Why AI gains are unevenly distributed in your engineering team | Asana’s Arnab Bose

0:00Welcome back to Dev Interrupted, brought to you by Linear B. My guest today is Arnab Bose, the Chief Product Officer at Asana. And this one's a treat for me because Dev Interrupted's production runs on Asana. Three quarters of knowledge workers use AI on the job now, but few companies can claim distributed productivity gains from it. And Arnab has a theory about why individuals get faster while the company simply doesn't. And we get into all of that and more, and how Asana measures its own AI success. Now, here's my conversation with Arna. I'm really excited about today's episode because we're diving deep into the next evolution of workplace productivity with a tool that we all know and love, I especially know and love, and that's Asana.

0:47And joining us today is Arna Bose, the Chief Product Officer at Asana. And it's amazing to have the opportunity to chat with you because here on Dev Interrupted, we run on Asana. In fact, I've ended up building our whole podcast production pipeline, our content calendars, and the cross-team dependencies for how Linear B interacts with our content, all using Asana. It's kind of like the central nervous system for how we keep this show running and communicating. So, you know, I'm really deeply invested in, I guess, the Asana cinematic universe, you could say. But, you know, looking at some of the recent things that have been coming up for Asana, like there's been a recent acquisition.

1:29In fact, the first one in Asana's 18-year history, which we're going to talk about, as well as announcements around agentic work management. it's really clear how your team is building towards a future that's going beyond just typical and traditional task tracking and management and kind of helping us navigate this agentic maze of ephemeral and real tasks and owners and non-owners and all of the real messy reality of how we do knowledge work now. And Asana is really going to become one of the forefront places where that develops. So Arnab, I'm really excited to have you here. Welcome to Dev Interrupted.

2:07Arnab Bose:Thank you for having me, Andrew. And amazing to hear how you're using Asana today and how Dev Interrupted the podcast is like built on Asana. Amazing. Yeah, we're going to dive into it a little bit more too. And so as we kind of crawl along on some of the latest developments with Asana, and maybe you can kind of help clue me in as well, just as like a developer, someone developing minded interest with the platform of like, what are the ways that I can be thinking of plugging in in the future. But, you know, I want to start at the top, which is a data point from BCG about a really frustrating reality for knowledge workers.

2:41And that's that 75 % of them are using AI in some capacity on the job at this point. We're talking about knowledge workers who, you know, are typically in like a more like a white collar industry where they're working with information likely on a computer. And, you know, most folks at this stage are using AI to do that. But only about 5 % of companies on an aggregate overall are seeing meaningful productivity gains. And that stat was really interesting to me because we talk about the same kinds of problems that are happening right now in just even inside of those orgs, the engineering orgs themselves, of how you get all of these really like spread out success in agentic engineering.

3:24You get the 10x engineer you get the the mythical 1000x engineer and everyone's trying to figure out how they're working and we're all trying to identify these players um but the reality is is that yeah sure you got these like really crazy uh like empowered in a devs but on the aggregate are these engineering orgs delivering more and higher quality engineering products are they getting more cost of finish line that doesn't need a refactor or cause an incident those questions are still really fuzzy. And that's what we say as like a productivity gain. So why do you think that like there is such a big gap between picking up and using the tools to actually getting like on an org level meaningful gains?

4:07Arnab Bose:So a great question. I think the research from BCG is quite precise in terms of highlighting that the individual within companies is now able to produce more work and probably higher quality work. But overall, when you take a look at a workflow that a group of knowledge workers have to build out end to end, those workflows are not actually being accelerated by AI as yet. And that's why companies are not seeing true productivity gains on the aggregate. And the thesis that I have, and this is the reason why I was drawn to Asana, is AI has been really easy to use on a one-on-one basis where if you're using a chat-based LLM product and you have a thought or an idea or a document you want to write or a document you want to refine or an image or a graphic you want to create, that interaction pattern is something that grew like wildfire three and a half years ago and has constantly gotten better as the quality of the models and the reasoning capabilities of the models have improved.

