How Atlassian Is Driving the AI Agent-Assisted Workflow
BARCELONA — The results of the same AI-based activities will vary widely by organization, according to Laura Tacho, CTO of DX.
“For quality, I’ve seen some organizations experiencing twice as many failures in production, while others have cut their number in half,” Tacho wrote in a LinkedIn post this month. “In terms of speed, some developers are shipping 3x more changes, while other developers are shipping fewer in order to make it through code reviews. Some orgs see a bump in code maintainability with AI, while it’s rapidly degrading within other organisations, sometimes up to 30%.
“Crunching the numbers into an average evens out this variance and gives a false impression that AI is low risk, low impact,” she concluded.
Three years into this new world of AI, it’s mostly the AI providers, after dogfooding their own products for years, that can boast — and then actually back up — headline-catching, AI-driven productivity gains. AI is, after all, based on data, which should make every organization’s AI strategy unique — and difficult to copy.
In April, at Atlassian’s Team 25 conference in Anaheim, Calif., an enthusiastic launch of Rovo-integrated AI tools for Atlassian Cloud products led with the prediction that any knowledge worker, even those outside technical roles, would soon be launching swarms of AI agents.
About six months later, at Teams 25 Europe, Atlassian’s leaders projected more caution.
They’ve spent this time trying to help their enterprise customers understand and adopt an increasingly larger suite of Rovo AI-backed, Teamwork Graph-backed tools. Because, when planning to train large language models (LLMs) on the internal systems of organizations — which often have thousands of employees — approvals and data access take time.
And even as AI planning demands emerge, Atlassian is still hand-holding many customers trying to bring more workloads to the cloud.
When I spoke with them in Barcelona last week, Atlassian leaders were very open and resolved about this slower, but steadier path to AI success. The collaboration and productivity tooling company now argues that enterprise AI adoption hinges on understanding workflows within and across teams.
On top of this more intentional path, Atlassian is looking to acquire a couple of partners who can help its customers speed up the path to AI return on investment — all while reorganizing the company’s broad tooling suite for clarity. Read on for more about Atlassian’s updated AI strategy.
What’s Holding Organizations Back from AI Adoption?
According to the more than 5,000 respondents to the 2025 DORA report, organized by Google, AI adoption in tech alone has hit 90%. But for enterprises, the AI dots are just not connecting to add significant value — yet.
The Atlassian AI Collaboration Index found that, while a lot is going on in individual perceived productivity, 96% of respondents said they don’t see AI facilitating collaboration among teams. This tracks with an MIT study published in July, which found that 95% of AI pilots fall flat. The MIT study pointed blame at a lack of contextual learning and continued experimentation inside data silos.
A successful enterprise AI strategy rests on the centralization of data within each organization.
“It takes more than individual productivity to get an entire organization working around AI,” Zeynep Inanoglu Ozdemir, Atlassian’s chief marketing officer, said in a press briefing in Barcelona. “One of the reasons the experimental initiatives are not really succeeding is they’re happening in silos.”
This organization-wide context, she contends, is the key to unlocking true productivity gains, but that context is constantly evolving.
“In order to unlock this potential of AI we’re talking about, you need to have your organization’s knowledge connected. And, every day, that needs to help your AI perform better,” Inanoglu Ozdemir said. To accomplish this, the Atlassian suite relies on its Teamwork Graph platform, the machine learning (ML) connective tissue under the software that aims to break down AI silos.
One of the big announcements at Team 25 Europe was that Teamwork Graph not only feeds on data from within the Atlassian suite — including flagship products Jira, Confluence and Loom — but now third-party integrations too, increasing that breadth of internal data.
Atlassian Ascend: Enabling Cloud Access
What hadn’t changed from the previous conference is Atlassian’s continued emphasis that any AI success can only be unlocked in the cloud. In fact, several Atlassian customers I had spoken with last April said they felt left behind from the big Rovo announcement from that event, because they couldn’t stay compliant with a move to the cloud.
Last week, the company announced Atlassian Ascend, which brings clear plans to deprecate its Data Center product. It also comes with a tiered strategy to continue to help every customer get to the cloud via Amazon Web Services and/or now, Google Cloud.
This cloud-forward strategy includes an aggressive compliance roadmap, an expanded data residency and new cloud offerings — FedRAMP Moderate-authorized Atlassian Government Cloud and the single-tenanted Atlassian Isolated Cloud.
