I'm very open. I'll say to employees all the time, "I haven't fully transformed the way I work." When I admit that, it gives people license to recognize that it's okay, and we all need to do better and they need to do better. The advice that I would give to CEOs or executives listening to this that are trying to get their company to adopt AI would be to...
You're 18 months into the job. So, right as AI starts to pick up, you took this job. How much of a factor was what was going on in AI a determining factor in you actually taking this job and coming out of retirement?
It was a small factor. At the time, 18 months ago, it feels like a million years ago, you're right. AI was starting to be baked into our products, but we hadn't yet had the explosion of internal productivity. AI coding assistants were not widely adopted like they are now. The magic that our employees feel because of the AI that they're using every day, that wasn't a thing 18 months ago. So I could see that there was enormous potential for AI to impact our products, but I did not forecast the incredible effect that it had on employee productivity. That's new and exciting and energizing.
And when did that kick in? Obviously December was a big inflection. You were probably doing it before that, I would assume.
We were. I mean, we're recording this now in August of 2026. It feels like probably about five, six months ago. So early 2026 was when Match Group had this magic moment. I think every company is on a different stage of their journey in terms of AI adoption. What we've been doing as a company was that first stage was just give everybody everything. We gave every employee Claude, ChatGPT, Glean, Cursor. It was just like buy all the tools and make it accessible. Then the second phase was do the proper training and onboarding to try to drive adoption.
Then we sort of took a step back and checked what the bill was, and the bill started to tick up pretty significantly pretty quickly. Instead of trying to really rein it in, we decided to just create some controls and a little bit more process, a little bit more structure, a little bit more education about which models to use when. During each of these chapters, there were really critical moments. I think the most interesting one for companies that are maybe a little bit behind was the critical moment in terms of driving adoption was we had an AI Day where we just canceled all meetings. Everyone just build, just mess around. You're all at different stages of your journey, and just build cool stuff. And then the next day, everybody demoed what they built.
And even for me as the CEO, that was a really useful, critical moment. I hadn't actually taken the time to really rethink my productivity and how do I manage my day. So I built for myself during that. I built AI tools that prepared me for my day, that sent me a Slack message 15 minutes before every meeting. For example, I got a Slack message 15 minutes before this podcast with all the preparatory notes, all the questions that we're going to go over, all the things I need to know. So now before every meeting, I get a little dossier sent from AI to myself to prepare me for the meeting.
I vibe coded a fun little chess app for me and my family during that AI Day where I can play all three of my kids at chess. I make a move, it texts all three of them that I've moved, and any of the three of them can make a move back. So it's kind of a one-on-three chess thing, which was something that I've always thought about doing. It was just a little test for myself: can I vibe code this? I'm not a software engineer. I've never written a line of code, and I managed to build that with Claude Code in an hour or two. So once we got to the other side of that, people like me who don't code every day were starting to see what's possible, and it really just kicked AI adoption into the next gear companywide.
And just organizationally, who are the overseers of all this?
Yeah, sorting that out was a bit of a challenge and still sort of work in progress. So the Tinder CTO has played a huge part, and then we created a product team, basically a pod at Tinder of about six people that we call Velocity. All they do is improve the throughput and speed of the other 500 engineers with lots of internal tooling. A lot of that is AI focused, but some of it is things like how do we speed up how long it takes to do a new build of code. If that takes 30 minutes, then an engineer is sitting there for 30 minutes. So how do we reduce that time? Some of it is AI related, some of it's not. That team is very involved. And then Hinge has kind of the same structure.
And then we created a lot of rules within Slack also to just share knowledge about AI. So we have a thing called the Brand AI Council, which is a Slack channel with like 10 or 15 people. They're the ones, with the CFO's guidance, that figure out the right token limits and the right guardrails, and when we allow certain people to use different models. For example, when Claude's Opus came out, it was really powerful but also really expensive. So we did a very small proof of concept, just let a couple dozen employees use it for a couple weeks. Then we did some internal benchmarking to compare Opus with other Claude models, and then ultimately we decided not to make Opus available to every employee. Now only a handful are able to use it, and they have different guardrails.
So this Brand AI Council establishes all of those rules of the road, but it took time. Figuring all this out—where does it live, what role does the CFO have, what role did the CTO have, should we stand up a separate function doing the training and onboarding and upleveling everybody, driving adoption—it was, and still is, a work in progress.
Not to go too deep in coding, but when you look at this Velocity group, how much of it is, "Man, now we can do new, cool things," or how much of it is the management of the legacy code is now easier? Where is that? Because I feel like certain companies have so much tech debt nowadays.
