When Anyone Can Build Anything: The New AI-Native Stack

Summary

A panel of founders from Decagon, Windsurf, and Glean, alongside Databricks co-founder Patrick Wendell, discusses how natural language is becoming the primary programming interface, collapsing the distance between human intent and running software.
The panelists agree that base models alone are insufficient for enterprise AI applications; successful deployments require governed context layers, specialized training, and multi-model routing strategies that select across multiple providers for cost and performance.
The panel examines the empirical evidence on AI's impact on software engineering employment and productivity, noting that the primary bottleneck is shifting from technical skill—which AI increasingly provides—to creative vision and judgment.

When Anyone Can Build Anything: The New AI-Native Stack

The act of building, whether software, knowledge systems or workflows, is being fundamentally democratized. Natural language is becoming the new interface between human intent and outcome. Natural language is becoming the new programming language. Ideas are becoming applications in minutes, not months. This panel brings together the founders of AI-native tools that are collapsing the distance between human intent and running software. Together, they'll explore a near-future where dramatically more people can ship products, and where the primary bottleneck shifts from technical skill to creative vision. The big questions: What happens when the cost of building software drops by orders of magnitude? Are we heading toward a world with far more software, fewer traditional developers or both? When AI agents can build, test, and deploy autonomously, how does engineering culture evolve? What's still hard and what will remain uniquely human in the development process? Where do AI agents break down and are we moving fast enough to find out before it matters Join the founders reshaping the most fundamental act in technology for a candid look at where this lands in the next 2–3 years.
Talk By: Ashwin Sreenivas, Co-Founder, Decagon AI ; Jeff Wang, Windsurf ; Patrick Wendell, Co-Founder and VP of Engineering, Databricks (E-Staff Care Internal Speaker); TR Vishwanath, Co-founder & CTO, Glean, Inc ;

Chapters

FAQs

What is Decagon and how does it use AI agents?

Decagon builds AI agents for enterprise customer support and customer experiences, allowing agents to talk directly to end customers to resolve problems and handle cross-sells, upsells, and operational workflows. A core concept is agent operating procedures, which lets both technical and non-technical enterprise teams build and deploy customer-facing agents quickly.

Why aren't base models sufficient for enterprise AI applications?

Enterprise AI applications require context, governance, and domain-specific knowledge that base models do not provide on their own. Panelists describe the need for governed layers over enterprise systems—such as Glean's enterprise knowledge layer—and multi-model routing strategies that direct queries to the best model for each task based on cost and performance requirements.

What happens to software engineers when AI can write code?

The panel explores evidence showing that AI tools are dramatically increasing what individual engineers can accomplish, but the impact on employment is still emerging. The primary bottleneck is shifting from technical skill, which AI increasingly provides, to creative vision, judgment, and the ability to define what to build.

How does Glean describe its role in the enterprise AI ecosystem?

Glean co-founder and CTO TR Vishwanath describes Glean as providing two products: a governed layer over all enterprise systems that supplies company knowledge and context, and an enterprise AI assistant that can answer questions, coach employees, and take AI-powered actions over all the company systems an employee has access to.

