A Cheeky Pint with Intercom Cofounder Des Traynor | CEO OS

tldr

Des Traynor walks through Intercom's transformation from live chat widget to AI-native customer service platform. The company made a hard bet on AI the weekend after ChatGPT launched — spinning up a 50-person AI lab and shipping Fin by March 2023. Fin now resolves 1 million customer conversations per week at a 65% resolution rate. The conversation covers why selling AI is closer to selling infrastructure than traditional SaaS, how Intercom thinks about usage-based pricing, and Des's "four horsemen" framework for evaluating whether an AI company is real.

Key Takeaways

Intercom bet on AI the day after ChatGPT. Decision made Sunday, AI lab assembled Monday. The internal mandate was: "Fin just has to win at all costs."
Fin is not a chatbot wrapper. It has 27+ components — retrieval, reranking, summarization, escalation detection, context layers. Most competitors don't go this deep.
Selling AI is infrastructure sales, not SaaS sales. You're not demoing a UI. You're competing on resolution rate, response time, and reliability. "Everyone's chatbots look the same."
The "torture test" works. Prospects who want to build in-house get handed 100s of real customer scenarios. Most come back and buy.
Usage-based pricing for AI is inevitable — and complex. Intercom charges $0.99/resolution on top of Seat Pricing, now on Stripe. Migration off Zuora was worth it.
The "four horsemen" of real AI companies: revenue backed by usage, mission-critical use case, deep/differentiated AI, clear path to positive unit margins.
Fin standalone was a strategic reversal. Instead of requiring customers to move off Zendesk/Salesforce, Intercom now runs Fin on top of them. Prioritizing AI adoption over help desk lock-in.
AI pricing has to follow value, not cost. The conversation you resolve at $0.99 may have prevented a churned customer worth $50K. Pricing anchored to cost misses this entirely.

Detailed Breakdown

[00:00] Intro

John Collison sets up the conversation noting Intercom's radical transformation and Des's reputation for sharp product thinking.

[02:58] Reinventing Intercom

Intercom has reinvented itself twice: (1) Live chat on websites → (2) Full customer service platform → (3) AI-first operation. Each reinvention required betting against the existing revenue model. Des describes the cultural difficulty of asking a team to cannibalize what they built. The key was internal clarity: this isn't optional, it's existential.

[06:31] Fin

Fin launched March 2023 — 3-4 months after the internal pivot decision. The speed required shutting down other work. Early Fin had a 25% resolution rate. Current rate: 65%, with the ambition of reaching 90%+ for structured query types.

The architectural depth is what separates Fin: 27+ subsystems including:

Knowledge retrieval + reranking
Summarization and abstraction layers
Escalation detection
Multi-turn context management
Tone and persona layers

Most "AI customer service" products are a single LLM call. Fin is not.

[18:06] 1M resolutions per week

6,000+ customers. 40 million conversations handled. $50M ARR at $0.99/resolution. Des notes the math: at 1M resolutions/week × $0.99, that's roughly $51M annualized just from per-resolution fees.

The benchmark matters because it grounds the AI claims in economic reality. Not pilots. Not evals. Paying customers, real dollars.

[24:22] Selling AI

Selling AI is fundamentally different from selling SaaS:

Old SaaS: "Look at this UI, here are the features, here's the checklist."
AI: "What's your resolution rate? What's your response latency? What happens at edge cases?"

Buyers have developed learned skepticism from vendor overselling. One procurement leader told Des his inbox was "relentless, absolute nonsense." The only counter is proof, not promises.

The "torture test" strategy: hand them 200+ real customer scenarios. Most in-house build advocates fail this. Many come back within a few months as buyers.

[29:34] Product marketing

Des's view: product marketing is the most undervalued function in B2B. The job isn't to describe features — it's to crystallize the category the product owns, and make the competition look like the wrong answer.

Intercom's $1M guarantee — "find a competitor that beats Fin, we'll pay you $1 million" — isn't about the money. It's a positioning statement. It signals: we're confident enough to put it in writing. Few competitors can or will match it.

[37:18] Listening to users

A caution against over-indexing on user feedback: customers will ask for features that make the existing product more comfortable, not features that make the category better. The danger is optimizing for "less bad" instead of "genuinely different."

Des distinguishes between:

Pain signals (valid, act on them)
Solution suggestions (extract the pain, ignore the prescription)
Competitive feature requests ("our competitor has X") — dangerous because it turns you into a follower

[44:14] Usage-based billing

Intercom migrated from Zuora to Stripe for billing. The combination of seat-based pricing (predictable) + per-resolution pricing (usage-based) required infrastructure Zuora couldn't cleanly support. Stripe handled both models natively.

The migration was painful but correct. The hybrid model is the right structure for AI products where value scales with usage.

[45:14] Advice for startups

On AI-native companies vs. AI-augmented companies — the distinction matters:

AI-native: AI is the product. Value delivery is computation, not human labor. AI-augmented: AI makes existing human workflows faster.

Startups building AI-native products face different unit economics, different sales motions, and different competitive dynamics. Know which one you are before optimizing for the wrong thing.

[52:09] AI pricing

Key insight: pricing AI by cost is wrong. You should price by value delivered.

A support ticket resolved by Fin at $0.99 may represent a customer relationship worth $50K in LTV. The resolution prevented a churn event or escalation. The value is asymmetric relative to the cost.

Des's framework:

Don't price off model cost (it will fluctuate, it's not your IP)
Anchor to outcome value (resolution = retained relationship = revenue)
Be willing to charge more than feels comfortable — buyers will respect it if the proof is there

[01:07:27] Cofounder dynamics

Des and Eoghan McCarthy have been cofounders for 15+ years. Observations:

The early division of labor (product vs. GTM) doesn't hold at scale — both have to be able to do both
The biggest tension point in any cofounder relationship is prioritization disagreements, not personality conflicts
Healthy conflict requires that both sides trust the other's intentions. When that erodes, the working relationship breaks faster than the friendship

[01:11:04] Predictors of company success

Des's view on what separates the companies that become real:

Taste — Does the leadership team actually care about quality in the product, the copy, the UX, the brand? Taste is rare and irreplaceable.
Conviction without brittleness — Can you hold a strong position and update it when evidence demands? Dogmatism kills companies too.
Monetization clarity early — Companies that treat revenue as an afterthought build products that are hard to price. The best products have a clear answer to "why does this cost money?" from day one.
The "four horsemen" for AI specifically: real usage → mission-critical use case → deep/differentiated AI → healthy unit economics

[01:15:56] How AI-native is Intercom?

Des's honest answer: Intercom is "AI-transformed," not AI-native. It was built as a SaaS product and has been rebuilt around AI. The distinction matters because it shapes the technical debt landscape, the sales muscle, and the org design.

Full AI-nativity would mean rearchitecting from first principles assuming AI exists. Most incumbents can't do this — but Intercom has gotten closer than most.

Notable Quotes

"We made the decision on the Sunday and started working on the AI version on the Monday."

"Fin just has to win at all costs."

"People compare AI with this perfect human that doesn't exist."

"It's often not a lack of intelligence; it's better architecture and tailored models."

"Selling AI is closer to selling infrastructure than traditional SaaS."

"A lot of people just don't realize how deep you have to go."

"Most people compare it with password resets — that's 0.4% of actual customer service scenarios."

One Thing to Act On

For SupportWire: The "torture test" is the move. When a prospect says "we're going to build this ourselves" or "we have an AI vendor," hand them 100 real support tickets and ask them to run their solution on all of them. Most can't. That's your sales conversation — not a deck, not a demo. Proof beats pitch every time. Build this test framework now, before you need it.

Tags

#intercom #ai-customer-service #fin #product-strategy #usage-based-pricing #cofounder-dynamics #ai-native #saas-sales #john-collison #stripe #cheeky-pint

Raw Transcript

Auto-captions from YouTube. Folded by default — expand if you need to grep the source or pull an exact quote.

