovo Dev
How Automation uses AI
Automation using AI is powered by large language models developed by OpenAI, Google, and Anthropic, as well as a combination of open-source large language models (including the Llama series and Mixtral series) and other machine learning models. These large language models include OpenAI's GPT series of models, Google's Gemini series of models, and Anthropic's Claude series of models.
Rovo uses these models to analyze natural language input and generate an automation rule for you within Jira and Confluence.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated by predicting the most probable next word or text, based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models, Google’s models and Anthropic’s models. For more information on open-source language models, see information on the Llama series.
Use cases for Automation using AI
Creating automation rules is at the core of the everyday automation experience, and we want to make this even easier for you by adding Rovo to the automation rule builder in Jira and Confluence. Now, you can easily create automation rules by simply typing in and describing what you wish to automate, and let Rovo handle all the heavy lifting of creating the rule for you. Find out more about Automation using Rovo for Jira and for Confluence.
We believe that Automation using Rovo for Jira and Confluence works best in scenarios when you are not sure how to get started or want to accelerate the rule creation process.
Not sure how best to create an automation rule?
Automation rules are created by a combination of different types of components: triggers, actions, conditions, and branches. Think of components as the building blocks of a rule. To successfully create a rule with Rovo, your rule must at least contain both a trigger and an action. For example:
In Jira:
Every Monday, find all the tasks with a due date in the next 7 days, and send the assignee a reminder email.
When a ticket moves to Testing, assign the ticket to John Smith.
In Confluence:
- Every Monday, find all the tasks with a due date in the next 7 days, and send the assignee a reminder email.
- Every 6 months, archive any pages that haven’t been updated in that time. After archiving, send an email to the page author letting them know.
- When a page is published with Product Spec in the title, create a Jira ticket to review the page with a link to the page.
In addition, for a rule to be successfully created, all its components must be supported by Automation using Rovo. This means that any triggers, actions, conditions, or branches in your rule must be compatible with Automation in Jira and/or Confluence.
{ "type":"imkt.components.ContainerAnnouncementBanner", "params": { "cookieName":"banner-0123c6ee-34c8-4434-b876-7d00fc36e837" } }
Considerations for Automation using AI
It’s important to remember that because of the way that the models used to power Automation using Rovo work, these models can sometimes behave in ways that are inaccurate, incomplete or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We’ve found that Automation using Rovo is less useful in scenarios where:
- You need to give Automation using Rovo access to information that isn’t readily available to you (for example, a restricted page or project) to properly answer your request.
- You need to perform one-off tasks.
- You need to query information from within your knowledge base.
For this reason, we encourage you to think about the situations when you use Rovo and review the quality of the responses you receive before sharing them with others.
Automation using Rovo will only work with the existing set of available automation components in Jira and Confluence.
You might also want to think about being as specific as possible in what you ask Rovo to do, as described above.
Your data and Automation using AI
We understand you may have questions about how Automation using Rovo uses your data. This section supplements the information available on our FAQ page.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as a Jira project or a Confluence page.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, using Rovo for Confluence automation applies the following measures:
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third party LLM provider other than OpenAI, Google, or Anthropic on AWS Bedrock
- Are not stored by any LLM vendor.
- Are not used to improve LLM models.
All third-party LLM providers are subprocessors and listed as so on our Subprocessors page. They do not use your inputs and outputs for any purpose besides processing your request.
This feature follows the permissions in your instance. For example, if you do not have access to a specific project or page, you will not be suggested content from those assets in the response you receive. If you do not want your content to be available in responses to other users in your instance, work with your org admin to ensure your permissions are set appropriately.
How Alert classification uses AI
Alert classification is powered by large language models developed by OpenAI, as well as other machine learning models. These models include machine learning models designed to identify patterns in alert data, and OpenAI’s GPT series of models.
Rovo uses these machine learning models to analyze alerts and classify them as either important or informational within our products based on patterns and heuristics identified in alert data. Rovo then uses large language models to analyze and generate natural language within our products.
These large language models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models.
Use cases for Alert classification
Alert classification uses AI to classify alerts and alert groups into “Signal” and “Likely noise”. This helps operations teams quickly distinguish alerts and groups that are more likely to indicate meaningful impact from those that are more likely to be routine or low‑impact.
We believe that Alert classification works best in scenarios where customers:
- You maintain consistent and descriptive alert naming conventions and tags across their monitoring and observability tools.
- You have sufficient historical alert data for the same services or components, so the model can learn meaningful patterns unique to your business over time, as described in the Your data section below.
- You configure clear routing, priorities, and escalation policies in Jira Service Management so alert attributes reliably reflect severity and ownership.
- You will use the AI-generated classification and its explanation as one input into their operational decisions (for example, triage, escalation, or suppression), rather than treating it as an authoritative source or the sole basis for action. AI-generated classifications work best when they are validated and confirmed by an experienced human before being actioned.
Considerations when using Alert classification
It’s important to remember that because of the way that the models used to power Alert Classification, these models can sometimes behave in ways that are inaccurate, incomplete, or unreliable. For example, the responses that you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We’ve found that Alert Classification is less useful in scenarios where:
- You need current and accurate information about people, places and facts.
- Your organization’s alert history is sparse or patterns change rapidly, reducing the usefulness of past data.
- Your organization’s alert titles and attributes are inconsistent or uninformative.
For this reason, we encourage you to think about situations when you use Alert classification and review the quality of the responses you receive before sharing them with others. Treat classifications as a way to provide support for decisions and apply human judgment before taking action.
Your data and Alert classification
We understand you may have questions about how Alert classification uses your data. This section supplements the information available on our Trust Center.