Huge problem. It's a great insight. Frankly, the AI coding assistants are awesome for building new products from scratch. If you're a startup, you can go so fast now because of coding assistants, but they stink at dealing with legacy code. We really have a pretty big tax to pay because Tinder is a 12-, 15-year-old company. Match.com is a 30-year-old company. There's a lot of code that's been built over a long period of time. At some point, AI coding assistants will become amazing at that too, but for now they're just not great at that.
So for example, a brand new feature from scratch that we built on Tinder called Missed Connections—which basically, if you've opted in to give us your location, then once a week we give you a drop that says who did you walk past. Here are a couple people that you crossed paths with over the past week. We don't say where it happened for privacy and security reasons, but these are people that you crossed paths with. That went from idea to shipping it in the app in a couple countries in probably a month thanks to AI coding because it was a brand new feature. In the old world, a year or two ago, that would have taken probably six months. So much, much faster to ship something like that because it's kind of brand new. But if it's something that's deeply integrated into the existing codebase, maybe it speeds things up by 2x, not 10x.
And it can't go retrofit the old codebase yet. You're saying it's not...
No, not yet. I keep dreaming of that. It's funny, we just had a meeting on something kind of related to that, which was, "Should we go..." I think it was around billing infrastructure or something. I forget what it was exactly, but it was like, is now the time for us to go back and rewire this whole thing? And I actually said, "Hey, maybe we just wait a year, because probably a year from now, the coding assistants will be so good that we'll be able to press a button or just make a prompt and it'll go rebuild that without us having to do it." So maybe we just kind of wait.
We just finished a three-year tech migration putting about a dozen of our older apps onto the same backend. So right now, Chispa, which is our app for Hispanic daters, OkCupid, Plenty of Fish, and Match.com, for example, all of those are on the same shared backend. That took about three years to put all of them and another dozen or so apps onto the same platform. I asked engineers a couple weeks ago, "If you were starting that project today using the powers that we have, instead of three years, how long would it have taken?" They said about six months. But if we started it in about a year, it would probably take three weeks.
So there's like a little game theory here of, well, maybe you actually should drag your feet on some of these integrations because it's only getting easier.
Well, that begs a bigger question. If you're a bigger company that doesn't have the technical leadership that Match does, do you just punt and wait it all out then?
I don't think you can, because your startup competitors are moving faster because they don't have the tech debt. I remember when I was at Expedia, and this was a long time ago, more than 15 years ago, they had to basically pay down all their tech debt at once and rearchitect everything. It took like two years, and they couldn't launch any new features during those two years, that whole time. And they fell behind Airbnb and a lot of competitors during that period. On the other side of it, they then were able to move much more quickly. But it's risky. If you were facing that decision today and you're like, "Well, maybe I'll wait a year or two and then it'll be faster and easier," in the meantime, you're falling further behind. So every company has to deal with it themselves, obviously, with different circumstances.
It's a wild time. It's an exciting time to be alive in tech obviously, because we're just moving fast with that. I'll give you one other quick example. Launching events on Tinder as a separate tab was an idea that we had. I came back from the Christmas holiday, so first week of January, I came back and said, "I'm tired of reading that Gen Z doesn't want to use dating apps and they just want to go to run clubs. Why don't we just put run clubs in the app?" We have all this research that says that Gen Z wants to meet in person at events. Okay, fine. Let's bear hug this observation instead of trying to fight against it.
And so, on January 2nd, we decided to try to bring events into the second tab of Tinder. So the first tab would be the swipe tab where you're looking at people, swipe right, swipe left, and the second tab would be events. March 12th, so just two months later, we shipped it in LA. And today, here we are in August, we're in about 25 cities. By the end of the year, we'll be in 75 cities. That sprint, that two-month sprint, was only possible because of AI, because of how quickly that team was able to move because of primarily Cursor and Claude Code. And had we pressed that starting gun or fired that starting gun a year earlier pre-coding assistance, it would have taken 10 times as long.
What's the overall tech stack looking like right now for you? Is it just still everybody's in there, or did you pick winners at this point?
We kind of let a thousand flowers bloom to see what people would use, and the clear winner for now at our company is Claude. I look at it every week because I can literally see spend and token usage across each of these different AI tools. Claude is number one, and then Cursor is number two, and then ChatGPT—well, really Codex, which is the coding assistant for ChatGPT—is number three, and then Glean is number four. Glean is an interesting one. Glean is enterprise search. It had really widespread adoption here internally early on, about a year ago, and now a lot of people are asking Claude the types of questions they would have asked Glean. So it's going to be interesting to see if Glean can keep up.
The data integrity was the best in Glean, or because Glean was the best at answering the question?