“The hype is real. It will just take longer to get there, for users to get there, for technology to get there, for products to get there,” said Mike Cannon-Brookes, CEO and co-founder of Atlassian, in a press briefing. “We’re trying to help the customers have the least crash and the most value, as quickly as possible.”
It’s a slower process than for Atlassian, he’s realized; his company had been training its own Teamwork Graph for almost two years before adopting Rovo. But for every “problem we solve for this customer, there are 50,000 other customers that probably have that same problem, right?” Cannon-Brookes said. “And we are constantly chipping away at that.”
What’s Working for AI Now
This doesn’t mean there hasn’t been AI progress among Atlassian users. There have been 2.4 million Rovo AI agent use cases injected into existing workflows.
“When I sit down with customers, the first thing I always tell them is: Let’s just start with something simple and get you moving,” said Cannon-Brookes. However, “they always try to come right out the gate with a really complex problem to solve.”
Instead, he recommends to them, “We have tens of millions, hundreds of millions, lots of business processes and workflows that run through our platform already. Let’s just find a piece of that that we can make quicker or better with AI.”
“Humans are orchestrating with AI how to work together at scale in a predictable, repeated, constrained way, where humans are in the loop.”
– Sherif Mansour, Atlassian
Starting off smaller with AI leads to earlier success cases, Cannon-Brookes said: “They build up, they get expertise, they get experience, they get confidence.”
At a time when everyone is promising magical solutions, he said, his company is trying to discern how to stand out. It’s been leading with customer success stories; in a majority of those stories, at least one step of a workflow is delegated to an agent, according to Sherif Mansour, distinguished product manager at Atlassian.
Mansour gave the example of someone contacting their mobile provider about a bill. The provider delegates an AI agent to pull up the relevant bill, check the logs and review the data, feeding back to the human customer agent with a link to the bill and any highlights.
Jira and Confluence both include agentic AI templates. But, 90% of the time, Mansour told me, they are being used as a jumping-off point. Each organization has a different context and workflow, which means that most organizations are assigning a step within workflows to an agent, not a whole chain of agentic actions — at least so far.
“Jira really is a way for humans to orchestrate workflow with other humans. What’s changed in the last two years is you can orchestrate that work with coding and noncoding agents,” Mansour said, human and AI. “Automation was just a way to string together repetitive tasks and do it at scale. Turns out, to make these little AI teammates that we have work at a repeated loop, you basically need all the same capabilities of automation platform engineering.”
He described a Rovo agent he set up for Atlassian dev tool sales teams. When they get an inbound request, the AI agent reads the ticket and tells the human agent: “Hey, the opportunity to talk to this customer is in these areas, because they haven’t got these things configured.
“Something that should take hours before, now an agent picks up and does that,” Mansour said. “That’s an agent-assisted workflow.”
Internal Documentation: ‘Explain It Like I’m 5’
Unsurprisingly, the top “boring,” cross-cutting customer use case so far is for internal documentation. Atlassian itself has an AI agent dubbed Eli5 — “explain it like I’m 5” — that allows employees to interact with all internal documentation via a chat interface, to understand any policy updates and how they affect them.
A large banking customer is using it for compliance, which Cannon-Brookes called a really interesting use case because “what the AI magicians will tell you is: It’ll do all your compliance for you. That’s bullshit.”
AI, instead, is great at ingesting information, which is why this customer is using it to read new financial regulations when they’re released. The AI agents can consume all the data, and then tell an individual or a team that pages X to Y contain the updates, he said, “and [page] 458 seems like it might be problematic for us.”
It’s not perfect, he said, but it’s about 90% correct and saves everyone from rereading hundreds of pages.
The Leadership Circle, for instance, uses Rovo agents similarly to pull documentation for sales and customer support to serve the consulting company’s more than 3 million customers.
The AI-Assisted Workflow for Developers
Unsurprisingly, many of the early AI success cases have been within internal software development teams, which tend to be more willing to experiment than teams that serve external customers — but also more able to absorb failure.
These first use cases aren’t necessarily for code generation, but rather applied to repeat tasks or repeat patterns in workflows. For example, accessibility fixes, which are very prescriptive around how to identify these bugs and how to fix them.
