Databricks co-founder Arsalan Tavakoli: 95% of enterprises still can't get AI agents into production — data governance is the real blocker

Have a good weekend. Let me tell you about Crowd Strike. Your business is AI. Their business is securing it. Crowd Strike secures AI and stops breaches. And without further ado,

uh we're going over to the co co-founder and SVP of field engineering from Data Bricks, Arcelon. Welcome to the show. How are you doing?

What's happening?

Hey guys, thanks for having me.

Good to see you. Great to meet you.

First time on the show, I believe. Uh please introduce yourself. Tell us what you do.

This is when we say uh first time uh first time tuning in, longtime watcher, lurker or something like that.

I appreciate it. We enjoy talking to your co-founder.

Awesome. Well, yeah guys, um Arcelon Tavakoli, one of the co-founders here at Data Bricks. Yeah.

Um you know, run what's known as field engineering. So, think about

anything technical not in core R&D. So it's about a third of the company um roughly give or take and then uh before that like all the other founders came out of Berkeley and as my adviser like to say I joined the dark side in between and meant I went to a consulting company McKenzie spent about four years there before we started about 12.

There we go. We love we loved you know we love to hear that.

Put in put in your time. We were we were on a flight the other day and and uh John and I were asking this guy uh to switch seats with us cuz we like to we like to we we uh some people are surprised, but even after 3 hours of recording, we still want to keep uh talking. And uh the guy had a UBS shirt on and John John was like, "Do you work at UBS?" And he's like, "Yeah." And John goes, "Thank you for your service." And the guy just started laughing.

He thought I was making fun of

Thank you. Thank you for your service at McKenzie. You got to thank your bankers, your Swiss bankers specifically. Uh, I want your reaction to this Tyler Cowan post. Today will go down as some kind of turning point somewhat arbitrarily, but it's okay if journalists and historians have to present things in that manner. He posted this at 11:00 a.m. on February 5th. That of course was the day that both Codeex 5.3 and Opus 4.6 launched. It feels like he's talking about software engineering, automated software engineering, AI agents, sort of everything coming to a head. How have you been processing this week? Does it feel like there's a turning point? Is this all part of some smooth curve? What's been your view on the ground?

Yeah, look, I think uh I've lost track now based on those definitions. I think we've had like 3,000 watershed moments in the past year that everybody said is like the turn

more than Yeah, more than one. Way more than one a day. We have three hours of content every day. It works perfectly.

I feel like every day I tune into X and somebody's like, "This is the day. This is it.

The singularity is here. It's over. We're back."

Look, I I mean, I I think it's more of a smoother curve, right, as we go through. And clearly every day, I think it's a combination of the technology is getting better, but I think a lot of it right now, what holds people back, what they can do is less the technology, but it's more people learning how to use it, how to deploy it, all the things around it to to put value. Like that's going to be the long tale. I think um we're not even close to harnessing the technology that we have today, much less saying, "Hey, it got a little bit better. Now everything's going to change." At least from what we're seeing.

Yeah. So talk about your forward deployed engineers, whatever you call them, like implementation, how you like to work with companies, and then I want to uh compare and understand what it might look like at the labs as they sort of dip their toe into the FDE model.

Yeah. Look, I think um so I think that there's two things. A lot of places when they talk about forward deployed engineer so this is the cool term right now I guess to use is this concept that says go to ask people what is it that you want what's the outcome that you want to drive you want to do you know revenue forecast optimization you want to do demand forecasting and then I'll build something for you

the problem is many of the organizations just came the talk about that are kind of like okay I'll build something custom for you um I'll get it fast but then the hard part is is that scalable is it open how does that work so data bricks is probably a a little bit different than the rest of them cuz we've got an actual platform and product that's used by over 20,000 people. But I think right now when you look at it in the AI world,

yeah,

what works for us at least is um we go in and some organizations right now are saying help me with a specific problem like I want to do you know let me get my data in order which is important for AI or let me kind of do a migration or let me set this up and then we have the same set of somebody says um come help me like transform like Fox new if you know Fox Cletus the sports app that they have that you can ask questions like that one they were like come build something

Paul at Fox is a good buddy I don't know if Yeah.