0:00 Do you ever work in a bar? 0:01 >> I've never worked in a bar. I have a 0:02 Guinness C in my house. 0:03 >> Actually, do you want another pint? Then 0:04 you 0:04 >> Yeah, sure. Sure. I was talking to a guy 0:05 who runs procurement and uh he was 0:07 saying like, "You have to understand, 0:08 I'm just getting bullshitted to all 0:09 day." You know, it's just all day 0:11 relentless. Absolute nonsense in my 0:13 face. 0:14 >> People do not have enough empathy for 0:15 the procurement person who just has to 0:17 endure non-stop nonsense. 0:19 >> Absolute garbage. 0:21 >> Chat GPT launched I think on a Thursday. 0:23 I had a call with Fergle, our head of AI 0:24 on a Friday. home made the decision like 0:26 on the Sunday I think it was and we 0:28 started working on like the AI version 0:30 of intercom on the Monday. It was a lot 0:32 easier to invest when being a founder 0:33 was uncool. 0:34 >> Mhm. 0:34 >> I blame genuinely the social network. I 0:37 blame just kind of the entrepreneurial 0:39 lifestyle. I blame like Tik Tok. I blame 0:41 all these things. 0:42 >> So, 0:45 >> you just show me your technique and I 0:47 will uh learn from you because I'm 0:48 pretty sure a point to Guinness in 0:50 America. My fellow Irishman, Dez 0:52 Trainer, is the co-founder of Intercom, 0:54 the customer service giant turned AI 0:55 company. He's also a prolific blogger 0:57 and one of the most respected voices on 0:58 product strategy. 0:59 >> Cheers. 1:03 >> Okay. 1:05 >> You are my the first Irish guest and so 1:08 I actually have a critically important 1:10 question that has been burning uh this 1:12 entire time. Whenever we release these 1:14 episodes, all the YouTube commenters are 1:16 obsessed with people splitting the G. 1:18 This was never a thing for me growing 1:21 up. A thing I never saw in Ireland. It's 1:22 like an invasive species maybe from Tik 1:24 Tok. Is this a a thing? 1:26 >> No. It's not amongst anyone who you 1:29 would respect. It's very it's very much 1:31 a an actual Tik Tok thing. It's a slight 1:33 bit of a tourist thing because of that. 1:34 >> Yeah. 1:35 >> But it basically means like you're 1:36 drinking I think it's a quarter to point 1:37 in your first mouthful and it's like I 1:39 don't know. I don't 1:39 >> Waste of goodness. 1:40 >> Yes. 1:41 >> Yeah. Okay. Um it's a little like when I 1:42 first came to America drinking games. 1:44 It's like back home drinking is a very 1:45 serious activity. Wouldn't make a game 1:47 out of it. Exactly. Yeah. Yeah. Yeah. I 1:49 was very confused by that in America. 1:50 Okay. And then my other stout related 1:52 question is um I saw you apping on like 1:55 Beish Murphy's everything like this. I 1:58 don't know what is your what is your 1:59 view on the stout landscape in 2:01 particular? 2:02 >> I I probably default to Guinness. Uh 2:04 there's been like there's been moments 2:06 when I go Kenny Bish Murphy kind of 2:08 usually it's like whenever you're in 2:09 their respective hometowns but for the 2:11 most part I default to Guinness. And for 2:13 for like one brief day I tried Island's 2:15 Edge. which I don't know if you remember 2:16 when they Hinekin he was Hinekin 2:19 the Guinness killer like and it lost it 2:20 all like 6 months or something like that 2:22 they l like I think literally they were 2:24 giving it away at the end and still 2:25 couldn't get rid of it actually. Yeah. 2:26 >> Have you heard all the Guinness stats 2:27 about you know the Guinness used to be 2:29 be a majority of Ireland's stock market. 2:31 Obviously, the canal system was built 2:33 for Guinness distribution, but um 2:35 sometimes when you tell people that 2:36 Guinness used to be a very significant 2:37 part of Ireland's economy, you know, 2:39 they don't believe you, but the stats 2:41 are really there. 2:41 >> And it's weird like it's it trickles 2:43 into modern day like where I live by 2:45 Castle Knock like the a lot chunks of 2:47 the Phoenix Park like are owned still 2:49 around exactly where the houses used to 2:51 be there and it's and it's like a they 2:53 don't and then maybe they donated it to 2:54 OPW or something like that. But yeah, 2:56 it's very much still kind of carries 2:57 forward. 2:58 >> Yes. Okay. And uh we're not actually 2:59 going to talk about Guinness the whole 3:00 time. We should also talk about 3:01 intercom. Uh and I was thinking in 3:03 preparing for this, I'm very impressed 3:05 by businesses that 3:09 can reinvent 3:11 themselves or maybe even reinvent 3:12 themselves multiple times. You think 3:13 about Netflix. They started with the 3:15 original DVD by mail business and then 3:18 oh my god, the internet's coming and 3:19 they had a few abortive attempts at 3:22 streaming. Like remember the whole 3:23 quickster uh debacle, but then really 3:25 cracked streaming movies. Uh, and so, 3:27 you know, you can watch The Godfather or 3:29 you can watch um whatever movie you want 3:31 on Netflix. And then, of course, as they 3:33 got more squeezed by the rights holders, 3:35 you actually can't go watch The 3:36 Godfather on Netflix anymore. Or you 3:38 type any movie into Netflix and it's not 3:40 there because now Netflix is all first 3:42 party content that they have developed 3:44 themselves. And so, they've reinvented 3:45 the platform once again to be rather 3:47 than watching other people's content, 3:48 watching Netflix content. They've twice 3:50 reinvented the company from DVD by mail 3:52 to streaming third party content to 3:54 streaming first party content. Intercom 3:57 strikes me as a business that has 3:58 similarly been reinvented twice where 4:00 you guys got started with the intercom 4:03 feature of you know you can talk to your 4:04 customer through the website and then 4:06 you guys became a customer service 4:07 company which is actually different for 4:09 reasons we can talk about and now you're 4:11 becoming an AI customer service company 4:13 that's my theory of intercom is that 4:15 actually an accurate theory 4:17 >> yeah it's it's that's true there's a bit 4:19 of like extra context I give it so when 4:22 we started it wasn't really it was like 4:23 about like this is like literally when 4:25 we started Ed and our initial plan was 4:28 like, hey, like you you remember the 4:30 internet cuz like Stripe was in its 4:31 early days back then as well, but like 4:32 there was no tooling to run a a SAS 4:34 business. There was like literally 4:35 nothing like you were using you were 4:37 kind of abusing like PayPal for payments 4:39 and you were using like Mailchimp for 4:40 like talking to your customers and 4:42 stuff, you know. We had this idea of 4:43 like talking to your customers is really 4:44 important, 4:45 >> so someone should work on that, you 4:47 know. And uh and like there was so much 4:48 stuff in the early releases of intercom 4:50 like it was the first live CDP like you 4:52 actually see who's live in your product 4:53 right now and what they're doing and you 4:54 could store data against them. You could 4:56 say show me all my premium customers and 4:57 weirdly like use cases like that still 4:59 exist today like I was in Denmark last 5:00 week and I was like oh I should message 5:01 all my Copenhagen customers and see you 5:03 and like all that stuff is like you're 5:04 still kind of like who else is doing 5:05 this? Yeah. So we start started out like 5:08 with a very very general purpose idea 5:10 which was like let internet businesses 5:12 talk to their customers and then we kind 5:14 of fell in love with this jobs to be 5:16 done methodology and like as one of the 5:18 things you do in that is you look at how 5:19 your product is actually used and then 5:20 you iterate on it from that point of 5:22 view and that kind of led us down this 5:23 path of sales marketing and support and 5:26 then I guess to skip a load of years 5:28 here let's just say 2020 things were 5:30 great 2022 things weren't as great 5:32 >> and uh and then it was like we need to 5:34 >> precise pre of the period Exactly. Let's 5:36 just cut a lot of the reasons why or 5:38 whatever. Some sort of disease as well 5:40 along the way. Um but yeah, so 2022 like 5:42 it was like hey uh you know um own 5:44 returned owner left in 2020 I think. Uh 5:47 returned you know business had been like 5:49 in declining net revenue. He said like 5:50 we need to focus we're going to focus on 5:52 customer service. And then a short time 5:54 later I think maybe like 10 weeks later 5:56 AI happened. Chat GPT launched I think 5:58 on a Thursday. I think we I had a call 6:00 with Fergle our head of AI on a Friday. 6:02 I spoke with all through the weekend and 6:04 like we made the decision on made the 6:06 decision like on the Sunday I think it 6:09 was and we started working on like the 6:10 AI version of intercom on the Monday 6:12 that was like 2022 and I think like were 6:14 it not for that it's hard to say exactly 6:16 where things would have gone but 6:17 certainly that's the reason I'm sitting 6:18 here. Yeah. And obviously people are 6:20 naturally wired to be skeptical. You 6:21 know, when they hear the AI version of 6:22 Intercom, you have, you know, everyone 6:24 and their mother out there saying we're 6:25 an AI company now. But you actually are 6:27 an AI company now. And so describe what 6:29 the AI version of intercom actually 6:31 means. 6:32 >> The biggest thing it means to us is our 6:33 product Finn. So we launched Finn in 6:36 March, I think 2023. 6:38 >> We launched a few little AI feature. We 6:40 were the first people to actually build 6:41 anything on like the the GPT3.5 and then 6:44 we launched Finn in March GP4 launch 6:46 day. And Finn was like basically like 6:49 the first chatbot that worked. It's the 6:50 best way you could think about it. What 6:52 that really meant was we could actually 6:53 have conversations and answer questions. 6:54 And when we launched it, it was doing I 6:55 think 25% resolution rate. And that was 6:57 like crazy numbers. 6:59 >> Today it's like 65%. And today Finn's 7:01 resolving about I think it's over a 7:03 million conversations a week. It's 7:04 handled about 40 million actual 7:06 endto-end customer service scenarios to 7:08 date. It's you know growing like over 7:10 300% year-over-year. It's like you know 7:12 it's we charge a dollar per answer so 7:14 you can work the revenue out or 99 cent 7:16 even. it's becoming like just this you 7:18 know AI sort of growth story inside 7:20 intercom which is already like a sort of 7:22 you know a mature SAS business in 7:24 hundreds of millions of revenue but I 7:25 think when we think about like what does 7:27 it mean to be AI it's like first of all 7:28 what is the future growth of your 7:29 business and the answer is AI and then 7:31 over the last say 6 months we've been 7:33 going hard on being like a kind of 7:34 properly deep AI company we're now at a 7:36 point where we're like you know we're 7:37 using our own models inside FIM we're 7:39 using custom reanker custom retrieval 7:41 summarization etc 7:43 >> and we're doing a lot of this work we 7:44 have like an AI lab of 50 people and We 7:46 really just kind of have gone all in on 7:47 the idea of like, you know, like 7:49 obviously intercom still has a help desk 7:50 product, but like the entire future of 7:52 CS is clearly going to be AI and that's 7:54 what we're all in. 7:54 >> Yes. Yes. Maybe there's the repeated 7:56 pattern in tech where the enthusiasm for 7:59 technologies comes before the technology 8:01 being ready and so people are excited 8:02 about computer games before the wave of 8:04 like good computer games or you know 8:06 people were pitching mobile internet and 8:08 you know you'll buy cinema tickets via 8:09 WAP. It's like J2ME and all that. 8:11 >> Exactly. Yeah. Yeah. 8:12 >> We actually had our version of that like 8:13 we had a product called resolution bot. 8:15 It was originally called Answerbot, but 8:16 I think Zenesk tried to sue us because 8:18 they had a competitive product, but like 8:19 resolution mod was actually a good AI 8:20 product at the time, but it's just AI 8:22 wasn't there, right? And so that was the 8:24 actual reason why we had such a head 8:25 start because we actually already had a 8:26 little AI group ready to go. We'd 8:28 already built a rag engine ready to go, 8:29 so we were able to jump a lot quicker 8:31 than a lot of folks. But yeah, like it's 8:32 I think a lot of these products, you 8:34 write like have kind of two or three 8:35 stabs before they go mainstreaming. 8:36 >> Yeah. And there was this whole 8:37 enthusiasm cycle for bots in 2017, I 8:40 want to say, and the tech just wasn't 8:42 there at all. 8:42 >> It was a horrible experience for 8:44 customers. which was also quite clunky 8:45 to set up for businesses and at some 8:47 point I think everyone looked you know 8:48 there was a genuine question at times 8:50 where it was just like is a web form not 8:51 just better like and I think a lot of 8:53 cases it was 8:54 >> yeah yeah yeah whereas I think now 8:55 people are starting to have the um 8:57 experience of you know the classic thing 8:59 is you know you're talking to a bot it's 9:00 like please will you please just connect 9:01 me to a human whereas now it's like can 9:03 you just connect me to a bot and like 9:05 >> we see that a lot we see like oh hey 9:07 Jenny sorry to bother you can you put 9:08 back on to Finn it was actually doing a 9:09 great job it just thought it wasn't you 9:11 know cuz I asked a few too many 9:12 questions but yeah I think like There's 9:14 a general pattern we're noticing which 9:15 is like a lot of experiences are just 9:17 better digitized cuz partially because 9:19 of like human considerations like one of 9:21 the reasons people go to like the the 9:22 kiosk in McDonald's I suppose cuz as 9:24 opposed to their their actual counter is 9:27 because they don't have to think out 9:28 loud in front of a human they're like oh 9:29 give me a second do I want fries 9:31 >> a lot of the reason why people prefer 9:32 Whimo for some people it's like I just 9:34 don't want to have the conversation I 9:35 don't know what the awkwardness around 9:36 tipping or whatever it might be and I 9:38 just think what um what we see a lot is 9:40 like once Finn answers one question well 9:42 people are like oh this thing's paying 9:44 Let me now that you're doing that, I'm 9:45 going to ask you all of the things I was 9:46 wondering about. 9:47 >> Whereas I think they'd feel probably 9:49 nearly weird unloading all that on one 9:51 per CS rep, you know. 9:52 >> Ah, okay. So, you're seeing a lot of 9:53 induced demand where people use 9:54 interactive customer service, 9:56 >> which is great, which is interesting 9:57 because it turns Finn into not just 9:59 being a kind of a, you know, a cost 10:01 takeout, but also it's like how much 10:03 better would your business be if 10:04 everyone knew how to do everything they 10:05 wanted to do? And like the answer is, 10:07 you know, a lot of times a lot better. 10:08 And it's not just the whole like I don't 10:10 want to burn any burning human. Also 10:11 like people often one of the biggest 10:13 fallacies in AI is people compare it 10:14 with like with this perfect human that 10:16 does not exist like the the driver that 10:18 never crashes or like uh in our case 10:20 it's like well a human artal handcrafted 10:22 answer would and I'm like yeah like 10:24 let's pretend that would be there in 7 10:25 seconds it won't be it'll probably be 18 10:27 minutes it's also not going to be 10:28 perfect it might not be that long you're 10:30 doing those handcrafted answers which 10:31 you're not someone who's like busily 10:33 trying to close the case before they 10:34 move on to the next one. So like I think 10:36 yeah comparing like I don't know we have 10:37 this thing where we expect our AI 10:38 products to be flawless and we're 10:40 totally tolerant of like humans jungle 10:41 pong over only speaking one language and 10:43 only working six hours or whatever you 10:44 know it's just it's a funny contrast. 