Scaling Enterprise AI with the Teamwork Graph and Rovo | Team '26 Europe | Atlassian

0:08 Please welcome Atlassian head of product information protection and public sector Rob Betett. Good morning or should I say good afternoon. Overall, I'm Rob Bisset, head of product for information protecting and public sector here at Atlassian. I'm going to be joined by Jamil Valani, our head of product for AI. He'll join us a little bit later. This morning, we saw a view of the art of the possible for human and AI collaboration.

0:43 Our mission for the next 45 minutes or so, we'll see how that goes, is to show you how you translate that vision into practical reality. So, let's start because it's fun with an uncomfortable truth. Most of the AI tools that we're using today are really designed to supercharge individuals, but they can sometimes leave teams behind. when your teams produce 10 times more content in parallel in vacuum, people can get left behind, right? Trying to keep up, they get overwhelmed.

1:14 So, while individual work may move faster, handoffs sometimes slow to a crawl. Our state of teams report shows this pretty clearly. If you look at the data, only 6% of executives have been able to point to orwide ROI on their AI investments and 87% of knowledge workers say they're trapped in tactical mode versus driving strategic value. So instead of eliminating friction, what we're end up doing is creating a fragmentation tax that is a drag on the organization. So this uncomfortable truth can often lead to an uncomfortable conversation. and I'm sure many of you have had this. We're spending millions of dollars on AI. What do we have to show for it? Right? Which is a challenging question to answer. So, let's dig into that a little bit.

2:00 AMP, our Agent multiplayer protocol that Mike talked about this morning, was designed to solve this problem, specifically this problem. For over 20 years, we've studied how work actually happens in teams and across organizations. We understand the real world constraints that you face every day. And honestly, as a as a multinational corporation, we live them, too. That experience translates into the tools and technologies that we build to help you solve for these challenges. At the foundation of AMP is the Atlassian platform. It delivers the key components for safe and effective human and a AI collaboration. There's a couple of pillars that come with this. The first is governance. It's the basis for safe adoption. If you can't trust your platform, it's difficult to put your data there. The second is context, the teamwork graph that fuels every agent with over 250 billion enterprise relationships. And the third are the workflows powered by ro by rovo, the AI layer embedded in the platform that drives the work that happens every day.

3:03 From a single line of code in a commit all the way to a multi-year executive road map, we help everyone be more effective. And true to our open DNA, we're not trying to force anybody into one box. Your teams get the benefit of the governance, the context, and the workflows regardless of where they work, which tool, which harness. In fact, our teamwork, our latest teamwork labs research found a 46% reduction in that same fragmentation tax amongst top performing teams. And those same teams are nine times more likely to increase collaboration and six times more likely to plan and prioritize faster. This is the fuel that powers teamwork. When you bring all these things together, the silos disappear. So, let's kind of take a take a look at what that looks like.

3:50 So, to get the real value from AI, the value we've been talking about, you need to have access to all of your organizational context. I think we've talked about that pretty extensively. But you can only unlock that context if you know that your data is safe, secure, and in your control. Right? That's a really foundational important component of this. We've enabled this with our enterprisegra grade platform designed specifically to meet the needs of highly regulated customers. Right? We're focusing on those highly regulated customers to make sure that we can support everybody. And this gives you the trust basis on which to build your AI powered system of work.

4:24 That's why today over 85% of Fortune 500 companies including many of the world's most heavily regulated firms already trusted Lassian. That means that we meet their requirements and I'm very very confident that we can meet your requirements as a result of that. Why? capabilities like SSO, data residency, customer managed encryption keys, audit logs, they're all baked into the platform from day one. You don't have to build them. You don't have to add them on. You don't have to integrate them.

4:50 With Atlassian, you're not having to choose between the AI innovation that you're trying to deliver to your teams or enterprise trust. Right? Ultimately, we can give you both. This isn't a claim, right? We can deliver it with proof. We have a long list of certifications that represent the depth and capabilities of the platforms. We've called out a few here that are most relevant today in Europe and we're continuously investing to expand this, including things like complying with the EU AI act and we're happy to announce that we've achieved ISO 4201 as part of that process. But we're not done with that work. We'll never be done with that work. We are on track to deliver ACN ENS and 27 2017 before the end of the year.