So he's like come build it soup to nuts for that. And that was one that we did everything end to end and deliver something that works um that they want. So we span the gamut from helping you work with products to kind of actually delivering outcomes. We do both.

Okay. And then uh best practices pretend I'm a CEO of AI lab who will remain on a nameless list and I'm trying to hire 700 a couple thousand uh consulting type four deployed sales engineering roles. uh what are best practices? How fast can you actually spin someone up to become a you know an effective for deployed engineer? What's what are the pitfalls?

Yeah. Um lots of questions embedded in there. So um look, I think first and foremost what you're looking for today for people is they need to do two things. Um one they actually have to be able to build hands-on. I think you end up in the solution engineer space getting a lot of what I call like the slideware demos like hey I can I'm a walking dictionary I can tell you about things I can build a deck um it just doesn't really cut it anymore. So one more that what people say is come show me rapidly prototype. So they have to be able to build and they have to be able to use AI tools to build otherwise they're going to be slow. The second thing is um you can't just go get engineers. They have to be able to talk to customers because like the most important thing is if I go walk to your basically CEO or CIO, I have to be able to say what's your business problem, understand why it matters and decompose that into technical pieces. So you need people who have both of those. Um so you have to screen from it and it's different than what the what I would call the traditional solution engineering interviews are in terms of how fast to get somebody ramped up. Look, you're hiring smart people. Generally, to get familiar with all the tools, we'd say 3 months is fast. 6 months is when they're operating at full steam. It doesn't mean that you can't drop them in front of customers, you know, in the first couple of weeks, but that's the amount of time that they get set like what are the most common problems? What's the most efficient way? How do I navigate an organization? And you can hire a lot, but absorbing too many people too fast just means you're also going to candidly degrade culture and skill set.

Yeah. Um, we were just talking to Doug Olaflin about how he's using clog code. He's he's vibe coding all these things and a lot of what he was describing were uh sort of like oneoff data analyses or uh small smaller projects that could sort of run locally be compressed like spreadsheet level certainly not you know big data cross you know like data bricks level infrastructure required but I'm wondering if you see a world where an individual would vibe code something that's so big and processing so much data that the agent or the individual vibe coder is actually pulling data bricks off the shelf like do we are do we go to a world like that I've had ideas for vibe coding projects that have been like yeah sure I'd love a personal assistant that read the entire internet every day but that seems like not on the table right now but are we moving to a future where that's possible

uh yeah so two things I think one we are but second like I sometimes say from a data bricks perspective I wish I could walk in you know in like Men in Black where they have that button that causes people to forget.

Sure. Yeah. I think that there is this vision of data bricks of hey it's really really good if you have a bunch of data and stuff like that but but actually like a bunch of our users today what's really valuable for your use cases even if you say you want to work for a spreadsheet you're like

I don't keep that who keeps anything on their local laptop anymore right just like I work in an organization I just want a place where I can access all the data big or small but it's governed it's high quality so actually a lot of the people inside of data bricks use I mean called genie right because you can get asked for is mainly like can I get a world where I can do exactly what you said I can quickly b code what I need um

small amounts of data but it's governed I can visualize it I can iterate on it so that's the first thing I'd put

I think what you start seeing is every everything that you start with for like any project it starts small great I want my data then your next thing is oh but it would be really cool if I could merge it with this and it would be really cool if it could do so as you get it to work it naturally balloons into something that both leverages more data is much more computational much more comprehensive. So you very quickly end up in a world where you've outstripped the can I just do it on my laptop even for the average user.

Yeah, we were uh we were talking to Sam Alman yesterday. We were about different bottlenecks and he was saying that data was a bottleneck for a while at least like a rumored perceived bottleneck. Uh it really hasn't seemed to bottleneck progress in any meaningful way. uh but I'm wondering if you see uh if you look at the growth of the data market or data is the new oil just is there a boom in in the scale or or importance of data with the Neol boom there's more and more companies that are training different models it feels like there's both more data being collected than ever but then also more data being generated than ever so can you just talk about zooming out like what is the what does the growth of just data look like in the US economy for me these days.

