Subscribe
Share
Share
Embed
Conversations with designers, founders, and builders behind some of the best work
Diego Zaks Full Podcast V2.txt
English (US)
00:00:00.000 — 00:03:05.480 · Speaker 1
This is a 90 minute conversation you won't find anywhere else. I mean, I was interacting with an engineer for a couple of weeks until I realized that it wasn't actually a human. The engineer that built Cody was like, no, I'm not giving you Cody's code. I was kind of upset because, hey, wait a minute. Like you built this thing.
Like I want to use it. You said look. Start from scratch. Tell Cody to help you. You'll thank me later. Three days in, I was completely hooked. I have to build knowledge, learning rituals. So a heartbeat, a purpose. And like all of these softer things that make a person. That's what he was trying to get me to do.
Yeah, without just giving me Cody. I flew to New York for this one. And I want to tell you why every company right now is trying to figure out the same thing. How do you actually install AI, not buy a ChatGPT license, or mandate the use of cloud code, but install it into the way a real company works cross functionally, and most organizations are still guessing, but ramp.
Already did it. Check this out. 99.5% of their company uses AI every day. 1500 internal apps shipped in just six weeks. 12% of production code is coming from people who are not engineers. And according to ramps own data on millions of U.S. businesses, heavy AI users grew revenue at 27% a year. Meanwhile, companies that didn't spend on AI grew at just 3%.
That's nominal GDP, and those numbers are why I got on a plane, because ramp is the clearest field report we have right now that the AI installed company is a completely different animal than the one who just bought a few licenses. And my guess is the reason I think this episode is different from any other AI related podcast interview.
Diego Zax runs design at ramp. He's not the CEO pitching a vision or the CTO pitching the next stack. He is the person redesigning what the organization feels like after they mass adopted agent harnesses and AI coding tools. In fact, they've been sharing all of their lessons in public, and today Diego is going to walk us through it.
He talks about what AI fluency really looks like and how it shows up at ramp. In fact, while it was there, I heard a story about an agent who realized one of the leaders was sick, ordered chicken soup to his house, and showed up to his meetings on his behalf to represent his interests, completely unprompted, and another that involves anthropic showing up at ramps office because their data showed them that ramp was using cloud code more than anthropic was.
If you run a team, lead a company where you're just trying to understand what the next few years might look like from inside a company that got there first. This is a 90 minute conversation you won't find anywhere else. This is state of play. Let's get into it.
00:03:12.440 — 00:03:24.520 · Speaker 1
When we did all these drops, opus 4.7 and all these other harnesses and products, you have like a deep rush to check them out or like, where do you stand with that? As soon as thing comes out, we usually.
00:03:24.520 — 00:03:44.000 · Speaker 2
Just roll it throughout the whole company and let people play with it. Um, we're a little more cautious as to what actually makes it into our, like, continuing work. I know, yeah, 4.7 just came out. Um, everyone was already looking at it. Yeah. All of our AI assistants are probably on it at the moment.
00:03:44.000 — 00:03:47.960 · Speaker 1
So do you have, like, a slack channel that that stuff just gets distributed like. Oh, this just dropped.
00:03:48.000 — 00:03:48.400 · Speaker 2
I think.
00:03:48.400 — 00:03:49.080 · Speaker 1
It's
00:03:50.280 — 00:03:50.600 · Speaker 1
it's.
00:03:50.600 — 00:04:06.220 · Speaker 2
Very organic. Like everyone's just looking and people want to try it. And then it's likely that we're in the early access, and we have some cohort of the company that's already been working with the labs and just has access to it, and then just rolled out to everyone else.
00:04:06.220 — 00:04:11.380 · Speaker 1
What are what are you building? And I don't mean like that ramp. I'm just curious, what are you anything on the side?
00:04:11.820 — 00:04:39.660 · Speaker 2
Um, well, besides a family, which is. It takes it takes takes all of my time outside of ramp. Yeah. Um, I'm building my side projects within ramp. So things that help me expand my thinking and my understanding of what AI is, where it's going. Uh, latest project is like a full AI assistant on slack. Um, and it's teaching me a lot about how
00:04:41.180 — 00:04:59.110 · Speaker 2
AI learns, how it remembers, how it builds knowledge, context, all that stuff. I'm very, very much not technical, so it's really difficult for me to, like, think about code and like session handling and terminal stuff. So it's this is my way of actually getting into this stuff.
00:04:59.110 — 00:05:17.390 · Speaker 1
I have slack setup with Open Claw. I have been using it this way since the end of January, and I've seen Karpathy drop the knowledge LLM. I've seen a lot of people talking about the ways to set this up. Um, was it Gary ten, who talked about big skills, little harness.
00:05:17.430 — 00:05:18.790 · Speaker 2
That skills in harness.
00:05:19.190 — 00:05:55.080 · Speaker 1
Which I thought was really smart because I've been using it. There's a it's a big black box. It's a project car and I don't know how to articulate how magical it is for me or why it's able to do that. Until I started thinking about it, like air traffic control, it really feels like the thing that some people will look at it and they're like, this is just a chat bot in Slack or Discord or whatever they're using.
And I've come to realize it kind of is, except the difference is it's ability query, um, to maintain context and to have knowledge to pull from. It's just traffic controlling. And because of that, I get these really rich outputs. How are you thinking about harnesses?
00:05:55.760 — 00:06:33.160 · Speaker 2
Well, I mean, over the last year or so, we've become a harness company. Like, we're, we've we're building our own harnesses on top of these models for, specifically for finance. So to be able to do that, we as a design team have had to understand what that means because like the traditional design of screens and flows and all that stuff doesn't really help you when you're thinking about designing the experience of a, of an lm, of interacting with a, with a model.
You know, for me, building this assistant was, uh, it's based off of, um, Tiny Claw, which was the
00:06:34.320 — 00:08:08.109 · Speaker 2
less dangerous version of Open Claw, because I definitely don't want to run this on a on a work computer, but building that out has been. I mean, it's a very naive approach. I would basically pick, uh, grab the link to the repo and say, hey, Claude, I want to build an assistant. Here's a repo that I think I want to base it off of.
And that was just a starting point. The sort of the source for that was one of our, um, in our engineering team. I mean, we have fantastic engineers and we've built multiple AI engineers at the company. So our slack, I mean, I was interacting with an AI engineer for a couple of weeks until I realized where I learned that it wasn't actually a human, that it was an LLM.
