The world of AI agents is still evolving fast, but Atlassian, a collaboration software company headquartered in Sydney, Australia, has been putting agents to work for more than a year, and offering agentic AI technology to customers since April. Atlassian, which had $4.4 billion in revenue last year, sells software including Jira, Confluence, and Trello.
Shihab Hamid, Head of Product for Rovo Agents at Atlassian, says that employees at the company have already created more than 8,000 agents for all kinds of task support. He says that while teams are empowered to build their own agents, “leadership plays a crucial role in fostering that culture of AI experimentation and play.”
Meanwhile, external customers using Atlassian software (from HarperCollins to FanDuel) have created their own bespoke agents based on Rovo’s new technology.
We recently sat down with Hamid to discuss how agents are built and used at Atlassian — and by its customers.
• • •
About a year ago, I was getting so many pitches and ideas from people talking about how agentic AI was the next wave. I can see that Atlassian has dedicated an entire arm of their company to this. Talk to me about how this came to be, including the why, ideation and the process of fine-tuning and delivering on the products.
We were one of the first companies to introduce agents over a year ago, but the journey for us really started more than 20 years ago when we started the company. We made it our business to understand how teams work — common patterns and anti-patterns in team behavior, organizational structures, lines of communication, really figuring out how the best teams work together and collaborate. All of that data has been poured into our data model, which we call the Teamwork Graph.
You can think of the Teamwork Graph as the connective tissue that understands how teams get work done. You can imagine in an organization, you’ve got a bunch of different documents that might be linked to projects which might be linked to whiteboards, brainstorming sessions all the way through to production code releases and support tickets. The Teamwork Graph really brings a lot of these content structures together into one navigable graph, and we can power our agents to query that graph in real time to help with real-time decision making.
The Teamwork Graph is our secret sauce that makes Rovo stand out. Rovo Agents [are] AI-powered, virtual teammates that operate with defined goals. They have character, they have access to knowledge, the ability to execute actions, and this really guides their behavior to accomplishing tasks as well as high-level goals. They can actually take real-world actions using their specialized skills to help teams move their work forward. At Atlassian, we’ve got a variety of Rovo Agents available out of the box to help teams get started and hit the ground running. We also have Rovo Studio, which helps teams craft their own custom agents tailored for their own unique workflows and business processes.
One of our common agents is helping teams understand what other teams are working on, like a team recap agent.
I’m sure this real-time decision making really depends on the context, but are there key themes of what this might look like in action?
One of our common agents is helping teams understand what other teams are working on, like a team recap agent. You can tap into the graph to find out who’s on that particular team, what they’re working on, what they’ve recently deployed.
Another common use case would be, if an incident happens, you can identify the affected services and all the code that has been linked to those services so you can quickly diagnose what went wrong. If you’ve got a support ticket coming in from a customer, you can easily identify the problems that they’re having, the products that they have access to, the historical behavior that they had in the past to accelerate that diagnosis of the problem.
I hear Atlassian has deployed thousands of agents. This sounds like a lot. Can you talk about where these agents live and function?
Well, our latest count actually shows that our employees have created more than 8,000 agents internally, so this number just keeps growing, and these agents have been successfully deployed across the company. It’s not just one particular department or one particular use case. We’ve got agents deployed from engineering to marketing, sales, even HR. [Agents are] really capable of executing complex workflows that are multi-step in nature.
Some examples include automatically organizing your backlog, getting a bunch of customer feedback and organizing them into themes, creating Confluence pages with project plans and figuring out how to get the right tone of voice and level of communication for certain stakeholders, reviewing design documents, helping you brainstorm with your team on a digital whiteboard, giving feedback and iterating. With more technical teams, handling incident management, generating post-incident reports and recommending follow-up actions.
People will just start with a particular agent, and you can really quickly tweak it to make it your own.
It sounds like something that goes into the creation of these AI agents is a deep psychological understanding of how teams work. I’m curious if that stems from people with expertise in that department, or if it’s just a product of the culture of Atlassian.
I’ll give you an example. Our onboarding buddy, which is our Rovo Agent for helping new hires onboard, was entirely built by our HR team. It’s not a team with developers or engineers. That really empowered them to think about their business process. You can imagine being on the receiving end of dealing with a lot of employee requests as they get started, questions about policy documents or company benefits. All of those basic onboarding questions that our HR team would have to deal with day to day, they creatively came up with this onboarding agent. Today, it handles nearly 100 of those inquiries per day. It’s a small agent, but I’m really proud of this because the HR team was able to construct it on their own, understand what would make sense for them and tweak it.
On the other side, our most popular agent that we’ve got, which we use internally day to day, is our Customer 360 Rovo Agent. That’s the number one fan favorite across Atlassian. It’s used by over 80 teams. In Atlassian, we talk to customers all the time. I do customer interviews and I might do a Loom recording. Some people might take notes in Confluence, [a collaborative workspace owned by Atlassian.] Other people would get feedback from support tickets. Sales teams would enter information in the CRM. And what this agent does is it can pull all of the resources through that connected graph and bring it to my fingertips. If I’m having a conversation with a customer, I don’t have to come in cold. I know what they’ve complained about, I know what they’re excited about. I know what domain they’re in and what problems and challenges they’re likely to have.