- We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as your alert data (alert titles, alert tags, priority, responder teams, description).
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
We process your alert data to train a version of the machine learning model to recognize patterns specific to your alerts.
- We store the identified patterns to provide you with insights.
- We do not use your alert data to train any LLM.
When it comes to your data, Alert classification applies the following measures:
- Your inputs and outputs:
- Are not available to other customers
- Are not sent to any third-party LLM provider other than OpenAI.
- Are not stored by OpenAI.
- Are not used to improve OpenAI models.
- OpenAI is a subprocessor on our List of Subprocessors. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows the permissions on your site. For instance, if Alert classification classifies 50 alerts (30 signal and 20 noise) and you have permission to view only 30 of them, you’ll only see those 30 in the list view. If you don't want your alerts to be available in response to other users in your site, please work with your org/site admin to ensure your permissions are set appropriately.
- We process:
How alert grouping uses AI
Alert grouping by AI is powered by large language models developed by OpenAI and Google. These models include an algorithm designed to identify patterns in alert data, and OpenAI's GPT series of models and Google's Gemini series of models.
Rovo uses these machine learning models to analyze and generate alert groups and give related suggestions (past alert groups and past alert responders) within our products based on the similarity of the alert content or the tags used. Rovo then uses large language models to analyze and generate natural language descriptions and content for these groups within our products.
These large language models generate responses based on your inputs and are probabilistic. This means that their responses are generated by predicting the most probable next word or text based on the data they have been trained on.
Read more about the capabilities of OpenAI’s models and Google’s models.
Use cases for alert grouping
Alert grouping uses AI to identify and group similar alerts together. It also helps you by identifying and recommending past similar alert groups and past alert responders (or teams of responders), based on the semantic similarity of the alert content or tags used.
When you want to escalate the alert group to an incident, alert grouping will also pre-populate all contextual information for you to review as part of the incident creation process.
We believe that alert grouping works best in scenarios where:
- Your organization frequently encounters patterns of similar or duplicate alerts occurring at a high volume, whether experienced over a short period or a more extended time frame.
- Your organization consistently categorizes alerts using tags.
- Your team often finds that similar or duplicate alerts should be escalated into incidents.
Considerations when using alert grouping
It’s important to remember that because of the way that the models used to power alert grouping work, these models can sometimes behave in ways that are inaccurate, incomplete, or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on or include content that sounds reasonable but is false or incomplete. In the case of the alert groups that you see, they might not precisely reflect the semantic similarity of their tags.
We’ve found that alert grouping is less useful in scenarios where:
- You need current and accurate information about people, places, and facts.
- You need alert grouping to have access to information that isn’t readily available to you to properly group the alerts. Alert grouping works within the boundaries of your team’s configured roles and permissions, so you’ll only have access to groups and insights for alerts you have permission to view.
- The alert tags used by your team are not consistent or well-maintained. Because alert grouping works by grouping alerts based on the semantic similarity of alert titles and tags, the quality of the alert groups it generates depends on the consistency and hygiene of the alert tags used in your team and your organisation.
For this reason, we encourage you to think about the situations when you use Rovo and review the quality of the responses you receive before sharing them with others.
You might also want to think about ensuring that you and your team follow consistent practices in using alert tags.
Your data and alert grouping
We understand you may have questions about how alert grouping uses your data. This section supplements the information available on our FAQ page.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as your alert data (alert titles, alert tags, priority, responder teams, description).
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
We process your alert data to train a version of the machine learning model to recognize patterns specific to your alerts.
- We store the identified patterns to provide you with insights.
- We do not use your alert data to train any third-party-hosted LLM.
When it comes to your data, alert grouping applies the following measures:
- Your inputs and outputs:
- Are not available to other customers.
- Are not sent to any third party LLM provider other than OpenAI or Google.
- Are not stored by any LLM vendor.
- Are not used to improve LLM models.
- All third-party LLM providers are subprocessors and listed as so on our Subprocessors page. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows your site's permissions. For instance, if Rovo groups 50 alerts based on their tags and semantic similarity and you have permission to view only 30 of them, you’ll only see those 30 in the group detail view. If you don't want your alerts to be available in response to other users in your site, please work with your org/site admin to ensure your permissions are set appropriately.
How AI summarizes pages and blogs in Confluence
Summarize pages and blogs using AI is powered by large language models developed by OpenAI and Google. These large language models include OpenAI's GPT series of models and Google's Gemini series of models.
Rovo uses these models to analyze and generate natural language within our products.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text, based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models and Google’s models.
Use cases for Confluence quick summary
Save time and get the details you need to do your work faster by generating a quick summary of a Confluence page or blog with AI. Find out more about using AI in Confluence.
We believe that summarizing pages and blogs using Rovo works best in scenarios when:
- There is a text-heavy page that takes a 5 minutes or more to read.
- There is a lot of written content, with limited visuals and/or other formatting like expands on a page.
Considerations when summarizing pages and blogs using AI
It’s important to remember that because of the way that the models used to power summarizing pages and blogs using AI work, they can sometimes behave in ways that are inaccurate, incomplete, or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on or include content that sounds reasonable but is false or incomplete.
While we continue to build better support for macros, tables, and expand in summaries, we’ve found that summarizing pages and blogs using Rovo is less useful in scenarios where:
- You need current and accurate information about people, places, and facts.
- You need a summary of a very short Confluence page where there is not enough content.
- You need a summary of a Confluence page where most of the content is in tables or expands.
- You need a summary of a Confluence page with most of the content in macros.
We encourage you to think about the situations when you use Rovo and review the quality of the responses you receive before sharing them with others.