10:46 >> Yes. Yes. That's interesting. And um 10:48 it's interesting you talk about product 10:50 on boarding here because I think of 10:52 intercom as you guys are very you guys 10:54 have a house view that product 10:55 onboarding should be much better. And I 10:57 remember a lot of the use cases you 10:58 would talk about for the original 10:59 intercom, talk to your customers through 11:01 the website was that you can have 11:03 personalized nurture tracks and like 11:04 it's weird that you drop people into SAS 11:06 products and just expect them to be able 11:07 to use them, right? And you should see 11:08 how people are using the product and 11:10 then give them kind of specific steers 11:11 based on their usage and it sounds like 11:13 you're coming to this vision again which 11:15 is people should have better onboarding 11:17 support. People should be nurtured along 11:19 based on their use case but now kind of 11:21 interactively AI powered. 11:22 >> Yeah. Yeah, I mean we we talk a bit 11:24 about like this idea of of like what is 11:26 ultimately a customer agent going to be 11:28 like that's what FIN will be when it as 11:29 it grows up it'll just become like this 11:32 way in which customer conversations are 11:33 handled and obviously the most direct 11:36 attack here is like customer service but 11:38 >> you know I think every single customer 11:40 touch point can be improved by like by 11:42 basically immediate accurate answers 11:44 available all the time 11:45 >> isn't an obvious limitation like right 11:47 now you require customers to come up 11:49 with a prompt and if you look at why Tik 11:51 Tok is so successful It's like I would 11:53 never prompt for, you know, I want to 11:55 see videos of planes landing low over 11:57 the beach in St. Martin, but like it 11:58 turns out that's what I want to see. 12:00 Yeah. Yeah. And similarly, people 12:03 probably have many more things they need 12:05 than they will actually come up with a 12:06 prompt for. And I think the product 12:08 today is still mostly prompt based like 12:10 reacting to what customers say. 12:11 >> Exactly. A customer has to come along 12:12 and like type things into the box. 12:14 >> I mean, today that's what customer 12:15 service is. It's still like kind of like 12:17 here's my problem and then we'll solve 12:18 it. I think you know for sure there's 12:20 obvious directions this will go as like 12:22 you know hey what does a good customer 12:24 look like and maybe we can honestly 12:25 infer that as well but but certainly 12:27 like people like you should do things 12:29 like this is definitely an 12:30 understandable domain and uh and then I 12:32 just think working out the right level 12:33 of interruptive help you know like you 12:35 don't want to be too naggy or too 12:36 pop-upy like you know it gets kind of 12:38 quite grating but I think if you can get 12:39 the first message right you can sort of 12:41 say hey if you come here you're always 12:42 going to get the thing you should do 12:43 next or like the thing it looks like 12:44 you're stuck on like if someone's on the 12:46 I don't know the renewal page and they 12:47 have an error message. We know they're 12:49 probably going to open the thing and we 12:50 know they're probably going to say 12:51 something that's got, you know, a lot of 12:52 the context already there. So, we can 12:54 work out the right things to say and do. 12:55 I think that's pretty doable. 12:56 >> It feels like you could do a lot around. 12:58 Yeah, you train a model on what uh the 13:02 customer is seeing on that web page at 13:04 that moment in time and use it to feed 13:05 the answer and things like that. 13:06 >> We already do a lot of that already like 13:08 in customer context. So like uh so 13:10 knowing that you knowing that it's John 13:11 and he's on the premium plan and he's on 13:13 the playlist page and there's an error 13:15 on the screen is all useful information 13:16 when it comes to cuz like I think people 13:19 a lot of like the you know YC I could 13:21 build out in the weekend type hacker 13:22 news crowd. I think one of the things 13:24 they often, you know, they're they're 13:26 thinking every customer support query 13:27 begins with like, "Hi there, my name is 13:29 blah and my username is blah." But 13:31 actually like most support conversations 13:32 begin with like this is broken and 13:34 you're like what's broken and like so to 13:36 solve that you need like a fatic reply 13:37 engine. It's just like hey let's chat 13:39 about what's going on here. But like we 13:40 realized quickly like people will kind 13:42 of disengage. So any amount of extra 13:43 context that when you say this is broken 13:45 and if if someone says this is broken 13:47 and there's a big red error box on the 13:48 screen we're like well it's probably 13:49 that thing that they're talking about. A 13:50 lot of people just don't realize how 13:51 deep you have to go to actually do a 13:53 great job. And like we're like we do 13:55 like say if you install fin today you 13:56 get like 65% resolution rate after 30 13:58 days. Like that's shocking but like we 14:00 have had to go really deep to actually 14:02 get to those numbers and it involves 14:03 like all sorts of every single smart 14:05 thing you can think of we've had to do 14:07 and then optimize and then find the 14:08 right model for it and all that. But one 14:09 of them is customer context. And that 14:11 obviously answers a lot of things. 14:12 >> Yes. What are the other smart things? 14:14 >> Well abstraction. So like you you know 14:16 like I guarantee you you've got like no 14:17 pages on your website that say stripes 14:19 works really well for a dentistry 14:20 office, right? You probably don't have 14:22 that in your docs. A very naive uh 14:24 ragbot will be basically like well 14:26 doesn't say dentist and we're told not 14:27 to hallucinate. So no we don't do 14:28 dentists. Sorry. 14:30 >> And like the abstraction is like you 14:31 know in that case is well what is a 14:33 dentist? It's a type of business. Does 14:34 Stripe work for businesses? Can dentist 14:36 can dentists be internet businesses? 14:37 Well we say we're great for internet 14:39 business you know. So you're kind of 14:40 working out what's the best 14:41 >> what's the best sort of risk risk 14:43 tolerant way to make grounded inferences 14:45 without going over the cliff. You know 14:47 that's one of like 27 different sort of 14:49 components of FIN. Then you've got 14:51 obviously your rag and then you've got 14:52 like you're like hey is this is an 14:54 escalation appropriate at this time like 14:55 hey have they have they threatened 14:57 something or if they you know like you 14:58 every single type of problem kind of 15:00 ends up you have to walk through it all 15:02 to actually recreate customer service. I 15:03 think a lot of times people will compare 15:05 it with like how do I reset my password? 15:07 Huh it found it. And you're like, 15:08 "Right, cool." That's like point4% of 15:11 the scenarios you deal with and you're 15:12 in customer service. Yeah. 15:13 >> There's like the there's a pattern. Did 15:14 you ever see the movie Armageddon where 15:16 like um it's like Bruce Willis and Ben 15:18 Affleck or whatever. But the gist of it 15:19 is they train a load of I think it's 15:20 like oil drillers to become astronauts. 15:22 And the comedy the joke at the time that 15:24 Ben Affleck always says he got drunk and 15:25 he recorded the voice over for the DVD 15:27 and he was like I always said why didn't 15:29 we just train the astronauts to drill 15:30 oil? Surely that's an easier problem. I 15:32 think the thing that we're realizing 15:34 with the AI movement is some version of 15:36 what's going to happen sooner. Will AI 15:37 people learn how to do CS or will CS 15:39 people learn how to do AI? Thankfully, 15:41 as I said, we kind of started off with 15:43 CS and AI in our DNA. 15:44 >> When you say people, you mean companies 15:45 in this case, like OpenAI get better at 15:48 customer service faster than Zenesk gets 15:49 better at. 15:50 >> Exactly. Exactly. That I think we just 15:52 we were lucky enough that we kind of had 15:53 already backed both horses somewhere 15:54 along the way. So, yeah. 15:55 >> One thing I find interesting about what 15:56 you do is every company is thinking 15:59 about AI right now. You know, every 16:00 company had a board meeting in 2023 16:02 where the board was like, can we do a 16:03 special deep dive on AI because it just 16:05 feels like it's a lot happening and we 16:07 need to be making sure we're, you know, 16:08 on the leading edge of AI. And uh then 16:10 every company was like, oh, we're 16:11 actually doing a lot in AI. For example, 16:13 we've seen great automation wins in 16:15 customer service. And so it's kind of 16:16 like, you know, the joke about the bike 16:17 shed, you know, versus the nuclear power 16:19 plant where everyone has opinions on how 16:20 to build a bike shed. H similarly, kind 16:22 of everyone has opinions on uh how to do 16:25 AI customer service. And so I'm curious 16:28 um how you sell given that I'm guessing 16:31 a lot of your customers think, "Oh, we 16:33 know how to do that. It's not that hard. 16:34 We hooked it up to a model and you know, 16:36 we're actually very smart on this topic 16:38 already." How you sell in that 16:39 environment where everyone has opinions. 16:40 >> Yeah. There's like um I've never seen 16:43 the like build versus buy thing play out 16:44 more often than we do today. Especially 16:46 with like certain like a lot of 16:48 customers are like, you know meme on 16:50 Reddit of I'm not like other girls or 16:51 guys, whatever. Like there's a lot of 16:52 that where it's like go you would never 16:53 possibly understand a B2C shopping 16:56 company and you're like really I've 16:57 never heard of such a thing. Sometimes 16:59 honestly we just be like hey look you 17:02 know godspeed you go and start building 17:04 this PS here's a torture test when you 17:06 think you've got something good run 17:07 these 100 questions to us let us know. 17:09 Often times that's where they're like 17:11 yeah okay we think we need to buy your 17:12 product. But like I think there is a um 17:16 everyone has this idea of like uh in a 17:18 move to AI what can we definitely do and 17:20 we can definitely answer questions like 17:22 how do I reset my password and again 17:23 this is the back to the whole that's 17:24 such a small amount what they can't do 17:26 is actually have conversations and all 17:28 that sort of stuff or like we're like 17:30 what is your opinion on the president 17:31 and how they're performing and like a 17:32 lot of times well you don't want you 17:34 know you don't want anyone to answer 17:35 that question on behalf of your company 17:37 but uh I think a lot of times people 17:40 like they dip their toes it's almost 17:41 like they fired a tracer bullet They're 17:43 like, "Yep, this seems like we're making 17:44 great progress." And every AI product 17:46 has this problem where you like make 17:48 epic progress in the first two weeks, 17:49 and then you hit this wall, this 17:51 plateau, 17:52 >> and then like, you know, 2 years later, 17:53 you're telling people, "Oh, Apple 17:54 intelligence is coming in 26 or 17:56 whatever, right?" So like in this case, 17:57 a lot of people start the project, feel 17:59 like they're definitely don't need to 18:00 buy 18:00 >> Finn. 18:01 >> We just help them understand the 18:03 difference between a good bot and a bad 18:04 bot, and then they come back and they 18:06 buy Finn. 18:07 >> So where is the Finn business these 18:09 days? I'm curious both just how it's 18:11 performing on revenue metrics and then 18:13 are you selling it to existing intercom 18:15 customers? Are you selling it to new 18:17 accounts? Just how does the whole thing 18:18 work? 18:18 >> Uh we're about 6,000 customers and 18:21 growing quickly. Finn does about a 18:23 million resolutions a week. Uh we're 18:25 charging a dollar per resolution. So you 18:26 can do 18:27 >> so 50 million revenue run 18:28 >> give or take. When we launched initially 18:30 we sold uh to just to our own customer 18:33 base 18:33 >> and then as we kind of progressed we 18:36 realized hang on moving help desk is a 18:38 nightmare. like you've probably done it 18:40 once or twice, right? If we done it's 18:41 big one, right? It's a whole ordeal and 18:45 Finn is brilliant. So, we're like loads 18:46 of people want this product but can't 18:48 buy it. So, we made the decision to 18:49 launch like what we internally call Finn 18:51 standalone or Finn for platforms. So, 18:53 now you can use Finn on top of Zenesk or 18:55 HubSpot or Salesforce or any of those as 18:58 well. So, basically Finn is available to 18:59 everyone and that's a relatively new 19:01 muscle that we've been growing but it's 19:02 actually, you know, that's kind of where 19:04 we see a lot of the future growth. And 19:05 so do you is Finn like um uh you can 19:11 connect your uh iPod to Windows for uh 19:14 you know iTunes for Windows but we hope 19:16 that one day you buy a Mac and it's part 19:18 of the whole digital hub strategy or 19:20 we're actually now all in on Fin the 19:22 engine and whatever customer service 19:24 platform you use is actually not a topic 19:26 of huge interest to us. 19:27 >> This is such a core question that we 19:28 kick back and forth quite a lot. Um 19:30 >> this is the offsite debate that is 19:31 currently 19:32 >> genuinely like uh at least it's 19:33 certainly one of them. The way like we 19:35 think about it first and foremost is 19:36 like the future is AI. So like Finn just 19:38 has to win kind of like at all costs 19:40 including our help desk. Weirdly our 19:42 customers are like they turn Finn on. 19:44 They're like damn this thing's good. Hey 19:46 now they're like 65% of our support 19:48 volume. Maybe we don't need 19:50 >> XY Z competitor and maybe we can go all 19:53 in on your on your help desk too. And 19:54 we're like okay cool. we that wasn't our 19:57 game plan, but but we're happy to help 19:58 if you know what I mean. Right. I think 19:59 like the actual the battleground we care 20:01 most about genuinely. It has to be the 20:03 AI agent. Um that's the one where like 20:05 we care about most. But it does produce 20:07 a lot of like you know demand for the 20:09 actual help desk product too. 20:12 >> Oh my god. 20:14 >> Oh dear. 20:15 >> Okay. So you've actually played darts. 20:17 >> Well I've been around at dartboard. I'm 20:19 curious what your AI stack looks like 20:22 where concretely, you know, what are the 20:25 models or collection of models and 20:26 prompts and everything that you are 20:28 using in production. How do you handle 20:29 model upgrades given that you know the 20:32 behavior is changing so much underneath 20:33 the hood? How deep do you go in terms of 20:36 developing the stack yourself? Can you 20:37 talk about the stack? 20:38 >> First thing I say is obviously Finn 20:39 isn't like one thing. It's like 27 20:41 different things or whatever, right? So 20:43 every one of those is like whether it's 20:44 the summarizer or whether it's like the 20:46 reanker or the retrieval engine or any 20:48 of these or the direct answer which is 20:50 where we actually go and formulate the 20:51 answer. Every one of those is paired 20:52 with the the like the fastest, cheapest, 20:55 lightest, most accurate uh LLM that can 20:57 actually do the job reliably like very 20:58 very high reliability. So like that 21:00 means like there's no one particular 21:02 model. Um so our primary partner will be 21:06 anthropic for uh for clouds on it. We've 21:08 architected it such that like we can 21:10 plug and play various different pieces 21:12 um whenever a new model comes out or 21:13 honestly a new idea for a new 21:14 architecture is in hey like we recently 21:16 launched the ability for fin to do 21:18 complex queries which would be like say 21:19 go and issue a refund and update the 21:21 name on the utility bill or something 21:22 like that. uh whenever like we have to 21:24 change the architecture, we have this 21:26 kind of like arduous torture test of 21:28 like thousands or at least a thousand uh 21:31 CS scenarios where like we have like 21:33 here's the question, here's the context 21:35 we're provided, here's what like the 21:37 current FIN answer to this question is, 21:39 here's the best available human answer 21:40 that we know of and then basically 21:42 here's what this new version uh would 21:43 offer us. And then so like whenever we 21:46 say like go GPT5 comes out or something 21:48 like that, we're like you know the 21:50 reason we're not just a lot of our 21:51 competitors are kind of quickly oh we 21:53 never run a GP5 and I'm like I take a 21:55 beat on that one you you know like you 21:57 shouldn't assume everything's going to 21:58 be great for your use case always right 22:00 >> and uh so we ultimately we have to run 22:02 it through this pretty like expensive 22:03 test to work out where the edges are if 22:06 it's scoring higher resolution right we 22:07 need to understand why because it could 22:08 be just that it's like trying more stuff 22:10 but that could also the shadow side of 22:12 that could be like excess hallucinations 22:13 or whatever So, whenever a model upgrade 22:15 comes, we have to trigger this whole 22:16 thing. But when we launched F, it was 22:18 like 25% resolution. Today, it's like 22:20 65. Uh, we've been increasing it roughly 22:22 a percentage point a month, give or 22:24 take, but like very little of that is 22:25 actually because of the upgrades or the 22:27 bumps from the models. 