D2C Revenue Rocket Podcast
EPISODE 13
Guest: Gil Hsu, Senior Product Manager, Klaviyo
Presented by Klaviyo & ECD Digital Strategy

ZAK: [0:04]
But now you're at Klaviyo, building a brand-new type of product that's never been invented before. I'm not a developer, and all of a sudden I built this super-complex dashboard with a couple of tools and a few prompts. Boom — click send, and it sends it out. And historically, as an agency, we're not going to go and do something like that, because it might only make the company…

GIL: [0:24]
We'll be able to take the intelligence of Composer from the AI and distribute it across the app. This is Klaviyo's intelligence highlighting things to you at any time, right when you're thinking about it.

ZAK: [0:39]
Gil, this happens every time that we talk — I feel like the time just flies by and we're just scratching the surface. This was very exciting, sort of what's coming.

ZAK: [0:59]
I am very excited for today's episode. We have Gil Hsu, Senior Product Manager from Klaviyo, joining us today. Gil is the man behind leading product on Klaviyo's biggest new release, Composer. Gil, welcome to the show.

GIL: [1:15]
Hey, Zak, that's quite the intro — thank you so much. Yeah, I'm excited to be here. It's always fun chatting with you, and I'm glad we can do it in podcast format this time.

ZAK: [1:24]
Yeah, I am as well. And I think a lot of the stuff you're able to share is going to be great for everybody listening, both on the brand side and the agency side, to really get a look into your mind in building Composer — what it took to get it to where it is today, and where you see it going in the future.

GIL: [1:43]
Yeah, really excited about this chat. I mean, obviously for all things Composer and Klaviyo, it's the next evolution of how we think about marketing. But generally, excited to have this conversation, do a little peek behind the curtain, and show a bit of what's coming up next. I think it's going to be a good chat.

ZAK: [2:03]
Awesome. Before we dive in too much, could you give us a little background about yourself — where you were before Klaviyo, and how you ended up leading product there?

GIL: [2:14]
Sure. I actually have a pretty winding resume, you could say. Prior to product, I was an engineer for several years, doing anything from identity tech to health tech, all in the Boston area. And even before that, I had a previous career in design, building mostly commercial architecture — so I was an architect.

The funny story I always tell people is that I actually designed the Klaviyo headquarters office in Boston as the architect — but at the time it was for a different company. Since then, Klaviyo has changed all the designs, so now that I work here, the views are the same, but all my designs have since been demoed. It's a bittersweet moment coming back into this office.

But yeah, I've been all over the map in terms of that trinity skill set — design, engineering, and now product. Very happy to be in this product space now. It's definitely the right fit for me.

ZAK: [3:18]
That is very interesting — the building you designed, and now you're at Klaviyo building a brand-new type of product that's never been invented before. It's kind of interesting, right? You designed the building, and a lot of what you designed, they knocked down. But now you're building something totally new.

GIL: [3:41]
Yeah, different kind of building for sure. It was funny, because as an architect I did mostly commercial architecture — I built a lot of offices, a lot of tech offices in Boston. I'd always jokingly say I was the bridesmaid, never the bride — I'd design these really cool tech offices, but then I'd move on to designing the next one. I got to really enjoy all the spoils that tech brings in terms of the working space.

And now I get to be in this space, building the next generation of really cool products, especially for marketing tech, which is a really interesting place to be.

ZAK: [4:19]
Yeah, it sure is. All right — for anybody who doesn't know what Composer is, could you give us the high-level overview of what it is and what the goals are for Klaviyo, and for you, in building it?

GIL: [4:34]
Sure. The high-level cliff notes on Composer is that it's a first-party, in-app AI built by Klaviyo. It's directly integrated into your account — it knows your account data, it knows context about marketing in general that we've trained the AI on, and it's basically a built-in consultant and knowledge base.

You can ask any question at any time — it never gets annoyed that you're asking so many questions about marketing or your program — and it's there for you as you build out anything related to marketing for your program, as well as specifically how Klaviyo does marketing actions. So: first-party, built by the Klaviyo product team.

And then in terms of the goals we're trying to accomplish with Composer — at the highest level, we know marketing is a complex process. You're an agency, you do this at scale, and people come to you because marketing is hard and you know how to do it well. For a lot of people who may not know marketing principles as well, all the way to people with very complex marketing programs, there's a lot of nuance in how the operations of marketing work — how to configure your automations or flows, all the way down to the best way to set up a subject line or the best times to send different campaigns. These are all opinionated insights that guide the success of a marketing program.

And it's very easy — I don't know if "mistake" is the right word — but there are so many ways you could optimize any of these marketing choices, from something as small as images and color all the way to what audience to select. So there's a lot of complexity within marketing, and we wanted to bring in AI to give guidance and expertise to entrepreneurs and enterprise customers alike, and further optimize programs of all sizes.

