Jira AI work breakdownNov 2023 – Apr 2024
Jira helps 300k+ teams plan, organize and track their work. Together they create 10M+ work items each month.
During a design secondment, I designed a pilot experience that used AI to help teams break large pieces of work into actionable work items.
Context
Team leads use project requirements to identify several major chunks of work within a project called Epics, which are further divided into actionable work items.
- Task
- Story
- Bug
Problem
Defining every work item was slow, manual work. How might we make Epic breakdowns faster without taking control away from team leads?
Primary research
We surveyed 20 teams to understand how they broke down Epics. Our sample included 10 internal Atlassian teams and 10 external customer teams such as Deutsche Bank, H&M and Coles.
We found that:
- Projects typically contained 6 to 8 Epics.
- Teams spent an average of 154 minutes breaking down each Epic.
- Each Epic contained an average of 12 work items.
Solution
We used AI to generate an initial breakdown of an Epic into child work items.
I began designing the experience around the assumption that users would want to guide the AI through prompts. The AI used the Epic description field as its source of context.
We tested the feature preview on Atlassian’s internal Jira instance. We observed that users often captured relevant information about an Epic in linked Confluence documents in addition to the description field. So we added Confluence as an additional context source.
We observed that most user prompts specified the types of work items they wanted. So we replaced the prompt with explicit options to generate Stories, Tasks, or Bugs.
This meant the models could no longer depend on detailed prompts to produce reliable breakdowns. Drawing on my engineering background, I curated quality examples of Epic breakdowns from our internal Jira instance. Once I had validated the dataset with the team, the AI team used it to fine-tune the models for the task.
We shared the refined experience with internal and external teams. Their feedback showed that users wanted minimal friction without losing the ability to customize the output.
In response, the final design automatically broke the Epic down into work items. Users could then refine it through prompts or edit individual work items directly.
We also added controls for users to like or dislike the generated output. Their responses provided a signal on output quality and additional data for training the models. We tested the updated experience again, received positive feedback, and then shipped it to all customers.
Impact
The shipped experience met both work-item acceptance thresholds while reducing the time required to break down an Epic.
- 68% reduction
- Epic breakdown time for test participants, from 154 minutes to 49 minutes on average
- ≥50%
- Acceptance rate for generated child work items without modification
- ≥80%
- Acceptance rate for generated child work items with some modification
Atlassian’s president presented this experience at the annual company keynote in May 2024.
Future improvements
This was a pilot project, so some improvements were left out of scope for future iterations.
Token usage
Improve prompt design, retrieve only relevant context, and use lightweight models where appropriate.
External documentation sources
Support third-party documentation sources, such as Google Docs and Microsoft Word, alongside first-party Confluence.
Semantic search
Automatically find relevant documentation and work items to enable more accurate work-item breakdowns.
Learnings
- Continuous user feedback helped us refine the experience as we learned how people actually wanted to work with AI.
- Listing my assumptions explicitly made it easier to challenge them and adjust the design as new evidence emerged.
- Collaborating with teammates across disciplines brought perspectives that helped us balance user needs, technical constraints, and product goals.