22:29 >> Genuinely, I I actually think and I and 22:30 I say this with a lot of respect and 22:32 love for the CS craft. I actually think 22:33 we've had enough intelligence for CS for 22:35 quite a while. In fact, we published uh 22:37 some material on this on our research 22:39 blog repost recently. Like when you look 22:40 at like you know you know people are 22:42 saying things like oh like the latest 22:43 whenever you know Grock can compete at 22:45 mathemat mathematical olympiad like 22:47 level seven or whatever 22:48 >> we're like right I think you can 22:50 probably do most CS you know like h so 22:53 it's not it's often not a lack of 22:54 intelligence uh is the reason why we're 22:56 not 100% 22:57 >> it's cuz they're too distracted by any 22:59 >> exactly the a lot of the wins come from 23:02 like honestly better architecture uh 23:04 better like tailored models or like 23:06 changing in the UI can change exactly 23:08 how things work and then sometimes times 23:10 you will get an occasional bump here and 23:11 error from a model swap. 23:12 >> It strikes me that a lot of the how you 23:15 include the amount the account context 23:18 and the amount of account context you 23:20 include is a big part of the secret 23:22 sauce 23:23 >> perhaps but there like it kind of it 23:24 varies customer to customer. It really 23:26 is one of these areas where like it's a 23:28 thousand lead bullets. It's not like a 23:29 single silver one. Like it's not if you 23:30 look at our resolutionary craft there's 23:32 no pop like 23:33 >> give or take one or two little tweaks. 23:36 like it's mostly just hey we ground out 23:38 through like you know optimizing this 23:40 prompt and changing this handover we 23:42 ground out another 7% and like you know 23:44 and you see the AI team celebrate that 23:46 on the balcony on the Friday being like 23:47 yay 7 up or whatever it's hard to work 23:50 out exactly uh what bits and then then 23:52 there's obviously multiplicative 23:53 benefits like you might have a win over 23:54 here to cost you something over here as 23:55 well with 23:56 >> yeah how deep down the stack will you go 23:58 like what's intercom's version of Apple 23:59 silicon 24:00 >> that I I don't know for sure I mean like 24:01 we're going to chase any edge we can get 24:04 um right now I think uh custom models is 24:06 definitely where we're going and that's 24:07 like a large investment from the AI 24:09 group which is like the most contested 24:10 resource that we have. Every bit of work 24:12 they're doing is finding a new edge in 24:14 resolution rate or resolution quality. 24:16 So right now it's paying out pretty 24:18 well. So I think we're going to kind of 24:20 place all our chips there until 24:21 something changes. 24:22 >> What does selling AI look like? 24:24 >> It's quite difficult. It's difficult in 24:25 marketing and selling cuz I think um 24:27 like 24:27 >> cuz it's so crowded and noisy. Well, 24:28 there's that, but it's also like it used 24:31 to be the case and for sure intercom 24:33 used to be one of these companies where 24:34 we are a product looked the nicest. So, 24:36 all we had to do was the age-old, you 24:39 know, blah blah blah reinvented and in a 24:41 big sexy screenshot. And uh and you can 24:43 still get away with that in certain 24:44 domains. Linear can get away with that 24:45 cuz their product is the sexiest, right? 24:47 I think with AI, everyone's chat bots 24:48 look the same. Everyone's kind of copied 24:50 our messenger. Everyone, you know, 24:51 everyone's kind of like roughly like 24:53 converging on the certain UI paradigm. 24:55 And so you have to ask like then when we 24:57 say like you know we are the best AI 24:59 agent like what do you think all the 25:00 rest of them are saying? We're the 25:01 worst. Like no of course. So they're all 25:03 saying this and then and then everyone 25:04 has the same screenshots cuz it's like 25:06 look what we do inside the chat window. 25:07 So you're like all right how do you 25:08 actually out market and then how do you 25:09 out sell? And one of the reasons we 25:11 launched like the Fing guarantee, this 25:12 idea that we'll we'll like we'll pay you 25:13 a million dollars if if you find 25:15 somebody who outperforms us is cuz like 25:16 we're trying to stress to the market 25:17 this idea that um that like we actually 25:20 believe in our product to a ludicrous 25:22 degree such that you should engage us on 25:24 any on any sort of bake off you're doing 25:26 >> but I think from a marketing perspective 25:28 it's really hard to stand out. So all 25:29 you can really do is like rely on on 25:31 like backing up your claims as hard as 25:33 you can and obviously customer 25:34 testimonials. selling is harder cuz I 25:37 think again in the olden days like 25:38 selling SAS in the olden days being like 25:40 pre2022 um 25:42 >> it was kind of like our UI is nice 25:44 theirs is ugly here's a feature grid 25:46 checkbox we've got 24 checks they've got 25:48 17 these seven matter we're in you know 25:50 like and that was like that like 25:52 obviously I'm skipping over several 25:53 steps of course sales and would kill me 25:55 but you get the the the basic idea right 25:57 and I think uh now selling out is closer 26:00 to selling like infrastructure in a 26:01 sense it's more like 26:03 >> our cloud is better than their cloud and 26:04 our performance criteria are better. 26:06 It's like at times it might feel like 26:08 you know Intel AMD or at times it's like 26:10 you know it's our response times versus 26:11 theirs or whatever but like you're 26:13 ultimately coming into it like with a 26:14 battle of metrics and stuff like our 26:15 resolution rate and our seesat versus 26:17 theirs. 26:18 >> But then people say well why and then 26:19 you have to then explain what's actually 26:22 happening beneath the surface a little 26:23 bit so that people can actually get a 26:24 bit of conviction other than just like 26:26 trust us or like please just go and try 26:28 our product cuz like it's not that easy 26:29 to try. You have to still have to like 26:31 turn a lot of keys and open a lot of 26:33 APIs and stuff. Um, so like the 26:36 challenge genuinely becomes like how do 26:38 you have a sales team that's actually 26:39 able to like speak with a good degree of 26:41 familiarity about like AI? It's funny 26:43 you mentioned this. We have this 26:45 specific problem at Stripe which is 26:46 invariably when people switch to Stripe 26:49 from a legacy processor, they see a 26:51 significant revenue uplift. And you 26:53 think like businesses are in the 26:55 business of uh, you know, finding ways 26:57 to get more revenue, you think they'd be 26:58 really interested in this. And we have 27:01 this thing that sounds shockingly good, 27:03 which is if you just move over to 27:04 Stripe, you start immediately getting 27:07 more revenue. And it basically comes 27:09 from two places. One is conversion on 27:13 the actual point of payment that if your 27:15 mobile app or if your web flow is kind 27:17 of janky or doesn't offer the customer's 27:18 preferred payment method or something, 27:20 they will abandon. And if you just look 27:22 at the abandonment rates, you know, if 27:24 you're seeing a kind of only 85% 27:26 conversion rate on that form, then 27:28 obviously getting it up to 90%, that's, 27:29 you know, that's a huge deal. And those 27:31 would be very high numbers. Most 27:32 businesses would not see anything close 27:33 to a 90% conversion rate on that form. 27:35 And so there's huge improvement possible 27:36 there to make the customer kind of 27:38 checkout experience as smooth as 27:40 possible. And obviously things like link 27:42 then where you're not asking people to 27:44 reenter their payment details. That 27:45 delivers a big offlift. The second one 27:47 is even crazier is after people enter 27:50 their credit card details frequently 27:53 charges are denied kind of spuriously 27:55 and so your bank thinks that it's 27:57 fraudulent because you know they haven't 27:59 heard of this merchant or whatever and 28:00 so they'll deny it h or they'll think 28:02 it's fraud or whatever like that and we 28:05 through many many years of optimization 28:08 uh have gotten good at ensuring that if 28:10 it is a valid transaction that that is 28:12 not. So what all that adds up to is that 28:15 we can make the claim and we've seen it 28:16 play out again and again. You know, we 28:17 just saw Herz switch for all their 28:18 e-commerce uh payments to Stripe that 28:20 when people move to Stripe, they see a 28:22 significant uplift in uh in revenue. 28:24 That's surprisingly hard to sell because 28:26 everyone is out there saying, you know, 28:27 we are the thing that gives you more 28:28 revenue. And we've had the exact same 28:30 thing where despite you can have all the 28:31 numbers and all the case studies in the 28:33 world, it's just 28:34 >> it's hard to sell because it's 28:36 undifferiated as a message. I I remember 28:38 even when we switched back to Stripe uh 28:40 either you or Patrick was saying like oh 28:41 well don't forget to do link and I was 28:43 like really I was like is this really a 28:45 thing that like you know people have 28:46 what some business has forgotten it 28:47 credit card or something like that and 28:49 you're going to be able to renew it 28:50 right it feels 28:50 >> unlikely it just feels implausible but 28:52 like at the same time the the data is 28:54 not like not the really debatable I 28:56 think people like to be able to explain 28:58 it to themselves and like not like you 29:00 know I think you wouldn't I was talking 29:02 to a guy who runs procurement and uh he 29:04 was saying like you have to understand 29:05 I'm just getting bullshitted to all Hey, 29:07 you know, it's just all day relentless 29:09 absolute nonsense in my face. So, like 29:11 if you think that like you're like, 29:13 "Ooh, 65% thing is going to stick." It's 29:15 not. It's just I take I divide it by 10 29:17 at this stage, you know, and you tell 29:18 me, "Oh, you're going to save me 2 29:19 million in CS salaries or whatever." I'm 29:21 like, "Yep, maybe in 3 years time we'll 29:23 see 200 grand, you know, like that's the 29:25 kind of the default posture for a lot of 29:27 these people." I think it is just it's 29:28 like they've developed quite an adverse 29:30 reaction to like marketing. People do 29:32 not have enough empathy for the 29:33 procurement person who just has to 29:34 endure non-stop nonsense. 29:36 >> Absolute garbage. 29:39 >> That's funny. I mean, this kind of gets 29:40 to a topic you and I have discussed a 29:42 lot, which is product marketing. How do 29:44 you effectively product market in a 29:47 world of everyone making claims? Like 29:49 one is the guarantee that you your guys 29:51 million dollar guarantee. Has that 29:52 worked? 29:53 >> It's certainly it's worked from a point 29:54 of view of like I I don't actually know 29:55 how many people are are in the program 29:57 right now, but I could say what has 29:58 worked is like 29:59 >> it's landed the message of we stand 30:00 behind the program. being able to say 30:01 like here's the reason why you can buy 30:03 you can buy us. I think that's a strong 30:05 message. The other I mean obviously like 30:07 being able to point to real customers 30:08 with real results sort of like and you 30:10 can say hey go talk to Natalie and newly 30:13 or who you know whatever company you 30:15 want go talk to that person and ask them 30:17 because like that's their name that's 30:18 their job title they work there. They're 30:19 saying this you know with with 6,000 30:21 customers it's kind of it gets it gets 30:23 more believable as the numbers kind of 30:24 tick up. But I so like I guess either 30:27 you know you can make crazy guarantee 30:29 claims, you can like just point to a lot 30:31 of successful customers. For us, it 30:33 might be different depending on your 30:34 domain, but for us like you know it's 30:35 not like we can show you U you can show 30:37 you like hey here's a beautiful backend 30:38 product and here's fancy reporting and 30:40 all that but that doesn't speak to the 30:42 the core thing someone's buying when 30:43 they're buying AI off you is to some 30:44 degree a replacement of work that they 30:46 had to do. And the two things they care 30:47 about are how much work are you going to 30:49 do for me and how well are you going to 30:50 do that work and you basically need to 30:52 product market both of those things. And 30:54 it's very easy to say, "We're going to 30:56 do all the work and we're doing really 30:57 well." Uh, so you have to actually 30:59 really help them understand how to 31:00 appraise the scenario. Like sometimes we 31:02 we put time into actually helping people 31:04 um helping people identify when they're 31:06 being lied to in a sense. So you you 31:08 we'll say like, "Hey, try this type of 31:10 question or ask them about this, you 31:12 know, help helping you almost kind of 31:13 teach them to be much more conscientious 31:15 buyers cuz we know the more informed 31:17 buyer like that suits us. It doesn't 31:19 suit people who are just kind of like 31:20 jazz handheling their way to an AI 31:21 product." But like, yeah, it's it's a 31:23 difficult one. It ultimately like 5 31:25 years ago, it would have been like, 31:26 well, the trick, John, is gifts. You 31:27 know, have you ever considered using 31:29 movies on your homepage, you know, 31:30 that's really engaging. None of that 31:31 works anymore, right? Cuz I just think 31:33 like what you're selling is basically 31:34 it's like an iceberg. Like you're saying 31:36 this little bit of like upfront UI of 31:38 like here's what actually happens for 31:39 your customers. Doesn't that look nice? 31:40 >> And you're selling this like gargantuan 31:42 pile of work beneath the surface that is 31:44 like, hey, all of the human toil goes 31:46 away if you make the switch. 31:47 >> Yeah, I can see that. What are your 31:49 other pet peeves when it comes to 31:50 product marketing? Actually, do you want 31:51 another pint? Then you have 31:52 >> Yeah, sure. Sure. Sure. 31:58 >> See, again, this one, did I not let it 31:59 settle for long enough? And then you 32:00 have the small head. Like, is the set 32:02 >> I think everyone's going to work out 32:03 perfectly. 32:04 >> Is the settling time loadbearing? Okay. 32:06 So, what are your pet peeves when it 32:07 comes to product marketing? 32:08 >> I think the thing that still kills me 32:10 that's still very common is marketers 32:12 that love marketing. So, like you rather 32:14 than actually saying anything useful or 32:15 specific, you'll get like, you know, 32:16 forget everything you know about email. 32:18 You're like, "Okay, what am I buying?" 