As artificial intelligence reshapes software development, the role of design has never been more critical. It has also never been more complex. During Atlassian’s Team ’26 Conference, we sat down with Charlie Sutton, the company’s Chief Design Officer to talk about how enterprise software companies navigate the shift from deterministic to generative experiences.
As Atlassian’s Chief Design Officer, Sutton has responsibilities across the user experience ecosystem. There’s design at the product interface side but also on the application developer side. The role carries both practical and symbolic weight. “If a company values the experience of its products, there’s a powerful practical benefit of having somebody who’s ultimately responsible for the quality of that experience,” Sutton explained. “And also there’s a bit of a symbolic benefit that it shows that design is important.”
AI as both tool and material
Sutton frames AI’s impact on design through a historical lens. “The whole history of design is a relationship with tools and materials,” he notes. Just as graphic designers transitioned from physical paste-up to Photoshop to Figma, product designers now incorporate AI tools into their workflows.
During his conversation with us, Sutton also acknowledges that AI is more than a new design tool. It fundamentally changes the material designers work with. “For a long time we had very deterministic software. We wrote the C++, we wrote the shaders, we knew exactly what we’re getting,” Sutton said. “With AI you move to somewhat non-deterministic experiences.”
This shift from predictable to probabilistic outputs creates some design challenges. When the same input can produce different outputs, traditional design principles must evolve too.
Agency versus control in non-deterministic systems
In our conversation with Sutton, we get into the tension between user control and AI autonomy. Sutton sees two things happening there. On the one hand there’s direct control, or the ability to delete or modify something. On the other hand, there’s the feeling of being in control. These aren’t identical, and the difference matters for AI products.
Drawing an analogy to social media feeds, Sutton notes that users don’t control the order in which content appears, yet can still feel a sense of agency through their ability to hide, like, or interact with posts. “What matters is not just your tangible control, but your sense of agency. Do I feel like I’m in control,” he explains.
The challenge gets harder when AI makes errors. In creative applications, a bad recommendation might be merely annoying. In work tools, it can undermine productivity and trust. Context becomes critical for reducing these errors.
The advantage of work context over consumer apps
Atlassian’s focus on work tools provides advantages that consumer applications lack. “When you’re on Netflix, it’s like, what have I watched before? And what might I interact with, it’s actually a pretty small pool for knowing what someone likes,” Sutton observes.
Work environments generate far richer context. “If you did the Netflix analogy, it would be like, Netflix was almost everywhere in your life for 8 hours a day, and you were doing lots of things with it,” he said. This abundance of implicit and explicit signals, such as project histories, team structures, communication patterns, and document relationships, helps AI models make more accurate predictions.
The Teamwork Graph that Atlassian has developed connects these diverse data sources. This makes it possible to add AI features that understand individual preferences and organizational context at the same time, according to Sutton.
Managing token costs through structured data
As companies deploy AI features at scale, token costs have become a big concern. Some organizations have discovered that AI-assisted coding, while productivity-enhancing, carries higher costs than hiring additional junior developers.
According to Sutton, Atlassian addresses this through strategic use of structured data. “One of the reasons that blows up token costs is when it’s having to traverse unstructured data,” Sutton explains. “So if I ask a model a very general question and it has to look in 100 different sources, it’s going to consume a bunch of tokens.”
At Atlassian, the design team creates structured resources specifically for AI consumption, Sutton says. Among these resources are design tokens that codify how the system works, documentation optimized for agent parsing, and templates that guide AI tools toward specific outputs rather than general searches. “We don’t go and say create a version of the Jira interface because it’s going to guess that. We’re actually giving it a specific set of instructions,” Sutton says.
As far as we’re concerned, this is an interesting approach. It demonstrates that effective AI implementation requires thoughtful information architecture, not just powerful models.
From experimentation to value-driven adoption
Like many organizations, Atlassian has progressed through phases of AI adoption. The initial phase emphasized exploration: “Please just use it. So there’s a lot of encouragement and enabling.”
The current phase Atlassian’s in focuses on strategic deployment. “We’re now in the stage where we want people to make the best use of it,” Sutton says. This includes encouraging designers to work more directly with code repositories, submitting pull requests rather than just creating prototypes.
An important distinction Sutton makes is the one between output metrics to outcome metrics. “You can get a lot of output. I could generate 10,000 prototypes for you, but that’s not an outcome. That’s just a throughput measure,” Sutton notes. “We’re ultimately measured by when you pick up an Atlassian product, do you feel like that was a good experience?”
Finding specific high-value use cases accelerates adoption too. For Atlassian’s designers, animation emerged as an area where AI delivers clear benefits. “It’s actually very complex, involves a lot of maths, and models are very good at doing that type of work,” Sutton says.
Principles for AI product design
To navigate the overwhelming number of potential AI applications, Atlassian developed guiding principles. Sutton shares one example: dynamic content forms. “We believe that the content that you see and what you see, its form should be dynamic,” he said.
This principle suggests that users should be able to transform content between formats, i.e. text to audio, audio to video, video back to text. Once established as a principle, opportunities become more apparent. “When you go and look at all the products, you’re like, oh, there are so many places where I could do that,” Sutton notes.
This principle-driven approach helps teams evaluate the hundreds of possible AI features and prioritize those that align with core values rather than chasing every trend.
The full spectrum of design approaches
Atlassian designers employ different methods depending on the task. Redesigning icons remains traditional design work in Figma. Building Rovo prototypes might use generative prototyping tools for rapid iteration. Making specific production changes might involve designers working directly in Bitbucket or GitHub. This flexibility reflects the reality that AI hasn’t replaced existing design methods. It has mostly expanded the toolkit.
The infrastructure challenge
Rapid innovation creates architectural complexity. Effective AI products require flexible infrastructure that enables experimentation while maintaining governance, visibility, and observability.
However, Sutton sees reasons for optimism. Unlike previous eras dominated by proprietary ecosystems, the AI era builds on open, interoperable standards, he notes. The emergence of protocols like Model Context Protocol (MCP) for agent-to-agent communication is an example of industry-wide recognition that interoperability serves everyone. “There’s no one who’s a completely self-contained universe,” he says.
AI as assumed infrastructure
Looking forward, Sutton expects AI to follow the path of previous transformative technologies. “Maybe for those of you who are my age, do you remember the early days of the internet when everything was internet-enabled. In the end, the internet became assumed,” he says. He expects the same will happen to AI: “AI will become assumed, but we still need to make sure that you feel in control.”
Also read: Atlassian takes Teamwork Graph off its leash for even more impact