Full transcript

[00:08] All right, welcome everyone. I know we uh still have folks filing in. Uh so no worries. Yeah, please fill in available seats. Um we'll do some quick introductions and then get into our conversation. Um so yeah, I'm Patrick Wendell, co-founder of Data Bricks. Um we're super happy to be uh working here
[00:21] with founders and and key leaders from uh several uh AI companies that are building really cool and interesting products. You all may be already customers. Um, but today we're going to just talk a little bit about some high
[00:33] level trends in AI. Uh, some of the reflections from these folks who are really at the frontier working with applying AI models into different business problems. And we'll hopefully also have uh some audience Q&A as well. Um, so to get started, I am actually I'd
[00:47] love to invite each of you to maybe briefly introduce yourself and probably folks are familiar with your company and product, but feel free to uh to clue us in and maybe we'll start at the end. Go ahead, Ash. >> All right. Yeah, thank you for having me. I'm uh excited to be here. My name
[00:59] is Ashwin. Uh I'm one of the co-founders and president at Decagon. Um at DecaGon, we build AI agents for enterprises. Um most focused on things like enterprise customer support and customer experiences. Um you can essentially
[01:12] think of Decagon as an AI agent that talks directly to your end customers. It uh resolves their problems, helps with things like crossells, upsell, sales, operational workflows, things like that.
[01:24] Um and you know one of the core things uh that we built out when building out this product was this concept of agent operating procedures which is the ability for teams across the enterprise both technical and non-technical to be
[01:36] able to quickly build and deploy these customerf facing agents. >> Awesome. Uh next up Vish let's hear >> Hey everyone I'm Vish. I'm co-founder and CTO at Glean. Glean is an enterprise
[01:48] work AI platform. You can think of it as your you know we provide two products. one is a governed you know uh layer over over all your enterprise systems that
[02:00] can provide all the company knowledge and context and AI powered actions over those systems and then on top of that we provide an enterprise AI assistant for everyone at work and you can think of
[02:13] this as this assistant for you who can answer and uh any questions you have over all the company knowledge you have access to it can it is your coach as well. It has a deep understanding of
[02:26] what you work on, what your team works on and is able to proactively help you get accomplish your goals. And then the final thing is it's your assistant in the sense that it can help automate, you
[02:39] know, your uh day-to-day workflows and make you much more productive. >> Cool. And I'm Jeff uh president of New Enterprise at Cognition. And at Cognition, we've built Devon. It's a agent you could run in the cloud. You
[02:51] hand off work. It verifies the work end to end, gives it back to you and what we're seeing now is people are handing lots of work in parallel to these agents and uh very excited to talk about this with you all. >> And Jeeoff, your story about how you uh
[03:04] became part of Cognition. >> Uh yes, I was previously the Windsurf CEO and um and it was like a very weird uh dual acquisition situation last year. Uh and uh but now we're all one big happy family. >> Nice. Yeah. Amazing. Um, one thing that
[03:17] I think is so interesting about this group is that um, you know, when the when chatbt first came out and we first had these kind of foundation models, I think no one was quite sure um, how the separation of concerns would would arise
[03:29] in the industry between like the producers of the models, higher level applications and then end users. And I think one thing all three of you represent is application level companies that are doing really interesting and valuable things like you know each of
[03:41] you has hundreds of millions or more in ARR and extremely fast growth. So uh I think the question has been answered in terms of like where the value capture will be. There's certainly a lot of value to be captured at the application level but you all cater to slightly
[03:53] different end users. You know you're really focused on automating software engineering end to end. Uh you're obviously focused on everyday kind of office workers and empowering them and then you're working with in many cases your end user I assume is like a consumer that's interacting with the
[04:05] brand or something like that. So I think my first question would be at a high level um you know how come I can't just take my base model and my problem is solved you know like where is you know there's clearly value acrewing at this
[04:17] higher layer what are the difficult problems that you guys are solving and I'm kind of happy to hear from anyone I'm sure each of you may have an interesting perspective on it um but why don't I call on you first Jeff >> yeah I mean everybody's probably using a lot of coding tools out there um we're
[04:30] always trying to go where the puck is going and um a lot of work goes into the infrastructure so if you're spinning spinning up agents in the cloud, you need certain types of VMs and different ways to deploy applications and
[04:42] different ways to test the applications on the cloud. Um, there's also way for example like we have like Android emulators like Windows, you can do almost any type of use case uh that you want to verify the work is being done and I think this is very key because you
[04:55] can't really do this with just a base model. you do need a lot of tools to have the agent understand the codebase, understand what the data and the logging data is showing and then going out there and solving the problems. Uh we also
[05:07] have infrastructure to go proactively look for things to solve as well. So we have a thing called Devon automations. It's literally just like monitoring a Slack channel and actively helping you as the conversation is going. Uh so this