32:20 Or like, you know, transformation 32:22 reinvented. And you're like, "Cool. 32:23 Sounds like I'm going to reinvent some 32:25 Transformers like but like what's 32:26 actually happening here?" I think there 32:27 is a general still type a type of thing 32:29 where my screen saver on my my laptop is 32:32 like literally a typewriter where 32:33 somebody said, "What are you actually 32:34 trying to say?" And I keep that there as 32:35 a reminder of like just 32:37 >> nine times out of 10, the best marketing 32:38 comes from just writing the thing you 32:40 want to say. Cuz like I do a docs, do 32:41 you ever get into a Google document like 32:42 our goal that is that by reading this 32:44 document the reader will know the 32:46 following? I'm like cool. Can we just 32:47 say that instead of think why is there 32:49 2,000 more words here? Yeah. 32:50 >> Um so like I think I I guess like 32:53 speaking in a way that sounds like great 32:55 to marketers uh is like probably the 32:58 thing that kills me most because they 33:00 don't marketers especially in the AI era 33:01 they don't really necessarily understand 33:03 the depth of what's actually happening 33:04 with the AI or whatever. There's a funny 33:06 stat that Auggle used to quote which 33:07 would say something like of all the 33:10 winners of the can awards every year 33:12 something like twothirds of them would 33:13 lose their contract that year because 33:14 the thing they won the award for was not 33:16 actually effective in market at all and 33:18 I think there's such a a repetitive 33:20 pattern there you you like where a lot 33:22 of people like they will look at say a 33:25 stripe or a linear and they'll be like 33:26 all right we should just do that and you 33:28 don't really like I think things that 33:29 people don't get is like everything 33:31 means something and this is like where 33:32 our CEO is like so differentiated is 33:34 just every single decision we we we pick 33:37 here whether what what photo what icon 33:38 what type face whatever it all sends a 33:41 message are we conventional or not are 33:43 we futuristic or not and I think 33:44 whenever I see folks just uh like copy 33:47 paste somebody else's branding in even 33:49 even in like sort of like oh we'll we'll 33:50 we'll change your homework along the way 33:52 I think what they're really doing is 33:53 saying we don't really understand what 33:55 we're doing here and and like that's why 33:57 a classic of this is you know whenever 33:58 an incubator spits out a new batch of 34:00 startups and they all basically have 34:01 right hand side screenshot left hand 34:03 side three bullets sign up button 34:04 whatever you're kind of like, "Okay, 34:05 cool." But have you actually thought 34:06 about what you're trying to say to the 34:07 world? There's also a thing for startups 34:09 where they probably shouldn't look at 34:10 what established companies are doing 34:12 because like Stripe for so long we clung 34:14 to uh making sure that we had code on 34:16 the homepage and people like you know if 34:17 you want to accept credit card payments 34:19 for your website like we're the place to 34:20 come and at a certain point most of the 34:23 relevant people coming to your page 34:25 actually know that you could do that and 34:26 you can experiment a little bit more and 34:28 you know uh Salesforce doesn't have to 34:30 hit CRM so hard on the homepage because 34:31 after 20 years they've earned the right 34:33 to talk about Einstein a bit 34:34 >> often times like a startup has a great 34:36 idea for a great product and They pitch 34:37 it and then like 6 months later they 34:39 work on some new feature and in their 34:41 heads the new feature is the big thing 34:42 that they're so impressed with not 34:44 realizing that 99.49s of the world have 34:46 not even heard about the original thing 34:47 yet but there they go destroying their 34:49 original pitch by being like and now 34:50 we've got blah blah. I'm like dude no 34:52 one's even heard of the original thing 34:53 yet and here you're pitching some 34:54 nuanced take on some extra feature. 34:56 >> Yeah. Yeah. Yeah. Speaking of David 34:57 Oggov you've read on advertising I 34:59 presume. I uh what is it? It was just 35:01 like that's such a beautiful book where 35:03 all the marketing copy in it is so good 35:05 like the the Rolls-Royce ad of you know 35:07 the only sound you'll hear at 60 mph is 35:08 the ticking of the clock. Um but somehow 35:11 that yeah it should be mandatory reading 35:13 for all product marketers. 35:14 >> Absolutely. 35:15 >> What kind of person succeeds in product 35:17 at intercom? 35:18 >> When we set up intercom originally like 35:21 we were kind of like building one thing 35:23 once we forked out to building many 35:25 different areas like sales, marketing 35:26 and support. I think we gave a lot of 35:28 freedom to product leaders and sort of 35:30 say you own the sales product or the 35:31 marketing product and I think the folks 35:33 who uh who succeed there are like they 35:36 have to have like decent taste and I 35:39 don't mean that in some lofty abstract 35:40 way but I mean they have to like you use 35:42 good software identify good software 35:43 ultimately know how to like pick one out 35:45 of the bunch in a sense right like a 35:47 very common interview question I ask 35:48 people is like what apps are on your 35:49 phone what's your favorite app and the 35:50 amount of times someone's like oh you 35:51 know I never really thought about that 35:52 I'm like so like like what's your 35:54 favorite song like you know like I'm 35:56 sure you care at some things. 35:57 >> Yeah, 35:57 >> certainly if you're interviewing 35:58 musicians, they should have a favorite. 36:00 >> Yeah, you'd like to think that, but they 36:01 also probably have a favorite app, too. 36:02 So, I think like taste is is a kind of 36:03 prerequisite. And then I just think like 36:05 the confidence to pick a direction and 36:07 then the you know, we say like at we 36:09 often say shipping is an act of like um 36:12 like confidence and humility. And what 36:14 that means is like you have to be 36:15 confident enough to put it live and then 36:16 humble enough to take the slap in the 36:18 face when you got it wrong totally and 36:19 react to that slap. Don't like kind of 36:21 be like no it's not me it's the 36:22 customers are don't get it right. So I 36:24 think like we need like high taste and 36:26 then like confidence and then like 36:28 ultimately understanding that like in 36:30 the Mark and Dre sense like a product is 36:31 a conversation with the market. Your 36:33 launch is like your opening bit and then 36:35 you have to basically adapt and react to 36:37 what gets thrown back at you which might 36:38 drag you in different directions and 36:39 then you need to have the again the 36:41 confidence to prune certain things and 36:42 like no we're not building an 36:43 attribution engine. Yes, we'll take on 36:45 some feedback on the CRM side, but I 36:47 think uh like the, you know, a lot of 36:50 product managers who don't work out for 36:51 us are like a lot more spreadsheety and 36:54 like, you know, um they won't take a 36:56 bet. They won't take a gamble. They 36:58 won't take a stance. They'll just be 36:59 like forever mired in like well the data 37:00 suggests and they're just trying to 37:02 hedge their bets whereas it's just not 37:03 the sort of company we are. I think we 37:05 kind of believe in having an opinion 37:06 about a space. Well, and the second part 37:08 of what you're saying, if I'm hearing 37:11 you right, is the good product managers 37:14 actually can listen to the market. 37:16 >> They have to be able to. Yeah. 37:17 >> And hear what 37:19 >> Yeah. 37:20 >> Um, like I think about this a lot in the 37:22 context of tech companies where Stripe's 37:25 first operating principle is users 37:27 first. We think that actually paying 37:29 attention to what users tell us tell us 37:32 in every sense you know via revealed 37:34 preferences in the data via just like 37:36 when we actually have conversations with 37:37 them um we start every week with uh 37:40 Monday morning meeting the first thing 37:41 we do is we actually host you know the 37:43 intercom guys there recently uh we host 37:44 a customer to uh tell us and give us a 37:46 report card and you know it's not an a 37:48 you know it's seldom an a you know they 37:50 always have things they want to fix and 37:52 they're very pragmatic things that they 37:53 want us to uh to improve there's no over 37:55 complicating them when we do our weekly 37:56 the all hands fireside we also uh bring 37:59 a customer to that but I feel like 38:01 there's a a problem of overcomplexifying 38:04 things and under talking to users in 38:07 Silicon Valley where yeah it's a bit too 38:10 much celebration of the individual kind 38:14 of product vision or uh a bit too much 38:17 as you say trying to data your way out 38:20 of it and if you're a product manager 38:22 and you're not talking to many customers 38:24 each week something's probably wrong. I 38:27 bring that up because like the whole 38:28 original intercom product was a way to 38:30 talk to customers. Like this is this is 38:32 kind of your guys. But but would you 38:34 agree that diagnosis that a lot of tech 38:37 products would be better if people 38:39 simply talk to customers more? 38:41 >> Yeah. I mean like one of the ideas that 38:44 like stuck with me very early on. It was 38:46 like 2009 2010. I'm going this is like a 38:48 deep cut or whatever but there's a guy 38:49 called Jared Spool who's like a famous 38:51 UX guy and I was on this tread of like 38:53 interaction design association type 38:55 people and somebody wrote this really 38:57 long like you know hey I've shipped X 38:58 and I've shipped Y and I can't work this 39:00 out does anyone have any speculation as 39:02 to why people aren't doing this thing 39:03 even though I make it really obvious on 39:04 the screen and he just like replied all 39:06 he's like have you tried asking them and 39:08 I remember like at the time I was like 39:10 right on like it was like it felt like a 39:12 revolutionary thing to say but I find 39:14 like you know I I shared this piece a 39:16 while ago um which was like the 39:17 questions I ask in every single product 39:18 review kind of like so you can kind of 39:20 either get ready to meet me or like just 39:22 ideally other people can replicate but 39:24 question one is basically what did our 39:25 users say about this when you showed it 39:27 to them 39:27 >> and everyone has to have an answer to 39:29 that question when uh when like when we 39:31 go in like I'm like hey well what did 39:32 the user say and like I need to like 39:33 understand that cuz if you're not 39:35 actually asking your users what what are 39:36 you doing like you know the only 39:37 validation we have is the market I do 39:39 think in the valley does and like well I 39:41 say the valley but like that just 39:42 basically means in the tech industry my 39:44 opinion there is this uh epidemic of 39:46 like uh hiding behind your data and like 39:48 you know what what can we instrument and 39:49 how many different you know mixed panel 39:51 dashboards can prove to me that this 39:52 product should be working just ignore 39:53 the fact that it isn't or whatever. I 39:54 think there's something kind of just 39:56 fundamentally broken there. Yes. 39:58 >> And interesting you you say like stripes 40:00 value is like users first. I'm just 40:02 curious is that is that deliberately the 40:04 word mean every word you guys say is 40:05 deliberate but that's deliberately users 40:07 as in do you mean like pointy clicky 40:08 users or do you mean customers or do you 40:10 mean prospects or do you mean like uh 40:12 >> yeah we we we deliberately chose users 40:14 because we just meant the people using 40:15 the product as opposed to um customers 40:19 you know if there's a buyer versus user 40:21 distinction we want to focus on the 40:22 people are actually using the product 40:23 the people who are you know managing 40:25 fraud within the business or actually 40:27 responsible for increasing conversion or 40:29 something like that so that was why who 40:31 chose that word. 40:31 >> Yeah. No, it it makes sense. I I often 40:34 because I I just I find often times 40:36 >> um if you want to perfect a product, you 40:38 talk to the users. Um if you want to 40:40 expand your market, you talk to like 40:41 prospective buyers. But whenever I like 40:43 even in my own portfolio when I talk 40:44 like what are you actually doing like 40:46 the businesses that like are how do you 40:47 say like prematurely talking to more 40:50 prospects when they have a load of 40:51 unhappy users are guaranteed this kind 40:53 of miles wide inches deep like messy 40:56 product that doesn't actually satisfy 40:57 anyone. Yes. But like they'll get there 40:59 kind of like one promise at a time. Oh, 41:01 we'll build that for Johnny and Johnny 41:02 will sign and we build this for Jenny 41:03 and Jenny will sign and like at no point 41:05 do they have one happy customer. What 41:07 they have is like a marauding churn bomb 41:09 of a of a user base. 41:10 >> It's funny you say that. It feels like 41:12 many tech companies overrotate on sales 41:16 feedback which will by definition be 41:19 from the marginal user. And they're 41:20 marginal in two senses. So you have all 41:22 your existing users, you know that you 41:24 dancing with the girl that br you over 41:26 here. And then uh you have this future 41:29 potential user who firstly by virtue of 41:32 the fact they're not already using you. 41:34 Maybe they're slightly outside your 41:35 wheelhouse or like the use case isn't 41:37 perfect or something like that. So maybe 41:38 they're not quite as good a fit as your 41:40 existing customers. And then also by 41:42 virtue of the fact that they have a 41:43 whole existing way of doing things when 41:46 they migrate over to your product 41:47 they'll do so in a bit worse shape of uh 41:51 integration where maybe they'll only use 41:54 one of the four features or uh you know 41:56 not everyone in the or will be 41:58 >> bought in versus the people who grew up 42:00 on your product. And so maybe just 42:01 restating what you're saying, I'm always 42:03 struck by people are way too focused on 42:06 we tried to win this big new shiny 42:08 enterprise account and they didn't have, 42:10 you know, we didn't have feature X and 42:11 so therefore we're going to develop 42:12 feature X as opposed to you've all these 42:14 users who grew up in your product and 42:16 really like it but they wish you had fix 42:18 A, B and C. 42:19 >> Yeah. 42:20 >> And just the nature of the fact that 42:22 sales gets more airtime than account 42:24 management essentially. Absolutely. 42:25 >> Means people really misprioritize where 42:27 they spend time. 42:27 >> Yeah. It's and people take an or for 42:30 granted and think that like the net new 42:31 revenue is that is hard right like and I 42:34 think one of the things that we see a 42:35 lot of is um like like in terms like you 42:38 know uh you know working out for your 42:39 current customers like we use the phrase 42:42 permission to innovate and permission to 42:43 expand in intercom which is basically 42:45 like you permission to innovate when 42:46 your product's pretty good like is in 42:48 yeah okay let's work on v3 but like is 42:50 v2 in good condition like or you know 42:52 and then permission to expand is like v3 42:54 is actually that exciting everyone's 42:55 happy with v2 now I think we can try to 42:58 do something new for customers like 42:59 expand our share of wallet or whatever. 43:01 But I think um a lot of people try to 43:03 solve revenue growth with like with like 43:06 um aimless product expansion to just try 43:08 and increase the share of wallet for the 43:09 people who are stuck with you. Yes. And 43:11 then they convince themselves they got 43:12 PMF or like you know that they have like 43:14 some sort of a good product because 43:16 they're kind of like fog grass style 43:18 force feeding new features down the 43:19 throats of their of their trapped users 43:21 and they're like like you know we're 43:22 doing great but they don't realize what 43:23 they're actually doing is making their 43:25 current product so messy that like 43:27 they're destroying the hope of future 43:28 revenue cuz like yeah you can force your 43:30 current customers into whatever upselles 43:32 you have or whatever. But your product 43:34 marketing along the way is getting 43:35 really difficult cuz all these features 43:36 don't make sense and they're just kind 43:37 of like you've tried to do this like 43:39 land and expand thing but you're 43:40 actually just expanding 43:41 >> and there's no landing happening in the 43:43 new product and uh and then you end up 43:45 twisting yourselves in knots and a lot 43:47 of startups like you know Jason Le used 43:48 this thing of like from zero to one's 43:50 impossible from 1 to 10 is hard and from 43:51 10 to 100 is inevitable like I think a 43:54 lot of people I I don't think that's 43:55 proven out to be true as as much as it 43:57 was back when he said it h I think a lot 43:59 of people get stuck in some sort of uh 44:01 glue around somewhere around the 10 44:02 million Mark where they don't know how 44:04 to like get the next 10,000 logos. So 44:06 they just try and milk the revenue out 44:07 of the existing customers 44:09 >> through like just forced product 44:10 adoption of new stuff like you know that 44:12 you see like here's your co-pilot I know 44:14 you didn't want it you know but here you 44:15 go like that type of thing. 