Are there any other ways that these agents tangibly impact the day-to-day experience of your internal employees and your leadership?
The number one thing that we’re seeing is enhanced productivity. Previously, you’d have to connect the dots yourself. You’d have to take on a lot of the work to identify what’s going on, and then cluster that information. That’s a lot of laborious work that people would have to do manually. Tasks that once took hours might take minutes.
The other thing that we’re noticing is improved collaboration. These agents have access to the Teamwork Graph, and they can identify people and teams from around the world, especially in remote or distributed teams, that might be working on similar concepts who may have expertise in certain areas. That can be hard to navigate and Rovo really brings the power of connection, even in a human context, so you don’t have to sift through information.
Atlassian’s AI agents are unique because they’re used internally and externally. How has the use of these AI agents impacted your customers?
The process for how these customers get on board is they start with a variety of out-of-the-box agents that we’ve got. We’ve got a handful of agents for different use cases, and we see them start to use them and then clone the agents themselves. We’ve got the ability to duplicate agents so that they can tweak the instructions and make it more tailored for their particular use case, for their particular department.
HarperCollins, a leading book publisher, started utilizing Rovo Agents to streamline their operations. Now they’ve reported a four-fold reduction in manual project work. They’ve used Rovo for drafting requirements; tasks that used to take 50 hours are now reduced to one hour. Project management work that previously took an hour after a meeting now only takes 15 minutes. These time savings really stack up.
We’ve got other customers like Pythian, a leading global services company, using Rovo to help identify and distribute company knowledge. This saves about 20 minutes per day. I know that doesn’t sound like a lot, but this translates to hundreds of thousands of dollars for this particular customer. We’ve got other customers using it for a variety of different use cases, like customer support or internal service teams…
…What you don’t want to do when you apply an AI system is to invent new metrics.
Hearing all of these stories about things getting better, I’m curious, how do you measure success, and on the flip side, how do you measure failure of agentic AI solutions?
We take a holistic approach to measuring AI agent success. It’s not one metric or one dimension. It’s important to consider the efficiency gains that automation can provide, but also the user adoption and customer satisfaction — you put in a process that looks faster on paper, but actually produces worse customer satisfaction. Then, thinking about the overall business impact of AI on workflows and productivity. The challenge is that every business process probably already has measures of success, so what you don’t want to do when you apply an AI system is to invent new metrics. You really want to measure the impact of AI on your existing business processes, and then you can open it up to creative ways of thinking about how you might restructure your day-to-day operations. Now you’ve got this new superpower.
An example that we use internally is trying to take AI in bite-sized chunks. We utilize high-touch AI pilot programs in a particular domain. So one of the things that we’ve kicked off is a pilot program for product managers using AI. We’ve started with a small pilot group, and then we’ve identified some tools that we think would be useful for the business processes that product managers run into day to day, and then monitor that group, ask questions, regularly check in to find out how they’re going, and then we expand it over time. In this particular pilot program, we’ve had participants report significant benefits, like 88 percent saving at least an hour a week. One of the stats I’m super excited about is that 94 percent, almost all of them, were gaining confidence with these AI tools and were hungry for more. That’s a way to start and iterate out.
We don’t think that it’s a thing that should be relegated to just developers or engineering teams.
You’re talking about iterating and improving over time. Who is responsible for identifying areas of improvement in AI agents? Is it the teams using them or developers and engineers?
One of the things we’re a strong believer in is enabling anyone to create their own agent. We don’t think that it’s a thing that should be relegated to just developers or engineering teams. The teams themselves understand their business process, and getting a third party to craft something can sometimes introduce friction. [Non-team members] might take more time to understand the process, let alone improve it. We’ve talked to a variety of different customers and the teams that create that space to tinker and to play with a lot of these technologies are the ones that see the most ROI.
There are a variety of agents that exist out of the box — theme analyzer, project recap, issue organizer, brainstorm agents — and then what we find is customers get inspired from these agents. Because we make the natural language instructions available, they can see how we’ve constructed our agents, and they can create their own and that kind of creates the viral loop. That onboarding agent that I described that can be cloned by anyone else in our organization. It’s not just new Atlassian onboarding. People can create onboarding guides for their own team. Just starting out in finance, you might have a specialized set of rituals that you want to go through. I think that makes it pretty amenable to iterating.
What’s been the most important thing to becoming a company dedicated to success in the AI agent space?
…Leadership plays a crucial role in fostering that culture of AI experimentation and play. I’ve talked to a variety of different leadership groups, CIOs, VPs. Everyone’s busy and no one has time to play with the latest tech. I think it’s very important to create space to experiment and play. Our leadership team really loves exchanging ideas on what’s working and not working. We use quick Loom snippets of a prototype created with AI so people can really learn from each other. Setting that tone top-down really makes a difference, because it encourages others to participate.
It’s not just the wins. It’s very easy to talk about all of the amazing stuff that’s happening today, but also highlighting what’s not working so as a team, you can work together. There may be other teams within the organization that have found creative solutions to that particular situation.
You must be logged in to post a comment.