[05:19] is like something you you need to work on at the application layer to plug it in and then tell the the agents what to do. Um it's almost it does feel like there's another co-orker just sitting next to you and and helping you solve
[05:31] problems. >> Yeah. So it's fair to say it's kind of a it's like a lot of infrastructure knowhow and uh and resources that are like beyond just kind of the generating tokens. Is that fair? I'm actually curious to maybe go down the line like V
[05:43] same question for you guys. >> Yeah. And for us you know our our goal is to actually deeply understand the enterprise. You know what are the enterprise uh systems of knowledge you're using? what is what is the or
[05:56] structure what are the relationships so step one is actually deeply connecting to these systems with full fidelity in real time and understanding the governance and policies. So that's a big
[06:09] thing that the model doesn't provide like you know how do you actually respect the governance that you already have in place for your company and um and then uh you know not having a parallel you know governance uh
[06:22] mechanism in gleam. So that's that's a lot of like you know the security making sure that you know the AI can actually operate on on knowledge that is not sensitive. uh it's the wild west in most
[06:36] companies often you know there are very sensitive documents that are overexposed and so you need that um you know governance and permissions layer also to ensure that AI doesn't accidentally
[06:48] reveal things that it's not supposed to do so that's the first layer like where we invest a lot of time then on top of that is the notion of how do we actually provide the information that's most
[07:01] relevant to to you when you ask the question that's a function of you as a person as well as the type of question you're asking and it's a very very uh deep problem for um which requires you
[07:14] to build the enterprise knowledge graph as well as a deep understanding of you and what work you're doing. Uh and that's important because in the days of like you know token efficiency and
[07:26] accuracy you need to provide the right information uh to to do this task well. So that's the next layer and then the third layer is the harness as we we have talked a lot about like you know it's
[07:38] not just the model it's the harness and we invest uh invest a lot in the harness and the ability for you to bring u different kinds of models for different tasks. So we have we have RLD our own
[07:50] model for certain tasks but we also like you know use the appropriate model for the task and that's also um one of the thing one of the additional things we provide. >> Yeah. Yeah. Fantastic. I mean similar
[08:02] themes I would say here. It's all about like the the data of your customers and that data is secure. They may not actually want that data to be in the hands of you know a bunch of different uh vendors. So so in some sense you can
[08:14] be like the trusted vendor there around the metadata. Um and actually in I'm curious in your in your context I mean it's a slightly different context.
[08:28] experiences with AI, right? And in a way, um, the I'm going to build a chatbot that talks to my customers has at times been the introduction to AI 101 project, right? So, the the question
[08:40] that often I get asked is, well, Decon is expensive software. what do you do on top of me just you know feeding everything to you know GPG55 or claude opus etc and the answer is actually in
[08:52] two parts one is on the AI side and one is on the nonAI side right so most of our customers um tend to be you know some of the largest enterprises in the world and on the AI side the answer is around two parts one is around accuracy
[09:06] and performance right so we work a lot uh over the phone now over the phone latency is extremely sensitive Right? You can't in in a lot of cases use the biggest model with the most thinking
[09:19] time, but you also cannot sacrifice on accuracy. Right? So the the question becomes, okay, how do how do I now create models that perform with the same uh uh level of accuracy and performance as a frontier model but at much lower
[09:32] latency. Cost becomes another concern because if I'm having millions of interactions a day, uh I need to make sure that I can make this economical for me as a business. Security becomes another one which is I I might want to
[09:44] have strict governance over where these models run, what kinds of models run. So being able to have a an AI stack that can do meet all of these constraints is something that we spend a lot of time building in house. Hallucinations are
[09:56] another one, right? So even if you have 99% uh performance on not hallucinating, 1% at enterprise scale is 10,000 times a day. So that's a lot of things that we've built on the AI side to make these
[10:09] things performant. On the nonAI side, it's all around making this deployable at enterprise scale, right? Because if I'm an enterprise building these end consumer experiences, I need the ability
[10:21] to have governance over how I build these agents, who's making changes, who's allowed to deploy things to production. I need strong versioning and controls and tests. I need to make sure that when one uh when one agent builder
[10:34] makes a change, it does not break and regress on a number of other things because ultimately this is representing my brand uh uh to all of my end customers. So building just software that I can ensure that this is
[10:47] deployable safely uh where I can test, I can iterate, I can QA um at an enterprise scale is things that we've spent a long time kind of building in this kind of agent world. >> Yeah, makes total sense. And I think I'll quickly ask you uh before we wrap
[11:00] this question, you know, uh AI for customer support has a bad reputation. Prior to Genai, uh you know, all of us are just hitting zero trying to talk to a human as of let's say 20 uh 2020.