Um, so the thing that was interesting there was the it wasn't it didn't feel like a tool. It felt like a person that I was talking to, and it was little nuanced details of the way it would react with emojis and some like slack, like emoji reactions, or let me think about it and then come back and like, make even make edits to its first message when it's response to stuff.
So these little moments that just felt Very human. And my goal was to try and replicate that and and see what are the actual gestures that you can give these llms whatever media, whatever medium or channel that you're interacting with them to make them feel
00:08:09.150 — 00:08:44.349 · Speaker 2
just human. Because we don't we don't like talking to bots. We like talking to other people. Right. And the thing that makes it, whatever you can do to make it feel like you're talking to someone else, like a real person is, is like almost intangible. It's like they made a little mistake. They added an emoji reaction and then took it off when they answered.
And they're tracking things and they communicate. There was this engineer, I said, oh, I want this code. Like, just give me the source code for this engineer. I want to run my own version and build my chief of staff and
00:08:46.110 — 00:09:10.290 · Speaker 2
the engineer that built the engineers. Cody. The engineer that built Cody said, no. He's like, no, I'm not giving you Cody's code because I was kind of upset. I was like, wait a minute. Like, you built this thing. Like, I want to use it. Why won't you give it to me? He said, look. Start from scratch. Tell Cody to help you.
00:09:10.330 — 00:09:10.810 · Speaker 1
Yeah.
00:09:11.050 — 00:09:44.250 · Speaker 2
And you'll thank me later. Three days in, I was completely hooked. I had, like, I had to build knowledge, learning rituals. So a heartbeat, a purpose. And like, all of these softer things that make a person just having to define that and think about those things as a, as something to build into design. That's what he was trying to get me to do without just giving me Cody.
00:09:44.290 — 00:10:16.820 · Speaker 1
The there's these abstraction layers you think about in the late 90s. Those of us like we built our computers and we kind of learned how they work. And you have kids? I have kids, and I see them be able to take that for granted. And the thing just works now. And it sounds like that's the same kind of comparable here.
Understand how this agent underneath is working, just this harness. We hear about a lot of companies who it's like, well, we all use cloud code or cursor. I haven't had an opportunity to talk to a lot of companies that have this like harness wide adoption. And so did that experience influence how you are onboarding the rest of the team?
00:10:16.820 — 00:11:55.610 · Speaker 2
It's definitely changed how I think what the role of design is, what we work on, and just fundamentally what we do. There's multiple levels of AI fluency from like, I use ChatGPT to answer questions and like chicken recipes, and instead of Google, I use Llms to write some code and prototype some stuff. I use cursor to like prototype my ideas.
Then there's I build my own tools to serve the like to do the work that I need to do. And then this one where I want to get everyone at the company at is I build agents that build and use tools to do a lot of my job. And I think there's there's a definite before and after for when you realize what the actual like, medium that you're working in is.
I think real from cursor he he was he wrote about how it's like like sculpting and clay and the medium is actually code and the medium is intelligence and knowledge. And like to me that that really stuck. And I made it a point to make sure that I experienced that. And by experiencing it, it completely changed my approach to design, my approach to building the team, my approach to leveling up the team, to what performance looks like an A and a designer, and how as we all converge into builders, what actually matters and what does a designer bring that an engineer doesn't bring, that a product person doesn't bring?
if we're all builders. Like where's the overlap and where's the. The things that the uniqueness that each function brings to the table.
00:11:55.810 — 00:12:26.370 · Speaker 1
This idea of AI fluency has, has gotten a lot of attention. Um, and some people say, you know, it's it's kind of a strange concept to try and outline in competency, because at the end of the day, it'll eventually just become very ambient and it'll be in everything. But it does seem like right now it is worth evaluating because we just ran a study and it seemed like there's three groups of designers.
There's those who are full on board. They're using all these tools, those that are dabbling and those who haven't even touched it yet. It's almost split into thirds.
00:12:26.370 — 00:13:08.140 · Speaker 2
I find myself going back to like, pen and paper, uh, for things, for my what ifs. Back to that fourth level. Once you like, get past the I use AI tools to do my job. There's like, I don't need AI tools to do my job. I need AI to do the work. So I need to be very clear about what it is that I want this entity to produce for me.
And the clearest way to do thinking is writing is drawing is directly like brain to hand. It used to start in Figma, which was my easiest way to mock stuff up. Now I literally like make little wires and drawings of what I want to make. And then I go to Claude.
00:13:08.180 — 00:13:27.940 · Speaker 1
There was a job posting for, I think, a design role here, and it talked about essentially, you know, it went viral for for being sort of a reverse approach, build the thing and then kind of back into Figma. How much weight on that kind of approach are you guys putting on the process? How are you thinking about the word process right now?
00:13:27.980 — 00:13:35.699 · Speaker 2
The way I'm thinking about the role of design at Ramp is you have two tracks of things. You have
00:13:36.820 — 00:15:51.860 · Speaker 2
the design infrastructure track, which means our goal is to make everyone at ramp every single employee A designer. Someone who can build and they can build something that would, if brought to crit, have very, very little feedback. That's the goal of design systems specifically. We've gone from like components and react library and like engineers can copy paste the code and whatever.
And like it's in Figma like that. We've sort of already turned into anyone that ramp can come up with an idea, build the thing that good UX, good copy uses the components correctly, has like fundamentals of design covered that allows the design team to not be the designer on everything. They can focus on the things that the team believes.
Actually, the the team believes that good design, good experience, how something feels is truly critical to the success of that product in that feature, and then designers will have the time to actually do a ten out of ten experience here because they've unblocked everything else. And once we once we're building everything, once everything is being built by agents, then whatever's working, you can pull into the attention of a designer and say, this idea works.
Let's turn it into a ten out of ten and 11 out of ten. This idea is not working. Great designer didn't spend a month on it. Just kill it. Get rid of it. Right? Everyone can build. That's that's the process. Really. That's the part of the process that matters. I don't care if you're drawing or if you're using Figma or paper or whatever.
Are you building as a designer? Ten out of ten experiences and everyone else like, are you able to build your ideas in ways that are not brand Destructive or like customer delight. Destructive, right? I think once you get to that, it's a fundamentally different design team and design philosophy.