5:22Arnab Bose:But that is not the way in which work actually gets done end to end in a business, whether it's a small business or a large enterprise. In any of these cases, it is typically a workflow which requires multiple human beings to get on the same page, sign off on the work, and then agree when something is reaching a quality bar that indicates that it's done. And so while AI tools are capable of taking actions inside of shared workflows across teams and projects, this is not the way in which most companies have been able to deploy AI. What our thesis at Asana is, as you were calling out yourself and you probably see on a day-to-day basis, Asana is a very interesting canvas where you can define who does what by when and how.

6:12Arnab Bose:And leveraging the work graph, which is a data structure and a data model we've been investing in for over 18 years now, you're not just creating tasks, you're creating tasks that are connected to teams of people, that are connected to projects, that are connected to portfolios and goals. And so a larger group of people can align on strategic outcomes, exactly how the production of the podcast should be done, set up approval workflows and things like that. And then if you can add in enterprise-grade AI agents that have the context of how these workflows were completed by human beings historically, that have the checkpoints where multiple human beings can provide these AI agents with feedback and there's clarity around how an artifact was created, then you can actually achieve end-to-end outcomes at the workflow level that truly move your company forward.

7:16Arnab Bose:So I'm not yet answering your question about the engineering workflows, but if you just simply take the knowledge worker workflows of aligning on a brief of a document or signing off on the production values of a video, So if the AI agent is not able to work with the context of your company inside the shared workflows across teams and projects, that's what results in that stat from BCG that 95 % of companies are not seeing real productivity gains. Yeah, absolutely. I think there's, you've rightly called out that there's a big difference in being individually enabled with AI and being on an organizational level enabled with AI.

8:00And those skills don't directly transfer. Instead, it's more like an accumulation of those individual AI abilities and those workflows and the things that they need and then figuring out how to harden them into states that are durable for the long term. And that's where you get that long term productivity gains, especially the larger the org, the more hardened and the more secure and the more specified that kind of deployment needs to be. And frankly, there's a lot of really large and sensitive information orgs that build and are powered by things on Asana. So when you think about building in that problem space and interacting with that messy context, you have to think about how to make it safe for everyone and all the different levels of the work that they do, which is a taller order than even most orgs have of trying to figure out how to just be productive with AI for ourselves, period.

8:54Correct. And so you've also called out that Asana becomes this place where all of the traditional history of how work was done and where the identities of the humans and the identities of the agents come together to queue up and pick work and give feedback to each other. It becomes this context, I guess, plane in which these two things are communicating with each other. And really what that starts to get to is it helps us figure out what's the new way of working. A lot of us are trying to figure out what are the new loops, what are the new ways that we operate together as a team. And so you've recently unveiled agentic work management, which is kind of like an operating system level answer to this question.

9:42It's acknowledging that computing up until now has been writing code to create things, to push things through pipes in and out through different destinations for different results. And now we can do that with intelligence and with natural language inputs and outputs. And so now knowledge work itself becomes something we can build with pipes and send around. So it becomes like an operating system. How do you think about that in your product's role at Asana and the stuff that you see?

10:10Arnab Bose:We've built agentic work management on a premise that came directly from our history. And, you know, you've been an Asana user for a long time. You can see that for over 18 years now, Asana has been building this work graph, which is a structured representation of who is doing what by when, towards what goal, and in coordination with whom. And there are some principles behind the work graph for human beings around shared visibility, clear ownership, permission-aware access, structured communication that can also be applied to AI agents. And the way in which we are bringing AI agents into this agent-tip work management product is we think of AI agents as true teammates, as actors within the system that stand alone.

10:57Arnab Bose:And those actors have the same kinds of shared visibility, clear ownership, permission-aware access, and structured communication as human beings do. When you combine these things, what ends up happening is you get some significant value that is directly plugged in to that shared workflow. So the first thing is, and I spoke about it a little bit already, because you're adding this AI agent as an actor into a part of the work graph and you're giving it access to it, it can see the activities that have happened in the past that human beings have accomplished. The second thing is because there's a permissioning system we can build based on which human being is interacting with that AI agent and in what part of the work graph, what we can do is instead of an AI agent being like Arnab's AI agent or Andrew's AI agent and only learning from feedback from me, we've built in a concept called shared memory.