[11:12] >> Um how, you know, how much better are we now? Like and I just would love an honest assessment of where the industry is like what works, what doesn't work. >> Is it is it realistic that in some scenarios it's more desirable to talk to an AI than a human? Like where do you
[11:25] think that line is right now? Yeah, I I think that's a great question. Um, the way that uh we view it and I I think the way to view it is that customers at the end of the day just want their problem solved right now. Whether an AI agent
[11:38] solves my problem or human solves my problem or the software solves it out of the box doesn't matter, right? Like if I knew that an app on my phone could solve the problem, you know, I just go there and check the app directly. I don't want to talk to anybody. Um, the nice thing about AI agents and the version that we
[11:52] have today is that compared to let's say the traditional chatbots from a long time ago is that they're a lot more flexible, right? If I had a static decision tree, decision trees are actually great if you are along the happy path, right? But you go one degree
[12:06] off a happy path and everything breaks completely, right? And I think this is what caused people to say, well, my problem is I was always trying to get out of this tree. I I don't want to be anywhere on this tree. I want to go talk to a human who can do a much more sort
[12:18] of diverse reasoning about what I need. >> Exactly. And I think that is exactly what the LLM today have given us, right? Which is if you can solve the problem, it's solved immediately for you, right? And this as as a customer, this is a great experience for me because I called
[12:30] in, I explained my problem. It's able to personalize the response to my situation. It's able to solve the problem for me immediately. No wait times, no holes. I don't have to repeat myself six times over. And the added benefit is if it can't solve the
[12:43] problem, it now takes all that context and finds the right team within the org to solve this. Gives them all the context. I don't have to repeat myself and it kind of streamlines that process for me a lot. >> Yeah. Because and one more on this cuz I just find it fascinating. It seems like
[12:56] and this probably is a theme that you all have to deal with, but it seems like one one question is just having the AIs know when they can solve the problem versus when they can't solve the problem. And that's something we deal with at data bricks also like you know people are asking analytical questions
[13:08] and and sometimes the hardest thing is actually to make sure that the AI will not give them an incorrect answer. It will just say I actually don't have sufficient information to deal with this. So I do do you have to deal with that problem like >> exactly and in fact you know um a little
[13:20] bit ago when I talked about hey we need to build all these like custom models to do things. One of the models that we built is to detect hey did I hallucinate? Another one is to detect did I know the answer to this? And if not we just say hey we you know generated this response I'm going to
[13:32] check it. Oh, you know, it's not actually doesn't actually have the answer. Offer to connect to to human. >> It's like you've got to know what you don't know, you know, which is by the way, humans, some humans have problems with that, too. I wish more humans would just say I don't know when they don't
[13:45] know. Um, so the next topic I thought would be interesting for this group. Uh, one one unique thing about each of you is you are buyers at scale of token capacity. So you are some of the biggest
[13:57] customers of the foundation model labs or or of you know biggest users of open source models. Um how and in a parallel way many businesses now are this is starting to become one of their biggest expense either AI uh higher level AI
[14:11] products like you all are building or even just raw token capacity they're using internally. So uh I think you are in some sense ahead of the curve on how you think about um you know how much diversity you want in your suppliers. Do you what's the right strategy? go allin
[14:24] with one, you know, um um foundation model company and get the benefits of a deep relationship uh with maybe the cost of some exclusivity. How do you all think about, you know, your token supply, let's put it, and and what are
[14:36] the the types of strategies you're employing? And and anything that could be, you know, helpful for the average business that's starting to see these costs get higher and higher and has to think long term about who they work with. And I'm actually happy for any of you to start. I'm curious, uh what you
[14:48] guys think. >> Yeah, I mean, um there's Frontier Models and Frontier Model Companies, but there's also open- source. So you have actually a lot of options of what models to choose. So what we've done is we've created a new eval called frontier code
[15:00] and it's really breaking down what the models are good at that make it to production and if it's formatting it correctly if it's creating the right test coverage. We have all these dimensions of these models and we try to
[15:12] route to the models that are the best and most cost optimal as well. Um we're thinking about probably two modes. One is going to be like performance optimize versus cost optimized. And I can tell you token cost is like the number one topic. Every exact conversation I'm
[15:25] having, this is what they're asking for. And we've gone as far as offering an ROI guarantee. >> So if you are not getting your money's worth, we will give you back up to $10 million. >> Wow. >> Um and and we are confident of this
[15:38] because we are actively measuring how much code is making it to production, how much human level effort it would have taken. And this way a lot of our customers can feel a lot better uh that they're not token maxing or choosing the