ZAK: [6:60]
And you guys have what, about 200,000 e-com brands on your platform? That's a lot of data, which is amazingly powerful. Right now, if I'm a brand using Composer inside my Klaviyo account, how unique are the suggestions it gives me — to me, my brand, my industry?

GIL: [7:30]
Yeah, there are a few different layers of context. If you've played around with AI at all, you know that context is the game — you want to give AI a particular lens of information so it can provide you with something that's bespoke and useful.

For Klaviyo, there are a few layers. One is brand context — information we either gathered during onboarding, or that you've provided in one of the settings areas where you can describe your brand and what you care about. That's first-party information. Then there's account-level context — all the marketing objects you have, all your brand assets like imagery, colors, fonts, and so on. These are pieces of marketing data that live specifically in your account.

And then the last layer is aggregate, anonymized data across the roughly 200,000 brands we serve. We collectively understand benchmarking and usage patterns across those brands, sliced by industry. As Composer works on getting you a recommendation or an optimization, it pulls through each of these layers to give you context and recommendations that are specific to you.

ZAK: [9:02]
That's a very clear way of explaining it. How are you seeing brands using it right now? Is there anything about how they're using it that's surprised you — where you thought, "I wasn't expecting that, but that's pretty cool"?

GIL: [9:19]
Yeah — now that we're in public beta, we're super excited about the growth in usage since we launched on June 30th. We're in the tens of thousands of brands that have used Composer within the first 30 days, which is amazing. We've seen a lot of different patterns — we'd been testing this with a private beta group for several months before that, and a lot of those patterns have largely held true.

We know flows, our automation tool, is a pretty complex feature, and understanding exactly what every automation does, and how each stacks on top of the others, is an area of a lot of opacity in terms of understanding how the program works. So we built tooling into Composer called flows auditing, which gives you a flexible understanding of what your automation program is doing — whether you have flows that overlap, people going down multiple flows at once, or configuration issues, like almost everyone going down the "yes" path and nobody going down the "no" path, which is probably a logic problem. It's able to triage all of that signal across your entire automation program with a single click and a few seconds of processing.

That's been the most popular feature since the early private-beta days, and it still is. But what's been really exciting is what debuted at public beta: the ability to generate campaigns. That was a little nascent before June 30th, but at public beta we announced it and people got to use it. Brands are using natural language to say, "I want a campaign that does this, leverages this segment, and give me two different variations for email versus SMS" — and getting real outputs from that. We knew people would want to do it, but we didn't know to what capacity. Now that it's available, seeing people actually generate, send, and—

GIL: [11:43]
—make money from campaigns via Composer has been pretty exciting.

ZAK: [11:48]
Yeah, it is awesome. Are you able to share any revenue lift you're seeing brands make?

GIL: [11:58]
I can't share the exact revenue lift. What I will say is that in these early days, it's largely about value for time. A lot of these decisions and processes — creating a campaign, creating multiple variations for different segments, and so on — historically took a lot of thought, coordination, and manual effort.

Now, because we can understand the prompt context and build different variations for each channel type and each pre-configured audience, there's a lot of time savings for equivalent or greater output. So partly it's more output for your time. And then, on the raw-numbers side, people are genuinely making money from campaigns Composer created.

We also have an internal metric we call "no edits, no notes" — Composer produces an output, and the user doesn't ask for any further changes; they just save it, send it, and it makes money. There's a growing number of these "no-notes" campaigns, which is pretty cool — it means we understood the job, we understood the context of your account, we understood what you were trying to accomplish, and it was good enough that you felt comfortable sending it. That's ultimately where we want to get to — as many of these no-notes campaigns as possible.

ZAK: [13:39]
As an agency, getting no-notes feedback from clients is always a wonderful thing too, so I can relate to that. A couple of the things you're talking about — coming from the agency perspective, I have really strong opinions. I've been talking to a lot of other agencies in this space, and there are really a couple of different ways to look at it, but—

GIL: [13:49]
Good to go.

ZAK: [14:09]
I'm personally super excited about everything you guys are creating with Composer, with Customer Agent, with all the AI tools you're building. But it's changing drastically what e-com agencies look like, and it will keep changing. For instance, the flow audits you mentioned — for anybody who's never audited flows, that can be a huge task. For one client we manage on an ongoing basis, it might be a quarterly task where one of our strategists goes in and audits all their flows, and it takes a lot of time. It's impactful work, but it's not high-level work.

So there are two ways to look at it from the agency perspective. One is, "hey, you're taking away this work." The other is, "wow, this is awesome — you're freeing up the work so it can happen continuously, and I can spend more time on higher-level, revenue-impacting strategy."

GIL: [15:13]
Yeah, I think that's exactly right, and I think most people fall into one of those two categories. Some see it as a one-for-one replacement of the work they're doing. Others see it as, "this is work I was never that interested in doing anyway — now I don't have to, so what else could I spend my time on?"