Look, I think the the growth of data is significant, right? I mean, there's no question about that. And look, you said data is the new oil. I think one, you've got to define you got to separate out the world of the consumer world, which I think about where many of many of where many people's familiarity with AI is, right? I want to ask a question. Help me understand, you know, is this a cold or an allergy or plan my travel vacation? And that's the I need the whole corpus of the web and that people that amount of data is kind of finite. If I if that makes sense and new one people are generating new ones through AI but that's one

I think that there's a separate question of why you said most people focus on data in the enterprise world it usually goes the following I need AI I have to pick a model let me pick which model is it open 4.6 six is it the new basically GPT model is it Gemini etc great I started crap I now I have to figure out how to deploy this and that's a lot about security evaluations governance great now I did this and my accuracy is 60% 60% is kind of crappy accuracy how do I get to 100% yeah

the answer is you need to have much better data so almost everything comes back to you know especially in enterprise lots of data spread out in a bunch of places how do you actually get it together know what you Make sure it's high quality and relevant and recent because that's the other thing that's happening. Data that you had from 6 months ago may no longer be valid right now. So, how do you make tell the system what's valid, what's recent? That's where we spend a lot of time on and I think enterprises spend a lot of time on figuring out like how do we actually make these agents work? And it's a different problem than it is for like general um consumer search.

Yeah.

What is uh what's your software buying framework these days? Huh?

What do you

what do you

like what do you decide like what when when you're evaluating hey we there's a job to be done or should we should we go and find a vendor that can do this? Should we do it ourselves? How do you what's what's your line?

Yeah, look I know that the cool thing is to say this following. It's like guys why would we buy software anymore? Like you can just v code everything. I think anybody who says that generally is not in the business of actually using software and production for most like

it's like build is easy maintaining and sustaining it is hard. I think um the short answer is in a world if I can I much rather buy and build right like that's just the reality of it. Um the the problem though is you need you have such unique needs right now and so if I look at it and see that the gap between what I can get off the shelf and what I need is really huge then I have to look at do I really want to depend on another vendor um for kind of being able to like how business critical is it for me. So, as an example, much of the things that we're doing to overhaul field engineering at data bricks to say, how do we use AI to accelerate it, there's nothing off the shelf that gives us what we need there, right? If there's a piece, I'll use it as a component, but we have to go build it because for me, it's like I'm iterating really, really fast to figure out what the job to be done is, how to deliver it. So, we have our own development team. But on other things like my CRM or something like that, that is not something that if I can help it, I'm looking to build internally. Like, I'm happy to use something externally. um given that I imagine my needs are similar to many other ones and there should be things that could meet that needs pretty quickly.

Yeah. How big is uh AI agent adoption broadly? I I want to know the details or the high level of the 2026 state of AI agents and specifically like how much of this is like yes they check the box we're using agents but it's just a test or it's very narrow or it's the one vibe coding intern that builds some fun workflow that's like yeah it's a bot it replied it's it's agentic but it's not the core business. Uh talk to me about the roll out and adoption of AI agents in real enterprises real businesses. Yeah, and I know I'm supposed to give like the party line and the popular thing, but look, uh, I think you can look that report that you cited talks about something like 95% of people still have a hard time getting it in production or at least getting value out of it the way that they wanted. I I think you have still an element of people who are like beating their chest. You go like, I have a thousand agents out and I'm like, okay, how much how many people use them? And they're like, why does that matter? I was like, why does anything else matter? Right. Um,

yeah. It's the same thing as like bragging. Yeah. CEO bragging about headcount and it's like, "Wait, what's your revenue per employee?" They're like, "30 grand." You're like, "Okay, maybe maybe you're focusing on the wrong thing."

Yeah. Yeah.