44:18 Dez is describing here how they've 44:19 transformed intercom from a SAS product 44:21 to a frontier AI business. And to do so, 44:24 they had to pivot not just the product, 44:25 but the monetization model as well. 44:28 Because inference costs are so 44:29 significant, AI powered companies tend 44:31 to charge based on usage rather than 44:33 just allowing for unlimited plans. It 44:35 gets complicated and really 44:36 multi-dimensional very quickly. Now, 44:39 fortunately, complicated and 44:41 multi-dimensional is what Stripe billing 44:43 specializes in. Our usagebased billing 44:46 engine can ingest up to 100,000 events a 44:48 second. 100,000 events a second. So AI 44:51 companies can monetize products based on 44:53 realtime customer usage. We're powering 44:56 consumption billing for companies like 44:57 Figma, Cognition, and tons of other 44:59 leading AI applications. Our usage based 45:01 billing platform has grown 145% so far 45:04 this year. So whether you're changing 45:05 your business model like Intercom or 45:07 starting a new product from scratch, 45:08 your business strategy should dictate 45:10 the billing system and not the other way 45:12 around. for usage based billing. Check 45:14 out Stripe Billing. 45:17 >> As you think about the prototypical $10 45:19 million revenue B2B company, yeah, what 45:22 are the common mistakes you see and what 45:24 do you think the actual path that more 45:26 of them should follow is? 45:27 >> I mean, the biggest problem mistake is 45:30 um is like not aligning your fundraising 45:32 with your TAM. Uh I think a lot of folks 45:35 like we got a little bit overconvinced 45:38 during the era of cloud that every 45:40 business had a right to be like a 45:41 unicorn. 45:42 >> And so there's a lot of businesses whose 45:44 idea was like totally fine but actually 45:46 they should have gone and and base 45:47 camped it more so than they did. Uh 45:49 because they've raised on the assumption 45:51 they're e there's an easy path to like 45:52 hundreds of millions of 45:53 >> be more small profitable $30 million 45:55 revenue companies. Well, yeah, exactly. 45:57 And and I think a lot of these 45:59 businesses would be great if only they 46:02 didn't raise 20 and tell their investors 46:03 that you're going to like easily be 46:05 worth a billion or whatever. So, think 46:06 like there's there's a genuine mismatch 46:08 there where I think people have like 46:10 overstated like how big this idea could 46:12 get, you know, as in hey, I know we all 46:14 we do is like time tracking for dentists 46:15 in Delaware, but like believe me, we're 46:17 going to be a billion dollar company. 46:18 And you're like, okay, well, one of your 46:19 restrictions is going to have to break 46:20 here. So, that's one problem which is 46:22 more like kind of business model like 46:24 and venture ambition. The other stuff I 46:26 see is like it is kind of like not 46:28 focusing enough on the thing your the 46:30 majority of your customers value. Like 46:31 it's easy to say I the best business in 46:34 the world is like one line of code that 46:35 all users execute and you sell it to all 46:37 users, right? They're like you know the 46:39 sweet spot. It's hard to do that in a 46:40 differentiated way. So because you know 46:42 obviously people learn that line of code 46:44 and that's where I think a lot of these 46:45 like horizontal products they say 46:47 something like a loom or whatever. 46:48 They're like a brilliant they're a piece 46:50 in everyone's workflow but they're no 46:52 one's like endto-end workflow. I think 46:54 they can do well too. But I think the 46:56 challenge is like when people rather 46:58 than nailing a specific small thing come 47:01 back to kind of the earlier point like 47:02 rather than like saying hey let's get 47:03 really good at X before we go beyond 47:05 when they kind of prematurely expand I 47:08 think they forgo all opportunity of of 47:10 like kind of being the best and if they 47:11 picked a really important area first 47:14 then they don't say it out loud what 47:15 they're saying is it's okay to not be 47:17 the best at the most important thing we 47:18 do 47:19 >> and I remember like I remember I think 47:20 it was 2012 I was in your office on you 47:22 know in FIA and I remember at the time 47:25 it wasn't it wasn't obvious to me that 47:26 you weren't going to expand and do some 47:28 sort of like you know peerto-peer 47:29 transfers and compet PayPal and I 47:31 remember like you guys had the 47:32 discipline like absolutely not we care 47:34 about helping businesses charge and like 47:36 there's a real harsh discipline you need 47:38 to have to like just basically say no to 47:40 all of the surrounding opportunities and 47:42 I think a lot of people that discipline 47:44 is the first thing to go when you hear 47:45 about competitors you have you're going 47:47 to hear somebody else encroaching on 47:48 your space you're going to have this 47:49 really weird broad view of all the 47:51 things you do like I know we just do 47:52 like whatever it is gifts and 47:53 screenshots But actually when you think 47:55 about we're a global creativity platform 47:57 and they have this premature view of 47:59 themselves as being massive and then 48:00 they feel then they go and raise off 48:02 that and they need to expand into that. 48:03 >> But I think like at the core of every 48:06 great business, 48:07 >> every great spouse business but in the 48:10 future AI business is something that 48:11 they're just truly world class at. And 48:13 it's like it's not some sort of 8020 48:15 trade-off. They've just basically said 48:17 we'll be better than anyone at this, 48:18 right? Like if you take like linear, 48:20 it's basically like they have literally 48:22 the most world's most efficient UI for 48:24 like for product management and they all 48:25 all sort of project management and 48:26 they've just gone really deep into like 48:28 all of the surrounding adjacencies you 48:30 would need to actually do that job 48:31 really well. Figma is just an amazing 48:34 creative collaboration machine. Like 48:35 everyone who like has done really well. 48:38 They've picked one thing and just gone 48:40 really hard, really deep, really far on 48:42 it. They haven't like prematurely blown 48:44 up and gone in seven different 48:45 directions. And I also think there's an 48:48 a weird celebration in the valley of 48:52 act two. Like the valley is obsessed 48:55 with finding 48:57 second acts that are totally unrelated 48:59 to the the the first business. Like the 49:01 number of people who bring up like you 49:02 know oh and we like invent like an AWS. 49:04 It's like okay you need to use a non 49:06 cliche example if you're going to make 49:07 that argument. And the flip side is, you 49:09 know, you're mentioning Figma Mau, which 49:10 I think is a great example where that 49:12 market proved to be way bigger than 49:14 people might originally have have 49:16 thought. Uh, you know, my favorite 49:18 example of this is, um, Nvidia, where 49:20 they're the world's largest companies, 49:22 and they started in the 1990s, uh, 49:24 making GPUs. And if you're an investment 49:25 banker trying to, you know, make a case 49:28 for how Nvidia can be a really big 49:30 company, maybe you'd say, "Oh, well, we 49:32 can expand into maybe we'll actually, 49:34 you know, make our own gaming rigs or, 49:35 you know, because it was all gaming at 49:36 the time originally. you know, maybe 49:37 we'll make gaming consoles or, you know, 49:39 we'll expand to some larger markets. 49:40 Whereas, actually, what transpired is it 49:43 turns out the GPU market is quite a lot 49:45 bigger than people thought. And being, 49:47 you know, the best at GPUs is a really 49:49 valuable prize. 49:50 >> Yeah. 49:51 >> And you you can't rush it, you know, 49:53 it'll emerge at it. But yeah, 49:55 >> they could have killed themselves if 49:56 they had gone in every other direction 49:57 like Right. And they would have lost 49:59 their edge in some sense. Figma was a 50:00 great example for me of like permission 50:02 to expand in that like they literally 50:04 nailed to a point of like no credible 50:06 competition uh this idea of like just 50:09 you know the Photoshop killer basically 50:11 let's just say 50:12 >> and now they can talk about like slides 50:13 and text to app builders and and like 50:15 every other dimension they want to go 50:17 and everyone's like yep that's that's 50:18 great cuz you guys make great software I 50:20 think you have to first be known for 50:21 like I'm trying to think like if Stripe 50:23 launched a payroll product it would 50:25 carry the brand of Stripe in the sense 50:27 of like being well it's probably really 50:29 good really reliable, really fast. It's 50:31 probably has really nice APIs, probably 50:32 works really well, work, you can almost 50:34 impute like all the ideas that would be 50:35 carried into it. And I just think like 50:37 you have to get to that point before you 50:38 have permission to make that that bet. 50:40 Like obviously it's a lot easier with 50:41 some like stable coin or whatever, but 50:42 like what kills me is when I'm when I'm 50:44 like, you know, I I don't even want to 50:45 like name a weak SAS company, but like 50:48 pick your favorite like, you know, 50:49 mediocre SAS company and anything like 50:51 is there any direction you you would 50:53 allow them expand in your head? No. Like 50:55 it's the short answer like 50:56 >> Yeah. Yeah. That's interesting. who are 50:58 you really excited to adopt new products 50:59 for versus who are you steering clear of 51:01 the new products 51:02 >> like if linear launched a I don't know 51:03 let's just say a source control to them 51:05 like yeah it's probably going to be 51:06 really really good 51:07 >> and like Seth Goden has this hilarious 51:09 uh point where he he talks about like 51:10 the the value of brand once it's once 51:13 it's like weaponized and he he describes 51:15 like uh Nike and Hayatt hotels and he 51:17 says if Nike opened a hotel you can 51:19 close your eyes and see it you know 51:20 exactly what the corridors are going to 51:21 look like you know the vibe of the whole 51:23 place you know everything it's going to 51:25 be if Hyatt launched a sneaker 51:27 you're like what? And it's just that's 51:29 the difference because Hyatt has a logo 51:31 and Nike has a brand and that's the 51:33 difference. Yeah. 51:33 >> Um a version of this actually maybe 51:35 quite literally is and I saw Equinox 51:37 launched the hotel was a pretty good 51:39 idea because the design center they had 51:42 the for the hotel is you just want to be 51:44 able to get a good night's sleep. And 51:46 it's funny how that's like a 51:47 differentiated product pitch in the 51:50 hotel space of like don't give me any of 51:52 that other shite. I just want to be able 51:54 to go to my room and not have like super 51:57 loud noise outside the window or like 51:59 weird light coming into the room. You 52:00 just want to be able to sleep precise. 52:02 Yeah. 52:02 >> Yeah. I thought that was funny. And 52:04 you're you're mentioning kind of Stripes 52:05 expansions and so this may be a good 52:08 segue into your pricing model change. 52:11 You guys are the poster child for the 52:14 move from per seat SAS pricing, the old 52:17 way of doing things to usage based 52:20 pricing. Maybe you can describe a little 52:22 bit about that and then how you 52:23 implemented it and what you're doing 52:24 with stripe. 52:25 >> Yeah, sure. Uh our pricing journey is 52:27 long and complex and and a lot of your 52:30 listeners or viewers will know intercom 52:32 pricing is a charge topic. 52:34 >> Uh not anymore. You know, we've turned a 52:36 corner. Um let me just back up a bit. So 52:38 like when when we had like two diverse 52:41 product strategy, we're trying to do 52:42 sales software, marketing software, 52:43 support software, and like sales 52:45 software is typically sold based on 52:46 leads created. marketing was charged by 52:48 like how many contacted people you want 52:49 to send and support was sold by seats. 52:51 So, we had this extremely uh let's just 52:54 say detailed but like unnecessarily 52:56 complex like pricing setup and we kind 52:57 of we lied to ourselves and said don't 52:59 worry because there's always be a human 53:00 to help people navigate this cuz you're 53:02 never going to have to self-service this 53:03 but like ultimately people are just like 53:04 I have been refreshing this for like 7 53:06 minutes and I can't understand a word of 53:07 it and that was just one of the few 53:08 things we got wrong in our kind of first 53:09 move up market. when own returned one of 53:11 the decisions he made was just like hey 53:13 we need to like you know sort out 53:14 pricing 53:15 >> and we handed back like truly handed 53:18 back I think about 50 million of revenue 53:20 >> uh I think that it's 53:22 >> was that controversial like with the 53:23 board with investors 53:25 >> like we had support for it I think it 53:27 was like uh people don't really underest 53:29 people massively underestimate uh what 53:32 it means to have a really happy customer 53:33 base it's because word of mouth doesn't 53:35 have like an attribution or UTM code if 53:37 you know what I mean so like they don't 53:39 understand how to think like happy 53:40 customers. So making the decision to 53:42 basically like you know kind of 53:43 standardize on an easy to understand 53:45 pricing that's like 53:47 >> fair transparent predictable etc. That 53:49 was the first decision that we made. 53:51 This is before AI right and that was 53:52 like and us returning to Stripe was a 53:54 large part of that. Um in fact as a 53:56 small segue like I think like a couple 53:59 years prior I had said to you or to 54:00 Patrick like hey you guys should 54:02 actually build as part of your product 54:03 offering a pricing page creator. Um, I 54:06 think at the time that you I probably 54:08 got one of those like thumbs up replies 54:10 or something like that and I was like, 54:10 "Yeah, whatever it is." 54:13 >> But I think on the list I think you've 54:15 done it since. Yeah. Yeah. Um, but like 54:17 my thinking at the time was basically 54:19 some version of this, right? You need to 54:21 not let your customers go wild with 54:22 pricing, right? You need to like 54:23 actually put some sort of guardrails 54:26 onto how they think about pricing. 54:27 Otherwise, they're going to go and 54:28 invent stuff that you guys don't support 54:30 and then you're going to move all your 54:31 business logic is writing checks. 