You might also want to think about:
- Asking Rovo to summarise pages that you know are heavy on text-based content.
Your data and summarizing pages and blogs using AI
We understand you may have questions about how using AI for Confluence automation uses your data. This section supplements the information available on our FAQ page.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as content from the Confluence page you want to summarize.
- Data about how you interact with our features, such as clickstream data and the people you work with
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, summarizing pages and blogs using Rovo applies the following measures:
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third party LLM provider other than OpenAI or Google.
- Are not stored by any LLM vendor.
- Are not used to improve LLM models.
- All third-party LLM providers are subprocessors and listed as so on our Subprocessors page. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows the permissions in your instance. For example, if you do not have access to a Confluence page, you will not be shown this feature or be able to summarize a page using Rovo. If you do not want your content to be available to other users in your instance, work with your org admin to ensure your permissions are set appropriately.
How Rovo defines terms
Defining terms using Rovo in Confluence and Jira is powered by large language models developed by OpenAI, Google, and Anthropic, as well as a combination of open-source large language models (including the Llama series and Mixtral series) and other machine learning models. These large language models include OpenAI's GPT series of models, Google's Gemini series of models, and Anthropic's Claude series of models.
Rovo uses these models to analyze and generate natural language answers within Confluence and Jira.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text, based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models, Google’s models and Anthropic’s models. For more information on open-source language models, see information on the Llama series.
Use cases for defining terms using Rovo
One of the most challenging things about consuming content in Confluence and Jira can be getting the context you need to understand what you’re reading. Abbreviations, acronyms, unfamiliar terms, and team or project-specific names can lead to a lengthy search for the information you need.
Defining terms using Rovo will provide the definition of company-specific terms (such as acronyms, project, system, or team names) on a page in Confluence or in a work item description in Jira. This gives users the information they need, when they need it - all whilst helping teams work better together.
Rovo can save you time by defining these things for you, without navigating away from what you’re reading.
If you encounter a definition that you feel is inaccurate, you can edit or add a new definition, then set the visibility to be for that page or work item, the whole space or project, or access your entire organization.
We believe that defining terms using Rovo in Confluence and Jira works best in scenarios when:
- A company has multiple pages in their Confluence instance that mention, describe, or explain what a specific term is for Rovo to reference.
Considerations when defining terms using Rovo
It’s important to remember that because of the way that the models used to define terms using Rovo in Confluence work, these models can sometimes behave in ways that are inaccurate, incomplete or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We’ve found that defining terms using AI in Confluence is less useful in scenarios where:
- You need current and accurate information about people, places and facts.
- You don’t have ample context on the term inside that Confluence instance (for example, if there are no pages that mention the specific term, that term’s definition will not be accurately generated).
- The definition requires access to Confluence content that you don't have permission to view
- You attempt to define multiple terms instead of one individual term at a time.
In addition, in Jira, we've also found that because defining terms using Rovo relies on search in Confluence, the feature will only work in Jira if you have permission to view a Confluence instance on the same site as your Jira instance.
It's also possible that you might find that defining terms using Rovo doesn't perform as expected in Confluence spaces or Jira instances that have content written in multiple languages.
Your data and defining terms using Rovo
We understand you may have questions about how defining terms using Rovo uses your data. This section supplements the information available on our FAQ page.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as the term you want to define.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, defining terms using Rovo applies the following measures:
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third party LLM provider other than OpenAI, Google, or Anthropic on AWS Bedrock
- Are not stored by any LLM provider.
- Are not used to improve LLM models.
- All third-party LLM providers are subprocessors and listed as so on our Subprocessors page. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows existing user access permissions, so users won’t be shown a definition from content they do not have access to. Instead, the feature pulls content and definitions only from pages and projects that the user has permission to view in the instance. If you do not want your content to be available in responses to other users in your instance, work with your org admin to ensure your permissions are set appropriately.
- If a user edits or updates a definition manually, the definition is stored and retained for 1 year.
How AI in editing experiences works
AI in editing experiences is powered by large language models developed by OpenAI and Google. These large language models include OpenAI's GPT series of models and Google's Gemini series of models.
Rovo uses these models to analyze and generate natural language within our products.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text, based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models and Google’s models.
Use cases for generative AI in the editor
Rovo helps drive effective communication across all teams in an organization to improve efficiency, decision-making, and processes.
We believe that using Rovo in editing experiences works best in scenarios like:
- Transforming existing content for different audiences. Rovo helps with changing tone, improving writing, and making technical information easier for other teams to understand. This works best for teams who want to make their writing more professional and concise.
- Summarizing existing content. With Rovo, you can turn rough notes into useful strategy documentation, knowledge base articles, campaign plans, and more. You can also use it to analyze existing information to define action plans and items. This works best for text-heavy pages where there is a lot of context to pull from.
- Generating new content. Rovo helps you draft new content such as strategy pages, project overviews, release notes, or user stories. This works best when teams use clear, specific prompts, with a specific goal in mind.
Considerations when using AI in editing experiences
It’s important to remember that because of the way that the models used to power AI in editing experiences work, they can sometimes behave in ways that are inaccurate, incomplete, or unreliable.
For example, the responses that you receive might not accurately reflect the content they’re based on or include content that sounds reasonable but is false or incomplete.
We’ve found that using Rovo in editing experiences is less useful in scenarios where:
- You need current and accurate, up-to-date information about people, places, and facts.
- You need to have access to information that isn’t readily available to you (for example, in your instance) to properly answer your request.
- You need to generate content in a format beyond standard markdown (for example, generating an info panel from scratch).
- You need to reference information that isn’t already present in the document being edited (for example, content present in another document, or another product).
- You need to generate and transform content in languages other than English.