5:34 All right, that's a lot. So now we know our platform is secure and compliant. We can move ahead and start to think about the data layer. Right? We have a safe platform. So now let's move into data. Many of you have 10, 15, sometimes 20 years of data inside your Atlassian estate. And that data can be very very messy. Hands up if you think your hands if your data is a little bit messy. Right? I I think it's probably true for everybody, right? It holds legacy credentials, unmasked data like credit card numbers, proprietary source code, tokens, confidential data, all kinds of things.

6:11 This is where Atlassian Guard comes in, right? Many of you know Guard today for identity and access management. Many of you are probably using it. But Guard has evolved to protect the full data layer and that's our mission. It helps you discover, classify, safeguard sensitive data across the whole Atlassian estate. So the context that makes AI powerful can be deployed responsibly. So picture a scenario. Two years ago, one of your customer service agents was working on a customer issue. They copy and pasted some passwords and tokens into a Confluence page to share with the team as they work through it. Okay. The good news is you don't have to worry about this. They got the customer issue resolved. The bad news is they forgot about the page.

6:52 Okay. That page became dormant in your corporate history. The minute you start to enable agentic workflows and you start to enable search across or AI and AI powered search across your estate that goes from dormant data to active risk. Okay, so how do we start to think about that? This starts with a with a content scan. I started this a couple of days ago because when you're scanning across a full site, it can take a little while to run. So you create your scans and you run them against your estate, part of your estate, all of your estate. This is powered by our full-sight scanning engine that was released a little bit earlier this year. It works across Jira.

7:28 It works across Confluence. It works across attachments and it's powered by an expanded set of detections that we are continuously growing including to expand our regional footprint to support various countries specific NHI PHI types of data. So you can choose which detections to scan, you can choose which data estate to scan, and then we can take a look at the results. We know that sometimes these scans produce large amounts of data. So we're constantly trying to improve the quality of our detections. We're also launching semantic scanning and exact data match a little bit later this year to help you very precisely narrow in on the types of data that you want to search for. And that's over and above our existing ability to create custom detection. So you can tweak these any way you want.

8:16 Ultimately, once your scan runs, we can go look at our findings page. Findings is a new page that we've just recently launched that brings all of those results back together into a single view. We can see in this case 8380 findings. This must be a newly migrated customer with 20 years of historical data that they're trying to clean up. We can go through, we can see the type of detection, the object that we found the detection on, the category, so on and so forth. We can click into any particular detection. We can start to see the context of where the detection happened, the details of that particular detection and then we can take action on that. In the case of a content detection like a confluence page, you can just go ahead and redact that, right? Just mask that data out. If it's an attachment, you can go ahead and delete the attachment so that it's no longer a risk factor across your entire state. Now, nobody really wants to go through and manually deal with all of the 8,233.

9:19 So, we've created a new type of policy called an information protection policy. And these allow you to move to create proactive policies to deal with these things at scale and effectively in real time. So, we can create a token policy, right? Right? If we reme remember if we remember back to our example where the customer service agent posted some tokens into a page, we can go create a policy to deal with that particular type of issue. So we select our information protection policy. We create that.

9:51 We can scope that policy to the whole organization or to specific sites and projects. We can add conditions to that policy. In this case, we select the detections that we want to fire off. We know that we're interested in tokens. So, we can search for the token detections and we don't want any kind of tokens. So, we can add the condition and then ultimately we can add a control. So, what control do we want to take? Do we want to create a violation which sends an alert and the signal out to your sock operations teams or do we want to go back and redact? In our case, we want to go ahead and redact because we're wild and fancy free. Okay.

10:31 Ultimately, we create the policy and you can see we can put all kinds of conditions and factors and limitations in that. And then we can go ahead and save that policy. We can save it as a draft if you want to have a peer review it or sign off. because these policies can have data changes. You can set it to a monitor to observe the behavior before you make it live or we can live again fast and fancy free and make that live. So, we'll go ahead and activate that policy. It goes in and creates the policy and then that will automatically start enforcing. So ultimately what does this look like from a user perspective? Very simple to see, right? We go back to our Confluence page. obviously you on a day-to-day basis don't want to do this. but it's super fun for the purposes of a demo.