The analogy I'd draw is what we're seeing more broadly with these large frontier models in engineering: a lot of builders at heart, now that we have tooling that can code on its own, get the ability to produce a lot more output, be more strategic, and say, "these are the things I'd really want to build if I had unlimited resources" — and yesterday they couldn't, and today they can.

GIL: [17:37]
Email variations are effectively free, so maybe I can think more about segmentation, customizing content, and being narrower in who I target. That requires a next level of strategic thinking you couldn't afford before, when you could only pay for one or two email variants per campaign. Now you can give clients ten or twenty and dream up whatever experiments you want to try with different groups.

I think agencies are going to find a lot more freedom to activate strategies they always wished they could do. The normal, day-to-day, lever-pulling grunt work changes and frees you up to do much more interesting things.

ZAK: [18:44]
Yeah, I think it's important for agencies, and for everyone, to understand what's happening. Whatever your opinion of AI is, broadly, it doesn't really matter — it's happening. As a business, whatever business you're in — marketing agency, manufacturing, whatever — if you put your blinders on and ignore it, a year or two from now, you're done. There won't be a playground left for you to play in.

GIL: [19:21]
Yeah, I definitely agree at a high level. The other side of it is: you've been given this unlimited, incredibly flexible cheat code. Do you want it, or do you not? I think you should take it — figure out how to use it. It's a really exciting time for people to lean in and do things you otherwise couldn't do before.

ZAK: [19:55]
Yeah — the way I described it before probably sounded like doom and gloom, and that's not what I meant at all. There are two ways to look at it: doom and gloom, or exactly what you said — "holy cow, here's this cheat code I have access to, where I can do things I could never do before."

I built a sales dashboard about a month ago — I pulled in data from about ten different sources for our internal sales dashboard, and I built this awesome, super-powerful dashboard. I'm not a developer, and I built it inside Lovable using Claude to do it. Now it's the first thing I look at when I open my computer. But I was thinking — if I'd built this a year ago, I would've paid twenty-five, thirty, thirty-five grand to get it built, and it probably would've taken three times as long, going back and forth with developers and fixing things that didn't come out how I wanted.

That was my first "aha" moment — being able to create that thing was mind-blowing. I'm not a developer, and all of a sudden I built this super-complex dashboard with a couple of tools and a few prompts.

GIL: [21:24]
Yeah, for sure. There are a few different ways to think about the economics here. There's the economics of outputs, which have clearly changed — whether it's an app, a dashboard, a report, an audit, these are all outputs, and AI basically pushes that cost to almost zero; it's just the token cost. Then there's the economics of thinking, which has also changed, because now Claude can think through all the steps a senior developer would think through to build the thing — that's also compressed in cost.

So when you talk about economics, where's the real value? It's in thinking about how you activate the AI, or multiple AIs, to do these things at scale, agentically, letting them run on their own. That's really where the thought-value comes from now. And for e-com marketing agencies, there's a lot of value in thinking about how you'd build out all the automations for all of this.

ZAK: [22:40]
Yeah — you definitely had this moment, but do you remember what your "aha" moment was with agentic coding and AI, when you thought, "holy cow, we're in a brand-new space"?

GIL: [22:59]
Yeah, I think mine was probably a little different from others'. I'm a big voice-to-text, dictation person, so I think it was in the early days when ChatGPT introduced the voice feature that let it respond in voice. It was funny — there were a few different voice personalities you could choose, so it was almost like, "what friend do I want in this process?"

That was the first "aha." In the beginning, chat felt more like a different way of doing a Google search, or like I could give it some of my writing and it would kind of mirror it, but it wasn't that good yet. The first time I was actually able to speak to the AI and have it speak back to me, it felt like, okay, there's something really novel here — like it's bringing its own thoughts and considerations. As a person, I felt like I was engaging with something, not just an intake tool like a Word doc. It felt like something was being given back to me. That was the first real "aha" of how AI could change not just how you work, but how you interact with technology.

Now I mostly interact with Claude through dictation, using products like Whisper Flow. I'm just waiting for the nexus of remote control to my server, plus ElevenLabs and all these things, to come together into a super-AI I can just talk to — or for Siri to get local AI so it can happen really fast. The biggest problem right now is latency.

ZAK: [24:56]
I often think about tasks I want to hand off to somebody else on my team, or my EA, and I don't have time to dictate them. So I think, man, it'd be great if I could just take this little thread of a thought, export it, have it expanded on, and then have it assigned out.

GIL: [25:27]
There's all kinds of thought-to-dictation stuff being experimented with right now — one day soon. But I'm curious about you — I think this is a pretty personal kind of story. What was your "aha" moment related to AI?

ZAK: [25:47]
One of my first "aha" moments — well, actually the most recent one — was when I started building things in Lovable, because I was able to take concepts from my mind and build them into an actual, usable product I could interact with, that was correct and really valuable. That was a big "aha" for me.