54:32 >> Yes. Yeah. Exactly. And uh once like I 54:35 remember like sitting at in a stripe 54:37 mini all hands or whatever explaining 54:38 like that like hey now none of our 54:39 business logic runs through Stripe and 54:41 either you or Patrick was saying like 54:42 some version like that's not a good 54:43 thing. I think you know generally 54:45 speaking my advice to any software 54:46 company is like don't afford yourself 54:47 too many degrees of freedom here because 54:49 you'll actually [ __ ] yourselves in a 54:50 quagmire of complexity that you'll take 54:52 you many years and and ultimately like 54:54 many tens of millions of dollars to get 54:55 out of. It's it's a weird failure mode 54:57 that every single company falls into 54:59 which is you start signing deals that 55:02 have some super creative pricing 55:04 structure and the customer negotiates 55:05 ABCD and E and it's like it is built in 55:08 Microsoft Word but it is actually it's 55:11 just not built practically in code 55:13 because it just exists for this one 55:14 customer and it may not even be possible 55:16 to build in code like sometimes the you 55:18 know it's kind of ambiguous or it's like 55:20 and if in the subsequent year this 55:22 happens then we go back to the prior 55:24 year and we do an adjustment 55:27 there's like a time travel component to 55:29 the whole thing and then they obviously 55:30 have this um again we see all customers 55:33 running into it these kind of manual 55:34 billing issues where there is a guy who 55:36 has to deal with all these like 55:37 contracts that uh were agreed during the 55:39 sales process and so as you were saying 55:42 an opinionated billing engine is 55:45 actually pretty important assuming you 55:47 believe that billing should be 55:48 automated. If you're happy like manually 55:50 getting out the calculator for every 55:52 single customer month, then that's fine. 55:54 >> And having like a large deal desk 55:55 function and doing all the work behind 55:57 the scenes. Yeah. So, so that was like 55:58 the first piece of our pricing. And then 55:59 the second piece was obviously when we 56:00 launched Finn then it was like, "Hey, 56:02 how do we charge for this?" because 56:03 we're we're replacing seats and like at 56:05 the time it hasn't proved out this way 56:07 fully but like at the time Finn looked 56:08 like it was going to be pretty 56:09 cannibalistic to intercom because it was 56:10 like hey if we're automating at the time 56:12 what we thought was like 25% of your 56:13 revenue we assume that means like 25% 56:16 less seats in the future or at the very 56:18 least what it would likely mean is 56:19 >> the growth rate or the nore of the seats 56:21 model will will be affected by the fact 56:23 that doing all the work and now at 65% 56:25 you'd expect it to be even further true. 56:27 So it was like hey how do we charge in a 56:28 way that makes sense and we're and then 56:30 also like how do we be aggressive like 56:32 as in we want really wanted to like um 56:35 like put a mark on the market uh and 56:37 that sort of says like you know we're 56:38 like very AI forward 56:40 >> and I think what uh what and then Darra 56:42 came up with was just like hey like 56:43 let's just charge per literal resolution 56:45 like every time we do work we charge 56:46 every time we don't we don't and this is 56:48 at a time when like AI was like you know 56:49 margin negative and all that sort of 56:51 stuff we're still working out how the 56:52 whole world plays out today we're you 56:54 know we're we're very happy with our 56:56 margins but at the It was like, "Hey, 56:57 did this good for me?" 56:58 >> It was a bet. 56:59 >> Yes, it was definitely a bet. We made 57:00 the decision and uh and I think the 57:03 market responded really well because I 57:05 think it was very clear that like the 57:07 state the single statement of like we 57:09 only get paid when we do work. We don't 57:11 get paid when we don't do work. 57:12 >> It's from the same vein as the guarantee 57:14 which is just like 57:15 >> that's how you know that we believe in 57:17 our product and that's how you know our 57:18 product works. Yeah, 57:19 >> it's been copied a million times since, 57:20 but I think like the actual decisions 57:23 that we made in the runup there were 57:24 really really important from a point of 57:25 view of backing up our our claims. And 57:27 then obviously for a lot of our 57:29 competitors, they were like, you know, 57:30 we don't really have a great way to 57:31 respond to that cuz either our product 57:33 doesn't work or we're kind of hooked on 57:34 these really expensive resolutions and 57:36 we were totally throwing catamox to 57:38 pigeons there, which has been like 57:39 really well received by the market and 57:41 our customers generally do love it. It 57:42 is funny though like you still get 57:43 people being like 99 cents ridiculously 57:45 expensive and then you know you're like 57:47 why do you think this and the answer is 57:49 because some version of we don't know 57:50 how to calculate cogs you know. 57:53 >> Yeah. Yeah. Have you how much are the 57:54 humans costing you? 57:55 >> Yeah. Exactly. And how much is your 57:56 office costing you and all the other 57:57 stuff you 57:58 >> Yeah. Yeah. But people like certainty. 58:00 How do you get them okay with the 58:01 variable component? You know, you can 58:03 obviously you can contract out whatever 58:04 you want, right? Like um 58:06 >> but what we offer to people is like hey 58:08 like you know most of the time people 58:10 have like at least one or two years look 58:12 back on like what you know like we have 58:14 customers who spike for tax season or 58:15 customers who spike for Christmas or 58:17 whatever and we can basically say like 58:18 hey let's like let's contract your base 58:20 rate and let's talk about overages for 58:22 the months you needed and that's like 58:23 that totally works. What we're basically 58:25 saying is like yes, you don't have 58:26 predictability in the sense of it not 58:27 being fixed, but you have like you can 58:29 model it based on what's happened in 58:30 your history. And it's only really like 58:32 brand new startups that don't have a 58:33 clue what's going to happen, but like 58:34 they're they're not usually worried 58:35 about this. 58:36 >> Yeah. Yeah. 58:37 >> So, you're using a relatively new 58:38 product usage based billing for this. 58:40 How is that you migrated from Zora for 58:41 that? How has that process been? 58:44 >> Yeah. I mean I would say just go back to 58:46 the earlier like um we afforded 58:48 ourselves too much complexity and we 58:51 kind of codified that complexity in 58:53 Zora. I guess the best way to describe 58:54 it is we just twisted ourselves in 58:56 knots, you know, and it got to a place 58:57 where um we actually we ended up like 58:59 Kieran Lee who was our CTO uh here, he 59:02 ended up actually returning to the 59:04 company with one mission which was like 59:05 I am going to un 59:08 Yeah, exactly. To to fix this, right? 59:09 And like it worked, right? But it was it 59:11 was like it was a substantial amount of 59:13 work to like unwind so much and then 59:15 like they were like 59:17 >> they kind of like deconstruct so many of 59:18 these like alakart Microsoft Word style 59:21 deals into something that was a go 59:23 forward acceptable or whatever and then 59:25 obviously moving towards like a clean 59:27 transparent seatbased pricing and then 59:29 just layering on a usage base on top was 59:31 actually pretty simple uh in the 59:33 greatest scheme of things like and like 59:34 all the stuff that you know we needed 59:36 you guys were ahead of us on like you 59:37 know discounts for volume etc. all the 59:39 sort of obvious stuff people would push 59:41 for. 59:41 >> Yeah. Yeah. Is this just where pricing 59:44 in this new world goes because obviously 59:47 no one buys labor on an unlimited uh you 59:51 know basis and at least for the moment 59:54 the inputs of AI do actually scale with 59:56 usage you know for a significant basis 59:58 and so it feels like you have to have 60:00 some usage based pricing. This is 60:01 certainly the bet we're making where 60:02 again the reason that billing um kind of 60:05 the top thing they're thinking about is 60:06 kind of making billing work well in a 60:08 usage based world is it just feels like 60:10 many products are becoming much more 60:11 expensive to serve and therefore have to 60:13 have a usage based component but is this 60:16 permanent or I don't know does the AI 60:19 get cheap enough that maybe we go back 60:20 to unlimited plans or I don't know 60:22 >> I don't know if unlimited plans will 60:24 ever well I don't know I here's how I 60:27 think about it like I think ultimately 60:28 all AI has like two vectors there's like 60:30 how much work you're doing and how well 60:31 you're doing it. 60:32 >> Yes. 60:32 >> And the volume of work you're doing, 60:35 it's almost and well actually both of 60:37 them are going to be like proportional 60:38 to how many tokens you're burning or 60:39 whatever. So like you're going to want 60:41 to factor that in especially if you're a 60:43 consumer app as well just going to go 60:44 nuts. So I think like you have to like 60:46 have some you know I'm not a fan of like 60:48 cost plus pricing but like but it does 60:50 place a kind of a lower bound on what 60:52 you can do here which is just like hey 60:54 you know unlike SAS you are actually 60:55 sending money out the back door 60:57 >> as well. So I think you have to have 60:59 something that's proportionate to how 61:00 much work you're doing. And then I think 61:02 as aside from that you have to charge 61:04 consistently like how much work are you 61:05 displacing. I think that's where you can 61:07 sort of say hey like you know for us 61:08 anyway like if you take an average 61:10 person who sits in a seat to do customer 61:13 service if they do like let's just say 61:15 they do 20 conversations a day that's 61:16 what 400 conversations a month. Mhm. 61:19 >> Um, when we were thinking about how we 61:20 charge, we're like, "Hey, well, if that 61:22 person does 400 a month and you know, 61:24 Finn does 65% of that seat, we're still 61:26 up because we're only charging like 61:27 whatever $90 for the seat." So, from our 61:29 point of view, it was like an obvious an 61:32 easy swap. I think for a lot of 61:33 businesses, it might not be if your AI 61:35 doesn't work or it's furious or like its 61:36 value can't be articulated. Like, isn't 61:38 it cool that you can now dynamically 61:40 summarize a GitHub issue or something 61:41 that you're like, cool, I don't know how 61:42 much people will pay for that cuz they 61:44 don't they don't know either, right? 61:45 like, hey, you can now generate random 61:47 graphics in your like in your newsletter 61:49 tool. You're like, 61:50 >> vitamins versus painkillers uh AI 61:52 pricing. 61:53 >> Yeah. And like and it's specifically um 61:55 like in this case, the painkillers have 61:57 a very strict like if we don't do it, a 61:59 human's going to do it and we know 62:00 exactly what they cost 62:01 >> and the vitamins doesn't have anything 62:03 approximating that. So, not only is it a 62:05 nice to have, it's like I don't even 62:06 know what it's worth. Like you know I 62:07 saw a while ago like someone said like 62:09 >> when you know when Studio Giblly came 62:10 out and everyone was like using uh uh 62:12 like that uh someone said like hey the 62:14 fiverr.com equivalent of of all these 62:16 things would have been like trillions of 62:17 dollars and you're like right but like 62:18 no one was ever going to spend that. So 62:20 there's no sane way to actually talk 62:21 about what what actually happens here. 62:23 >> I think it was Burn Hubbert who said 62:24 that you know when you're tied to 62:25 business outcome that business outcome 62:27 is usually done by humans. I think it's 62:28 going to be really really easy to join 62:30 to make a business case for saying swap 62:31 this over to AI it's better faster 62:33 cheaper. I think when your AI is not 62:35 tied to business impat 62:43 you know so it's like you can have a 62:44 normal seat or an AI seat and then 62:46 you're kind of like I hope no one uses 62:47 the AI too much 62:49 >> you're you're permitting yourself to 62:51 build weak AI stuff if you do that cuz 62:53 you're not pushing yourselves to say hey 62:54 we need to articulate the value of each 62:56 incremental usage here. Well, when you 62:58 talk about this um AI pricing dynamic, 63:00 one thing that really strikes me is just 63:02 how fast AI companies grow from a 63:04 revenue perspective. So, I just saw 63:06 Matty uh from 11 Labs. Um we actually 63:09 had a great session at our customer 63:11 event in uh in London, but he tweeted 63:13 that they've just passed 200 million in 63:15 AR and that's 2 years after founding, 63:18 maybe three years after founding. But, 63:20 you know, in my day, businesses didn't 63:22 do that. And it's really striking for me 63:24 how somehow 63:25 >> they seem to climb the revenue ramps 63:27 much quicker. 63:28 >> I know. 63:28 >> I mean, you guys with Finn is another 63:29 example. 63:30 >> Yeah, for sure. Um I we forecast like 63:33 Finn will be 100 million probably early 63:35 next year or whatever. And like back at 63:37 >> Yeah. From from when like starting from 63:39 >> I don't know it probably about two years 63:40 some like 63:41 >> Yeah. Yeah. So two years to 100 million 63:43 >> like when we started and probably when 63:45 you guys started like it was like that 63:46 was the threshold to go public. 63:48 >> Exactly. Yeah, it used to take a long 63:49 time to get to 100 million. 63:51 >> It was like seven years a Exactly. back 63:54 in the day. But yeah, it's the the 63:56 acceleration is is important there. 63:57 There's like 63:59 >> Matt's you know 11's a fantastic 64:01 product, right? Like and and it's a 64:02 great it's a great example of like 64:04 there's like four kind of if you like 64:06 horsemen of AI products that I observe 64:08 whenever I'm investing like it's rare 64:09 you see all four, but the things you 64:11 want in an AI startup is kind of one is 64:13 like is the revenue backed by usage like 64:15 and that's why I love usage based 64:17 revenue, right? as opposed to like uh 64:18 you know shelfware or pilot wear you 64:20 know like hey we sold it to two guys in 64:21 the corner and they're going to put it 64:22 live someday. So like you you want 64:24 revenue backed by usage you want the 64:26 usage tied to a real business impact. So 64:27 has to be mission critical you know as 64:29 in like if you're building a phone 64:31 product on top of 11 like if that 64:32 doesn't work that's really bad. So they 64:34 so it's it's critical. 64:36 >> The third one is obviously you want deep 64:37 AI deep differentiated AI can't be a 64:39 thin wrapper. And then the fourth one is 64:41 like you actually want positive unit 64:42 margins and they all just or at least a 64:44 clear path to positive margins if you're 64:45 not there already. And I think when you 64:47 look at at so much of the AI landscape, 64:49 you'll see so few businesses that have 64:51 all four. It's such a rare sort of air 64:53 to be in to be like actually okay, we're 64:55 doing a real thing with real 64:56 differentiated AI. It really matters to 64:58 businesses and we're making money off 64:59 it. Most of the time when you hear about 65:01 these, well, we went from 0 to 6 million 65:03 overnight. It's kind of like, you know, 65:04 to generate JPEGs of a smurf or 65:06 whatever, and you're like, "All right, 65:07 cool. I'm not sure that's going to 65:08 renew." You know, 65:09 >> that's the simplest AI investing 65:11 framework I've heard. I 65:12 >> I tell you why it's simple. because 65:13 you're going to make sure you write no 65:14 checks. 65:16 So like I I guess I'd say most of my uh 65:19 most of the AI companies I've invested 65:21 in are probably three or four. Three or 65:22 four. I'd say 65:23 >> the only one I might equivalent there. I 65:25 think that's very good for staying out 65:27 of trouble. 65:28 >> And this is where you know I tend to 65:29 push back when people are saying oh it's 65:30 an AI bubble. It's like I don't know. I 65:32 think people are happy with the tokens 65:33 they're buying. You know, I think like 65:35 there's a lot of tokens happening and 65:36 just generally they seem to be 65:38 delivering useful outcomes to customers 65:40 cuz they're actually delivering value on 65:41 the customer service side or people 65:43 enjoy their midjourney adventure, you 65:44 know, but like people are getting value 65:46 from the products. 65:46 >> I used to say sorry to push back like 65:48 doesn't like 65:49 >> I was going to push back on number four 65:50 which is positive unit margins because 65:53 just aren't the underlying costs ch like 65:55 when you guys started Finn it sounds 65:57 like it was underwater, 65:59 >> right? But then just pretty quickly it 66:01 right sizes as you optimize it. And so 66:04 couldn't one be too focused on the 66:06 current implementation? 