11:14 And we can paste a token in there. And you can see that that gets redacted in real time. So we have real-time near realtime redaction depending on whether it's a live edit page or a page that you edit and save. Right? We can do both. Ultimately we can mouse over that redaction and it will tell us why it was redacted based on a guard security policy. So if a user has a piece of data that gets redacted they know why and they can reach out to the team if they want to deal with that. Even more interestingly, with our Robo Chat security capability, we can go ahead and actually do this in chat. So, in this case, I'm saying, "Hey, let's just go check if this token is live." And if I if I send that, you'll see that Robo actually intercepts that. And what's really interesting about this is that it's not the LLM that intercepts it. We have pre-processing on the prompt that does the redaction before it gets to the LLM. So that token never actually reaches the LLM and doesn't become part of the process corpus of knowledge.

12:12 Okay. So that's a prelim interruption of that data. Interestingly enough, if the LLM happens to give you a token back, it'll redact that as well. Right? So that's birectional. Okay. So this is some new features that we've delivered from a data governance data protection perspective. Let's move on to how we govern our agents. We have a secure platform. We've confident that our data is now prepared for AI. We're now going to unleash aic experiences across our teams. As part of their workflows, the teams are going to deploy native Atlassian and external agents. And we saw some great demonstrations of that this morning. Now you can manage those agents as if they were part of your team. Our core pillar of AMP is treating agents as first class teammates, which Mike talked a lot about. So we've built a centralized control plane for all nonhuman actors.

13:02 So we can walk back over here and take a quick look at what this looks like in real life. So again we'll come over into Atlassian administration. I built a page to send an executive update or I had robo build a page to send an executive update to my leadership on a project that we're working on. As you do before you send such an update, take a quick look through it. Make sure that nobody put a comment on there. Change the data before you send it. And let's just say for example, I've noticed that JL's page assistant posted an update. Now, his feedback may be good, might not be, I don't know, it's hard to say, but I don't want to send this page to my executive leadership because this agent's posting stuff on it. I don't know where it came from, right? So, this sort of a bit of a contrived example of an agent that's behaving in a way that's unexpected. So, as a user, I might log a ticket with my admin, right?

13:53 Fortunately, I'm also an admin. So, I can show you how we'll deal with that. Ultimately, we have a nonhuman identity dashboard in the system. Now, this shows us all of the agents, and it doesn't matter whether it's an internal agent or an external agent, right? Because that agent has an identity, it's registered with the system, and it shows up here. I can go search for them. we can filter by status or type, whether it's delegated or autonomous. we have all of those capabilities. but in this case, I happen to know that it's Jamil's agent. Okay. So, we can look at this here. We can see what's going on.

14:28 and I can click into this to look at the agent audit log. And now we can start to see the activities that that agent has undertaken. Again, because we're treating this like a human, we have the same user behavior and user action logging that you would have for your human users. So, I can start to look at this and I can say, okay, like it it's actually spamming a bunch of pages. This is problematic. So, in one world, I could just slack Jamil. I mean, he's backstage. He shouldn't be on Slack, but that's okay. We could slack your mill and ask him what's going on, right? But as an administrator, this might be an active security incident. So now we can just go ahead and deactivate them. The same way that you would deactivate an active user, we can go and deactivate it. That removes the permission from the agent to be able to go and act on J's behalf, on anybody's behalf, so that we can now go troubleshoot it. And perhaps there was just a condition that wasn't correct, or we just need to to tweak that agent a little bit. And once we're confident that we fixed that agent and it's properly scoped, we have the ability to reactivate that agent same way you would with a malicious user. Okay, so this is a foundationally new capability to allow you to control your agents in a very granular way to see what they're doing so that you can have trust in the system as we're going along and that's all within Atlassian administration. It's not a new capability. It's in the core platform there. So what you see here is the benefit of a single integrated platform where enterprise trust, compliance, security controls are tightly integrated and brought together.

15:53 Right? With the Atlassian cloud platform, you get one control plane for every human and and agent actor across your entire estate. You have one policy framework to enforce data boundaries across your entire estate. And ultimately that gives you a single security posture spanning all of your tools and experiences that are powering your work. That centralized platform allows you to stay ahead and upto-date as the as the world evolves. We're excited about the capabilities we've already delivered and we demoed a few of them here live in this session. But Ging our isolated cloud, our single tenant SAS solution a little bit earlier this year. delivering the full-sight scanning, expanded detection, semantic scanning, exact data match and guard that we just talked about, and launching the new non-human agentic controls, which we just took a very quick look at. We're also pleased to announce, and Mike talked about this this morning, that we're taking regional options even further by announcing our EU based AI inferencing, ensuring that your LLM inferencing, particularly here in Europe, stays in Europe, so that you have the confidence to run the AI workflows that you need.