[15:50] the best models for everything. Um so these are very important to us to um mix the models uh get the right model for the job and and show the customer they're getting the ROI. >> I'm curious if you were to say you know roughly how many different sources of
[16:03] tokens do you have? Let's call a model vendor a source or an open source particular model family. Is it one or two? Is it a lot? Like roughly. >> Yeah. For us it's probably about a dozen. >> A dozen. But um each customer has their
[16:15] own restrictions. Some customers might only be able to use uh from a region. Some might only use from a specific vendor. And we do have very sophisticated controls to um encapsulate what they're what they have access to. >> Okay. Okay. Got it. Um either of you
[16:28] two? I'm curious. Same thing. >> Yeah. So we we strongly believe in like model choice like model choice is a key feature of the glean product and part of that so as a result we support you know all the frontier models as well as a
[16:41] bunch of open-source models and we offer the choice to the customer and to the admin like we believe that the admin should be able to have current you know have controls so that you know not every user is using you know the latest opus
[16:54] model for the most trivial task and we have a default uh auto option which tries to which does the same thing which is like tries to figure pick the best model for for the particular task. Um,
[17:08] >> is your view long-term users are the right people to choose or or it's going to be more and more of the automatic routing or like >> it's absolutely the automatic routing. Okay. >> But but we do want the but you know we do want the control because if you're
[17:20] building a very special purpose agent at least till we learn till the agent learns what's the right thing like we want to give users the control on that. Um and internally as well when and so we
[17:33] offer like you know we offer customers they can either bring their own key or set of keys for models or we provide um you know the glean key which encapsulates all the models we offer.
[17:46] Got it. >> And and there we you know and based on the usage we sort of like negotiate between the different model providers. >> Yeah. And I'm sure Ash you're such a a higher level product you know I'm sure there's no model picker when I when I'm
[17:58] calling the phone. put cloud on please. Um, so how do you think about that? >> Our our opinion on this is actually a little different in that we think there are kind of best-in-class models for a
[18:10] specific task. Okay? >> Right? So we have uh in the tens of models. So when you call in and say hey I have a problem with X, it's not a single model that does something. It's actually a whole group of models, right? Like one is figuring out, okay, what
[18:22] additional data do I need? One is figuring out, okay, I'm going to start drafting a response. Another one is checking it to see did it make any mistakes. Another one like I said is checking to see oh did you actually respond or did you say I don't know right so it's a team of models and we
[18:34] believe each model we're trying to make it as small as possible again for cost efficiency uh second we're trying to make it as performant as possible um because you know you you can often if
[18:46] you're training models on very specific tasks make them extremely performant uh uh you know just at that very narrow task more so than let's say a frontier model that's 100 times the size. So we think of it less as routing between
[18:59] different models and saying you have a team of models each of which are very small and extremely performant at its specific task. Now um in terms of token costs like answer the previous question maybe in in two ways because I think it
[19:11] might be interesting to this group. We think of uh token costs both as token costs for our product right which is um here right which is okay which is a group of models that we use can we make them smaller can we make them more
[19:23] performant but we also think of token costs internally right like us as a business as I'm sure a lot of people here are and you know there are all the >> how many employees do you guys have now >> uh we're a little over 500 people
[19:35] >> so those people can run up some bills >> yeah people can run up bills um and we found that there's actually a lot of work that goes into making sure that people are using these tools well, right? And we've we've thought about how go to market teams use tools well, how
[19:49] GNA teams use well, how post sales teams use well, and how EPD uses uh these tools well. And we found that you actually have to do a lot of work internally as an org to help this get set up to to work well. I'll let I'll let uh I'll let you talk about the
[20:02] coding agents, but let's say even for non-coding agents, being able to figure out, okay, one, are we measuring token usage well was actually the first start. We found that actually uh for us and for a lot of companies that uh that we've
[20:14] spoken with, it's actually uh a longtail distribution, right? It's it's exponentially. You have a small percentage of people using the vast majority of tokens. So all we did was once we had visibility into who is using tokens and how we actually just went to
[20:27] the top 10% and said you you're using a lot of tokens. What are you what are you doing here? And some people were you know not using it particularly well but most were doing it in a particularly interesting way. So then >> yeah, like I'm curious, were they doing
[20:39] high ROI things or were they, you know, I left the for loop running while I was uh
[20:51] to show up at a at a sales meeting, I'm going to have the most personalized deck for this company. I have briefs for everybody that's attending on who is attending. Here are all the calls that we've had. Here's a here's a follow-up email right after the call. So, it's
[21:03] it's all about on on the go to market side, efficiency, personalization, ensuring that we're kind of showing up very well prepared for all of our customers. On the post sales side, it's um you know, we're an enterprise product, so we have a team that goes on
