How Dropbox Is Rethinking Work With AI And Dropbox Dash

Welcome back to the Tech Talks daily podcast and today I want to start with a question. How often do you catch yourself drowning in tabs, files, chats and random scraps of information? They're all seem to scatter themselves, the very moment you need them or a colleague as she for one of them. I think most of us live in that digital fog every day and it feels impossible to keep track of anything with any kind of clarity. What my guest today, he spent the last 20 years building products that solve these kind of problems. His name is Josh Clem, he's the VP of engineering at Dropbox and he's leading the company's AI efforts and the push behind Dropbox Dash. What is it? What problems does it solve? Well we'll talk about all that today but he's also going to bring a mix of scale engineering, product curiosity and a whole heap of hard earned lessons from LinkedIn and Uber. These are the kind of stories that move from early experience in personalized data products all the way to global outages triggered by free donors which is exactly the kind of chaos that shapes better systems. So my conversation today will dig into the real meaning of AI fluency, why context beats hype and how Dropbox is trying to make work feel lighter with knowledge management that adapts to you. So here's my question for you, what would your day look like if all your work context just surfaced itself without you having to hunt it down? And while you ponder that question, I think you're perfectly set for today's interview. Before I bring today's guest on, I just want to give a massive thank you to my friends at Denodo because after visiting over 25 different events in 2025, one of the phrase that I keep hearing is no data, no AI and agentic AI simply needs better data. Now agentic AI is here but it only works when the data behind it is complete, governed and in real time. And this is one of the areas that Denodo helps because Denodo gives you a logical data foundation that accelerates AI, boosts lake house performance and turn your information into reusable data products and for every team. So CIOs, architects and business owners each get the data that they need instantly. And their global partners help you get up and running faster than ever. So if you want AI that doesn't hallucinate but actually delivers real business outcomes, visit Denodo.com and start making your data work harder. But now let's get today's guest on. So a massive warm welcome to the show Josh, can you tell everyone listening a little about who you are and what you do? Hey everybody, I'm Josh Clem, I am currently the vice president of engineering at Dropbox. I'm in charge of our AI initiatives including building out a new product called Dropbox Dash. And I've previously been at Uber, I worked on Uber Eats for quite a number of years and before that I was at LinkedIn. Yeah. Awesome. Well, there's so much I want to talk with you about today, especially Dropbox Dash of course. But I always love to find out a little bit more about my guest origin story and you mentioned there, you was at Uber and then I think there was LinkedIn and when I was doing a little research I had that you were you witnessed some of the main key initiatives that shaped how Uber underlined architecture worked and how Uber managed to scale the processing of millions of trips a day and operating in over 70 countries. It feels phenomenal to be a part of that but especially from the outside looking in but tell me more about that journey. Yeah, so I was at Uber for almost eight years and I just particularly absolutely love history. I love understanding why are we doing the things we're doing? What was sort of the evolution and these companies like Uber just go through a tremendous amount of scale. As an engineering leader you want to really understand, okay what were the things that the early team did? What were those key decisions and I think it's just very helpful to kind of write stories like that. I had written a very similar story at my time at LinkedIn about 10 years ago called a brief history of scaling LinkedIn and so I thought I'd do a very similar one at Uber and of course I was on the Uber each side and so I saw a lot of those stories I really understood how that scaled. It's a lot of that story that you're describing was really a combination of things that I had to kind of personally look up from even before my time what things I was personally involved with and really kind of putting together a nice story overall and you know when you think about kind of scaling as an engineering leader you really want to learn you want to reflect on some of those decisions and a lot of times it's when things don't go right and you learn a lot and you have to really update your processes and think about your technology. I'll give you a quick story so back in 2017 Uber Eats was just getting going and we had all of our global operations team they were trying to make Uber Eats work and be very very successful in their country or their city and so our UK team decided let's do a promotion. They went in they talked with crispy cream donuts and they decided you know what we're going to hand out free donuts offered on Uber Eats and not just free donuts but free delivery as well and Neil it turns out people absolutely love free. We ended up getting about a hundred times more traffic during that promotion than we ever were expecting all hell broke loose everything went down the app went down nobody got their free donuts