And then, more recently, I got into Claude Code and created what I call an "ECD brain" and a "Zak brain." I imported all my historical emails, all my Google Drive information, and a lot of context about me individually, and created two different brains. Now I can ask questions about how I'd respond, or have the Zak brain craft messages to my team in my voice — that's pretty wild.

Something neat I'm starting to do too: I do a lot of stuff on LinkedIn, and most of what I talk about is current events or things happening now. But I have a pretty interesting background I don't talk about much — in my twenties I lived in about thirty-five different countries, my first business was a mountain-bike tour company in Costa Rica, I traveled the world on a sailboat, and I imported textiles from Thailand and Nepal to the U.S. I kept a blog through all of that, so I imported the blog into the Zak brain and had it start tying that historical background into what I currently talk about on LinkedIn. That's pretty cool too.

GIL: [28:11]
Do you draft new content based on that historic stuff?

ZAK: [28:15]
Yeah, exactly.

GIL: [28:39]
I think that's one of the highest-leverage, highest-impact things — if you're in marketing, marketing yourself is always valuable. I think about that too. What's interesting is I've set up my Claude with a similar context layer — I take a lot of the insights I learn from working here at Klaviyo, the problems I see, the way to interact with people or technical problems, and I'm able to extract some of those learnings from conversations with the AI and say, "this would actually be an interesting blog post — let's do a draft on it." It proactively scans conversations, meeting notes, all of that, to give me blog or LinkedIn post ideas, which is pretty—

ZAK: [29:14]
Nice — from stuff you're naturally talking about or thinking about.

GIL: [29:17]
Exactly. I had a post about token burn the other day — at the time, a lot of companies were saying, "unlimited budget, do whatever you want, we want to learn." I wrote that this isn't going to last forever, and you should think about it now. And sure enough, companies are really tightening their belts now because their AI spend went wild. That was a literal idea that came out of conversations I was having here at Klaviyo.

ZAK: [29:47]
Yeah, budgets went wild. But I think AI is still relatively cheap right now — I think it's artificially inexpensive. Maybe in a year or two, as the models get more powerful, that starts to change.

GIL: [30:11]
Yeah, exactly — more powerful, less subsidized, and your expectations for it go up. It's coming for you.

ZAK: [30:25]
AD READ: Want to see where your e-commerce brand is leaving money on the table? Book a free revenue audit with the team at ECD Digital Strategy and get a clear plan to increase revenue. Go to klaviyo.com/ecd to schedule yours today.

ZAK: [30:41]
I'd like to talk a little about the creative side of things — design, graphic design, and copy. Where is Composer today on the graphic-design level, and when do you think it's going to take that next big jump to being significantly better?

GIL: [31:14]
Yeah, there are really two pieces to the "when does it get better" question, which is a common question any AI product is going to keep facing, even as it improves. One piece is: how good is the context? The other is: how good is the model?

At Klaviyo, we don't build our own models — we sit on top of the large frontier LLM providers, whether their focus is intelligence, like Anthropic, or image generation, like Gemini, and we choose the best one for the right task. What we build on top of that is a harness — an opinion about how you should use all of these different models and tools for a particular goal, in this case marketing tech.

So there are two pieces: as the models get better, we get better for free, since we sit on top of them. And the other piece is improving context — some of it user-provided (things they explicitly tell us, like the metrics they care about or their brand colors), and some of it learned implicitly, from behavior — whether they accept recommendations, what percentage of campaign generations they accept, how often they edit flows versus campaigns per week. We learn from the user, and that also informs context. And then there's platform context — as the platform gets bigger and pulls in more integrations and tooling, the surface area of context grows too. As all of these layers grow, along with the model, that's how things get better — though it's not always perfectly timed.

ZAK: [33:42]
Do you have a favorite model you're using right now?

GIL: [33:50]
I don't know if I have a favorite. We're very lucky at Klaviyo — we get a fairly generous budget for models, though the most expensive tier has been too costly for us so far. Right now I'm doing everything on the newest release, which came out just a few days ago; before that I was on the previous generation. I think at some point one of the earlier versions felt like a really nice sweet spot, but now it feels dated. I've kept pushing up to the newest models. And this is such a nerdy thing to say, but there was a version that came out and felt like a step down from the one before it, and a lot of people were complaining about that — I felt it too, not sure if it was placebo. But the newest release, even after just a couple of days of use, feels—

ZAK: [35:04]
Your LinkedIn bio — I'm going to butcher it, but it says something like "burning tokens at Klaviyo," and I love that. Do you guys have a leaderboard of who's using the most, who's burning the most tokens?