For this reason, we encourage you to think about the situations when you use Rovo and review the quality of the responses you receive before sharing them with others.
You might also want to think about:
- Being as specific as possible in what you ask Rovo to do.
- Break down complex requests into smaller, more manageable tasks.
- Incorporate relevant keywords to improve the accuracy of generated content.
- Use proper grammar and punctuation in your input text.
- Proofread, review, and edit the output generated by the AI writing assistant for accuracy and clarity.
- Experiment with different prompts or variations of your input text to explore different ideas.
- Collaborate with others to gather feedback and improve the quality of your output.
Your data and AI in editing experiences
We understand you may have questions about how AI in editing experiences uses your data. This section supplements the information available on our FAQ page.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as the product you triggered Rovo from.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, Rovo in editing experiences applies the following measures:
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third party LLM provider other than OpenAI or Google.
- Are not stored by any LLM provider.
- Are not used to improve LLM models.
- All third-party LLM providers are subprocessors and listed as so on our Subprocessors page. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows the permissions in your instance. For example, if you do not have access to a certain Confluence page, you will not be suggested content from that page in the response you receive. If you don't want your content to be available in responses to other users in your instance, work with your org admin to ensure your permissions are set appropriately.
How work item reformatter uses AI
Work item reformatter is powered by large language models developed by OpenAI, including OpenAI’s GPT series of models.
Rovo uses this model to analyze and generate natural language within Jira.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models.
Use cases for work item reformatter
Work item reformatter helps improve the clarity of your Jira work item descriptions by reformatting them using a template developed by Atlassian. This template covers the types of information that we usually expect to see in a Jira work item description, such as a user story, context for the work, and acceptance criteria.
We believe work item reformatter works best in scenarios where your work item descriptions already contain useful information (such as acceptance criteria or links to sources) but that information is not formatted using a clear or consistent structure.
Considerations when using work item reformatter
It’s important to remember that, because of the way they work, the models that power work item reformatter can sometimes behave in ways that are inaccurate, incomplete or unreliable.
For example, your reformatted description might not accurately reflect the content that it was based on, or it might include details that sound reasonable but are false or incomplete.
We’ve found work item reformatter is less useful in scenarios where:
- You need current and accurate information about people, places and facts.
- Your work item descriptions don’t include much information, because there’s a higher chance that the models powering work item reformatter might add information that wasn’t included in your original description.
For this reason, we encourage you to think about the situations when you use Rovo, and always review the quality of the responses you get before sharing them with others. You might also want to think about reviewing and confirming that your work item descriptions include all relevant information before you start using work item reformatter to reformat them.
Your data and work item reformatter
We understand you may have questions about how work item reformatter uses your data. This section supplements the information available on our Trust Center.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as the Jira work item description and summary.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, work item reformatter applies the following measures:
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers
- Are not sent to any third party LLM provider other than OpenAI
- Are not stored by Open AI.
- OpenAI is a subprocessor on our list of subprocessors. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature only uses the information available within the Jira Work item and is invoked from the Jira Work item. Therefore, it can only be invoked by a user who has permission to view the work item and will not access any other information.
Read more about Rovo
How Rovo summarizes work item details in Jira Service Management
Summarize work item details using AI is powered by large language models developed by OpenAI and Google. These large language models include OpenAI's GPT series of models and Google's Gemini series of models.
Rovo uses these models to analyze and generate natural language within our products.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text, based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models and Google’s models.
Use cases for summarizing work item details in Jira Service Management
Instead of reading through long descriptions and numerous comments on a Jira Service Management work item, you can use Rovo to quickly summarize this information for you. This helps agents quickly understand the context of the work item and any progress made, enabling them to take swift action and provide timely assistance.
We believe that summarizing work item details using Rovo works best for:
- Work items with description and comments in English.
- Work items with a large number of comments and/or lengthy comments and descriptions.
Considerations when summarizing work item details using AI
It’s important to remember that because of the way that the models used to power summarizing work item details using AI work, these models can sometimes behave in ways that are inaccurate, incomplete or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We’ve found that summarizing work item details using Rovo is less useful in scenarios when:
- You are using a language other than English
- The work item has no history or details
For this reason, we encourage you to think about the situations when you use Rovo and review the quality of the responses you receive before sharing them with others.
Your data and summarizing work item details using AI
We understand you may have questions about how summarizing work item details using AI uses your data. This section supplements the information available on our FAQ page.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt includes details about the Jira Service. Management work item, such as the work item description, comments, and users involved in the ticket.
- Data about how you interact with our features, such as clickstream data and the people you work with
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, summarizing work item details using Rovo applies the following measures:
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third party LLM provider other than OpenAI, Google, or Anthropic on AWS Bedrock.
- Are not stored by any LLM provider.
- Are not used to improve LLM models.
- All third-party LLM providers are subprocessors and listed as so on our Subprocessors page. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows the permissions in your instance. Only agents and project admins have the visibility to the Summarize button.
How Atlassian Intelligence summarizes Smart Links
Summarize Smart Links with Atlassian Intelligence (AI) is powered by large language models developed by OpenAI, Google, and Anthropic, as well as a combination of open-source large language models (including the Llama series, Phi series, and Mixtral series) and other machine learning models. These large language models include OpenAI's GPT series of models, Google's Gemini series of models, and Anthropic's Claude series of models.
Atlassian Intelligence uses these models to analyze and generate natural language within our products.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text, based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models, Google’s models and Anthropic’s models. For more information on open-source language models, see information on the Llama series and the Phi series.