Well, look, I I look, so I think that there's everybody we talk to is trying to and there's some statistics you saw in there that make sense. Like most people are no longer doing like, "Hey, let me get a chatbot." It's like multi- aent orchestration, right? Like they're using multiple things. Second,

nobody is wedded to models anymore. I think like 77% of our customers use multiple models as well to pull it together. Um and I think that the most popular use case is um you know beyond internal use case something called like market intelligence. So like how do I figure out what my customers are doing? What's the next best action I should take and the like. Um I do think it's still early like most people um like it's rare to see somebody say I have a lot of agents in production really effectively as I mentioned the most the place where we spend a lot of time on is how do you govern it and get security how do you get really really high quality out of it how do you balance cost and quality and that's where like our whole offering around agent bricks is mainly around how do you actually get this thing in production um and so early days we're seeing we're seeing at least people turn from I spun up a chat GP PT interface I have AI to like what's the ROI I can drive from it and realizing what are the blocking factors data or some of the things I talked about to get there. uh given given your experience at McKenzie, there's been some news recently that uh Enthropics working with I believe Accenture uh OpenAI is hiring consultants internally. Obviously, the big consulting firms want to get in on the action. They've been doing a good job making money on generative AI, maybe making more profit than a lot of companies. Um but uh when when you think about swarms of consultants going out into the world to unlock the power of agents like what are you going to be looking for to see uh kind of their success? Look, one, I think we've gone through these iterations before and um in general, one, I I'm hardressed to believe that every enterprise is going to all of a sudden have an army of their own really, really effective vibe coders, right, that are going to build everything themselves. I think that there's still going to be the dependence on external parties to help make something real. Um I do think that there is a element that the SI firms right now are trying to figure out how do they reinvent themselves a non-trivial number of what it was before was um how do I do some more let's say a little bit more of a commoditized task where

SI is software integrator

oh sorry system integrator sorry like when you think about the asentures the TCS the cognizants and the stuff

um you know so it's like how do I throw a lot of people to do it I think the expectation from people now is going to be great. I'm going to hire you, but in the ele element of AI, I expect you to be able to do things faster and I also expect to be able to get out of more uh out of each of the individuals. So, one, go build applications for me, two, do it faster, and three, you know, your cost, your economics should be better. So, I think that there's going to be a push on doing things like that. But otherwise, we're still going to need um Worlds of Folks. They're also trying to figure out how do they expand their like teams with like many many more agents that can work faster. So, it's more of a shift in the model than economics, but you will still need consulting organizations to help.

Well, thank you so much for taking the time. Jordy, do you have anything else?

No, great to meet you. Come back come back on more this year. The really helpful context.

It's uh Yeah, you guys have such insane visibility into how

uh into the at real diffusion. So, it's

or we can just wait a week. We can wait a week. It's a new day. Yeah, it's going to change in a week. Changing pointed watershed moment every hour.

Every hour of every day.

We'll we'll chat with you about it. Have a great rest of your day. Great to meet you. Thanks for coming on.

We'll talk to you soon.

Let me tell you about Label Box, RL Environments, voice robotics, evals, and expert human data. Labelbox is the data factory behind the world's leading AI teams. We got a bunch of breaking news show out with.

Breaking news. Hobbit inspired startup becomes first new bank greenlighted 2.0. One of the most insane headlines. Boom. That one is for Trevor Palmer and the whole team.

Arabore founder Palmer Lucky was one of the tech industry's early Trump supporters and he's known for pension for Hawaiian shirts. They're really focusing on Palmer's like non-banking relationships here.

Palmer is a banker now. everyone.

Uh, Arabore will will cater to startups and high- net worth individuals on Friday. Uh, it became the first newly created bank to receive a national charter under the second Trump administration, launching with $635 million in capital. It says it will occupy a hole in the market left by the collapse of Silicon Valley Bank. Um, the bank is the brainchild of Palmer Lucky, one of the tech uh industry's early supporters of Donald Trump. Its founders include its funders include Lux Capital and Horwitz 8VC. Uh Allah Gil and founders fund is lined up a handful of potential defense and industrial tech focused clients including ones with ties of teal and other investors who say they are eager to do business with a bank that understands their needs.