11:55Arnab Bose:And what shared memory means is if you have an AI teammate that is a launch planner or a podcast production specialist, and it's getting feedback on the work that it's doing from you, Andrew, or somebody else on your team, when a third person comes ahead and uses that AI agent, it will remember all of the nudges and the feedback that it's received from everybody and filter down based on the particular project or task it's working on. And it can accomplish that task that much better. It's like onboarding a human teammate onto your team and then having multiple team members mentor and coach the person.

12:36Arnab Bose:And then the person remembers how the Dev Interrupted podcast is produced and the nuances and the quality bar that you're achieving. So I'm just layering on the concept. So the first concept is the ability to access that enterprise work graph to see how historical work has been completed. The second is the ability to get to shared memory so that multiple people can train the AI agent and it constantly gets better with use and not just for one person. And the third thing is having a full audit trail of the activity that it's done in a way which is shared and visible by anybody who's an administrator or a project lead.

13:18Arnab Bose:And what this achieves, this last part is, again, if you're accustomed to using AI agents as like your private personal chatbot, actions are happening in a private thread or behind interfaces where the AI's involvement is invisible. And like all of the prompts that I put into leveraging a personal chatbot to create a presentation or to generate a video, that back and forth is something that is invisible if I simply download the presentation or video and put it back onto a file share system, right? The file share system doesn't have a way to represent the interactions. Whereas if you have that same interaction with the AI teammate inside Asana, all of that is visible to your team.

14:05Arnab Bose:And if your manager or somebody else who's an editor on the production team disagrees with something, they can also nudge the AI agent to achieve the best outcome for Dev Interrupted. And it'll be fully transparent to everybody on the team. Right. Right. So the way that you've broken this down into three concepts is actually really profound. We've talked about how these are so game-changing for large organizations with a lot of other really smart folks here on the show. The second thing that you called out there in particular, beyond the first thing, which is the context graph. Yes, we have this nice interconnected graph of who, what, when, where, why within the world of work.

14:49And then the second part of this shared memory is so critical. We learned this too from, we had Karthik Ramgopol. He's a distinguished engineer at LinkedIn, talking about how they transformed their engineering org and about how they were working with AI. And it all started with unlocking shared memory for workflows across folks that were using these tools, allowing the gains to be aggregated as a team, not just like on an individual basis. And then also to this ability that you've very smartly called out for admins, project leads, folks that are on the leading side of like building and deploying these agents to achieve certain outcomes for the business.

15:34they need to have that visibility too into how they're used and what that data looks like to the point where, you know, that whole creation, curation, learning process needs to happen in one closed space so that even the continuous conversations and training of it are, it's all together in one thing. Like you can go to an agent and you know the sessions and all the traces that resulted in this agent being the way it is and producing what it is. And honestly, I think that even goes back to what I think the future of what CodeStore is going to look like for engineering teams. Checking in and opening a PR and checking in your code is one thing.

16:14We need to also be checking in the session transcripts that resulted in the code. What did you prompt? What was your harness? What were your skills? Did you use an MCP server? Was this reviewed? There's so much that's actually missing. And then in an iterative process, which engineering is, people want to pick up that PR and they want to work on it and improve it. And now we can't because we don't have the work history of how we got there. We have to throw it away. And the thing is that Asana can swerve and avoid this problem because you have the knowledge graph, you have the shared memory of everything in that work and how it got there.

16:51And if the AI agent is created and interfaced all within Asana, then you can always pick back up where you left off in terms of training or fixing or aligning it to outcomes. And it becomes a truly iterative process, which knowledge work deeply is. I think that's what we need in order to really trust and fall in love with this kind of tool.