[21:15] the ground and helps customers. It's around how can we uh make that process much more seamless and smooth. So, we spent a lot of time figuring out, okay, what are people doing manually in house and helping them build uh uh internal
[21:27] tooling to kind of do this well. So it was all about measurement of every of every step along the way to make sure people weren't token maxing. >> Yeah, there's the famous saying of like you can't improve anything that you're not measuring, you know, so I guess measurement is the first step. Um so one
[21:41] other topic I I thought this group would be uh really interesting to hear from on before we get to audience questions is you know there's this big question about uh the the benefits of AI automation and to what extent they're aug they're
[21:53] augmenting existing you know uh humans and in some cases there's maybe substitutive you know they might be things that are are going to replace what a human is doing now um and uh it's you know people often ask Sam Alman or
[22:05] Daario that that question but I actually think they're asking the wrong the wrong folks. I think the real folks that will have the most insight on that is you all because you're building the application level companies and you're going in and
[22:18] looking at how AI is being used and you know where is it letting me where is it letting existing people do more where is it in some cases able to fully automate what a human is doing and how quickly is that happening such that they can those folks can be reused in other areas of
[22:30] the company. So obviously this is a a sensitive topic, but I do think it's really important uh uh writ large and I'm kind of curious where you see now you you're all you're all dealing with very different types of end users.
[22:42] You're dealing with consumers and maybe uh helping what would be a a classic customer support team. Uh you're dealing with everyday knowledge workers and and you're dealing with mostly software engineers. I know there's some overlap and it's not exclusively that, but where
[22:55] are we right now as you look at deploying your products in the the spectrum between, you know, augmenting versus fully replacing and substituting? And I I'm actually any of you can start on that, but I do want to hear from all
[23:08] of you. >> I mean, right now it's high ROI or stuff that nobody wants to work on. So high ROI is um for example, actually the combination of them is vulnerability detection and and remediation. So right
[23:20] now cyber security again one of the hot topics uh it's like tokens and and cyber security um and we are now automating um vulnerabilities as detected a devon will automatically fix it and then in fact
[23:32] there's like backlogs of vulnerabilities that you can just throw devons at and these are things that again nobody wants to like drop what they're doing to go fix these but they have to like the most important thing if there's a day zero vulnerability um and that's one example
[23:45] where you're augmenting and you're taking things off someone's plate and it's something that's very important higher. Um, but then there's tedious work. There's like uh bug fixing. A bug is reported. Someone has to replicate it. Nobody's raising their hands to do
[23:57] this, right? So, you can use Devon to go investigate, open up the app, replicate the bug. Um, and then there's things like modernization, migrations. Um, these are huge uh budget projects and
[24:09] again very highly repetitive work. So, very good for agents to work on. So, these are the things like we we try to target first. Um, when you talk about like Greenfield and you're making some new apps, you you you want some human taste and human judgment. That's where
[24:22] you kind of get into something interesting. It's like you're unlocking the things that humans want to do and taking away the things that they don't want to do. >> Yeah. So it's maybe more of a rebalancing of sort of who's doing what >> rather and it is I mean empirically uh
[24:36] even though software engineering in some sense seems if you know I'm a software engineer and others are and you just see oh wow this model just seems to be doing so much of what I used to do but um but empirically you know unemployment for software engineers is very low right now
[24:48] like it's actually been a hiring blitz in many ways for software engineers across the industry in the last 12 months as the coding models have gotten really good so there's some interesting sort of economic uh um dynamics mix here of like does this tilt you towards maybe
[25:01] your software engineers are so productive now that the leverage you have on each individual is so high that you want to hire more of them like I'm curious you know both I'm sure you're hiring a bunch of engineers in your company and when you go in and work with the company and you you know finish your
[25:14] your kind of setup how does the labor uh rebalance basically >> yeah I mean there's two categories right now I'm seeing in in the field which is the category where you have unlimited ambition like a road map is is so
[25:26] stacked uh or you have a lot of work that you've scoped And then there's this other category of like you don't really know what the team's going to do if they finish the work. And basically what we're trying to hopefully unlock in the world is to be more ambitious to pull in
[25:40] your road map. Um with startups all everyone here on stage is like it's a battle like we are hiring more engineers because they're more productive and they're getting our road map ahead. But there's some larger companies who don't
[25:52] really sense the urgency yet and they are doing their work much faster but they don't necessarily know what to fill the work with. And we are there to help be their partner to identify that work to scope that work with them. >> And I guess I was going to say one point