everyone in the UK was incredibly upset but even worse is when you know the UK had that issue it actually took down all of Uber Eats across the globe and it really kind of kicked off this this initiative we've got to improve the resiliency of this product. We have to figure out how to make sure you know local city can do these sort of promotions but the same time building in resiliency patterns so that it can't necessarily affect the rest of the world. So we did a lot of both very short term fixes as well as think about our underlying architecture and ended up being very very successful I don't think we had another outage for at least another three years. What a great story absolutely love that and one of the reasons I wanted to mention it is there's so much hype around AI at the moment a lot of people forget I think that when it comes to disruption around technology we've been here a few times before from the arrival of cloud to mobile and obviously AI now and I think it's also important to point out that you've been working in AI long before the hype so tell me more about your work in AI and how that would lead you to Dropbox. Absolutely yeah so I would say my first real exposure to building personalized data products was probably back at LinkedIn you know they were an early pioneer they had some phenomenal data sets and a lot of the product experiences that you see both on LinkedIn and really any social network really came from LinkedIn things like people you may know you know I was part of the profile team so we would show similar profiles we would show users skills a lot of times inferred skills that you could endorse later I worked on this initiative we were trying to get more students and universities on the on the platform and we really leaned into data inside so better school search we ended up building our own school rankings and one of my personal favorite features was something called notable alumni and you know you can go to any school and we were able to really mind the LinkedIn data set and fine hey these are very successful alumni from these universities the universities actually love this because it gave them a really great a list of folks that hey maybe they should know about and engage with and actually get them to come speak with them more of the university and things like that and so you really learn hey there's a lot of power in building these AI products and so then when I went over to Uber Eats I understood hey we need to really think about bringing in more personalized data insights there and I really leaned into investing in our search and discovery team when you think about Uber like Uber rides when you open up the app you know where you're going Uber Eats very very different you don't necessarily have a particular restaurant in mind you know you're hungry in fact about I think 20% of users at open up Uber Eats knew exactly what restaurant they wanted most of them were open to discovery they maybe had no idea and you could really help them maybe that a cuisine type I want pizza great here's some options and so Uber Eats ended up being a very powerful machine learned product very similar to how I spotify my work or a Netflix and all those recommendations that you're going to see it was really important for us to take in all these different insights whether it was the time of day your past orders sometimes even the weather we were taking that into account to find the right personalized recommendation towards the end of my experience of Uber Eats we started to use a lot more with natural language conversational AI even large language models and I started to see how impactful that technology could be and so for me you know coming to Dropbox really felt like this natural continuation of everything that I had been building for the last 20 years it's AI but it's grounded in reality you're solving universal problems around just information overload you're trying to help people get their job their task done faster and now we have that technology that can that matches that ambition I think we are at a time now if we walk fast forward to present day where every business is coming to terms we've not just being a tech business but also evolving into an AI business so obviously with Dropbox that has continued to grow over the last decade so how is Dropbox dealing with this latest shift absolutely I mean Dropbox recognizes that like like many companies AI is reducing a lot of the busy work and it allows employees to really free up their focus on what matters so there's just a lot of excitement going on right now in the company at Dropbox around AI we'll do things like hack weeks we had one earlier this summer and we didn't even necessarily identify AI as the key topic or key theme but almost 95% or more of the projects that ended up coming out of the team were all about leveraging AI leveraging AI tools trying to figure out how to do things in just more novel ways and that was really encouraging a lot of those even ideas have been added to our product roadmap and we continue to do this within the company we highly encourage a lot of show and tells demo days etc to tap into that that excitement that curiosity and make sure that other employees know that it is okay not to use these types of tools now some of the tools that we do deploy within Dropbox things around you know better ways of coding or better ways of trying to build prototypes it really does change the nature of the work the approach to work is very different you're almost having to think about