At the Work Evolution Summit, Atlassian’s Sven Peters took the keynote stage and made a case that ran against the usual AI pitch: the technology is the easy part. He would know. He joined Atlassian in 2011 when the company had 300 people and he was the developer advocate. He evolved into the team advocate as the tools spread from engineering into marketing, legal, and everywhere else. Now, at 15,000 employees and counting, he is the AI evangelist. One thing stayed constant across all three roles, and it shaped everything he said next: people still have to collaborate, and now they collaborate with AI as well as each other.
If you run AI programs and you keep hearing that the models will do the work for you, this talk is the corrective you need.
Intelligence Is Cheap. Context Is the Hard Part.
Sven opened with the thing everyone celebrates and nobody questions. The frontier models get smarter every week. He admitted he cannot keep up with which one leads on any given Tuesday. Then he made the point that reframes the whole game.
“So intelligence is not context, right? We need to add context to those models.”
Those models trained on the internet. They know nothing about how your organization makes decisions, runs projects, or connects one team’s work to another’s. Sven cited a figure that should stop you: 69 percent of workers believe their data is not ready for AI. When your data sits in silos and you point AI at it anyway, you get faster silos. You speed up the exact fragmentation you were trying to fix.
Atlassian’s answer is what they call the teamwork graph, a live map of how your organization actually works. It knows which teams sit close together, which project a work item belongs to, which Confluence page sparked the idea, who contributed, which meeting produced the decision. Connect Google Docs, SharePoint, and Slack, and the graph gets richer. The formula Sven kept returning to is acceleration equals intelligence multiplied by context. Rovo, Atlassian’s AI layer, adds the intelligence through an AI gateway that routes each request to whichever vendor performs best for that task, Gemini for documents, Claude for code, and so on. The context comes from the graph. Neither one accelerates anything alone.
Here is the uncomfortable data he paired with it: 84 percent of leaders prioritize technology over employee enablement. His point was that you cannot solve this with technology alone. The human plus the technology is what makes it work.
The 10,000 Agents Almost Nobody Uses
Sven built more than 80 agents last year. He uses four or five of them, and some only once a week. Then he scaled the confession up. Atlassian has roughly 10,000 Rovo agents across the company. The number of people actually using them sits well below 1,000. Legal would not let him share the real figure, so he rounded generously.
Did building 10,000 agents and using fewer than 10 percent of them mean they failed? He argued the opposite. Every dead agent taught them something concrete. This one was a machine learning problem, not a generative AI problem. That one needed access to a system the teamwork graph could not reach. Another produced flaky results, working yesterday and breaking today. They learned all of this by trying, without standing up a formal project to sanctify each idea first. And they are in crowded company. Sven pointed to the MIT study reporting that 95 percent of AI pilots fail, then welcomed everyone to the club.
His mental model for this gap is worth stealing. The vendor demos promise a robot that folds your laundry and empties your dishwasher. What you get when you try to use it looks like that same robot held together with duct tape, working until it doesn’t. What you should build instead is the boring, purpose-built machine that solves one real problem completely, the way a robot vacuum nails cleaning the floor because you actually have a pet at home shedding on it.
—
About Life in Codes
Life in Codes is on a mission: to support organizations of all kinds to work in a more productive way. That means smart tools, healthy practices, and training the people. As an Atlassian Solutions Partner active in Romania, Estonia, Belgium, UK and the UAE, with our team spread across Europe.
Our client roster includes start-ups, SMEs, large financial institutions like SWIFT, government organizations like the European Commission, and logistics providers such as DHL.Our expertise spans a wide range of solutions, including ITSM, Agile Project Management, Digitalization, Knowledge Management, next-level customer support, DevOps, cloud migration and Rovo AI agents.
We firmly believe that working smart is universal – regardless of industry, company size, or team composition. – our diverse client base is a testament to this philosophy. Whether you’re a tech-focused team or not, Atlassian tools, coupled with our expertise, can significantly enhance your productivity and collaboration.
At the end of the day, we believe in the power of teamwork and we aspire to help people reach their full potential. By partnering with Life in Codes, you’re not just adopting new tools – you’re embracing a more efficient, collaborative, and successful way of working. Schedule an appointment
Four Phases You Cannot Skip
Sven laid out the arc of the talk as four phases every organization has to move through in order: adopt AI, amplify its usage, drive real impact, and design human-AI workflows. You do not get to jump ahead. Most teams stall, and he was precise about where.