You can think of us like a farmers bank for tech said Lucky who will serve on Airborg's board but will not have an operating role. I think most farmers banks won't claim they are the best bankers in the world, but they do understand farmers. Uh Chris P in here. Uh he said, "Is a lo is a local bank going to lend against you?" No. He said, "If uh if you have less than $10 million in revenue and you're struggling to secure a loan from a traditional bank, investor expectations are high." Arabore was valued at about $2 billion in a funding round last year, over seven times its book value according to investor pitch deck. a subsequent round uh valued airbor at four billion. So off to the races,

the timeline, the pace from the team getting this done. Obviously there's, you know, the regulatory side, but there's actually the execution to get here. I remember talking to Trevor uh last year, uh Trevor over at Airborline and I was like, "Yeah, that's sounds great. Like let's see, let's see how it actually turns out." And

uh that you know, incredible pace. So

yeah, we'll have to have Paul on the show to talk more about it.

Real bank,

we have another we have another gong hit for Jennifer Garner's Once Upon a Farm.

IPOed today. Shares popped 17% in public market debut. It's on the New York Stock Exchange, baby.

There we go.

Interesting how I wonder how they uh priced it.

Yes. Uh so they raised almost $200 million, $197.9 at a valuation of $724 million. Um and uh the company's IPO comes as shoppers and policy makers alike have pushed back on ultrarocessed foods, particularly when consumed by children. Uh so it's on the New York Stock Exchange now under the ticker OFRM. Cavoo Venture Partners, Consumer Partners, Cavoo Consumer Partners, uh was a big uh backer of this company and probably has done quite well in this IPO. And also, Techno Chief had a little bit of an update here. Super random, but Kavu Consumer Partners just raised a new $325 million fund for better

good timing. Hit it again.

So, congrats over there. Uh, he's been investing in the category, worked on vitamin water. He's an absolute dog and an industry legend and we're happy to see success in this industry. Certainly AI resistant. It feels like something uh Tyler won't get out of bed for. He's he's knocking it. But uh you know certainly opportunity.

No, I mean in a post scarcity world every every child in the world could be

enjoying a fun day at the New York Stock Exchange for everyone.

Do we have anything else? Um, not not too much. Uh, there's a bunch of stuff. Uh, I think we went through most of this stuff. There's there's new data from ramp rarian shares the fastest growing companies on ramp. AI infrastructure means open source is likely ticking up. AI dev tools are not going away. And AEO is the future of marketing. He calls that try profound. So ramp's top software vendors the fastest growing. were anthropic lovable cursor 11 labs superbase uh replet interesting uh cursor has been seeing like some there's a bunch of FUD on the timeline people are saying I don't need it anymore but it's still growing according to ramp um little bit of a narrative violation there and then uh trending breakout growth relative to size paper for software design cerebras for AI infrastructure juicebox for recruiting AI runear for AI infrastructure clarify for AI infrastructure nova I AI for AI infrastructure crew Brusso for AI infrastructure, modal for AI infrastructure and then uh peak AI for SEO and AEO intelligence and also profound. So this is from the ramp economics lab ramp.com/data. You can go check it out for yourself. We love chat chatting with Ara about what you see.

Well uh it was a brutal week for tech.

But uh at least we had an opportunity to put on the white suits and I can't wait for Sunday to see the ads. Yes, it's gonna be a great one. The Super Bowl.

John and I will be there.

Yeah.

With with uh with Ramp and some other friends. And we will be glued to our phones.

Yes.

In the back of the of the of the box just

just watching the ads. So,

yes, for sure.

We hope you guys

plant the bomb. In other news, after 12 years at the Wall Street Journal, Joanna Stern is launching her own consumer tech media company. Congratulations to Joanna Stern. hasn't shared exactly what it is yet. Uh, but there is a link you can go sign up. This is my next thing.com. Good name. I like it. She has been writing for the Wall Street Journal for quite a long time. I've always enjoyed her consumer tech coverage. And of course, we'll have her on the show when she launches. Leave us five stars on Apple Podcast and Spotify. Sign up for our newsletter at tbn.com.

Goodbye.

Nice work, brothers. I'll see you on the next one.