00:15:52.260 — 00:16:04.820 · Speaker 1
I want to ask about the design system in a second, but there is a question in there that that says you're leaning into these design adjacent roles starting to come into a sort of design function of their own.
00:16:06.460 — 00:16:12.220 · Speaker 1
Do you? Is there a world where Ramp no longer is hiring a brand new designer out of school?
00:16:12.220 — 00:16:29.820 · Speaker 2
I think those designers are probably the ones that are least indoctrinated into the old way of doing things and are the most curious, and they don't even think about this being new. She's like, yeah, this is just the way it is. They're just playing with AI is like AI on everything. So kind of the opposite.
00:16:30.060 — 00:16:37.940 · Speaker 1
Interesting. Yeah. You anticipate then a increase in sort of newer career designers here?
00:16:38.580 — 00:16:42.759 · Speaker 2
I think they they do different things. There's
00:16:43.880 — 00:18:02.609 · Speaker 2
the more experience of a designer you are. It means just you've done more reps. You have better intuition as to what is going to be a good experience. So it's a delightful experience. You're able to project yourself into your customer very easily. You're able to switch between the full like 100,000 foot view of the whole system.
How does this decision that I'm making kind of fit into that? And you can switch between those altitudes really quickly. So you're working a year ahead and shipping next day. Those are the altitudes you can move through. A junior designer can't really do that. You know they're there. I don't expect them to be able to do that without any help.
So but I do expect them to bring a level of energy, intensity, curiosity that is different from a more senior designer. So we're actually hiring almost like my own design process. Like both ends of the spectrum is a brand new and super experienced. And then you put those two two together and they're learning from each other.
The the new designer is seeing how precise someone is with their decisions. And I'll speak for myself. I'm seeing all these new designers that are
00:18:03.650 — 00:18:13.530 · Speaker 2
kind of putting me to shame in some areas. And like, I got a fire under me, like I need to. I need to show them that they like that I deserve to be a leader for them.
00:18:13.530 — 00:18:36.210 · Speaker 1
The research process here at ramp. One thing I'm hearing from some teams is that build speed is is so fast now. And some of the research practices, the quarterly working groups, things like that are becoming, they're trying to figure out how to keep up with it when you're shipping every week or so. Right.
Um, what does the research process look like? Is it having any of those challenges?
00:18:36.250 — 00:18:48.349 · Speaker 2
I'll describe the research team a little bit. We have a very small research team. It is, uh, three people we have focused. We've tried to build a
00:18:49.990 — 00:19:21.470 · Speaker 2
research function similar to what we're doing with design systems. That makes everyone a researcher. As a designer, I have a question I will go ask ramp inspect, which is our coding agent that has access to snowflake and all our data and can actually find any customer that fits my profiles that I'm trying to talk to that uses features in the way that I want them, that I have questions about.
It'll help me think through. We have an AI research, um, assistant that will help you extract your questions and structure a research plan.
00:19:21.510 — 00:19:24.350 · Speaker 1
Is that an individual interface or is that in slack?
00:19:24.390 — 00:20:22.480 · Speaker 2
It's in slack. Slash in notion. Like it's been everywhere. Like every time. Every time the tool changes, we are like following it. So it's actually fairly chaotic. But the research process in a nutshell is I have a question. I have agents that help me find the right people to talk to. I have agents that help me structure the questions and the research session.
I then work with the research team to schedule everything. All of those sessions are recorded, all that context. It fed back to the system. All the insights are shared by the agents so that next time anybody asks a question, that research is accessible to them and we can just answer the question. Um, and now we follow up with customers and give them thank yous and tell them how their input actually helped inform what we're building.
This is a research team of of three.
00:20:22.480 — 00:20:23.000 · Speaker 1
Yeah.
00:20:23.480 — 00:20:29.280 · Speaker 2
That has enabled a thousand people to run research on a daily basis like we have.
00:20:30.680 — 00:20:40.650 · Speaker 2
Um, I think the design and product team did like 150 to 200 research sessions this month. Wow. Yeah.
00:20:40.650 — 00:21:02.730 · Speaker 1
So something like ramp Inspec gets built and you've got, you know, 1500 other tools that people are building. And some might just be for, you know, my individual workflow. Others might have, uh, use cases that should be shared. How do you approach adoption? Hey, researchers built this great way for everybody to do this.
How does that get introduced and onboard it?
00:21:02.770 — 00:21:05.530 · Speaker 2
I mean, it's a total meritocracy.
00:21:05.570 — 00:21:06.010 · Speaker 1
Yeah.
00:21:06.370 — 00:22:01.750 · Speaker 2
The right skills, the right tools survive. The ones that are not that good. Don't inspect. It was it didn't work for a while. And then when it started working, couple engineers picked up on it and realized, oh, this is a lot of value here. I'm going to start using it and got better. It got better, it got better.
And then like and then it caught. I think our research process similarly, you know, it was I get a lot of value from talking to customers. It's very painful to talk to customers. And then some designers started using the system product. Operations and product managers started getting on a lot of calls, and designers started to see like, wait a minute, how is my PM?
Getting all these calls like, this seems awesome. And they realized, oh, wait a minute, there's a system like, I can use this thing, and then I'm going to stop doing my thing. I'm going to do this. It's like you're just the same way as like how a product gets the product market fit. You iterate on it and then it just goes wild.
00:22:01.790 — 00:22:07.550 · Speaker 1
And just word of mouth really like, oh, so-and-so built this thing. Go check out how to use this tool. Yeah. Interesting.
00:22:07.590 — 00:22:16.630 · Speaker 2
I mean, you can you can tell people to like, use AI as much as you want. They're not going to do it until they have an aha moment of it changes your life.
00:22:17.430 — 00:22:44.120 · Speaker 1
Well, I think that's an interesting point because, you know, there's a lot of companies that are mandating, okay, we are now using cloud code. Some companies are mandating how to use it, when to use it, others are a little bit more loose. They say, hey, you have no token constraints, just go wild. Um, you know, and then some companies are like, we have a luncheon learn.
And that's about as far as we go. How did ramp get to 99.5% adoption of AI?
00:22:44.400 — 00:23:10.199 · Speaker 2
The first thing we do is make sure everyone understands that this is important. So we relentlessly communicate. This is important. This is important over and over. Then we tie this to actual performance. So there's a rubric in performance reviews which is AI fluency. It's a real thing that we talk about and how you're trending and how you're tracking.