17:12Arnab Bose:So I want to pick up from where you left off and highlight a couple of things that we're working on on the product strategy side at Asana. So agentic work management is what we announced. It's available today where you've got these AI teammates that are pre-built based off of our existing knowledge worker personas, like people in marketing or operations or IT, and they're working side by side with human beings. And they're already deployed at major companies like FedEx and Koss. And Koss is a European clothing company. And I actually interviewed their chief digital officer at an event in London.

17:48Arnab Bose:and they've totally transformed how they go from a runway production to having those products available on their e-commerce site in a matter of hours, leveraging AI teammates in Asana. And so that's the type of work that's already in production today and people are using us for. On the engineering side, we have a pretty interesting product we're working on called Command by Asana. It's in early access slash design preview. We've got a couple of design partners who are deploying it. And most interestingly, our internal R &D team is using this to build Asana products. And it's built on this idea that you're calling out around iteration and compounding improvement.

18:30Arnab Bose:And what we were seeing was in terms of leveraging the latest frontier model coding agents against our existing source code to build incremental features. Even in the calendar year 2026, there's been a tremendous amount of improvement and the ability to generate these PRs and code fast from a well-written ticket or spec has improved tremendously. But what was holding us back as a team was actually going from the ideation process. Like you have a three-person whiteboarding session between the product manager, the engineering lead, and the designer. going from that to well-written tickets in a nice iterative way.

19:18Arnab Bose:And then as you're calling out, tracking and logging all of the AI-powered coding sessions so that if there's a voice of the customer feedback that comes up or there's an issue we detect within our internal usage process, we can go back and automatically update the PRD, update the ticket with those learnings. So we build a self-learning loop end-to-end across all the knowledge work, the product work, the design work, and the engineering planning work and not just focus on the code base. That's what we are lacking. And so command is all about that. It's about helping the product builder stay in flow where you can have these cleanly built tickets based on all the context of meeting recordings and PRDs and historical tickets plus your existing code base, generating that ticket so that your coding agent needs less handholding and can get to a great outcome the first time.

20:17Arnab Bose:It's about ensuring that for product managers and engineering managers, they can look through their entire Kanban board of who is doing what by when. They can predictively find risks in the process and automatically figure out remediation. So somebody might be out on leave, there might be one person who's run into a significant issue because the PR previously generated has some like infrastructure or scaling challenges, identifying all of that in a fully agentic way. And then the third thing is more for like executives like myself, where because all this work is tracked in Asana, I now have an agentic interface where I can query the status of individual features all the way up to major product launches.

21:05Arnab Bose:And I'm getting like high quality results that I can trust out of the system versus only getting a particular slice of the pie, which is like, okay, what does my ticket status look like for coding, but not getting the insight into all of the planning and design milestones. So anyway, this is what I wanted to come back to, which is I also believe that we are only in the early stages of improving the looping for the product building lifecycle, just like we're in early stages for basic knowledge work. And a lot of improvements have been made in the coding agent slice of the work stream. But now I think everybody should open their aperture to what does it truly mean to go from idea to revenue when they're building a product?

21:52Arnab Bose:And then what is the tooling required to manage that entire lifecycle in an agentic engineered way where you're constantly compounding benefits? Yeah, I'm definitely subscribed to this narrative. I think this is something we're really obsessed, especially with building at Linear B and understanding developer productivity with an organization, but also tracking AI's adoption and then its downstream impact. And tracking whether or not your AI adoption is going well is really messy for most organizations because it's not just how many seats are getting used or how many tokens are getting consumed.

22:28It's about what are the tickets that are getting closed? How efficiently are we moving through our epics? And when code is getting delivered, is it getting reworked later, right? Is it causing incidents or outages? It becomes a deeply complex problem to measure and understand, oh, we are being successful or we are durably increasing our AI quality in a sustainable way. I'm curious now, just based on what you talked about, it sounds like that gets managed in Asana itself as a data management layer. Or, you know, what are the things that y 'all think about is measuring success and AI usage using these kinds of tools and looking at that kind of data?