16:59 This is a pretty a pretty important feature as we move forward. This is really just a snapshot of the work that we're doing, right? We're going to continue to invest in building and maintaining that trust across four core pillars. Okay, so what we're thinking about as we move forward and the first is going to be around sovereignty and compliance. And you've heard this word a couple of times already. We're going to build more capabilities in more regions and we're going to continue to support more regulatory frameworks. That work will never stop. The second is data governance. We're going to provide deeper coverage. We're going to continue to in invest in more dynamic threat detection and more controls and capabilities to help you manage your data estate. The third is with AI and agent governance. We're going to continue to invest in agentic inventory, expanding agent audit logs, and building dynamic threat modeling for agents so that you can have the confidence that you can control those agents as again the technology continues to evolve.

17:55 Finally, we're going to continue to invest in business resilience, trusted change controls, expanding backup restore, and business resilience capabilities, and expanding into data life cycle management. Ultimately, you can be assured that when you're choosing the Atlassian platform, you're building a on a proven platform, but b a platform that is also evolving dynamically to address the needs of the market as it continues to evolve. So with that I'll hand over the stage to Jamil head of product for AI and he's going to dig more deeply into the context and robo layers. Thank you very much.

18:42 >> Thank you, Rob. And I appreciate you getting my rogue agent under control. good afternoon. Thanks for coming. I hope you all are energized to be here. Rob spoke about how governance gives enterprise AI permission to operate. Now, let's talk about context, the rocket fuel for AI. Think about your most effective teammate. When a critical project lands, they hit the ground running because they have all the context. Who owns the decision? Which Slack thread overturned last week's meeting? And which customer dashboard is the actual source of truth?

19:16 Now contrast that with the most with most of the AI tools today. Even the best agents on the most powerful frontier models only get basic knowledge but not comprehension without the teamwork graph. As a result, agents are burning thousands of tokens as they try to piece together the most relevant information or they're just leaving it to humans to untangle the mess. The teamwork graph provides agents an essential living map connecting your team's goals, knowledge, communications, data, code, and more. It's not a static database you build once. It's a dynamic engine where your context continuously compounds in four stages.

19:57 First, it ingests data through over 80 connectors, enriching the graph as teams work. Second, it synthesizes context across tools, uncovering hidden connections between data that is easily missed. Third, it actively measures and improves graph health in real time, catching scale documentation and broken links before they impact answers. And fourth, it connects that unified context into your daily workflows wherever your teams are working via Atlassian MCP and the teamwork graph CLI. The entire system gets smarter the more you use it.

20:39 To give admins complete visibility into the health of their graph, we are introducing the context quality score. Here admins can visualize how apps across the company connect across the company, spot coverage gaps and get actionable recommendations. For example, on the screen here, you can see that the graph can be enhanced with more context. It immediately recommends adding a key data source and within seconds you have a clear path to enrich your team's context. That context can now power the tools your teams rely on every day.

21:14 The first place most of you will see this context in action is in search, the front door where millions of people begin their work. Every Atlassian customer has worldclass enterprise search powered by the teamwork graph at their fingertips. You can access it from every search box, be it in Jira, Confluence, Loom, the desktop app that you saw this morning, and the mobile apps on your phone. Search is your gateway to work and the foundation for every agentic workflow. We've made search 40% faster. So, let's take search for a spin. So, when you click into search, it doesn't just wait for the keywords. You can see here that it's proactively surfaced suggested actions, recent docs, and cross app activity.

22:03 Now, if I start typing Nova, my project name, I get personalized suggestions immediately. In this case, you can see relevant pull requests from GitHub, some Jira issues, and many more items. Now, if I keep typing my full query, Nova Launch Blockers, since I'm in Jira, I get the option to just search Jira, or I can do a full search across the Teamwork graph, which is what I'll do here. Now, while this is looking, you can look up at the top and you'll see all the different places we can filter by. Just like you saw in the Revo desktop app this morning, you can click through Confluence, Google Drive and you can see all the other items here, all the other sources you can filter by here.

22:51 Using these makes it easy to find context from across your apps. Now when I look at all apps here, I get a quick status update of the Nova project and its launch blockers pulled from Jira at the top right here. Now, it looks like there's a critical blocker up here, and I'm pretty sure we're working on it. So, I'm going to go ahead and do another search here to ask for progress on that issue. Progress on usage base building in Nova, which I tried earlier.