And then the third is
00:23:11.480 — 00:24:36.950 · Speaker 2
we just give people real access to it. And access is not enough. Like you can you can tell people, hey, I'm evaluating you. This is important. Here's the tool. I think the biggest effect, like the biggest change we saw, was we started highlighting the things that work, and we started having some people who were a little bit ahead of the curve start to sit with other people and say like, okay, show me, like, tell me what's painful for you.
Okay, great. Let's just open up, Claude, both of us, and let's just do this together. There's, um, there's a good friend. Uh, he's is an Am here at ramp. Um, he's a name manager at ramp. And I sat with him and said, great. Tell me what is like what do you do throughout the day? Like what takes time? He says, oh, well, like before, uh, after every meeting, I have a I, I read an email wrap up.
I spent like 30, 40 minutes on this email of everything we talked about, everything. We sat with them and said, okay, great, we have an hour. I put an hour on both our calendars. We're going to talk about this thing. We're going to build this thing. He kind of expected me to come in and like, tell him about Claude and tell him about what to do.
And then he would go do it. He's like, no, no. Share your screen. Open terminal type in Claude Dangerously. Skip permissions. Cool. All right. Um, and then he would ask me, okay, but how do I do x? Great. Ask Claude.
00:24:36.990 — 00:24:37.630 · Speaker 1
Yeah.
00:24:37.910 — 00:25:21.749 · Speaker 2
That was literally my job was like, okay, ask Claude. Okay. Ask Claude over and over until like 3 or 4 questions. Then he stop asking me and started asking Claude, okay, how do I do x? And within 40 minutes we had automated the email, like the summary email that he sends. So he has to do is like go on, gone. Take the transcript, paste that in, and then he gets the summary that he that he can send out.
And it matches all the summaries that he's written before. Because we just like that moment for him was the whoa, this is different. This is actually worth my time and my investment. And we just did that at scale. So now he
00:25:22.910 — 00:25:26.230 · Speaker 2
other people and I can say it's like the pay it forward method.
00:25:26.310 — 00:25:42.320 · Speaker 1
You've been with ramped for a while now since the early days, right before the $32 billion valuation before this AI boom. Walk me backwards to. Well, actually, the first question I have is how would you define like the seasons of Ramp and what season are you in right now?
00:25:42.560 — 00:25:50.120 · Speaker 2
Uh, we are in a true renaissance season, right? Like, it's it's it's a really interesting period. I think the,
00:25:51.480 — 00:27:39.550 · Speaker 2
the two most interesting times in ramp history were January 2020 launching the product. And now when we are completely rebuilding company and product and in this new world, ramp has not been foreign to AI ever. Like it was always a thing we always had. We've been using GPT in the app forever. Uh, since, I don't know, I think it was like probably like 22, 23 was like started making its way into the product that I'm aware of, but we were using the old version of machine learning and AI to match receipts to save time for people.
And I remember when ChatGPT when there was the ChatGPT moment, um, the pressure was out of chatbot to everything, right? Um, we very, very intentionally didn't do that because we didn't. We were already like in the, in the wagon. We kind of knew what AI was about. And we nobody wants to talk to their bank account.
Like, that wasn't the thing that we were building. Um, so I remember writing like AI principles when the ChatGPT moment came out, because you kind of had to, like, distill what you believe. Otherwise you get swept up. Um, so at the time, I remember it's it was we are building magic. Like, if you can build something that feels magical, it doesn't matter how you do it.
It could be an LLM. It could be a very well timed email. It could be a very good line of copy. As long as it feels like you've taken away work from someone in a delightful way,
00:27:40.590 — 00:28:26.430 · Speaker 2
then your building ramp. The beauty of it is that you can, for the first time in history, like throw money at a problem and get time out of it. It's the one resource that no matter you either scale headcount or you work longer hours, but now you can actually pay to have more work done with the same amount of people with the same amount of human time.
And for a company that's all about saving people time and money, this feels like crazy cheat code for us. I can just take all this work that you currently do on the app and just make it go away completely. Just tokens. My my my vision of
00:28:27.530 — 00:28:31.090 · Speaker 2
Nobody uses ramp is very possible.
00:28:32.450 — 00:28:53.930 · Speaker 1
The the 99% of the company using AI will walk me backwards in this season. This renaissance that you described of how you got there, what was was there a pivotal meeting that happened and then decisions that followed that led to that? What did that look like and how quickly did that occur?
00:28:54.410 — 00:29:34.500 · Speaker 2
I mean, it all started with our engineering team. Like AI got really good at writing code. Um, we've been using cloud and GPT and Gemini and all these things to write code for like multiple years now. Um, the code wasn't good enough for a long time. And then it started to get okay, and I started to get pretty good.
Um, at one point, anthropic came to visit ramp to look at our engineering team because we were using cloud code more than they were, like they need to. They had to. Okay. What's going on at ramp? Like what's happening? This is this is crazy. Like 80% of your engineers are using this product.
00:29:34.660 — 00:29:38.140 · Speaker 1
Really the top user of the product at one point, the top organization.
00:29:38.180 — 00:30:05.020 · Speaker 2
I don't I don't know that, um, I do know that by, like, density, it was probably up there. Yeah. Um, they were they were kind of shocked. They were like, okay, this we have something here. We should we should go talk to this. We should go talk to ramp. Um, so it started with an engineering organization. And from there we started realizing that, okay, code is getting cheaper.
Where is this going?
00:30:06.180 — 00:30:19.380 · Speaker 2
Where does this go? We started thinking about our internal operations. So two years ago, we, um, at a company event, we set one of our goals was to become the most efficient company in the world.
00:30:20.900 — 00:30:59.310 · Speaker 2
There's a company goal. So we had product goals and customer goals. And then the company one was become the most operationally, operationally efficient company in the world. And that meant every team, every person looking at what they do every day and thinking, do I need to do this? Is this valuable? Is this process in my way or not?
Is there a better way to to approach this problem? And that's when we rolled out GPT for everyone. We were all allowed access to cloud code for everyone. There was a discussion originally of, hey, like this is expensive. Should we give the team cloud code?
00:31:00.470 — 00:31:47.370 · Speaker 2
Should we give the sales team cloud code? And the initial reaction was probably not. This is a very expensive, but it really helps when your CTO is like completely AI pilled. Yeah. And just overrode everyone who said no. I said I don't care. Spend it, spend it. It's a it's a big bet. At the time it was an entitlement.