23:08This episode is brought to you by Linear B, the engineering productivity platform. There's a widening gap between engineering teams that have turned up AI adoption into delivered work and the teams that haven't. It's real, it's measurable, and it's growing every month. We sat down with Linear B CTO Yushai Beery to walk through new engineering benchmarks that are built upon 2.7 million pull requests from over 250 engineering organizations. These metrics cover why high AI usage correlates to a double increase in PR mergery and why more AI code doesn't automatically mean more shipped code. And the lowest effort win that's available right now, which is turning on AI code reviews.

23:50The full report is out. Get your copy over at linearb.io.

23:55Arnab Bose:Some of the things that we're looking at are, what does the cycle time look like? Because, I mean, I'll tell you the challenge. At a company like Asana, based in San Francisco, being like a technology company, there isn't any resistance, certainly within the R &D organization, of utilizing AI tools. In fact, people want to use everything all at once, right? So looking at adoption data is a bit of a red herring for us because that's also not helping us track true outcomes. And are we being efficient with our spend? Are we actually driving real productivity? Or are we ending up in a state where we're playing with a lot of tools, but we've actually not improved the craft?

24:36Arnab Bose:And the same thing I think happens if you take a look at other lower level metrics like ER velocity and things like that. So we are trying to look at it. And again, we're in early days. I don't think we've cracked the nut here. but we're looking at things like end-to-end cycle time for projects, right? Like, is the end-to-end project getting delivered faster, quarter over quarter, month over month, for projects of similar size and complexity? So cycle time is interesting versus this pure PR velocity. And then there are some other gates that we're taking a look at as well for, okay, like, you've built a particular update.

25:12Arnab Bose:Like, let's say you were on the AI teammates team and you shipped AI teammates. are you seeing the level of adoption from a quantitative perspective and customer happiness from a qualitative perspective that you set out for? Because we also don't want to be in a situation where everybody's patting themselves on the back because PR velocity is improved and you can say you shipped 40 features this month versus 20 features last month, but 20 of those features are sitting on the shelf or the ones that are actually being utilized are not meeting your quantitative or qualitative success metrics. So I think it's a combination of those three things, which is let's not try to degrade the analysis to simple things like are people just utilizing the tools or are you able to check in more things faster?

26:01Arnab Bose:Perhaps looking at slightly longer term, a wider aperture around cycle time, cycle time plus gates around adoption and quantitative and qualitative success is a better metric. At least that's where myself and my CTO counterpart have landed at. And I'm sure as we iterate through this, we'll come up with even more improvements going forward. Yeah, absolutely. That's a really great glimpse into how y 'all think about it. Cycle time and understanding the ultimate impact of, I love how you said it from idea to revenue. That's really what we're geared around as well is understanding how do you go from, You know, even what you described it as like this three-person tiger team whiteboarding about like, you know, they each come with their discipline, they each get it all laid out.

26:45Like, how do we go from those people in a room with, you know, two pizzas to then having the revenue impacting delivery on the other end? And, you know, the road to get there for organizations right now, it's really rocky, like especially one of the big rocks that orgs are navigating right now, especially in a world where tokens are, you know, relatively subsidized and compared to probably the future of what tokens will be for us, is people are getting big eyes for maybe building stuff instead of buying stuff. You know, Asana actually recently completed an acquisition of Stack AI. And this was to address cross-system execution.

27:23I'd love to learn more about that. So you were posed with the question of build versus buy. We're not going to build a Stack AI. We're going to buy it. You know, why? What was the thought process there? What does Stack AI bring? And what does that look like as like a durable moat for you?

27:37Arnab Bose:So I'll sort of start with the strategy first about why is that technology or functionality interesting to Asana? And then we could talk a bit about like build versus buy versus partner. So again, like having been a long time Asana user yourself, you've seen the product evolving from being a system that tracks work. And now with the power of AI and the fact that most systems that you're using as a knowledge worker have great APIs, have MCP servers, you can go from the idea of tracking work and breaking down your projects into this sort of collaborative task, into the task actually completing themselves and the journey being fully completed.