One popular early agentic AI use case is generating test cases. “You click on your Jira issue and drag it to: Generate test cases,” Mansour said. “If the customer has decided to implement an agent there, that AI picks it up, reads the ticket, reads the linked pull request, goes and generates a test case, then attaches the results back on the ticket.
“These are the things that humans are orchestrating with AI how to work together at scale in a predictable, repeated, constrained way, where humans are in the loop.”
Within dev tools like Jira and BitBucket, there can often be millions of lines of code,” which, Mansour said, demands context engineering — reducing the context the AI has to the smallest possible set of knowledge. Before directing an AI agent, he argues, a human agent needs to be able to answer:
- What I need to do, usually clarified within a technical and a functional specification or a Jira work item.
- Where I need to do it, with the architectural specification.
“Early on, they didn’t perform well,” Mansour observed. “But then they performed better when the context got more narrow and as the spec context got more defined.”
He has found that workflow proximity matters too, which he defined as: “How the user interacts with AI in the situation they’re in will dictate AI use.”
If it’s a small Jira ticket, he might assign it to an AI agent. If it’s more complex, he said, “I’m going to need to pair with the AI and go back and forth in chat in my IDE.”
Similarly, if a build fails, that conversation and pair programming with an AI agent should happen within a CI/CD tool, he said. If it turns out that a test failed, a developer could then assign an AI agent to fix it right there.
This month, Atlassian announced Rovo Dev will soon be generally available with AI agents for software development, including:
- Code planning
- Code generation
- Code review
- Documentation
- CI/CD
- Debugging
Rovo Dev is available in BitBucket and Jira, as well as now integrated with GitHub, GitLab, OpenAI and Anthropic, among others.
Atlassian Acquisitions to Advance AI Adoption
Another thing that might have slowed the original uptick in Atlassian AI usage was that customers were starting to show caution, too.
AI metrics vary so much by organization — and across organizations. And much of this massive AI investment goes unmeasured.
“AI is changing how developer productivity needs to be measured for two reasons. It’s increasing development productivity massively, but it’s also increasing bills quite quickly,” said Cannon-Brookes. “For a lot of customers, they’re like: ‘Whoa. Suddenly I have all these bills going up per developer, and I want to understand what I’m getting for that.’”
This is one of the reasons, he said, that this past September, Atlassian announced its intent to purchase DX developer insights platform, authors of the AI Measurement Framework, for $1 billion.
“You really want to know how the code is being written, what type of [pull request] cycles are happening, what type of build failures are happening and what type of overall speed and throughput you’re producing,” Chirag Shah, head of product for Atlassian Developer Solutions, told me. “And, at this point, quality and securities have become table stakes.”
“The hype is real. It will just take longer to get there, for users to get there, for technology to get there, for products to get there.”
– Mike Cannon-Brookes, Atlassian
“If you throw AI in the mix, it’s a massive amplifier of both opportunity and problem,” Shah said. “Now you’re trying to figure out that one of your small teams decided to start using an AI tool like a Cursor, for example. Another one was hot and heavy on Copilot because it was so adjacent to your GitHub, and then there are so many downhills across the entire software development journey that you’re constantly looking at.”
Since different teams have different metrics, Shah observed, it becomes subjective, and it’s hard to tell where AI is working or not. Only a mix of quantitative and qualitative data, he argues, can enable fine-tuning.
When CTO Rajeev Rajan joined Atlassian, the software organization began to internally measure “developer joy” over developer productivity, betting that only the first can drive the second.
Over the last three years, Atlassian had built its own developer survey tooling, but then, after realizing that about 90% of its Software Solutions customers were already using it, the software organization piloted DX. Then it made an offer.
DX isn’t even the only intent to buy that Atlassian made this September. Atlassian announced at the start of the month its offer of $610 million to acquire The Browser Company of New York, the creator of the Arc and Dia web browsers. Atlassian has learned more about its own customers over the last six months since the first big Rovo release.
The browser — the way employees are first accessing all of these new AI tools — is still a security and privacy question mark for most enterprises.
“We deeply believe that AI will become part of everybody’s platform,” Cannon-Brookes said. “It’s not going to be an add-on that you buy. It’s not going to be a separate thing. It’s just going to be part of the fabric of everything.”
Sign up now to be the first to receive my free new ebook: AI for the Enterprise: The Playbook for Developing and Scaling Your AI Strategy.