1. Adoption Is Table Stakes
Getting people past their fear comes first. Trainings on what a prompt is, what an LLM does, and how to use AI ethically set the ground rules. Hackathons and promptons let teams try, fail, and succeed together. Atlassian packaged its own hard-won lessons into open playbooks, an AI training playbook, an innovation day playbook, a teammate playbook, a fix-it Friday, each a step-by-step recipe for rolling AI beyond engineering into finance, legal, and marketing.
Rovo itself comes down to three things. Search runs across your teamwork graph. Ask for a project status and a product manager named Jamil sees Jira items, goals, Google Docs, Teams conversations, and GitHub pull requests. Ask the same question as Divya in marketing and Rovo returns the same underlying facts shaped for her role, because it knows who she is and what she needs. Same blocker, different framing. Atlassian made that search 60 percent faster over the past half year.
Chat is where you execute. Ask for last quarter’s results as a diagram and Rovo finds the data and builds it. These actions are called skills, and Rovo has many, creating Confluence pages, adding comments, turning a Zoom transcript into a Confluence whiteboard your team can open a week later. If you have ever stared at an empty chat window wondering what to type, Atlassian publishes prompt examples you can copy for things like meeting agendas.
Studio is where builders make agents for repetitive work. This is where context engineering earns its name. Sven walked through the anatomy of a good instruction: start with role and identity, set the goals, then give the details, the output format, examples, and guardrails. To spare people from learning all of that, Atlassian built an agent that builds agents from a plain-language prompt. On stage, Sven asked it to create a bug review agent that flags reports missing reproduction steps and assigns them back to the reporter. It assembled the instructions, knowledge, and tools in under a minute. The live execution failed, which he handled with a shrug and an offer to show it working later. That is roughly the honesty level the whole talk ran on.
2. Getting Agents Off the Shelf
A working agent that nobody uses is still a failure. So how do you move an agent from the shelf into daily work? Sven focused on trust and visibility.
Trust starts with confidence. AI sounds certain even when it is not, so teams simply ask it how confident it is, and it answers honestly. Sven has seen customers wire this into automations: when confidence drops below 90 percent, the task escalates to a human. To find which agents matter, Atlassian is rolling out an Insights dashboard that ranks agents by real usage across manual, chat, and automation, so the automation-driven ticket readiness checker does not drown out everything a person actually chooses to open.
Their long-running favorite is Nora, the onboarding agent pointed at the HR policies in Confluence. Ask Nora what to do when you feel sick and the teamwork graph tells it you are a German employee, so it pulls the German leave policy and points you to Workday. That one agent saved HR countless back-and-forth messages. Sven’s current top five at customers: a risk analyzer a German carmaker uses to flag higher-risk feature requests, meeting notes and tasks, epic-to-user-story generation, a duplication checker that cut one carmaker’s duplicate Jira tickets by 70 percent, and the ticket readiness checker.
He also tracks people, not just agents. Developer Joy surveys and adoption dashboards show which departments cross the 30 percent super-user line. The point is to find champions in sales, legal, and finance, then give them time to help others build. Adoption spreads through people who know the work, not through another mandate from the top.
3. Speed Is Not the Same as Impact
This is the section every executive should read twice. Ninety percent of companies say they have adopted AI. Workers report a 33 percent productivity increase, a number Sven has never seen the industry produce before. Then the floor drops out: only 3 percent of organizations report a significant organizational impact. Tokens burned, faster work, almost no change to the business.
So he asked the question that separates the 3 percent from the rest. Why are you using AI for this task at all? Sure, AI writes an email you would have spent an hour on. Was that hour worth it, when a bullet list would have done the job? Then he skewered the software pitch directly.
“As a software developer, I want AI to take over coding so I have more time writing great specs. Who said that?”
No developer ever said that. The AI vendors wrote that user story, not the people doing the work. His practical fix is to know your value streams and your bottlenecks. If your problem is code review and pull requests pile up waiting, you can generate 10 times more code and ship nothing, because the bottleneck never moved.
How do you find the bottleneck? Atlassian asks. Their Developer Joy surveys pose the same questions repeatedly: how satisfied are you with your speed to ship, your wait time, your access to tools. That surfaces a broad problem area. Then they define a signal they can actually measure, apply AI, and check the data on a short feedback cycle. One example ran end to end. Developers hated wait time, so they measured pull request cycle time, added an AI reviewer, discovered PR size drove much of the delay, cut that too, and reduced PR cycle time by 45 percent. When Developer Joy climbs, productivity in that area climbs with it, because developers want to be productive.