23:24 Now, in addition to the rich results that come up at the bottom that show even status updates and all, at the top, I get a smart answer powered by Robo. This is an actionable AI generated answer grounded in the apps my teams use every day. You can see here it's pulled context from Slack, Google Docs, Zoom and more. So if I look at the issues here, I see that there is an issue related to this bug here, retry log failure. And I can actually highlight this right from here and pretend like I have a computer science degree, which I do, just a little old, and go directly into our code search app to get more information.

24:13 That was the Rover browser extension, making any Robo search directly accessible through my browser. Now the code search app searches across your repos including Bitbucket, GitHub and soon GitLab. It gives you a single place to search and pinpoint the exact lines of code that are triggering this error. Now I know in this case that my team is using GitHub and I just want to straight up string match. So I'll do lexical search. And I can see here that not only is my project Nova, which is down here, have this error, but also project strawberry has several files that reference this error condition. So I can actually go and load some of these functions here and some of these files here. And I can see that there's a function here that looks interesting. And I can see that this function has is referenced in multiple places. That gives me a suggest suggestion that project strawberry has actually handled this error condition before and I might want to actually use it as an example to fix project nova.

25:23 Now normally after discovering this I begin digging through commits or setting up a call with the strawberry team to figure out how they approach the problem. But instead I can just use robo chat. I'll go ask here in just natural language to compare the two implementations between Nova and Strawberry, identify differences, and recommend a plan for how I can solve the issue. Now, this is super useful for me for people like me who aren't living in our codebase every day, but need to use code to make better decisions or get things quickly unblocked. Code search lets me be precise when I know what I'm looking for with controls filtering by repo path source code control system forks and more. But it also supports semantic search which you heard about this morning. So I can just search by concept. Under the hood, Rovo isn't just doing a surface level text match. It's actively reasoning across both services, inspecting the recent pull requests, analyzing the billing implementation, extracting a proven fix which might account for project Nova's specific dependencies. And here we go.

26:36 So I can see here that there is a full recommendation analysis of the bug root cause here, the comparison of the code and key differences. so I can learn about what Project Strawberry did better. And if I scroll all the way down, the recommendation plan and even a suggested prompt that I can toss into Claude to go fix this. Now, I can even try to hand this item directly over to Claude now just by mentioning it, but I won't get carried away right now.

27:12 software fixes are just one piece of the puzzle. What if you need help with broader multifaceted projects like spinning up a working prototype, designing a customer web page, or drafting an executive presentation? For that, we built Roork, which is now available in beta. Rovo work takes you from accelerating simple tasks to delegating substantial projects. You don't need to juggle 20 tabs or put together a coalition of the willing in your team. Robo can just help you drive the project forward together. It can produce sophisticated interactive sophisticated interactive reports, interactive media, extensive analyses, and even code. It's capable of delivering high-grade publish ready content from across content formats.

27:58 Now, you saw a simple example in the keynote this morning of what it can do, but let's try to give it a bigger challenge. So here this is a page where I jotted down some rough early ideas on expanding robo to frontline workers. I want to pressure test this before sharing it to the wider team. So I can open up ask rovo here. Select the work mode you see right here and issue a simple natural language prompt.

28:37 So, a proper pressure test involves a lot of steps. Building a business model, sifting through customer feedback, sizing up the market, getting lots of stakeholder input. It can be so overwhelming that people resort to building on gut instinct resulting in costly mistakes. But we're going to see here work can actually help. Let me go make this full screen just so we can kind of follow its progress.

29:08 And you can see here it's starting its work. Now, while that's running, give it a minute. Let's look at what's actually happening behind the scenes. Robo's gotten straight to work. First, it's mapping out an execution plan. It's spinning up a virtual secure sandbox in the Atlassian cloud. It's acquiring the exact tools and skills it needs for the job. And from there, it's writing and running code, doing its data analysis, and if it doesn't work or it runs into an error, it self-corrects on the fly, iterating in real time until it can deliver a publish ready artifact. So, Robo isn't just searching for documents that mention Frontline. It's following connections. It's connecting a P. It connects a painpoint mentioned in a recent customer call to an open request in Jira to a confluence proposal to Slack debates happening across the team.