GIL: [35:28]
Technically, yes, but it's actually more engineering-focused than product or design. I think, over time, we're thinking about that differently. I always say leaderboards predate AI — depending on what you're trying to create a leaderboard for, you're going to incentivize certain things. Now that there's a lot of automation built into these tools, it's pretty easy to gamify a leaderboard like that — you just run a bunch of cron jobs and explorations.

So I think leaderboards have become less valuable compared to the early days, when everyone was still figuring things out. But generally, you want to be above maybe the bottom thirty to forty percent, not to compare against anybody, but because you want the reps. You have an opportunity to use and flex this cheat code, and here at Klaviyo it's free for us as employees, so you really want to be using it and giving yourself the best chance to succeed.

ZAK: [36:55]
Yeah, I hadn't even thought about that — if there were leaderboards, those would be extremely easy to game.

GIL: [37:05]
Yeah — it was actually really funny. Here at headquarters in Boston, we're all on one Wi-Fi network, and there was a pattern where the Wi-Fi got really slow. I was complaining to our IT department, like, "hey, what's going on, I can't get any work done." They investigated, and it turned out to be one person running a massive cron job that was taking up literally all the bandwidth on the floor. We were like, okay, let's not do that — it can't possibly be valuable enough to knock out Wi-Fi for everyone else on the floor. But it was funny to see that, because the budget is unlimited, some people start doing these really wild things. If we did have a leaderboard, that person would be at the top, but it's unclear whether it was delivering any actual value.

ZAK: [38:15]
Cool. I want to shift a little — right now, on the customer-facing side, the other big AI product you've launched is Customer Agent, which has been live for a while and is a really cool product. So you have Composer, you have Customer Agent — down the road, what does the overlap between the two look like? What's the relationship going to be as they work together? Walk me through the vision there.

GIL: [38:60]
Sure, let me set the stage. Customer Agent is an AI product that lives primarily on our customers' customer-facing website — for the people who actually purchase from that seller or brand. The most obvious use cases are letting that end purchaser chat with an AI on the website to ask questions — status of an order, shipping, the ingredients list of a product — a knowledge base for that company that the AI can answer from. That's a fairly foundational, level-one use case, and there are a lot of chatbot providers that can do something similar.

What makes Customer Agent really unique, and why Klaviyo has invested a lot into it, is that it opens an entirely new avenue of signal on that customer. When a customer talks with Customer Agent, there's a lot of implied signal we can collect — if they say, "this is a purchase for my wife" or "my spouse," we can infer it's a gift purchase; if their questions are about ingredients, maybe there's a concern about an allergy. All of these are signals Customer Agent, as an AI, can write onto the profile as proposed attributes.

Pre-AI, you as the marketer had to be really deliberate about what attributes you were collecting and why — you'd say, okay, here's a sign-up form, we got this lead through a promo email, so there's some attribute we're interested in, plus the email opt-in, plus the product they might be interested in. These are—

GIL: [41:25]
—clear attributes you've decided to write onto the profile so you can use them downstream — you might build twenty different segments, each with a different favorite product type, and use that as content in your messages. So you had this whole loop in your mind about what attributes to author so you knew what marketing you could send. In this new world, where Customer Agent can agentically propose attributes on the profile on its own, you have an entirely new lens of possible signals from that customer. Instead of only the attributes you've authored, now the AI has authored a whole set of potential attributes, giving you far more opportunity for personalization and bespoke automations.

This is where Composer comes in. I always tell people — I don't write every sentence myself these days. I use Claude to help me write a proposal, some guidance doc, a product spec. All the ideas originate from me, but do I know the exact construction of every sentence? No. Similarly, profiles authored partly by AI may start to grow in ways the marketer doesn't fully track. That's where Composer — the AI that helps make sense of a complex marketing program — comes in: it can also make sense of profile attributes that have grown very large because Customer Agent is contributing to them. It might say, for example, fifteen percent of your customer base has ingredients-related questions — maybe you should build a knowledge-based, three-email newsletter series over the next week about how your ingredients are all natural. That's signal created by Customer Agent, triaged and made sense of by Composer, to help you build the marketing execution plan on the back end.

GIL: [43:51]
Ultimately, what happens is you get a feedback loop: you're collecting signal on the customer-facing surfaces, triaging that signal within Klaviyo through Composer, and then pushing output back out to customers, and continuing to learn. If both Customer Agent and Composer do their jobs well, you can see a world where they start handling that on their own, and you're really free to focus on what products to feature next season, what new promotional opportunities exist generally — without having to think hard about segmentation, personalization, or timing, because the system can experiment on your behalf. We're all waiting for it to hit that quality level, and once it does, it's going to be really interesting.