Use cases for summarizing Smart Links
After you hover over a Smart Link from Jira, Confluence, and Google Docs, Atlassian Intelligence can help you summarize the content, which allows you to determine the importance and value of the link and decide your next action. This reduces the need to leave the current page and switch contexts.
We believe that Summarize Smart Links with AI works best in scenarios where:
- You are viewing a page or work item with one or more Smart Links.
- You are viewing a page or work item with one or more Smart Links that contain a lot of information or dense content, which will take time and attention away from the main content you wanted to read.
Considerations when summarizing Smart Links using Atlassian Intelligence
It’s important to remember that because of the way that the models used to power Summarize Smart Links with AI work, these models can sometimes behave in ways that are inaccurate, incomplete, or unreliable.
For example, the summaries you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We’ve found that Summarize Smart Links with AI is less useful in scenarios where:
- You need current and accurate information about people, places, and facts.
- You need to summarise content that is incredibly short.
- You need to summarise all the metadata and the content in a link. For example, if you wanted to understand all the field values in a Jira ticket as well as its description and comment content.
For this reason, we encourage you to think about the situations when you use Atlassian Intelligence and review the quality of the responses you receive before sharing them with others.
Your data and summarizing Smart Links using Atlassian Intelligence
We understand you may have questions about how summarizing work item details using Atlassian Intelligence uses your data. This section supplements the information available on our FAQ page.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as the content of the link you want to summarize.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any responses you choose to share as part of your feedback.
When it comes to your data, Summarize Smart Links with AI applies the following measures.
- Your summaries:
- Are not available to other customers.
- Are not sent to any third party LLM provider other than OpenAI, Google, or Anthropic on AWS Bedrock.
- Are not stored by any LLM provider.
- Are not used to improve LLM models.
- All third-party LLM providers are subprocessors and listed as so on our Subprocessors page. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows the permissions in your instance. For example, if you do not have access to a page or ticket in Jira, Confluence, or Google, you will not be able to summarize content from that source. If you do not want your content to be available in responses to other users in your instance, please work with your org admin to ensure your permissions are set appropriately.
How Atlassian Intelligence summarizes work item details in Jira
AI summaries in Jira is powered by large language models developed by OpenAI. These models include the OpenAI models described here.
Atlassian Intelligence uses these models to analyze and generate natural language within our products.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated by predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI’s models, or about this approach in OpenAI’s research papers.
About summarizing work item details using Atlassian Intelligence
Instead of reading through long descriptions and numerous comments on a Jira work item, you can use Atlassian Intelligence to quickly summarize this information for you. This helps agents quickly understand the context of the work item and any progress made, enabling them to take swift action and provide timely assistance.
We believe that summarizing work item details using Atlassian Intelligence works best for work items with a large number of comments and/or lengthy comments and descriptions.
Limitations and considerations when summarizing work item details using Atlassian Intelligence
It’s important to remember that, because of the way they work, the models that power AI summaries in Jira can sometimes behave in ways that are inaccurate, incomplete, or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on or include content that sounds reasonable but is false or incomplete.
We’ve found that summarizing work item details using Atlassian Intelligence is less useful in scenarios when:
- the work item has no history or details
- the work item has more than 100 comments (we will summarise only the last 100 comments)
For this reason, we encourage you to consider the situations in which you use Atlassian Intelligence and review the quality of the responses you receive before sharing them with others.
Your data and summarizing work item details using Atlassian Intelligence
We understand you may have questions about how summarizing work item details using Atlassian Intelligence uses your data. This section supplements the information available on our Trust Center.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, including details about the Jira work item, such as the work item description, comments, and users involved in the ticket.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, summarizing work item details using Atlassian Intelligence applies the following measures:
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third-party LLM provider other than OpenAI.
- Are not stored by OpenAI.
- Are not used to improve OpenAI models.
- OpenAI is a subprocessor on our List of Subprocessors. They don't use your inputs and outputs for any purpose besides processing your request..
- This feature follows the permissions in your instance. This feature only uses the information available within the Jira Work item and is invoked from the Jira Work item. Therefore, it can only be invoked by a user who has permission to view the work item and will not access any other information.
How Jira Service Management's virtual service agent uses Atlassian Intelligence
Jira Service Management’s virtual service agent is powered by large language models developed by OpenAI and Google, as well as open-source large language models (including the Llama series) .
The virtual service agent uses these models as follows:
- Intents: Each intent represents a specific problem, question, or request that your virtual service agent can help resolve for your customers. Intents are trained using a machine learning model that’s tailored to your organization, based on a training data set (described in more detail below, under “Your data and the virtual service agent”). This model is used to understand questions from your customers in the context of your organization. Read more about intents.
- Atlassian Intelligence answers uses generative artificial intelligence to search across your linked Confluence knowledge base spaces, and then summarize that information in response to your customers' questions. Atlassian Intelligence answers are powered by large language models developed by OpenAI, and other machine learning models. Read more about Atlassian Intelligence answers.
- ‘From your data’ intent templates are generated using your project’s work items. We use machine learning to group the most common themes, suggest an intent for each group, then generate titles, descriptions, and training phrases for each suggested intent. ‘From your data’ intent templates are powered by a combination of machine learning models, including large language models. Read more about intent templates.
How large language models work: Large language models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text, based on the data that they’ve been trained on.
The large language models used to power the virtual service agent include OpenAI’s GPT series of models and Google’s Gemini series of models.
Read more about the capabilities of OpenAI’s models and Google’s models. For more information on open-source language models, see information on the Llama series.
Use cases for Jira Service Management's virtual service agent
The virtual service agent helps teams automate tier-1 support interactions, powered by a conversational Atlassian Intelligence engine that analyzes and understands intent, context, and permissions to personalize interactions.