30:02 This nuance usually gets lost without access to or or context. The teamwork graph is filling that gap. It knows how all this work connects, allowing Robo to follow all the relevant threads it discovers rather than guessing. And crucially, it only accesses what I'm permitted to see. So our governance model stays completely intact. So let's check back in, see how it's progressing. So here it's actually shared a plan. So I can actually analyze this plan, see all the different sources. It's it's going to read through how to synthesize and pressure test.

30:39 And I can actually steer its focus. That's real human and AI collaboration. Robo is doing the heavy lifting, but I stay in control. So let's say here I want to modify this plan and just ask it to break out the EU numbers separately. I've modified the plan and Rovo can go ahead and get to work. Now, this plan takes this project will take it a little bit of time to run. So, let's go and actually show you what has actually done when I ran it earlier.

31:14 This is what it delivered when we ran this same request just before Showtime today. It's an interactive app that my team can actually play around with. I got this clear executive summary at the top, a full opportunity model, so I can actually play with it. So, for example, I can change my assumptions for year one adoption and see how the rest of the model plays out. Same with deployment coverage. All these variables I can play with. That's normally something you'd have to go and manage an Excel spreadsheet somewhere. I can even see as I scroll down, you know, customer evidence and that EU breakout I requested in the modified plan.

31:52 All this is sourced from multiple places across the teamwork graph and is an artifact that's ready for my team to use and pull together. It would have normally taken me weeks to get this ready to go, but now within just a few minutes, I was ready to go straight into a planning review. Now, reviewing this artifact in the artifacts app, which was introduced just a few hours ago, a couple of things I'd like to point out. first is that the versioning I can actually see here the version history of this artifact.

32:27 So if somebody else wanted to work on it in my team through whatever favorite agent they have this versioning history would get upgraded and I can always look forward dial back as much as I'd like to see how work evolved. We've built strict human in the loop governance so agents can propose changes but they cannot overwrite or delete your work without your say so you stay completely in control. Now, the other cool thing is that I can take this URL and just paste it directly back into the original confluence page.

32:57 Just simple copy and paste. Each artifact is a smart link in the Atlassian platform, which means I can do what I just did, a simple copy and paste of the artifacts URL in any Atlassian app like this confluence page, and you'll see the full artifact render in line. It also is now searchable in robo search as well. So the deep organizational intelligence living in your teamwork graph isn't locked inside Atlassian. It's also completely portable. It is able to be with your teams wherever they already work. So to see how effortless this is, I'm going to show you how you can bring your teamwork graph into cloud.

33:38 Now with the teamwork graph, your organization's live work comes alive in cloud with just a few clicks. You can go to Claude's connectors or plugins screen. Simply search for Atlassian add and then click to enable. It's that simple. Instantly, Claude has access to your graph. So, if we flip to the terminal, you can see all the available connections, giving you an immediate gauge of your context depth across all the apps your teams use. Normally connecting external AI tools into your enterprise workflow requires configuring these connections many times over. But with the teamwork graph, you get the entire connected ecosystem in just a few clicks. And once that context is live, you can put it to work immediately. In fact, you just saw a recording of the actual session we ran just a few hours ago before Showtime. So you can see here my quad terminal my terminal here was cla running. we asked it to list we added the teamwork graph connector asked it to list all the sources it's connected to. So you can actually see all my sources are ready to go here and I can just easily hand it a job. so in this case I'm going to ask it to go and build a customer wall that I can share with my engineering team to capture customer takeaways from customer calls including Gong recordings and then upload it as an artifact.

34:55 So, I'll send this off to the races and I'll check back in a bit and it'll run. Now, Quad will be able to efficiently mine the teamwork graph and create a complete customer wall, breaking down key themes, sentiment trends, and direct customer quotes for me to share with my engineering team. Here we can take a quick look at an example. So that same query is that's running right now. We ran a little bit a little while ago just to get ahead of the presentation time here and you can see that claude generated this customer wall. It's an artifact just like I asked and all this customer feedback is on a wall sourced from all over the teamwork graph. So you can see that there's gone recording, zoom recordings, some Salesforce records and more all over here.

35:43 Now just how effective is this? To find out, our teams ran sideby-side comparisons using agents with and without the teamwork graph and put it to test in their daily workflows. Here on the screen, you're looking at real recordings from those tests. These aren't hypotheticals. They're real workflows and measurable gains. We'll be publishing more of these benchmarks so you can see the results firsthand. From cutting token usage up to 50% to catching critical bugs before production. We're working to bring this type of benchmarking to your teams. So come visit the teamwork graph booth on the expo floor to learn more about that.