ZAK: [44:51]
Yeah — you said Composer can look at a set of attributes and say, "hey, fifteen percent of your customers want to do X." I think you'd agree that's the old way of looking at it — fifteen percent, okay, that's probably worth doing. But now it can say, "point-one percent of your customers want to do this, and I've created an email campaign for those seven people — do you want to send it?" Boom, click send, and it sends it out, because historically, as an agency, we wouldn't go do something like that, since it might only make the company an extra thirty bucks a month. But if Composer and Customer Agent are working together to give you twenty, thirty, forty, fifty of these a day or a week, that thirty bucks starts adding up, and it gets that much more powerful. Everyone's always talking about personalization in digital marketing, and this is actually going to get us to true one-to-one emails — down the road, every single email could be unique to that specific person, based on what they've done, how they've interacted, how they've spoken. It could even be in their customer voice, using the words they've used historically in reviews or with Customer Agent, to give them a genuinely unique experience.

GIL: [46:29]
Yeah, the number of permutations is genuinely unlimited. There are so many just within the email example you described — content, voice, tone, timing. And then there's the segmentation of a very small slice of people. On the omnichannel side, which I still oversee, a lot of that thinking is about the relationship between channels — the timing of when an email goes out isn't the same as when someone engages with an SMS. So there's optimization on what channels people prefer, at what time, and what order of channels people would rationally engage with — does an email alert bring them to check SMS, or does an SMS alert bring them to check email? There are all these permutations we can think about that would be genuinely impossible for a human to manage. With one-to-one personalization and AI managing this, it's actually possible. And the thing that's hard to fathom is that the learning loop is really where the value is.

It's not that you get all this signal and output exactly the right thing — you get the signal, create a hypothesis of how it could work, run another loop, see what worked and what didn't, try the winners again, try something different for the rest. You keep looping until you've honed in on a success path for a subset of customers, while others haven't been figured out yet — so you've got "wins" and "warms." You keep looping through the people who haven't been warmed yet. This whole learning-loop process is genuinely unlocked through AI in a way that just wasn't possible for humans to do manually — that's where the gains are.

ZAK: [48:47]
It's so exciting — the more I talk about it, the more I nerd out on it. This example we were just talking about — when is that coming? When is that going to be a reality?

GIL: [48:57]
There are some non-AI answers to this. Obviously I'm biased, working here — but Klaviyo has made some very strategic investments in surfaces that aren't specifically lifecycle marketing. We're obviously invested in Composer and the products we're familiar with — campaigns, flows, forms, and so on — the standard lifecycle and retention marketing surfaces. But the real innovation, beyond a lot that's happening in those cores, is in other surfaces Klaviyo is investing in: Customer Agent, Customer Hub, which live on the customer's website, so we can collect signal there. There's also our social product, where we can start collecting information based on users' opt-ins with different social influencers, or signal from reviews, or a help-desk product where people file tickets, which tells us their concerns about a product. All of these add into the profile layer that enables these learning loops.

So in my view, the majority of e-com brands just won't be able to run learning loops at the fidelity Klaviyo can, because they're not invested in these surfaces. Either they have to integrate with third parties and build relationships to share data, or they can only get learning loops from opens, clicks, and conversions — standard retention-marketing signal. You could call that personalization, but it's going to be at much lower fidelity than if you had all this other signal coming in.

So when do we get there? For Klaviyo, I genuinely think it's going to be soon. For the broader e-com market, I'm not so sure. But Klaviyo has the infrastructure and investment in the right places, and I'm excited to deliver on that vision.

ZAK: [51:11]
I don't disagree with that at all. Klaviyo started talking about the D2C CRM maybe a year and a half ago, and I remember thinking at the time, "there's no way this makes sense to me" — because I didn't know what was coming. And then you had marketing analytics, Customer Agent, Customer Hub, Composer, and now more products layered in. Now it really makes sense — it is the D2C CRM, in the sense of unified data in one place, but it's more than a CRM, because it's a system where all this different data learns from each other and you come out with one cohesive strategy to market to your customers where and when they want.

GIL: [52:15]
Yeah, and I think that's the dream. You tell me — you're an agency, effectively passing along the value proposition to your customers about what you're providing. I think the dream for someone like you, as an operator, is to have a single source of truth — to build a marketing program off of everything you know from one profile that's collecting information across first-party-supported products, so you can iterate on strategy, especially now with access to AI. How do you think about a dedicated, bespoke, first-party data platform versus other products with a more scattered approach to data?

ZAK: [53:07]
I'm such a big believer, for e-com brands, that all of their marketing channels that are possible should be on Klaviyo — I talk to e-com brands about this all the time. It bewilders me when I see a company using Klaviyo for email and a different platform for SMS, for countless reasons.

Same thing with Customer Agent — we're implementing a ton of it, because it just makes so much sense. We used to use a separate customer-service tool before Customer Agent came out, and it was cool, it helped with customer service. But once Klaviyo's Customer Agent came out, the idea that you could get all this data written to the profiles — the power behind that is amazing.