the upfront task of planning what you want if I'm an engineer a lot of times it's okay I needed I know what I need to do let me just start coding and the script flips a little bit it's more let me take a breath let me really understand the requirements let me really understand what good looks like when I'm done with this this task how do I know it worked and almost defining that upfront and then leveraging AI to really help with some of that more busy work and so the this you are shifting in how you think about work and I think that's actually really exciting we've got designers who are mocking up examples of our new products all with live code that you can test out even put in front of customers and do some user research and get really really valuable insights overall and ultimately you know how are we deploying AI at Dropbox we are building a custom solution called Dropbox Dash of course it's four external customers but we are our it's number one customer as well so for people listening hearing about Dropbox Dash for the very first time how would you describe it what's it bring to them so Dropbox Dash is really the Dropbox version of 2025 yeah when Dropbox started many many years ago you had all of your files on you know your thumb drives or you're you're you know personal pc and it was very hard to keep everything synced and organized you know one file over here and and then it gets it gets lost and that problem just becomes even more apparent at work when you're trying to share physical files with one other one other in a different co-workers so what does Dropbox Dash do it recognizes that a lot of the files today whether it's in your personal life or at work they're all in the cloud or they're all a tab in your browser I don't know about you Neil but I've got probably about 50 open tabs right now in in my browser and you're always trying to jump from one place to the other and trying to find where you left off or where was that file that's what Dropbox Dash does we connect to all these different third party apps including a lot of your browser history we bring all that in in one place we apply AI to auto organize it and then we allow you to do extremely effective search and once you have that you can start doing AI answers and you really can start to collaborate with your co-workers in a much more efficient way and for people listening that already have a Dropbox account which I already imagine will be for most people is this an addon for them how does that work is it something else that they they need to subscribe to is it a part of their membership for example so I'd say there's two things here one Dropbox Dash is a standalone product yeah it is a completely separate URL it is a completely separate purchase because it provides that rich experience connecting to these third party apps beyond just your Dropbox content and so you can go to Dropbox.com/dash today we do have a self-service option and you can kind of get going right away a couple of weeks ago we did introduce more robust AI features within the core Dropbox app so if you do have just your Dropbox files you absolutely can start doing and leveraging more powerful search you can start using chat across your different files we have this new feature called stacks which is is a really a smart collection of any kind of content that you can then share with co-workers and so the Dropbox experience does have some of those features but if you want the the full comprehensive AI knowledge management across all of your cloud apps that's where you need to go and get dash awesome and before you came on the podcast I was doing a little research on your work especially adding AI into knowledge management and I was reading that your perspective that fluency with AI rather than hype how that is the true the true differentiator in your eyes especially for business success today but tell me more about that perspective in your experiences adding AI into this. The lot of hype out there right now Neil there's we you know you see these different studies that are always saying hey these these executives are green lighting AI projects because they feel they need an AI initiative a lot of the customers we've talked to the dash echo a lot of that same thing hey I keep hearing about AI do you have AI it's like yes you know we do have AI but let's let's go more into it because it's really important to understand what exactly are you looking for what outcomes are you hoping to accomplish and I think that's really really critical for executives and various leaders to understand they need to be fluent in this technology they need to understand what's possible what's not what does the security profile is I know there's a lot of questions there and there's a lot of buzzwords out there AI was the first one then you had terms oh MCP is that something I need and then later became agents I want that and unless you really sort of think through these these outcomes and understand what are the what what technology can potentially get you that you're going to have a potentially failed AI deployment and you're starting to see that you know there was that famous article a few months ago from MIT where 90 they said 95% of AI deployments are failing McKinsey was reporting 80 80% of companies aren't seeing tangible ROI from gen AI and there was that really fun article I loved about workslop I think the harbor business review talked about workslop at work and what is that you know and what what what are what's happening with all these projects that