36:24 Customers are also sharing success stories. When a leading European retailer grounded their workflows in the teamwork graph, they saved 15 minutes per action. Multiply those 15 minutes across 50,000 work items every quarter and that's over 12,000 hours returned directly to the business. Why? Because a teamwork graph delivers what siloed tools could not. Generic AI is great for speeding up isolated tasks, but an enterprise AI platform connects the dots across your entire organization, getting smarter as you work and transforming how your business operates.

36:59 That brings us to our final chapter. how humans, Robo, and other agents work together on your daily workflows. And that's where AMP really comes to life. Think about how your day typically starts. We've all had mornings where you spend your first two hours just piecing together what happened while you were offline. You're chasing updates across tickets, docs, DMs. Most AI sits there waiting for you to tell it what to do. But with Rovo, you can get proactive AI that understands what matters to you and can step in before you even ask. Let's check out what that looks like.

37:39 So, by now you've all probably seen this robo button. In fact, millions of you use it every single month. When you click on it, you will see this for you tab available all the time. It instantly surfaces what matters most from recent decisions, emerging risks, and critical team priorities. Here it's tracking several projects and epics that my team are actively working on. And I can click in here and see how it's going. For example, I identified this risk. I see the summary. I can see how it sourced that risk from these different sources in the teamwork graph.

38:15 And I can flip through, see decisions that were made and even highlights. And if I want to go and ask further questions or get help resolving any of these, there's a Rovo chat suggestion right here that fires Rovo chat and helps me get to work right on that issue. Now, work doesn't always show up neatly in a formally tracked project. Real work starts organically. It's messy. It's tracked in many different projects, Slacks, threads, etc. That's why we built suggested project projects.

38:47 Rovo uses the teamwork graph to spot active collaboration patterns across your tools and identify emerging work streams before anyone has even created the ticket. So if I click here for example I see a suggested project about the go to market launch for strawberry app and I can view details to learn about what the teamwork graph knows about it. The teamwork graph here is pieced together who's contributing what activity is connected potential risks and key stakeholders. So even though this project is not formally tracked, it still gets visibility, shared context, and actionable recommendations. We've seen Rovo and the teamwork graph help you get caught up and plan ahead. Now, let's look at how we're supercharging Revo chat to move from individual productivity to cross team action. Let's start with a simple question I ask all the time. Which enterprise accounts are blocked and why?

39:42 Answering that usually means digging across tools I rarely touch or have expertise in. CRM pipelines, finance systems, scattered call recordings. Let's look at this investigation which I recorded earlier today. Instead of just looking at pipeline stages, it analyzed unstructured deal notes, conversation history to understand what the customer actually cares about. With the blockers identified, let's use the data insight skill that you saw this morning to drill down into the accounts that are blocked by compliance and legal features.

40:13 Getting this data used to mean filing a data request with some data analysts and waiting several business days. Now, it's a simple prompt. Looks like this account few accounts make up the bulk of the total pipeline, which is why we need to reconsider some roadmap decisions. Now, I don't want insights like this on a one-off basis. I want to see it every week. So we turned that whole analysis of blockers and what they're worth into a custom skill and scheduled that skill so that it can run every Friday morning.

40:40 As you see here, there was no custom prompt script configuring automation rule from scratch or any of that. Robo just packaged the workflow into a reusable skill in our library and set up the recurring schedule. Gave it to me to review here and I can accept. That's how individual habits can turn into transformational ways for a team to operate. Now let's look at a real customer example. Centric software. They built a custom agent directly inside their Jira resource management queue. After they turned it on, the impact was immediate.

41:12 Resolution times dropped by 50% from 11 days to 5 1/2 days. High priority incident responses accelerated from 3 days to just 10 hours and their seat reach an all-time high of 4.7 out of five. That's not just a marginal productivity boost. That's a complete transformation of their service operations. And while Centric proved the power of a bespoke agent, our goal is to make agent building easy for everyone. With agent templates, we've packaged proven workflows into over 30 pre-built agent templates. From backlog grooming and triage to compliance checks and marketing briefs, preconfigured with the right tools and permissions out of the box. With one click, you can create an agent and add it to your workflow, making it easy to start seeing value right away. This gives non-technical teams a clear prototype of what a successful agent looks like and have it work in just a few minutes. Now, let's look at an example that's closer to our hearts. Assessing security risk.