Two years ago, that just meant that we were extremely well positioned for this December when ages actually started to work. So they were better code. They got good at code. They started getting kind of bad at planning and logic, and then they got good at logic and like figuring stuff out. And then by then the entire company had already been experimenting with this.
00:31:47.370 — 00:31:58.690 · Speaker 1
So you get something like glass that comes out and you're like, see, that's exactly the kind of thing we're looking at. What's the story behind glass? How did that come about? Maybe for people who don't know what glass is. How about you describe it.
00:31:58.730 — 00:32:02.810 · Speaker 2
So so glass is our internal
00:32:04.170 — 00:33:24.310 · Speaker 2
equivalent to cloud coachwork. If people are familiar with that it's essentially a desktop app that uses a file system. It's just a harness for cloud or any of them to be able to modify files and do things on your computer like you would. So we realized there was a real gap in what we wanted people to do and how capable they needed to be in order to do it.
We didn't want people spending time learning terminal learning how to set up mic connect. Everyone had to connect notion. Everyone had to connect slack. Everyone had to connect to to Datadog, snowflake. All these things are great. We should just package everything in one. Make it really easy to use.
Just download glass, open it up, sign in via Okta. Everything is connected. You can use it on slack, you can use like you can connect all of the Google Workspace into this. You can just talk to an AI and have it draft documents based on like, everything that you know about the company without having done any of the setup.
So really it was how easy the setup is. And glass just collapsed all of that into your download an app. Well, actually, I think we pushed it to everyone's computer so they didn't even have to download it. They just had it. Yeah, and it just worked.
00:33:24.350 — 00:33:52.350 · Speaker 1
And I noticed with glass got some beautiful branding on it. This is an internal tool. Now, in my past, internal tools were kind of second class citizens. It's like it needs to work. It needs to serve, you know, account management or somebody. Right. And I saw Simon Corey shared with me even a tool he's working on.
And you came back and he liked drop caps and you added this beautiful tapestry effect to it. And I find that really fascinating that the internal tools are getting so much love like this. Is that across the board or is that like the winners start to get more attention?
00:33:52.350 — 00:33:55.110 · Speaker 2
There's no criteria for it. The
00:33:56.230 — 00:34:01.710 · Speaker 2
it all goes back to is the person working on this thing passionate about it and having fun with it.
00:34:02.830 — 00:34:14.590 · Speaker 2
So there's a lot of craft. Like people will bring their best craft to the things they care about. And if you know, you, you know, like we we built a ramp CLI,
00:34:16.190 — 00:34:27.129 · Speaker 2
it has, I think three games in it that you can play in the terminal. There's no reason to do that other than it's fun and like,
00:34:28.210 — 00:35:01.370 · Speaker 2
I wanted it like, so I designed, like, we just designed, like, ski free in the terminal, you can play a little ramp that's kind of like skis around. And if you remember this game or like, a Yeti gets you. Yeah, it's just in there for no reason other than we like it. Somebody thought it would be fun and we just made it.
And then when you let people when you let fun and curiosity and like interest drive, you get a lot better outputs than if you just focus purely on functional things.
00:35:02.170 — 00:35:12.970 · Speaker 1
So glass has created some of these tools get created. People want to just continue and have fun with them. Who's governing glass, for example? Where does governance fall into this equation?
00:35:13.330 — 00:35:41.900 · Speaker 2
Um, there's I mean, ramp is a very high trust place. Um, we we very high agency, very high trust. We let everyone make those decisions. Um, when things work, they request they ask for. Hey, can you. Can I get some feedback on this? What do you think? And then they nerd snipe one of us, and it's like, oh, this seems fun.
I want to have fun with this because I'm going to go, like, work on it. Um, so, I mean, for internal tools, it's,
00:35:43.020 — 00:36:27.280 · Speaker 2
uh. Yeah, chaos. There's like, all the brands are all built into our internal tools. It's not. It's not just the ramp one. Uh, when something works, then we start to put more investment into it and chase momentum. Once the things start gathering, like like a snowball, they start like gathering people, and everyone brings their skills.
And then things start to look and feel more like the ramp brand as you as you progress. Um, for customer facing things, it's very different. We we definitely have more of a process for building, like getting to quality. We have more of a bar that you need to meet before you get to go to GA, but the early alpha is like.
00:36:27.360 — 00:36:27.640 · Speaker 1
Wide.
00:36:27.680 — 00:37:44.090 · Speaker 2
Open. Wide open. Yeah. You know, you find five customers that you that are friendly and you embarrass yourself in front of them as quickly as possible, and that's fine. Um, that took me a while to get used to, by the way, because as a designer, I'm like, oh, I want everything to be perfect. Uh, and I would polish some fundamentally misguided idea to show it to a customer, and then they'd just look at me and say, like, why would you build this?
This is you completely missed the point. It's pretty, but like, you've completely missed the point. So, like after doing that multiple times, I've learned, oh, it's I need to be vulnerable and uncomfortable here. And I'm really polishing this for me, not for them, because I don't want to look bad. So if I completely lose fear of looking bad, Then I get to learn a lot faster.
And it's the same thing with all this AI adoption. If I. I'm not afraid to look like I don't know, then I can admit that I don't know and just get the help that I need right away. It's it's really easy. Also, when you can just ask Claude, it doesn't judge you. It's like, I don't know how to use this thing. Can you tell me?
There's no judgment. It's just you get an answer and you start learning.
00:37:44.090 — 00:37:58.090 · Speaker 1
Now with glass. An interesting insight here is that a cloud coworker does a lot of this. Why? Why not use a vendor product that packages it all and allows and has these built in? Why build your own internal tool for it?
00:37:58.130 — 00:38:00.570 · Speaker 2
It needs to be really good at ramp.
00:38:00.890 — 00:38:01.250 · Speaker 1
Yeah.
00:38:01.290 — 00:38:08.770 · Speaker 2
And you know same thing with our like design infrastructure that we're building like an AI product designer for everyone.
00:38:09.970 — 00:38:14.410 · Speaker 2
It doesn't need to be good at it doesn't need to be a good global designer.
00:38:15.650 — 00:39:02.980 · Speaker 2
It just needs to be really good at designing ramp. And that is a scoped problem that is much easier to nail than a general purpose designer. That should work like needs to go work at Airbnb or Uber or some startup or whatever. It's the same thing. Like coworking is very good at like just like getting you like getting some work done.