Glean is great for internal company knowledge. Like, has anyone at—we've got a couple thousand employees all around the world—has anybody here ever built a feature before with Missed Connections? What Glean does is it goes and it checks every Slack message, every Google Doc, every Notion, every whatever, and it surfaces that. And the coolest part about it is it stays true to the permissioning. So if the CEO asks that question, it would access any Slack channel that I have access to or any Google Doc that I have access to. But if a different employee asks that question, it just honors their permissions—the Google folders that they have access to versus the ones that I have access to, etc. So that's really important obviously for internal confidentiality and things like that.
But increasingly, this feature set is basically coming to Claude Code and into Claude Co-work. So if I go ask Claude that question, it's going to be a pretty similar answer to if I ask Glean that question.
How afraid does this make you then to try other solutions and invest time and energy in it if you think it's all going to be...
Well, yeah. I'm very bullish on Anthropic. It has totally changed the company. And a year ago, this was an OpenAI shop, and I think increasingly you would call it an Anthropic shop. Now, it might change. In another year, who knows? Every...
Switching costs are going to be high to change.
No, I think the switching costs are pretty low. And in Gemini, Google's not done, right? Google's not giving up; they're just getting started. Meta is entering the fray. So it's all changing very quickly. SpaceX, which just bought Cursor, is going to rename Cursor Grok, I think. And obviously, they're not done either. So the beneficiary is companies like us, obviously, and startups that are using all these tools to great effect.
I want to jump back to Missed Connections, because that is on the development side, and you mentioned Velocity on the development side. But what about everybody else?
Marketing, legal, HR, for example, they are also fully adopting AI tooling in a pretty spectacular way. We just completed an employee engagement survey, for example, and the HR folks have done everything through Claude Co-work. So all the action plans and follow-ups and data analysis and data visualization, the types of things that used to be done in the Google Suite or in the Microsoft Suite—Excel, Outlook, Google Docs—or...
They pay somebody else. They pay...
Yeah, exactly. They're paying some software vendor or consulting firm to do that type of analysis. They're now completely doing it with Claude. We've been doing this forward-deployed engineer model with HR, legal, and marketing. By which I mean, let's take legal. Because everybody has all these tools, you have lawyers and paralegals and other people building different legal bots or contract analysis skills or artifacts inside of Claude, etc. And in addition to that, we now have put one or two engineers to literally sit with the lawyers...
...within our own groups, and they're just building stuff.
Marketing is, I think, maybe the even better example, right? We spend $600 million a year on marketing globally. We had a forward-deployed engineer build an AI tool that uses Claude to automatically generate different paid marketing creative, connects to Meta or Instagram's ad product, does the analysis, changes the copy, and changes the digital assets based on what's performing and not performing, all using Claude. That's something that employees used to do manually, and it's allowing us to create many dozens of more digital assets and do much more effective advertising because it's all AI-generated. Instead of just having marketers build that, which they are able to do, we actually had engineers do that for the marketers, and we think that has a huge unlock. I've seen that at other companies that I'm involved in, too, other startups that are doing this really effectively.
What is the harness that you're building around? What are the guardrails? Because we tried that and it went horribly wrong because we were like you in the early stage, just, "Ah, who cares? What's the worst that can happen?"
What we did during those first couple months, we just kind of let everyone use different things, and then we saw where costs settled. What we then did was we agreed upon token limits. We classified every employee as either an engineering or a technical employee, or a non-technical employee. And we gave different caps. We're still sort of tinkering with them, but I think it's around $1,000 a month for technical employees and around $200 a month for non-technical employees. And we selected those numbers based on the 80th percentile of usage. So the idea was to make it so four out of five employees would never hit the cap, would never even be aware that there was a cap because they're using less than that.
And then we keep it intentionally as a soft cap. So if you hit it—actually, this happened with my CMO, who's a direct report to me. Saturday morning, I got a Workday notification that says that Stephanie hit her AI token limit. Would I approve an overage for it as her manager? I shot her a quick note and was like, "Cool. I'm kind of surprised that you're using AI so much as the CMO. What are you doing?" And she's like, "Yeah, I'm doing this thing for this upcoming whatever, I'm doing this analysis, blah blah blah." I was like, "Okay, great." So then I went into Workday, I approved it. I gave her another hundred bucks for the month. She's non-technical, so her $200 got up to $300, and then she's back at work. That's how the approval sort of happens.
There are also actual governors on the model selection, and there are different harnesses or routers, and we're testing different ones. Stripe just bought OpenRouter for six or seven billion. I think they're a category leader. We're testing OpenRouter, and I guess Cursor has one kind of built into Cursor also. We haven't selected that router thing which helps the engineer choose which model to use. At some point, we need to select one of those and stick with it.
Go back to your marketing example, though. Do you have a human reviewing the ads as they come out still? Is it just, "Okay, I approve, I approve"? And is it recursive? Is it self-learning at this point as to what sounds good for brand and what doesn't?