Before, when we used a non-Klaviyo tool, it was essentially a chatbot — good for customer service, but that's not the exciting part. The exciting part is that it's a shopping agent — you can link it to your flows and automations to set up upsells and cross-sells, connect it to other apps you're using, like Smile Rewards, to pull reward balances and sync them into your flows, and then surface that data through Customer Hub for personalization on the website. It's amazing what you're able to do. I'd love a real look behind the curtain at what Klaviyo has coming next, because I'm sure it's even more mind-blowing.

GIL: [54:58]
Well, thank you, that's well said. There's a lot of excitement around Customer Agent, which is still a relatively young product but has already been doing amazing things — not just the chat function, but lifting the ceiling on the lifecycle marketing side.

As for what's coming next: it's a really interesting time. As of this podcast, we're in public beta for Composer, which is where I spend most of my time, and the next iterations are going to be super interesting. Right now, Composer is largely a conversational knowledge base — it knows everything about your account and how Klaviyo works, and gives you guidance and recommendations based on best practices and your account data, telling you what you can probably work on to make your program better. But it's really just conversational knowledge.

The next thing we're working on is having Composer take actions on your behalf. Instead of just saying, "this is what I think you should work on next," or, "this is probably a quick win," it will make a proposal — similar to a suggestion in a Google Doc that you can accept or reject. Composer will say, "it looks like you have a logic problem here — would you like me to fix that?" Accept or reject. That's a precursor to agentic work — giving you a proposal with a human still in the loop, so you still get to learn and then decide to accept or reject the guidance.

So Composer starts to take actions, not just guide you — in the nerdy way of putting it, it can write, not just read. That's going to be super exciting, and one of the things I'll be shipping very soon. There are a couple more things, but I'll stop there — what do you think about that?

ZAK: [57:13]
I think that's amazing. As an agency, let me back up a little — Composer is going to be making these suggestions, but to know whether you should accept or reject them, you still really need to know what you're doing. You need to understand the impact these suggestions will have.

I'm bringing that up because I think a lot of people, especially on the brand side managing their own Klaviyo account, can get scared that Klaviyo, or Composer, is taking over their job. That's not what's happening, because your context is still needed to know whether to accept a suggestion or not — it's not always going to be simple. Some of them may be, but—

That excites me a lot. Say we have our strategists open their Klaviyo dashboards on a Monday morning, and Composer has surfaced the top ten recommendations from over the weekend. Our strategists go through each account they manage and accept or reject those recommendations. Maybe they made an extra seventeen dollars for one brand that day, a hundred and seventeen dollars for another, a thousand extra dollars for a third — just by going through and accepting or rejecting what Composer surfaced. I think that's going to be a huge unlock, enabling strategists, whether agency-side or brand-side, to make the brands they work for a lot more money while doing more high-level work.

GIL: [59:03]
Yeah, that's a hundred percent what we're hoping for. As we talked about earlier, it's about layering on context — once you provide your bespoke opinions about how Klaviyo, and marketing, should work, the guidance for these recommendations becomes more bespoke and custom to you, and you're much more likely to accept the guidance, fixes, or new variants when they align with what you care about. That's number one.

I'll give you three things. Number one is what we call one-click actions — the ability for Composer to act on its own recommendations. Number two: at our June 30th public-beta launch, we shipped the ability to create campaigns. Creating flows is already out to a percentage of customers and will be out to everyone very shortly — you'll be able to use natural language to create entire flows, the same way you're describing for campaigns. You can say, "help me build a welcome flow, here's the content for each step," and it uses the same engine for generating email content, or you can specify the logic for delays, and Composer will also give you guidance on best practices for building these flows. Generating flows through natural language is going to ship really soon — that's number two.

ZAK: [1:00:42]
Let me jump in on that one — that's also really exciting for us, but I think it'll be really exciting for the Klaviyo sales team too. We do a ton of Klaviyo implementations — about ten a month — and we work closely with your sales team. Oftentimes, one of clients' biggest hesitations about migrating to Klaviyo is the work of actually building out their account, and they might not want to pay an agency an extra fee to build it out. Now, sales reps will be able to say, "you can build out your Klaviyo account with Composer" — that's going to be a big unlock for them, and they should be pumped about it.

GIL: [1:01:28]
Yeah — I won't go too deep into this particular part, but there's a whole opportunity around onboarding and getting people up to speed. You've got an absolute zero account, nothing in it, and you need to get it going. Composer's ability to understand "this is a zero account, here's what we know about the customer, here are the things we normally do for every brand to get them to a healthy starting state" gives you basically a personal tutor or consultant next to you, who knows exactly what state your account is in and what you need to do next, and why. Onboarding, and getting started with Klaviyo, is going to be better than ever, really soon.

ZAK: [1:02:24]
What about deliverability and warm-up? I'd bet that's coming too — Composer using best practices and what it's seeing happen to automate the warm-up or deliverability fixes for a brand.