But if you want to be good at ramp, you have to be able to access our data securely. You have to be able to access all of our knowledge base. You have to be able to access, like all the context of the company, all of the skills that we create and share and like the way that the way that I write our marketing, for example, if you're wanted, if you need to launch a product,
00:39:04.220 — 00:39:07.499 · Speaker 2
you need to write marketing, you need to write copy. You're going to come to me
00:39:08.580 — 00:39:18.480 · Speaker 2
in glass. You have a version of me that writes copy for you, and we can continue to invest in those things. So it's, um,
00:39:19.560 — 00:39:29.720 · Speaker 2
it's kind of a no brainer once you realize that it's just for you. And it just has to be very good at what you do, not at everything.
00:39:29.760 — 00:40:04.560 · Speaker 1
In our in our survey, what showed up was an interesting insight of the respondents. Uh, lead designers principles were the just underneath design engineers in the amount of AI coding and things they were doing there. Right underneath them were managers, surprisingly, if I recall correctly. But I know that underneath the managers were your general ICS, and I thought that that was fascinating that managers were doing more, uh, AI coding and tooling than general ICS.
As far as our respondents were concerned. How how does the manager role shaping up here?
00:40:04.680 — 00:40:35.449 · Speaker 2
We don't have a lot of those. Um, I don't think we have a pure manager role at Ramp. It is still a like hybrid player coach where if you're in the director track, you're probably spending like 40% of your time icing 60% of your time helping people understand, like how to grow and like what their goals are and like what products they're working on and all that.
But that 40% of the IC time is spent
00:40:36.690 — 00:41:49.910 · Speaker 2
nudging through the craft, like igniting new work, new ideas, making a sketch, giving it to someone, taking a junior designer who's struggling with something and just taking over the project and saying, okay, let me show you how I would solve this design over, like look over my shoulder. Will I do this thing at ramp?
In order to lead, you have to be able to make and we have very, um, reluctant managers is the kind of manager that we hire. I need to take on people, and I need to teach them and help them grow, because I want to make more things, and I see that being like the only way forward for me to actually build more stuff so that I can't do it all by myself.
The manager needs to be ahead in AI, in craft, in everything, which is kind of back to what I was saying before of like the more senior people are ahead in like product thinking and insight and ability to be empathetic to customers and just to represent that. Um, but the fire that is being lit under them by the juniors is a real thing.
There's a real pressure for me to deserve to lead these people.
00:41:49.950 — 00:42:17.840 · Speaker 1
Designers and different functions are creating agents to automate parts of their work. And there's a lot of questions right now about are we automating ourselves out of work, whereas other studies and anecdotes are saying, actually, it's making people work even harder. They have compressed this function and now they can go and work on this function.
What are you starting to see on the team, and what do you think about some of those concerns that ICS might have?
00:42:18.120 — 00:43:34.610 · Speaker 2
I don't think I don't think if a company is working out, they're going to run out of things to to go after. That's just if people are 30% more effective, you don't like go 30% of your people. It doesn't make any sense. You want more people because they're more effective. So it's actually the opposite. So I have the I guess, the view that the more optimistic view of of AI transforming business and helping companies grow, like we were seeing that in our data companies that are using AI are growing faster than companies that are not.
And this isn't just tech. I think we're I can't remember the industry, but it was like a like a manufacturing. And I think it was like an HVAC installer or like a plumbing supply that was like leaning into a I was growing 65% year over year versus the competitors that are growing like 4 to 8. So company grows.
You have job. More AI. More growth. More job. That's just what I'm seeing in the data. What I think is more interesting is once you get past the frantic part of AI, which I think we are going to start, we are getting there. You actually get to work on the things that are interesting for you, and
00:43:36.130 — 00:43:56.530 · Speaker 2
you just work on the stuff that's actually really important. So what I was saying before on making everyone a designer, I don't have to design some like toggle three clicks into a settings panel. Someone can just build it. Ship it. I don't even have to see it. I get to work on some like, really ridiculous
00:43:57.610 — 00:45:15.920 · Speaker 2
video game built into our CLI because it's a delight thing for the developer that's using it, and that they're going to that developer is going to remember that. So for me what's interesting is I get to focus on the emotional side of of design. How do I want to make someone feel? And I can actually take the time.
And because now code is basically free. I get to build whatever I want. So I think you reference this. You talked with the thing that Simon was sharing. Um, it's not out yet, but the like, if we write you a briefing on your AI token spend, which we have, you know, we just launched, when you log into this thing, we tell you, hey, your token spend is up or down based on these things here, the PR that drove this up, here's the engineer that's doing X, Y and Z.
So yeah, sure, you can just write it. I'm right now I'm just in a, in a mode of oh, I kind of like the I'm looking at like the little icons and um, I mean like religious icons and like Renaissance type things where you said, okay, why don't we have, like, this, like highly ultra complex, like drop cap with gold leaf and like little illustrations?
And then what if we animated it? And two hours later, I had everything. Like I have every letter, I have every single thing animated. And I can put it into this briefing,
00:45:17.000 — 00:45:46.880 · Speaker 2
and it's just a thing that I really wanted to do that I would have. It would have been impossible for me to do before that. It's going to be memorable. And when someone sees this thing, hopefully they'll just get a kick out of it and they'll get the information. They'll also just have a slightly better day because it was like a nice moment, completely unexpected and totally unnecessary.
Yeah, it's like it just allows you to care and put that into actual like product.
00:45:47.320 — 00:45:49.040 · Speaker 1
Now, I have heard
00:45:50.080 — 00:46:00.280 · Speaker 1
off the record from a lot of design leaders that I've talked to, that they think there is a potential future where companies need fewer designers.
00:46:01.400 — 00:46:06.560 · Speaker 1
Would that do you? It sounds like you probably would disagree with that based on your previous answer.
00:46:06.600 — 00:46:09.180 · Speaker 2
I would disagree with that, I think.
00:46:11.500 — 00:46:20.900 · Speaker 2
I think you can do a lot more with a lot less people. So if you have a team of 300 people that are currently doing 50 things.