aren't necessarily successful well again it goes to oh we need AI let's roll out a bunch of tools and if you don't explain the guardrails if you don't explain what success looks like employees might be generating a bunch of workslop it's low quality work they're not reviewing it they're not treating it as a first draft and it ends up creating a downstream negative effect where there's a lot of correction and so really understanding that AI isn't the problem it's really how people are using it that's I think going to be important overall and then again when we when we start to talk about dash with customers we do see some of these similar themes when customers say hey I want to add agents to my work it's like oh great you know describe your use cases and you get some blank stairs you know I'm not sure I'm not sure so then you really work with them and try to understand what's possible come up with some examples show some examples etc and even just the blank chat box box that you might see if you go to any of your your favorite chat providers today that's intimidating a lot of employees just don't know what to put in there they need examples they really need to understand how it can be used and I think a lot of that starts from the top that leaders really need to be fluent themselves to understand what works and what doesn't yeah I completely agree with you they another 10 or 25 tech events around the world this year and AI seems to be the topic of every situation every event and every brand is desperately trying to be part of that AI narrative but I'm curious so from your work what have you learned at Dropbox about AI tools is you build and scale your own AI products we've learned a lot especially building dash and when you build a tool and you deploy it to your company that's the best feedback you can get a lot of the best products are going to be what's considered you know your dog food in it your company's using it themselves every day day in and day out and you get amazing feedback very very fast feedback from you know you're different employees and so you know what are we learning well first AI tools really should get better the more you use it as you put in a a chat or as you interact the product if it doesn't work there should be a feedback loop that's really important for companies building AI tools what is that feedback loop am I doing anything with both positive or negative outcomes and am I making that system better in our case for for dash if you know somebody has a great chat experience they hit thumbs up and we get that information we can we can definitely improve it same thing with thumbs down okay clearly this this answer missed the mark we have to do a better job and just you know the more you can build feedback loops like that the faster your AI tool quality will improve I mentioned before we've learned a lot about where AI tools do hit limits that same thing that empty chat box very intimidating what do I need to do for dash we are moving more towards proactive suggestions the chat box will always be there but what are suggested searches what are suggested chats that you could do and the more you can move AI into kind of your normal work mode where it isn't just the chat box box I think you're going to see a lot more success and same thing where can AI work in the background I mentioned before about dash dash is connecting to all these third party apps bringing all this content in well we want to auto organize that information we want to sort of create pockets of you know these are topics or this might be your working set that is AI also and just being able to surface hey here's exactly what you're working on where you left off is incredibly powerful it doesn't require employees to have to type anything it's just there and it just works another thing we've learned is specifically from customers is there's still a lot of skepticism out there around how my data is used in AI are you training off my data are you sending it off-prem and obviously for Dropbox and and as we're building Dropbox dash privacy and security is a key principle that we're building within but it's also important to explain that and be very very transparent with the folks that you're deploying these these systems to there should be great help center articles explaining how you're using the information but even within the product I think it's important to highlight that hey your data is safe it's secure a lot of the the techniques we're doing to bring in your more personalized work context is done in a very very secure way another lesson we've learned a lot of AI today is what I would call single player I'm thinking about your favorite chatbot whatever it's a chat GPT or cloud it's pretty much you and this chatbot you're typing in some question and you get back an answer there's really no concept of you and a team your team's not seeing that answer you're not really getting any kind of collective wisdom or collective intelligence and we're started to think more more at Dropbox as we're building dash is how do we continue to make these AI products more multiplayer can I collaborate more effectively can I share project updates with a group my project team can I see potentially other example agents or chats and just make it feel a little bit more like AI is part of your team and not just oh it's my kind of assistant I think that's going to be a really interesting lesson and potential trend going forth and then lastly the most important thing we're learning is context is king what