Using Atlassian Intelligence, the virtual service agent helps teams scale their service desks and delight their customers with three key capabilities:
- Intents, which are created by your project admin and represent a specific problem, question or request that your virtual service agent can help to resolve for your customers.
- Atlassian Intelligence answers, which uses generative artificial intelligence to search across your linked knowledge base spaces and summarize answers to your customers' questions.
- ‘From your data’ intent templates, which are personalized intent template suggestions generated by grouping your project’s Jira Service Management work items – making it even faster for your project admins to create intents.
The virtual service agent is available in multiple channels, including Slack, Microsoft Teams, the Jira Service Management portal, and more. Read more about which channels are available for the virtual service agent.
We believe that the virtual service agent works best in scenarios where:
- Your service teams receive a high volume of Tier-1 questions.
- Your agents find themselves answering the same questions repeatedly.
- You have a complete, up-to-date linked knowledge base that the virtual service agent can use to offer answers to customer questions.
- Your service teams commonly receive requests that have multiple possible resolution outcomes, which are suited to using both key capabilities in project-level virtual service agents — intents and Atlassian Intelligence answers:
- Intents work best when an action needs to be taken to resolve customer requests (for example, creating a work item for an agent to resolve), or when a back and forth conversation is needed to gather information from a customer. Intents are also ideal for scenarios where you need a fixed, deterministic outcome for specific requests, since intents provide more control over what information is shared with your customers.
- Atlassian Intelligence answers work best when customer requests can be resolved by providing information and instructions, without needing any agent intervention.
- In the case of ‘from your data’ templates:
- Your project has received at least 50 work items, so that the intent template suggestions have access to enough information to be helpful.
- You’ve experienced a recent change in customer behavior, like an upcoming event or even a technical outage, since ‘from your data’ intent templates will identify and group recent, common customer queries.
Considerations when using Atlassian Intelligence answers in Jira Service Management
It’s important to remember that because of the way that the models used to power the virtual service agent, they can sometimes behave in ways that are inaccurate, incomplete, or unreliable.
For example, responses might not accurately reflect the content they’re based on, or they might include content that sounds reasonable, but is false or incomplete.
We’ve found that the virtual service agent is less useful in scenarios where:
- You have a low volume of Tier-1 requests from customers.
- You haven’t made use of all of the virtual service agent’s capabilities, for example by creating intents (or using ‘from your data’ intent templates), or enabling Atlassian Intelligence answers.
- In the case of intents:
- You’ve added training phrases that are too broad or diverse in meaning. Intent matching works best when intents are specific.
- You’ve added training phrases that don’t accurately reflect the way customers typically ask questions. Matching is semantic, and looks for similar meaning between training phrases and what customers are asking.
- You’ve created multiple intents that are too similar. Overlapping topics between intents can confuse the virtual service agent and lead to the wrong intent being suggested.
- In the case of Atlassian Intelligence answers:
- Customers need current and accurate information about people, places and facts.
- Customers need Atlassian Intelligence answers to access information that isn’t readily available to them (for example, in a linked knowledge base that they don’t have access to) to answer a question.
- Your knowledge base is outdated or incomplete, or its articles aren’t detailed enough to properly answer questions.
- In the case of ‘from your data’ intent templates, the data available in your Jira Service Management project work items is of poor quality (or you have less than 50 work items), so that no meaningful groups can be found.
We encourage you to think about the situations when you’d use Atlassian Intelligence, and review the performance of the virtual service agent before you turn it on for customers. Read more about improving your virtual service agent’s performance.
You might also want to think about:
- Encouraging your customers to be as specific as possible when they ask the virtual service agent for help.
- Continuously reviewing and updating your linked knowledge base to make sure it stays complete and up-to-date. Find out how to set up your knowledge base to improve the quality of Atlassian Intelligence answers.
- Proactively reviewing the permissions and restrictions applied to your linked knowledge base spaces so that the virtual service agent has access to the right information.
Your data and the virtual service agent
We understand that you might have questions about how Jira Service Management’s virtual service agent uses your data. This section supplements the information available on our FAQ page.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as a Confluence article from your linked knowledge base being summarized into an Atlassian Intelligence answer, or your historical Jira tickets to generate suggested intents.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, the virtual service agent applies the following measures:
- In the case of intent matching in Jira Service Management project channels:
- Intents (including ‘from your data’ intent templates) are trained via a machine learning model that's tailored to your organization, based on a training data set provided by you. This includes intents and training phrases created by your project admin, or suggested by us within standard intent templates or 'From your data’ intent templates.
- The training data and trained machine learning model developed for your organization won’t be shared with other Atlassian customers.
- Training is done at a project level only, meaning intents are trained on the training phrases provided by admins during the intent creation process, and admins have complete control over the training data. Data is not used across multiple projects on a single site.
- These models are stored against a Jira Service Management site and are deleted if and when you delete your site.
- In the case of intent matching in Jira Service Management help centers:
- We create and store vectors, in compliance with data residency.
- These models are stored against a Jira Service Management site and are deleted if and when you delete your site.
- In the case of Atlassian Intelligence answers:
- Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not available to any third party LLM provider other than OpenAI or Google.
- Are not stored by any LLM provider.
- Are not used to improve any large language models.
- This feature follows the permissions and restrictions applied to your linked knowledge base spaces. This means all pages available to customers on your Jira Service Management portal will be available for Atlassian Intelligence answers to use when responding to customers. For example, if access to a certain Confluence page is restricted and not generally available via Jira Service Management, the content from that page will not be used in an Atlassian Intelligence answer. If you don’t want your content to be available in answers for other users in your instance, work with your org admin to make sure your permissions are set appropriately.