GIL: [1:02:37]
Even conceptually, it's like, "you're a new account, we need to warm you up — here's what that means, here's the plan for how we do it, are you cool with this plan?" Accept it, and we get going. So that's two.

The third one is a little further out, but I'm really excited about it and wanted to share it with your audience: intelligence that isn't specific to the chat. Right now, the primary way you engage with Composer, and with AI generally in most products, is through chat. But with Klaviyo, the opportunity to optimize and create new things isn't really relegated to chat — the core marketing product inside Klaviyo, the UI we give marketers, has a whole set of opportunities to show you things directly. "Hey, this flow needs an optimization, here's the canvas highlighting the area with the logic issue." We can give you guidance today that happens inside chat, but there's a whole back-and-forth through chat that's really a proxy for the thing you actually want to do.

Because we own the Klaviyo UX, we'll be able to take the intelligence we've built and distribute it across the app — inline in the table of flows, in the canvas, when you're building a form, or looking at reporting and trying to find guidance. All of these little moments of intelligence become first-party, in-app. It's not, "this is AI, I have to go talk to it" — it's Klaviyo's intelligence highlighting things to you at any time, right when you're thinking about it.

ZAK: [1:04:55]
That's awesome — so if you're building out flows and setting something up in a certain way, Composer surfaces guidance directly, right there.

GIL: [1:05:07]
Yeah, it's completely contextual. It understands exactly where you're at and exactly what step it's describing to you. There are all these moments — anything from "I don't know what this button does" to "I'm not sure if this is the right configuration" to "here's a brand-new welcome flow, and I want help understanding each of its components."

If you wanted a consultant or tutor to help you fix or understand something, it probably wouldn't be through emailing back and forth — you'd want that consultant to bring it up on screen and walk you through it, point and click. That's what Composer is going to do, because we own not only the intelligence but the UX. That's going to be a meaningful differentiator as people use frontier AI natively and connect it to Klaviyo through an MCP to pull data out — you can get a lot of intelligence and reasoning, and some things to act on, through that. But once this is available, the experience of Composer inside Klaviyo is going to be a meaningful difference from going through a general frontier LLM.

ZAK: [1:06:48]
And that's really going to be a moat for you guys too — one of the moats.

GIL: [1:06:55]
One of the moats, yeah. We have the first-party data — we understand what good marketing looks like through the brands we serve. But we're also currently building the harness, the opinion of how AI should be applied to marketing, and how it should be applied to a UX surface for marketing. All of that combined is going to be a best-in-class experience for the integration of AI into marketing. My opinion is people will stop thinking about it as AI at all — they'll just think, "you were supposed to give me this guidance and intelligence all along; this is how I expect it to show up." It'll just be a natural part of the day-to-day working experience.

ZAK: [1:07:43]
That's awesome, super exciting — thanks for sharing all that. Gil, this happens every time we talk. I feel like the time just flies by and we're just scratching the surface. This was very exciting, sort of what's coming, and I'm going to have to have you back on the podcast in a couple of months to talk about some other things happening.

Before we hop off, I've got a couple of rapid-fire questions for you — are you ready? One metric you think brands over-index on?

GIL: [1:08:26]
That's a good question — maybe let's skip this one. I'm not the metrics guy.

ZAK: [1:08:35]
One tool outside of Klaviyo you love right now.

GIL: [1:08:41]
Outside of Klaviyo, it's got to be Claude — that one's easy. But if it wasn't Claude, it would be Whisper Flow, which I mentioned earlier. I save hours a day on that.

ZAK: [1:08:55]
It's amazing — I use it as well. At first I didn't understand it; people kept talking about it, and I thought, well, you have voice-to-text right inside Claude, right inside ChatGPT, why do you need it? But then once you start using it—

GIL: [1:09:11]
Yeah, there's something about it — auto-creating lists for you, numbered or bulleted, taking away the ums and ahs and half-formed thinking, and turning it into an actually clear statement. It's hard to beat right now.

ZAK: [1:09:29]
Gil, where can people follow what you're building?

GIL: [1:09:34]
For Composer, which is what I think about most these days, there's klaviyo.com/composer — you can see all the latest updates there. We'll also be publishing a change log very soon, since we're shipping new releases basically every other day. And if you want to follow me on LinkedIn, it's linkedin.com/in/gilhsu, where you can hear about what I've learned and worked on at Klaviyo.

ZAK: [1:10:08]
Awesome, Gil — thanks so much for coming on the podcast today.

GIL: [1:10:13]
Hey, Zak, appreciate it — let's do this again.

ZAK: [1:10:17]
AD READ: Before we wrap up — if you're serious about growing your e-commerce brand, go grab a free revenue audit. We'll find the revenue you're leaving on the table. We'll analyze your store, marketing, and retention channels to show you exactly where the biggest revenue opportunities are. It's completely free, and you'll walk away with a clear growth plan. Just go to klaviyo.com/ecd and book your audit today.