00:46:22.900 — 00:46:36.300 · Speaker 2
If you lean into this thing, you can have a team of 300 people doing 200 things. So what does that unlock for your business? What new revenue can you go after? What new areas can you enter? New ideas? Can you explore? So,
00:46:37.340 — 00:47:10.390 · Speaker 2
um, I'm very optimistic about the the future of design, because I love your thought on this, because there's there's this idea that everyone becomes a builder. Um, and the approach to that, I think, is we're not all the same. There's a lot of there's a lot of diversity in our thinking, specifically between the functions that we're bringing, things that are very unique to what we do.
Designers are very visual, very emotional, very empathetic.
00:47:11.990 — 00:47:27.510 · Speaker 2
We are able to put ourselves in someone's shoes. We actually feel things very, very deeply. Um, I would speak for myself, but like, I like I feel pain at using bad products.
00:47:28.710 — 00:48:44.440 · Speaker 2
I literally won't use them. Like, one of the reasons I'm at Ramp is because submitting expenses was so horrible that I wouldn't do it. I'm about thousands of dollars from previous jobs. And when Karim and I started, when he started telling me about this idea, I thought, amazing. Like, I can fix this problem.
I can never do this again. It's amazing. Um, so I think as a designer, you're bringing a sensitivity that product or engineering doesn't have, or an engineering builder will bring a sensitivity to scale how something is architected to precision, to like a system that doesn't fail. How a product person will bring a product builder will bring sensitivities of I know how to sell the crap out of this thing.
Like I will build the thing that will sell. And a designer comes in with like how do I want people to feel? So I think you have a lot more people problem matching than you would otherwise. And I think a problem can be handled by one person, and you can give it to someone else and give it the touch of the engineer, the touch of the designer, the touch of product, the touch, the thing.
I think it's a much more fluid and collaborative future than it is now.
00:48:44.520 — 00:49:30.660 · Speaker 1
Yeah, I think, I think I'm seeing a lot of that from from accelerating companies who are very optimistic about this. Um, there's I've heard people talk about this in a way that I that resonated with me, these special ops teams, where each initiative kind of has a leader and you have maybe 3 to 5 people who cover wide surface areas, somebody who goes out there and they're just.
They're a scouting function, and they're just kind of making a mess of things to try and figure out what's worth investing into, right? Once they kind of find those ideas, it passes the baton back to, okay, we're going to go ahead and establish that area. And what does that look like? Um, and I think that that makes a lot more sense.
And, you know, I have ideas on what I think this function turns into in five years. What do you what do you think it turns into in five years?
00:49:30.700 — 00:49:42.940 · Speaker 2
Five years is a long time. It is a long time. Like if you asked me four months ago, I think it would have been a very different. So I'll caveat that it's probably not. Not right. I think you become a builder that
00:49:44.300 — 00:50:28.670 · Speaker 2
just works directly in production is shaping intelligence, shaping knowledge and shaping work. And you're, I think, mostly focused on how you want something to feel. I don't think you have to worry about will it work? Will it not? Will it like, do the job. I don't think we're going to have to worry about that very soon.
You can kind of go back to like the fuzzier part of design, which is really why I got into it. Um, which is what do you want someone to think? How can you plant an idea in someone's head and, like, have them feel a certain way for the day or for the year and, like, change someone's life? Like, that's.
00:50:30.550 — 00:50:34.949 · Speaker 2
I think that's where it all goes. Like, you have an idea how to make something better
00:50:36.110 — 00:50:37.790 · Speaker 2
and then it just exists.
00:50:38.830 — 00:50:59.030 · Speaker 1
So parents, a parent, a lot of a lot of comparisons to how do you teach even late career folks how to adopt an entirely new way of thinking about this, too? How do you teach a child how to think? What are you preparing to teach your children about this new world that we're evolving into?
00:50:59.070 — 00:51:00.910 · Speaker 2
Fundamentals of what?
00:51:02.330 — 00:51:08.169 · Speaker 2
People need to know in order to be good people. I don't think change
00:51:09.770 — 00:51:25.009 · Speaker 2
being kind. Being honest. Being open. Taking risks. All those things are the same. Um, I think what changes is what access do they have? Do things. How possible. Things are
00:51:26.530 — 00:51:51.530 · Speaker 2
like they. You know. My kids, my oldest is four. So we're not really talking about computers yet, but I don't know what the technology looks like in five years when she's nine and she's like, maybe starting to like, she has a school project, she has to write something. Uh, I my main thing is to teach my children how to think, and
00:51:52.730 — 00:52:24.180 · Speaker 2
I, I don't know that you get a replacement for that. And that's maybe, like the trap of AI is that you replace thinking and you outsource it. That's my fear of of AI and I where I think it goes wrong. So that's probably what I would focus on is just own your thinking. Don't let it be influenced by anything else out there in the world.
I don't know, like I'm terrified of social media.
00:52:24.260 — 00:52:24.860 · Speaker 1
Yeah.
00:52:25.300 — 00:52:39.420 · Speaker 2
I don't know what that looks like. When Sora came out, I thought it was the end of the world. I'm glad it's no longer a thing, to be honest. It's just it's it's really scary. So there's. When there's no limit to what you can do. You just need to.
00:52:41.460 — 00:52:45.660 · Speaker 2
Learn how to wield that power without destroying yourself.
00:52:47.580 — 00:52:53.940 · Speaker 2
I got kind of intense, but that's probably where. Where I want to get my, like, what I would teach my kids.
00:52:53.980 — 00:53:42.350 · Speaker 1
Well, Sam Altman says these algorithms are one of the first examples of kind of misaligned eyes. I thought that was true. I agree with that. And I think we're seeing because we've seen a lot of the cognitive destruction that happens from this, and we're seeing a lot of cognitive impacts of, of AI from a number of places.
Right. Never mind people's fears in the economy. But even people like us very AI build using this stuff all day. I feel like I get in an F1 race car and I move so quickly that now the way I have to think about what's around the next corner is like what's around the next three corners, because I'm about to be around that corner by lunch.
Yeah, and and that is exhausting. I'm not used to thinking like that. Are you feeling any effects of things like that and how are you dealing with it?