- Your prompts (inputs) and responses (outputs):
- In the case of ‘from your data' intent templates:
- Your Jira Service Management ticket data (inputs) and the suggested intent templates created from that data (outputs):
- Are not available to other customers.
- Are not available to any third party LLM provider other than OpenAI or Google.
- Are not stored by any LLM provider.
- Are not used to improve large language models.
- Your Jira Service Management ticket data (inputs) and the suggested intent templates created from that data (outputs):
- OpenAI is a subprocessor on our Subprocessors page. They do not use your inputs and outputs for any purpose besides processing your request.
How AI Work Breakdown uses Atlassian Intelligence
AI Work Breakdown is powered by large language models developed by OpenAI.
Atlassian Intelligence uses these models to analyze and generate natural language within our products.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated through predicting the most probable next word or text, based on the data that they have been trained on.
Read about the capabilities of OpenAI’s models, or about this approach in OpenAI’s research papers.
About AI Work Breakdown
AI Work Breakdown suggests child work items based on a Jira work item you’ve created, making it easy to break down large pieces of work into smaller ones. Your work item is used as context to generate suggestions for child work item summaries and descriptions.
We believe that AI Work Breakdown works best in scenarios where:
- You start with a work item that has a detailed summary and description, including any constraints, instructions, and definitions of done.
- The starting work item doesn’t have existing child work items.
Limitations and considerations when using AI Work Breakdown
Your data and AI Work Breakdown
Read more about Atlassian Intelligence
How Meeting Agent uses AI
Use cases for Meeting Agent
Considerations when using Meeting Agent
Your data and Meeting Agent
Read more about Meeting Agent
How Agents in Jira uses AI
Use cases for Agents in Jira
Considerations when using Agents in Jira
Your data - Agents in Jira
Read more about Agents in Jira
How Jira Cloud for Slack uses AI
Use cases for Jira mentions in Slack
Considerations when using @Jira in Slack
Your data and Jira mentions in Slack
Read more about Rovo
How Loom Editor uses AI
About Loom Editor
Limitations and considerations when using Loom Editor
Your data and Loom Editor
Read more about Loom Editor
How Loom Screenshots uses AI
About Loom Screenshots
Limitations and considerations when using Loom Screenshots
Your data and Loom Screenshots
Read more about Loom Screenshots
How Create with Rovo uses AI
Use cases for Create with Rovo
Considerations when using Create with Rovo
Your data and Create with Rovo
Read more about Rovo
How Create with Rovo uses AI for Jira space creation
Use cases for creating Jira spaces with Rovo
Considerations when using Create with Rovo for Jira space creation
Your data and Create with Rovo
Read more about Rovo
How the Marketplace app review summary uses AI
Use cases for the Marketplace app review summary
Considerations when using the Marketplace app review summary
Your data and the Marketplace app review summary
How AI Generated Summaries for Work Items uses AI
Use cases for AI Generated Summaries for Work Items
Considerations when using AI Generated Summaries for Work Items
Your data and AI Generated Summaries for Work Items
Read more about Rovo
How change risk assessment with Rovo uses AI
Use cases for change risk assessment
Considerations when using Change Risk Assessment
Your data and Change Risk Assessment
Read more about Rovo
How Forge LLMs use AI
Use cases for Forge LLMs
Considerations when using Forge LLMs
Your data and Forge LLMs
Read more about Forge LLMs
How Request Resolver uses AI
Use cases for Request Resolver
Considerations when using Request Resolver
Your data and Request Resolver
Read more about Rovo
How AI assisted metadata for block templates uses AI
Use cases for AI assisted metadata for block templates
Considerations when using AI assisted metadata for block templates
Your data and AI assisted metadata for block templates
Read more
How Trello card summaries uses AI
Use cases for card summaries
Considerations when using card summaries
Your data and card summaries
How create incident uses AI
Create incident with AI using Rovo is powered by large language models developed by OpenAI, Google, and Anthropic, as well as a combination of open-source large language models (including the Llama series and Mixtral series) and other machine learning models. These large language models include OpenAI's GPT series of models, Google's Gemini series of models, and Anthropic's Claude series of models.
Rovo uses these models to analyze and generate natural language within our products.
These models generate responses based on your input and are probabilistic in nature. This means that their responses are generated by predicting the most probable next word or text based on the data that they've been trained on.
Read more about the capabilities of OpenAI’s models, Google’s models and Anthropic’s models. For more information on open-source language models, see information on the Llama series.
Use cases for create incident with AI
When escalating one or more alerts or alert groups to an incident in Jira Service Management, create incident with AI uses Rovo to quickly pre-populate all contextual information for you to review as part of the incident creation process. This allows users to quickly understand the context of the incident created from those alerts or alert groups, and review and confirm pre-populated information including the title, description and priority of the alert when escalating it to an incident.
We believe that create incident with AI works best in scenarios where:
- The alerts that you are escalating to an incident include title and descriptions in English.
- The alerts that you are escalating to an incident have lengthy descriptions.
- When you are creating an incident from more than one alert.
Considerations when using create incident with AI
It’s important to remember that because of the way that the models used to power create incident with AI work, these models can sometimes behave in ways that are inaccurate, incomplete or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We’ve found that create incident with AI is less useful in scenarios when:
- You need current and accurate information about people, places, and facts.
- The alerts that you are escalating to an incident include their title or description (or both) in a language other than English.
- The alerts that you are escalating to an incident contain only limited details.
For these reasons, we encourage you to think about the situations when you use Rovo and review the quality of the responses you receive before sharing them with others.
To get the most useful results we suggest being as specific as possible in what you ask Rovo to do.