[26:05] I was going to ask you next but go ahead. Yeah. >> One point on uh you know software engineering jobs being safe is you know the the cognition team is one of the most sophisticated users of coding agents and they are hiring more engineers. >> I know exactly by the way and and it's
[26:17] for instance um anthropic and opening are hiring on a tear hiring engineers. So like and sometimes I find the I just find the the the two things a little bit in contrast. It's like okay on the one hand you know we feel all software engineering has been automated on the
[26:29] other hand we have 500 job openings for software engineers like those two things seem slightly inongruous um but I think >> each each software engineer we hire is like hiring 10 software engineers >> exactly like I think this phenomenon does exist where like if if the if you
[26:42] are able to get huge marginal benefit from each like individual human because they're so levered now with coding tools um you still will hire you know it's just it's just a question of like can it impact your top line and continue to grow your business when you hire these
[26:54] people >> I'm actually curious so you're on sort of very other end of sort of I would say skill sophistication in some ways not to say that it's not a complex job but um yeah what are you seeing similarly >> yeah you know the the way that we think
[27:06] about it is um is through this kind of lump of labor u fallacy right which is there isn't a fixed amount of work to be done when it comes to businesses talking to their customers right so um if magically every
[27:21] business were like hey I'm looking at my existing queue of customer interactions and all of them disappear It's not like they would all sit back and be like, well, we're not talking to customers for the next 2 months, right? There's always a longer set of things that you could do. There's always deeper that you could
[27:33] go in with with the customer. So, the things that we've found with our customers is twofold. One, because we actually fundamentally change the economics of interacting with the customer. Turns out some of our
[27:45] customers want to do more, right? Let's take an example of data bricks. If let's say you were onboarding a new a new user onto data bricks somewhere within within a large org in an ideal world you'd say
[27:57] great we're going to send someone over they're going to sit down with you and they're going to show you everything about the data bricks platform but it's not economical to do that for every single new person but if you had an AI agent that could do this at like a hundth the cost then you'd be like
[28:10] amazing I I want to do more of it I want to do more customer interactions because it keeps my end customers happier and I can do this in a way that is economical for Right. And so that's the exact same thing that we're seeing, which is when
[28:22] we create these AI agents for customer experiences. All we're doing is we're changing the economics of being able to interact with these customers more at scale in a more personalized way. And companies are like, great, now I want to do more of this. I want to create more
[28:35] opportunities for my customers to tell me about their problems so that I can fix it before it gets to a point where these are now unhappy customers. >> Yeah. Yeah. Absolutely. So um in a minute we'll take um customer or I'm sorry audience questions but um yeah I'm
[28:47] curious Vish like similar in your experience or >> similar but I think the like for because it's uh everyday knowledge work like a lot of it is first of all you know uh it
[28:59] really helps uh we find that a AI really helps us turbocharge productivity. So you get things done that you know would take several hours you would get it done in the order of a few minutes right uh
[29:11] for example I you know there's a meeting I need to prepare a slide deck like it'll take me a lot of time to think through things or like write a design dock for sharing and I just summarize
[29:23] with bullets AI is my thought partner and I can generate that really quickly. >> Uh >> by the way I consider myself the glean power user and I still have a job so far so that's a good sign. I think it's made me better at my job rather than um been
[29:37] you know ready to replace me. But go ahead. That's and then the the other thing is you know classes of work that you would never do like for example you know all our uh product managers like they are able to for for a feature they
[29:51] can say you know go over thousand like hundreds of gong calls and then figure out what's what's working with this feature what's what's not working and uh you know what what what are customers
[30:03] telling us with respect to like you know what are the competitive gaps that are there this kind of analysis is is something that you can do and on a regular basis like clockwork and that's the sort of thing that would never
[30:16] happen before that. So the this is not replacing work it's actually making you much much more productive. >> Yeah. Um and so that's that's the general theme of things and the other part is like consistency in your
[30:28] business processes like if you're um like you know doing an interview debrief for example there's so much of inconsistency so many things you need to um you know figure out as to like you know is the feedback all structured
[30:40] properly and so on and and you can have like agents that actually make all of those things really really structured and consistent and that that really improves the way you work.
[30:52] So every day to Rick's employee, we have more than 10,000 employees. When they open Chrome, Glean is there. And that's just how I start my day now. You know, it's telling me what I need to work on. I'm like in there trying to figure out, you know, oh man, I have this I have to host this panel in in 10 minutes. You
[31:05] know, remind me what this panel is about. He >> He almost didn't make it, by the way.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.