00:53:42.870 — 00:54:05.810 · Speaker 2
So let's break it down into personal and then product. Yeah. Personally I've gone through multiple rounds of. This is exciting. This is noise. This is exciting. This is noise. So I'm like very energized, completely exhausted, very energized, completely exhausted every time I refine a little bit. Um,
00:54:07.610 — 00:54:14.850 · Speaker 2
so yes, that happens a lot. Um, what I've learned, what I've taken from that is that
00:54:16.490 — 00:54:19.610 · Speaker 2
the actual experience and like how you make people feel
00:54:20.690 — 00:55:07.140 · Speaker 2
by when interacting with eyes is extremely important. And like, everyone's talking about taste. And I think that to me, that's what that what that means. That's my, my, my take on it is in our product. In my product. I don't want people using Ramp every second of their on ramp. Second, they're not doing their jobs.
They're not selling. They're not doing something else. I want you in and out and as fast as possible when you're in there. I want it to be a delightful, calm experience that just makes your day better. Um, That might be through some hyper ornate, like gold leaf. Like little detail in the in the product that nobody asked for.
But it's just delightful or.
00:55:09.220 — 00:55:29.340 · Speaker 2
A very well placed text message that just tells you, hey, everything was done. Don't worry. So I think more and more the product goes toward don't worry, it's done. You can check it if you want versus what a dumb tool is, which is notification, notification, notification, notification. Come back, come back, come back.
00:55:30.420 — 00:55:33.299 · Speaker 2
So it's really if you understand
00:55:34.820 — 00:55:40.419 · Speaker 2
how learning and knowledge and all these things work, you can
00:55:41.620 — 00:56:11.800 · Speaker 2
build a system that matches how humans already work. Instead of people onboarding onto your tools like you're onboarding your tools onto the people. So you kind of need to learn how someone wants to work and match that. And I don't think we're there yet as a is an industry, but we're going there. And I'm like that's what I'm building here.
It's it's the future of that is how do you actually build products that understand who they work for?
00:56:12.000 — 00:56:13.120 · Speaker 1
I think, um,
00:56:14.240 — 00:56:18.920 · Speaker 1
AI can learn discernment and taste at some point.
00:56:18.960 — 00:56:19.360 · Speaker 2
I mean, we.
00:56:19.360 — 00:56:20.040 · Speaker 1
Have.
00:56:20.600 — 00:57:37.210 · Speaker 2
Agents internally that can build things that I think are good. So I've I've gone through the exercise of codifying my taste. Um, recently, you know, we were strapped for web design. So we went into glass, built a skill, built a tool that, like cloud code output. I can tweak it, I can give feedback and then goes back to the skill.
New output. Go back to the skill new output. Go. And then now we have a skill that just builds a website that I don't have any knit with. Like the spacing on everything is right, the eyebrows and the titles they have, like the 16 pixels and like the title is exactly the size and things are grouped visually. So again, like teaching and coding taste you can do and I think it's on us to do that.
But then it's a lot easier again when you're doing it for your company versus you're doing it for everyone, right? Yes. As we'll get good enough to do all this stuff. Um, I don't think they'll replace thinking they won't replace the, the idea and just the passion that you have for something. They just become accelerant for every human.
00:57:37.210 — 00:58:01.260 · Speaker 1
We saw in the in our survey, which was really interesting to me because I hear a lot of the loud minority on mine about AI who kind of push against it. But we saw that of those who were heavily vibe coding in their process, the satisfaction rating of their workflow is about a 7.5 out of ten, and those who were not, it was closer to about five out of ten.
Are you having fun?
00:58:02.420 — 00:58:09.660 · Speaker 2
I mean, I'm building video games in a terminal and show me it's like a million people. So. Yeah. Um,
00:58:10.740 — 00:58:59.200 · Speaker 2
like it? I think it it all. It's very simple. It all comes back to. Do you care about what you're doing? If you care about it, you're going to work hard. It's not going to feel like work. Both my parents are ballet dancers, so that is probably my biggest lesson from them is that if you love what you do, you don't work a day in your life.
I have been at ramp for seven ish years. I haven't worked a day. It's it's amazing. It's so fun. Um. And I want everyone in my team to feel that way, to feel empowered to do that and just to think, like, what makes my day great? Let me go do more of that and less of what doesn't. Let me build the workflows that take everything else away, so that I can focus on the things I actually enjoy.
00:59:00.240 — 00:59:11.840 · Speaker 2
And it just shows in the work. And AI is the biggest cheat code for that. Just give the work away. It's literally what it's meant to do.
00:59:12.120 — 00:59:29.040 · Speaker 1
Of those 37% who responded to our survey saying they weren't touching AI and, you know, for a number of reasons that they outlined. But what do you say to those folks who aren't participating in what's happening right now out of either fear or cynicism?
00:59:29.400 — 00:59:42.080 · Speaker 2
I mean, I have a lot of empathy for that. I was in that. I was in that camp at some point, I'm sure. Um, I've been afraid of it. I've been tired of it. I've been angry at it,
00:59:43.480 — 00:59:47.239 · Speaker 2
and I'm very happy with it. It's like helped me a lot. So
00:59:48.560 — 00:59:54.970 · Speaker 2
I think it's happening whether people like it or not, and that's kind of a tough pill to swallow.
00:59:56.010 — 01:00:07.529 · Speaker 2
Once you come to terms with that, then you can. The only thing you can control is how you react to it, whether you are going to make it fun for yourself or
01:00:08.730 — 01:00:28.050 · Speaker 2
you're just going to be bitter about it. And this is going back to children. There are ways that you can get them to do what you need them to do. You just have to be creative and kind of think about that for yourself. Like, what kind of life do you want? What kind of work do you want to be doing? And if you're seeing patterns,
01:00:29.610 — 01:00:41.770 · Speaker 2
what good is it to bury your head in the sand? Because it's it's happening no matter what. I think my job is going to change, no matter what I feel and what I like or don't like.
01:00:42.970 — 01:00:58.790 · Speaker 2
So I best, for my own sake, find what am I really good at? How does this actually help me? And then lean into those things and automate away the things that I don't want.
01:00:58.830 — 01:01:48.790 · Speaker 1
Most of the A.I. debate is two camps arguing about the tools. One calls it a miracle and the other calls it slop, and both are arguing about the wrong thing, usually from a place of small surface area. But Diego doesn't talk about the tools. He talks about the entity. It's possible that the companies that compound over the next few years will be the ones who realize their job changed from using AI to building something that looks more like a person with tokens, skills, and memory that works inside the company.
If you're actually running complex agent harnesses inside your company. And I'm not talking about pilot programs, find me. I'm mapping this pattern and I would love to hear from you. The Ramp Lab's links are in the description. Please read them for yourself. I'll see you next time.