Your data and create incident using AI
We understand you may have questions about how create incident with AI uses your data. This section supplements the information available on our FAQ page.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as the Jira Service Management alert description, title, and priority.
- Data about how you interact with our features, such as clickstream data and the people you work with
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, create incident with AI applies the following measures:
Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third party LLM provider other than OpenAI, Google, or Anthropic on AWS Bedrock.
- Are not stored by any LLM provider.
- Are not used to improve LLM models.
- All third-party LLM providers are subprocessors and listed as so on our Subprocessors page. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature follows the permissions for alerts in your instance. Only agents that have permission to view the alert and escalate it to an incident will see suggestions by Rovo for filling the details for the incident getting created.
How create post-incident review uses AI
Use cases for create post-incident review with AI
Considerations when using create post-incident review with AI
Your data and create post-incident review using AI
How Bitbucket Cloud uses Rovo to generate pull request descriptions
Use cases for generating pull request descriptions with AI
Considerations when generating pull request descriptions with AI
Your data and generating pull request descriptions with AI
How AI generates SQL queries in Atlassian Analytics
Use cases for generating SQL queries using AI
Considerations when generating SQL queries using AI
Your data and generating SQL queries using Rovo
How search answers in Confluence uses Rovo
Uses cases for searching answers in Confluence
Considerations when searching answers in Confluence using AI
Your data and searching answers in Confluence using Rovo
How Rovo searches work items in Jira
Use cases for searching work items in Jira
Considerations when searching work items using Rovo
Your data and searching work items using AI
How AI Work Breakdown uses AI
About AI Work Breakdown
Limitations and considerations when using AI Work Breakdown
Your data and AI Work Breakdown
Read more about Rovo
How suggested topics in knowledge base use Rovo
Use cases for suggested topics in knowledge base
Considerations when using suggested topics in knowledge base
Your data and suggested topics in knowledge base
Read more about Rovo
How Searching Confluence content uses AI
Use cases for Searching Confluence content using AI
Considerations when using Searching Confluence content using AI
Your data and Searching Confluence content using AI
How JQL error fixer uses AI
Use cases for JQL error fixer
Considerations when using JQL error fixer
Your data and JQL error fixer
How Chart insights uses AI
Chart insights is powered by large language models developed by OpenAI, Google, and Anthropic, as well as a combination of open-source large language models (including the Llama series and Mixtral series) and other machine learning models. These large language models include OpenAI's GPT series of models, Google's Gemini series of models, and Anthropic's Claude series of models.
Rovo uses these models to analyze and generate natural language within our products.
These models generate responses based on your inputs and are probabilistic in nature. This means that their responses are generated by predicting the most probable next word or text based on the data that they have been trained on.
Read more about the capabilities of OpenAI's models, Google's models and Anthropic's models. For more information on open-source language models, see information on the Llama series.
Use cases for Chart insights
Chart insights uses AI to help speed up your understanding of data in any chart in Atlassian Analytics. It does so by using the dashboard title, chart title, and chart data (including column headers and row values) to generate a natural language summary of that chart and its data. It will also aim to identify any trends or anomalies to provide you with certain insights into that chart.
We believe that Chart insights work best in scenarios where:
- Charts have many rows of data.
- Charts have a dashboard title.
- Charts have column headers.
- Charts have values in all of the rows and columns.
Bar charts, line charts, and bar-line charts work best with this feature since they typically have trends, dates, and many rows of data.
Considerations when using Chart insights
It's important to remember that because of the way that the models used to power Chart insights work, these models can sometimes behave in ways that are inaccurate, incomplete, or unreliable.
For example, the responses that you receive might not accurately reflect the content that they are based on, or include content that sounds reasonable but is false or incomplete.
We've found that Chart insights is less useful in scenarios where:
- You have charts with one or just a few rows of data.
- You have charts that are of the single value type.
- You have charts missing titles, axis-labels, and column headers.
For this reason, we encourage you to think about the situations when you use Rovo and review the quality of the responses you receive before sharing them with others.
You might also want to think about:
- Double checking the accuracy of the insights with other users who may have more context on the specific data displayed in the chart.
- Taking into account that Rovo is only using the context of a single chart and not the entire dashboard when providing a response.
Your data and Chart insights
We understand you may have questions about how Chart insights uses your data. This section supplements the information available on our FAQ page.
We process:
- Your prompts (inputs) and responses (outputs).
- Context from your instance relevant to your prompt, such as the data in your chart.
- Data about how you interact with our features, such as clickstream data and the people you work with.
- Feedback you choose to provide about this feature, including any prompts or responses you choose to share as part of your feedback.
When it comes to your data, Chart insights applies the following measures.
Your prompts (inputs) and responses (outputs):
- Are not available to other customers.
- Are not sent to any third party LLM provider other than OpenAI, Google, or Anthropic on AWS Bedrock.
- Are not stored by any LLM provider.
- Are not used to improve LLM models.
- All third-party LLM providers are subprocessors and listed as so on our Subprocessors page. They do not use your inputs and outputs for any purpose besides processing your request.
- This feature only uses information from the dashboard you have access to and have requested insights for.
How suggest request types in Jira Service Management uses AI
Use cases for suggesting request types in Jira Service Management
Considerations when using Rovo to suggest request types
Your data and using Rovo to suggest request types
How Focus Area Executive Summary uses AI
Use cases for Focus Area Executive Summary
Considerations when using Focus Area Executive Summary
Your data and Focus Area Executive Summary
How Rovo Survey Insights uses AI
Use cases for Rovo Survey Insights
Considerations when using Rovo Survey Insights
Your data and Rovo Survey Insights