do I mean by that AI and these large language models are incredibly incredibly knowledgeable but they're knowledgeable about stuff you don't need at work yeah I don't need to figure out a recipe for a omelet you know I want to understand how to draft a really amazing strategy document using all my kind of work context and so really providing where context using the the more proprietary private information that works have is a huge huge frontier that we're looking into and building towards but I think a lot of companies are going to start to look into more I'm glad you mentioned that because a few minutes ago we were talking about the dangers of AI slop and work slop and I think we've all seen examples of that but on the flip side of of this it doesn't need to be that way and as you said context is king so tell me more about why context aware AI how you think that could be the next step in the future of work and how it could enhance all your content and and ultimately achieve clarity because there's a feel good story here too isn't it's not all doom and it's not all doom and these are these large language models are unbelievably powerful if you can provide it the right context and for me it really starts with you have to kind of understand how large language models fail and there's actually a few very distinct ways and once you understand that then you can know how to provide and and counteract some of those failure modes so let me go through the few of those right now the first thing that large language models fail is if you give them too much information they can get overwhelmed with with knowledge this is a concept called context rot and our brain sort of fall into the same thing all right there's these various psychological studies these like theoretical use cases you know for example Neil if I said all right and then name in the next 10 seconds name as many red things as you can as you can do you're probably going to lock up and you're probably going to say well that's a lot of stuff versus if I said tell me all the red things in your fridge yes you're immediately going to be able to list a ton a lot more large language models are very very similar in that if I dumped every work document I ever had and said all right large language model write me my help me write my strategy it doesn't know where to start it's going to be completely overwhelmed you have to be able to provide a much more narrow specific set of documents or context and it's going to it's going to be a lot more successful so that's something called context rot it gets lost when it has too much context a second failure mode is the classic it hallucinates and what does that mean well if it doesn't know the answer it'll just make something up very very dangerous overall you have to ensure you're not overly trusting the the output like I said before do treat these as a first draft never turn in your first draft I don't know about you in school but when you write an essay and you wrote your first draft you didn't turn it in you made sure you you did a couple revs on it and yet people are still turning their first draft and that ends up being work slop now you again if you provide the right facts grounded facts to these large language models it will use those facts and when it uses those facts it doesn't hallucinate a third failure mode I like to say large language models are gullible whatever you tell it whatever you provide it it kind of just repeats it'll pair it back anything you say and so if you're giving it the right information fantastic it's going to tell you that but if you give it the wrong information it'll also tell you the wrong information and this this is actually quite problematic in the security use case this is where you you might see stories where LM's get tricked to reveal an exfiltrate people's data it's because they can kind of get tricked you can prompt it to hey why don't you go do this other thing and it will and so putting in the kind of safeguards around that is going to be really really important overall so those are just a few ways that they fail and the answer is this term that's been popularized as context engineering so I'm going to go search and retrieve the right context in a very narrow way and then provide it to a light language model and it's going to be far more accurate it's going to be far more reliable and frankly it's going to be magic to your point earlier it's not all work slop it can be absolutely magic and I think that's that's kind of the the step of kind of this context aware AI it's almost like if you know you're not going to ask Albert Einstein to be the magician at your kids party right they they they you know these language models are incredibly smart but they don't have everything they don't have sort of your work context and with that I'd love to kind of actually walk through how Dash works a little bit more on the technical side if you don't mind Neil so I mentioned before Dropbox Dash will connect to different third party apps will go and retrieve various documents or images media from all these different sources this could be you know different SAS apps this could be your HR apps it could be Google Doc you know Google Drive it could be you know your project management tickets etc and we'll bring that all together and we do something called content understanding on this material and what that is is you know if I get a let's say a Google Doc well that's a bunch of text that's easy enough to understand but what if I get an image how do I extract any