42:13 Assessing security risk across thousands of Jira work items is one of the biggest drains a security team faces today. manually checking permissions, user roles, and data boundaries on every single request simply does not scale. But what does scale is an agent. Let's take a look. So here's a example request from a from a support Q for security review. Now I have customized an agent from the agent library called security access reviewer. So I can bring it up here.

42:50 And we preconfigured as part of that agent template a starter prompt here which is my most common use case. Review this request for security policy violations. So I'm going to let that fire and it's starting to run. Now it's analyzing the prompt invoking the tools it's been configured with. It's referencing the context from the work item here. It's searching our knowledge base for relevant company policies. It's checking user permissions to flag any potential access violations and it's ultimately going to combine all that information into a rich analysis which will show any second now.

43:27 here we go. So you can see here that it's found some policy violations and it's assigned risk scores provided some descriptions and recommended some actions. So it goes through each of those findings, gives rationale, sites, policies, etc. Incredibly powerful, but also manual. So let's see how I can make that a bit easier for me. So they go, I can edit this agent.

44:00 Now I'm opening this agent up and I'm going to add an automation trigger. And I can do this in just plain language here. What should trigger the event? Well, let's do when a work item is created in this space perform a security review and let me know your findings. Super simple, but you can imagine a whole host of you know complex requests being put in here and the automation trigger will be created accordingly.

44:37 Now, with a few natural language phrases, we built an endto-end automation rule that executes autonomously the moment a work item lands in the queue. You're already familiar with automation build, the automation builder experience. So, you can see this project that's been built. Now, your agents can take autonomous action on these rules. Now, I also want to show you a few other things that we've been able to do with agent with agents. So I can go here and I can see for our security access reviewer agent first of all the access accounts that it has. We're excited to announce agent accounts which you can see it's using right here. So instead of agents borrowing the permissions of individual employees, this automated agent can run under a dedicated identity with lease privilege permissions scoped only to the exact spaces and tools that they need. Now there's other controls as well. For example, if I look here, I can actually see what reasoning mode it's using, which is our default think deeper mode. But if I wanted to say customize that, I can go here and I can actually decide first of all, does it run as think deeper or full-blown workquery? And I can even choose what model it's running. So here I've pre-selected CloudSonet 46.

45:55 That's actually an important element for any agent creator to have because they want to be able to first of all ma manage what models are used so they can predict what the output will be and also manage their token expense for that agent. Now let's take a look at what that finish output looks like. So here's an example work item where that agent ran. So you can see down here the access reviewer left its comments confirming that it ran and actually created a confluence page report outlining all the risks.

46:29 Now what used to take hours of manual back and forth across teams just happened in a few minutes. That is the power of agents. They are ready to handle a wide variety of tasks from routine workflows to complex projects. And as agents join your team, the excitement will be contagious because every agent can collaborate across the Atlassian platform through our agentic multiplayer protocol. They become secure, governed, and contextrich team members. And making that vision a reality through the platform, we let's recap the three ways that happened. One, establish governance that unifies every human and AI actor under a centralized control plane.

47:15 Connect context powered by the teamwork graph so everyone has that organizational knowledge they need to move faster, smarter, and cheaper. And use Rovo specialized agents to transform governance and context into trusted execution inside the tools your teams use every day. Together, these three investments to more than speed up individuals. They eliminate fragmentation, accelerate enterprise adoption, and deliver true business value. The best part is that you don't have to wait. You don't need to stitch together a patchwork of point solutions.

47:47 Everything we showed you today is available with your Atlassian subscription. And if you're looking at what your organization thinking, we want this level of transformation, but we don't want to build it alone. We've got you covered. Through our four deployed engineering program, our senior AI engineers embed directly in your business, and their impact has been great. We partner alongside your teams to design, code, and harden custom robo agents to tackle your highest impact workflows and deliver measurable production grade ROI from day one. To see what they're building, visit the FTE booth on the expo floor. So, here's what you can do when you return home. Work with your security and IT teams to extend enterprise controls across the experiences. Broaden your context. Check your content quality score. Context quality score. Connect those hight traffic sources like GitHub, Slack, and Google Drive and watch your answer quality and efficiency compound. And try out Rovo work to orchestrate those complex multi-step jobs and empower your teams to build custom agents using those out of the box templates. We cannot wait to see what you build. Thank you so much for all your all your time and attendance. Have a wonderful time at Team 26 and see you around.