Two more moves make this stick. Atlassian acquired DX to benchmark AI impact against industry standards. And they pair bottom-up measurement with top-down ownership, giving the CTO a goal on developer joy and the CPO a goal on changing how the company works. Out of that came a team hunting 14 hero use cases across departments, each with measurement built in from the start.
4. Designing Human-AI Workflows
The last phase is where the real transformation lives, and where Sven spent his best example. Atlassian embeds agents wherever you already work: assign a Jira ticket to an agent, mention one in a Confluence comment, tag one in Teams.
He was blunt about the agentic-workflow hype. Anthropic recently had 16 agents build a working C compiler in Rust over two weeks with no internet access. Impressive. Then he asked the room how many of them build C compilers for a living. Nobody. A C compiler is a clean demo precisely because the problem is fully specified and settled. Real product work needs feedback from users and coworkers along the way. Nobody writes one spec, walks away for two weeks, and accepts whatever agents produce. The industry already learned that waterfall does not work and switched to agile.
The shift he sees is in where your time goes. Work used to mean short planning, long execution, quick validation. With agents, execution becomes cheap and fast, so planning grows and validation stretches out. You move from producing the work to managing it. You give AI context, validate the output, hand off to the next agent, and design the flow.
His customer interview workflow shows what that looks like in practice. A Loom notetaker captures the transcript. A Customer Insights agent applies a framework Atlassian used for years before AI, observations, problems, opportunities, and generates a Confluence page. A human reviews it, because they refuse to ship hallucinations. A Slack summary goes out in five bullets. The result: product managers now run 10 times more customer interviews, because documenting them stopped being a chore.
Then came the honest catch. All those AI-generated pages pile up and nobody reads AI-generated text. So a weekly Insights Theme Analyzer agent scans every interview, finds the patterns, and drops them on a Confluence whiteboard for the product team’s weekly meeting. Now the discussion runs on data instead of the loudest anecdote from one customer. The team builds what most customers want, customer satisfaction rises, and that, not raw speed, is impact.
The 2011 Lesson That Still Applies
Sven closed by going back to where he started. In 2011, companies came to him proud that they had bought Jira, certified their Scrum masters, and started running sprints, yet saw no gains. His answer then: that is not agile. Installing a tool and running rituals is not agile. Agile is a mindset, a habit of questioning your processes constantly.
Fast forward to 2026 and the script repeats. Companies buy shiny AI tools, run some enablement trainings, host a hackathon, and declare themselves transformed. The hard work is changing how you actually work, rethinking your workflows, knowing your value streams, finding where you generate impact. He mapped it as a maturity curve: explore, optimize, enhance, transform. Adoption and optimization are table stakes. Most teams stop right there and call it done, exactly when the hard work begins, the part money cannot buy.
He ended on the boundary between what AI does and what you do. AI knows more than you could ever read, it reasons, proposes, and executes. It still needs you to validate, discuss, decide, and set direction.
“At the end, it’s my name on that.”
Whether it writes code, an article, or whatever you put out, your name goes on the output. You stay in the loop. You build the human-AI workflow together, and you do the hard work that technology was never going to do for you.
Ready to Work Like an AI-Native Team? Start With Cloud.
Everything Sven demonstrated on that stage lives in Atlassian Cloud. The teamwork graph, Rovo search and chat, Studio agents, the Insights dashboard, the human-AI workflows, none of it runs on server or data center deployments. If your organization still runs on legacy infrastructure, these capabilities stay locked away while other teams put them to work and pull ahead.
This is where Life in Codes comes in. We migrate you to Atlassian Cloud, and we push hard to make it the smoothest move your teams have ever gone through. We run the assessment, build the migration plan, execute the move, and clean up your instance afterward, all without the downtime and data loss you are worried about. Your people keep working while we handle the migration. The day you land in Cloud, you unlock the exact AI-native workflows this article walked through.
Want your teams on Cloud this year? Talk to Life in Codes and we will plan your migration from wherever you stand today. We have done this for organizations of every size, and we will do it for yours.
Follow Life in Codes for more insights from the Work Evolution Summit and practical guidance on turning your Atlassian tools into real impact. Do not let another quarter pass watching these capabilities from the outside.