sort of relevant information from that or what if I get a PDF PDFs are images there's text there's sort of a combination and we we do a bunch of work to really pull out all of the the right information from things like PDFs or imagine it's a video think about for a moment that scene from Jurassic Park where you know they they saw the dinosaurs for the first time and they sort of turn to the side and they turn take off their sunglasses and they have this look of dismay and awe as they see the dinosaurs for the first time what if that's a video that you want to find later how would you do that how would you retrieve that information well we use different multimodal large language models to extract what's happening in that scene so once you start to build all the the understanding of these different documents then we take it a step further we build a knowledge graph and we start to connect relevant information across apps so for example you know you're working on a project there's people involved there might be a document there might be a meeting transcript all that you could form as a a almost a an inside a bundle of knowledge then we go ahead and we will index all that information so all the content has been understood it's been indexed and we've even created these knowledge bundles and that's what makes Dash so powerful that we're able to do phenomenal search phenomenal retrieval once you do that chat becomes far more accurate agentic work becomes far far more accurate and this is the kind of the key this is why context aware AI is really the future of how we think about AI especially at work feels like an incredibly exciting time for you there and I appreciate you've already shared so much with us today but trying to get a few teasers out of you are there any other upcoming product to now announcements that might demonstrate how Dropbox is continuing to evolve beyond a file storage company and also the the grand vision for AI at Dropbox any teasers you can leave us with there yeah absolutely we are integrating a lot of the AI features the dash teams built into Dropbox effectively making the platform smarter and extending its capabilities from just file storage to more understanding your team's content so you have more you faster access to your information you have smarter search and of course you have the ability to act on content without switching tools and like I said a little bit before Dash is also now available as a self-serve option so if you're a small team out there you can go and sign up and start using Dash in minutes rather than just going through the sales team in general you know Dropbox we do want to create a world where we are the most intuitive place for work where content is easy to find and teams can focus on the bigger picture items rather than that busy work and that's really the key on where we think AI can be incredibly incredibly impactful and for anybody listening that would like to stay in touch with all the kind of announcements that we're going to be seeing over the months ahead and equally find out not just about Dropbox but Dropbox Dash any websites you just want to mention one more time just so people can go in and check those out and keep up to speed with everything absolutely so if you are interested in Dash head over to Dropbox.com/dash of course you can follow Dropbox and Dash on Instagram LinkedIn and X the handles at Dropbox and you can find me I'm on LinkedIn and X and I do provide quite a bit of updates as well oh okay well I will have links to everything you mentioned there including your X and LinkedIn channel and I think this year in particular we've seen more and more work slots but today it was great hearing about the cause of it the context rock that you mentioned there but equally the flipside where we're heading what we can do now and there's so many great things coming to be interesting to get you back on next year in 2026 and hopefully see how we're moving beyond these things and really unlocking new opportunities for increased productivity and better working etc but Josh thank you so much for shining a light on this today yeah thanks Neil for me I think today's conversation was one of those that leave you looking for your own workflow in a slightly new way and Josh broke down what the messy reality of information overload and digital clutter but showing how context aware AI can take the sting out of these things and we've all seen workslop AI slop and context rock and I think it reflects the honest tension that many teams are feeling right now but thankfully Josh offered a path forward one where smarter retrieval tighter context and steady feedback loops create something that feels useful rather than overwhelming and Dropbox Dash seems to sit at the right intersection of everything we're talking about here it's treating AI as a practical tool not a spectacle and I think Josh is thinking on AI fluency will also give leaders something very real to work with and I appreciate it just how much he shared about the lessons inside Dropbox and the shift from single player AI to multiplayer collaboration of sorts but I'd love to know what stood out for you in this conversation does context aware AI feel like the missing piece in your own workflow or do you see other changes coming first love to hear your thoughts tech talks network.com and you can also send me a DM on LinkedIn x Instagram just @neolcqs but that is it for today so thank you as always for listening and I'll return again tomorrow with another guest bye for now