66:06 >> Yeah, I mean that's this is a 66:08 conversation we have internally with our 66:09 our CFO quite a bit actually. Like we're 66:11 like cuz we're we're good. 66:12 >> You can imagine it be the kind of thing 66:13 a CFO would Hey De, do you have five 66:16 minutes? 66:16 >> That's exactly money. 66:18 >> Yeah. Hey, quick chat. I can't help but 66:20 notice the team have done this like 66:21 preemptive like loading or whatever it's 66:23 costing a shitload of money. Uh so like 66:25 what's my counter to that? I guess like 66:28 I prefer it if the path towards 66:31 profitability isn't just like the you 66:34 know open AAI is going to figure this 66:35 out for me right like an interesting way 66:36 I'd say this like um 66:38 >> with Finn for example obviously ARM our 66:40 profit goes up when we are firing less 66:43 dead tokens the dead token being we've 66:45 generated an answer and it wasn't right 66:46 we can't charge money for it like if 66:48 you're like say like guessing the next 66:49 line of code right or like tab to 66:51 autocomplete the next line of code if 66:52 like five of six of those is wrong 66:55 >> I don't know if you're ever going to get 66:56 bailed out by because you're basically 66:58 56s of your of your costs is is like you 67:00 know it's not something you can resell. 67:02 Yes. 67:02 >> So like there's there's a question there 67:04 of like how much of your tokens are 67:05 actually generating a thing that a user 67:06 wants. Uh independent of what you charge 67:09 for it. As long as the user wants it, I 67:10 think you're always in good condition. 67:11 Whereas if you're like burning a million 67:13 tokens to find one and that one you're 67:15 never going to be able to recoup your 67:16 cost or at least you know I'd love to 67:17 see your telemetry to make sure that you 67:19 actually have thought this through. I 67:20 suspect you haven't you know. 67:24 No, see 67:25 >> 45. That's not bad. 67:26 >> Yeah, but I was very impressed by the 67:29 multiple clearly deliberate 20s. So, I'm 67:31 curious about the co-founder dynamic you 67:33 guys have across all the co-founders 67:35 where, let me try this on. My sense is 67:37 that people want to have a subject 67:40 matter area based explanation for 67:43 co-founder, you know, collaboration 67:44 where, you know, I'm the technical guy 67:46 and they're the business guy, whatever. 67:47 And in my experience, or at least with 67:49 me and Patrick, it's much more 67:52 personality tension based where I would 67:54 say he's more visionary and expansionary 67:57 and I'm more well, you know, we have 67:58 food at home already. You know, you got 68:00 to finish the products that you're 68:01 already doing or, you know, uh I'm more 68:03 frugal and he always wants to spend all 68:04 our money or, you know, whatever the the 68:06 tension you're describing is. And then 68:08 there's a useful um well one it's it's 68:10 useful to have someone to be able you go 68:12 bad mad by yourself trying to solve all 68:14 these um fairly naughty problems but 68:16 also a good company strategy probably 68:19 exists at the intersection of those 68:21 tensions. Does that describe your 68:23 relationship with your co-founders and 68:25 what would you describe as the 68:26 personality tensions? 68:27 >> I mean we're definitely all different. 68:29 Uh there's a lot of like key things we 68:31 all agree on. Own would be like a like 68:34 first and foremost I mean he's like he's 68:36 a a very strong CEO. He's very decisive 68:38 and he's very uh brave is the best way I 68:40 could describe it. An interesting thing 68:42 like when he returned to Intercom, one 68:43 of the things he did was like basically 68:45 rebuild the culture and one of the 68:46 things he focused on was like resilience 68:47 and open-mindedness. 68:49 >> You know, we didn't know AI was coming. 68:51 He didn't know AI was coming. But like 68:53 to be able to like react to AI requires 68:55 a lot of manic pivots, zero certainty, 69:00 and ultimately conviction bets. And I 69:02 can't think of somebody better to do it. 69:04 Uh that wouldn't have been me. Like not 69:06 in a million years. I would be like like 69:08 even being as AI pilled as I am, I still 69:10 would have like and I can even I even 69:12 look back at my own performance in that 69:13 period. I'm like, you know what? I 69:15 wasn't brave enough. Like one of the 69:16 things I pushed for was this idea of 69:17 creating the team Finn, which is like 69:19 hey, let's just build a new startup. 69:21 Let's isolate them from everyone else. 69:22 Different floor, different section of 69:24 the office. 69:24 >> No one else is in there. It's just their 69:26 own Slack channels, their own 69:27 everything. They're entirely secluded. 69:29 And had he not pushed for that, I don't 69:30 know if we would have had the clarity 69:31 and focus that we needed. 69:32 >> People might be offended. 69:33 >> Yeah. Yeah. Of course. Like all of the 69:35 things all of the downsides you'd 69:36 possibly guess are all there. 69:38 >> I just I also think like that uh there's 69:40 no path to like there's no way, you 69:42 know, the phrase I've settled on when I 69:44 look back on this is like sometimes you 69:46 just you have to go too far to know 69:48 you've gone far enough, 69:49 >> you know? And I think a lot of the 69:50 mistakes I see in people who are trying 69:52 to adapt to AI for an example, I'll come 69:53 back to the co-founder here is like is 69:55 like they tell themselves that they've 69:57 they've done enough because they oh you 69:58 know a few sparkly buttons the merge 70:00 feature is AI and we're happy and we 70:02 >> we have an AI assistant. 70:03 >> Exactly. And we've updated our home page 70:05 to say we're AI first. So like we're 70:06 good. Yeah. Like and I think you need to 70:08 be willing genuinely willing to like 70:10 make brave hard to undo bets. And I 70:14 think you need like sort of obviously 70:16 you know uh having the sort of moral 70:18 authority of a founder and then being 70:19 CEO kind of h gives you some of that but 70:21 still it's a huge decision to make and I 70:23 think like I am much more of an uh my 70:26 default DNA is like I'm I'm like more of 70:28 an operator in the sense of like all 70:30 right what are we doing okay well I'll 70:31 make it work you know whatever it is and 70:32 I think if it was a company of like 70:34 people like me what you'd see is 70:36 >> probably like um you know predictable 70:39 reliable sustainable performance or 70:40 whatever but like but probably not 70:42 enough actual like sort Well, definitely 70:44 not enough kind of brave big swings 70:46 which is actually where you need to get 70:47 to but I mean it is a cocktail like it's 70:49 in somebody needs to go and actually do 70:50 the thing once we've decided what we're 70:51 doing as well. The way it ended up like 70:53 I was leading team Finn uh after owned 70:55 it and like that was like ultimately 70:57 what led to the creation of like the the 70:59 whole Finn initiative or whatever. Yeah. 71:00 >> You've now worked with so many different 71:02 companies um uh externally you've seen a 71:05 lot. What is predictive of success and 71:07 what is predictive of failure? The 71:08 biggest thing I'll always come back to 71:10 when I'm talking to anyone who's trying 71:11 to pitch me to invest or pitch me to 71:13 induce John to invest is it's always 71:15 some version of do you have a real 71:18 product that solves a real problem that 71:20 would that really exists and people are 71:22 really already trying to solve by paying 71:24 money or time somewhere. It sounds so 71:26 trivial, but you'd be shocked how many 71:27 times you'll fail or you'll get some 71:29 sort of jazz hands type routine 71:30 somewhere along the way where it's like 71:32 don't look too much at this but just 71:33 trust me. The areas that I I end up 71:35 being blind to in that is like you know 71:36 the ex extremely market expandy type 71:39 things like as in if someone said to you 71:41 hey like all companies are gonna have a 71:42 chat room and they're going to all hang 71:43 out in it all day and have unproductive 71:45 conversations it's going to be big I'd 71:47 be like oh I don't see it like you know 71:48 uh whereas like so you would have missed 71:50 out on Slack or whatever right but I 71:52 think like I can almost you know hear 71:54 from the like you know what are you 71:56 building and why and who's it for and 71:58 show me what the product does. If it's 71:59 not a real solution to a real problem 72:01 I'm kind of already out. Mh. 72:02 >> And then the other big I'd say 72:03 prediction is just um like there's one 72:06 of the things that's happened in the 72:07 last 10 years um I'm sure you've seen 72:09 this a load is like it was a lot easier 72:11 to invest when being a founder was 72:12 uncool. 72:13 >> Mhm. 72:14 >> And I think like combination of like I 72:17 blame genuinely the social network. I 72:19 blamed just kind of the entrepreneurial 72:21 lifestyle. I blame like Tik Tok. I blame 72:23 all these things. So yeah, 72:27 all all of that, right? To some degree 72:29 like remote working I throw into the mix 72:31 as well. But I think the amount of 72:33 people who are chasing the trinkets of 72:35 being a founder of a startup, even if 72:37 they're quite smart and they could 72:38 actually go and build something, if 72:40 their actual motivation isn't the 72:42 problem or isn't just some deep like 72:43 desire to be quite successful, but it is 72:45 instead to be perceived like the whole 72:47 kind of I could have been a contender 72:48 rather than I could have contended. Yes. 72:50 like if you don't really really want to 72:52 be uh like to actually play the game 72:55 with said you just want to be seen to be 72:56 playing the game. 72:57 >> I think that's probably the single 72:58 biggest thing that tells me like you 72:59 know you're probably you're like best 73:01 case scenario you'll sell a 5 million 73:02 but more likely you'll still be alive in 73:04 seven years all your investors will 73:05 wonder what you're doing and you'll be 73:06 basically sending one investor update 73:08 every now and then. Yeah. 73:09 >> Yes. I have noticed investor updates 73:11 with metrics 73:13 don't um predict success, but investor 73:16 updates without metrics 73:18 >> that tell a really fancy story 73:21 >> but don't have metrics 73:23 >> are actually quite predictive of 73:24 failure. 73:25 >> Yeah, 73:25 >> those companies always fail. 73:26 >> I basically 100% agree. And honestly, 73:29 you can even tell where the metrics are 73:30 in the update. 73:31 >> Mhm. 73:31 >> Yeah. Like cuz often times like my 73:33 favorite updates I mean there's a 73:34 company that actually probably should 73:35 >> and no investor updates are fine. 73:36 There's like a bunch of successful 73:37 companies that just never bought. We 73:39 never sent investor updates. Like I'm 73:40 sorry for all the investors, but we were 73:42 bad communicators. I was curious. 73:44 >> Yeah. Yeah. But if Exactly. But if you 73:46 go I owe you an email. If you go to the 73:50 trouble of writing an investor update 73:52 and then make a proactive decision to 73:54 not say how your business is doing, that 73:55 suggests some deep denial about what 73:57 running a business means. So, there's 73:58 one company I can't say, but like we're 73:59 both an investor in it, but like their 74:01 their updates just one of the most 74:02 recent ones was just like, uh, here's 74:04 performance, uh, AR plus 17%, blah, plus 74:08 this, blah, plus that. Uh, you know, 74:10 something like, you know, I hope you can 74:12 see from the numbers we're doing great. 74:13 Uh, best of luck. See you next see you 74:14 next quarter, SM. And I was like, yeah, 74:15 brilliant archive. I'll mark it up, you 74:17 know, like uh, so something like that. I 74:19 think in general, you know, the degree 74:22 to I think it was Paul Graeme said like 74:23 the ratio of numbers to words is usually 74:25 the actual thing you're looking for like 74:26 which is like if the numbers speak then 74:27 the words don't have to 74:28 >> Yes. Yes. Um what else is predictive of 74:31 success? Um numbers is one. 74:33 >> I almost kind of want to say the inverse 74:34 of all the things that I hate seeing. I 74:36 hate seeing founders who invest 74:37 massively in their personal brand 74:38 instead of their company brand. 74:40 >> Um I hate seeing like people who are 74:42 like obsessed about like if the first 74:43 three or four updates I get are like 74:44 begs for retweets and quote tweets and 74:46 all that sort of stuff. That's always 74:48 >> not a great sign because it sort of says 74:49 to me you haven't worked out how to 74:51 market or whatever like anything around 74:52 like you know what are the customers 74:54 saying like that's like you know 74:56 whenever I reply and say like what do 74:57 customers think of this feature they're 74:58 like oh we're going to ask them 74:59 >> maybe what you're describing is there's 75:01 a very boring playbook 75:03 >> uh boring and glamorous playbook for 75:05 making products work about writing code 75:07 talking to customers running that 75:09 iterative loop uh and avoiding 75:11 distractions and people who seem 75:13 incurious about that playbook or just 75:16 are failing to execute on it is is kind 75:18 of a warning sign. 75:19 >> Yeah, I would say that's definitely 75:21 true. Like so it's basically like if you 75:22 could got like a decent decent product, 75:24 decent area real solution 75:27 >> and are you just willing to like work on 75:29 a boring stuff that needs doing to 75:31 actually make that whole thing and then 75:33 will you get bored in a year or are you 75:34 still excited by it like and I think to 75:36 some degree 75:37 >> even a successful uh founders can get 75:40 distracted by glamorous opportunities 75:42 where whether it's like oh there's a new 75:44 wave of whatever like crypto or like 75:46 NFTTS or whatever you you'll see people 75:47 get their head turned quite quite a bit 75:49 and I think if you're like genuinely 75:51 married to the problem and married to 75:52 the solution, you'll tend to like sort 75:53 of not be as distractable and then like 75:55 so many of these businesses just need 75:57 time, you know, like they just need time 75:58 and execution. 75:59 >> Yeah, totally. Um, last question because 76:02 um I've uh um had a lot in what ways is 76:05 intercom itself AI native. 76:07 >> The biggest initiatives we've driven 76:10 recently has been like around um how we 76:12 actually do R&D. So I think uh we 76:14 launched this initiative DRA launched it 76:16 I think about 4 months ago called 2x 76:18 where we basically said hey like we're 76:20 going to double the productivity of R&D 76:21 before February 1st that means the and 76:24 we measure this and like everyone's 76:25 going to poke holes in this and that's 76:26 grand but it doesn't really matter. The 76:28 measurement is I think it's deployments 76:31 to production involving code that has to 76:34 execute regularly right so it's in it's 76:36 a hard thing to fake if you fake it it's 76:38 like we should probably fire you know 76:40 cuz like so now interestingly everyone's 76:43 like is that just the engineers going 76:44 really hard no like there's like so many 76:46 different elements to this but one of 76:47 the biggest ones that we saw recently 76:48 which really has been like awesome is 76:50 emit our our head of design he basically 76:52 said every designer by August 1st needs 76:55 to be shipping code this is like the end 76:57 to the discussion Should designers code, 76:58 right? Designers should code. Um, so we 77:01 basically said, hey, like uh all 77:02 designers can now can now ship code. 77:04 Weirdly, the win there is is that like 77:07 the amount of engineering distraction uh 77:08 has just gone away. So every paper cut 77:10 in your UI used to used to result in a 77:12 GitHub issue that get filed and triaged 77:14 and get picked up on a Friday and 77:16 between breaks or whatever. And now all 77:18 that just goes. So now you're like, "Oh, 77:19 I want to fix this button, fix that 77:20 padding, fix that thing, change that 77:21 radius, change that color." All that 77:23 [ __ ] just happens automatically. And 77:24 like it's getting to much meteor stuff 77:26 like redesign this entire UI, redesign 77:28 this flow, change this wizard. All of 77:30 that's now being managed entirely by our 77:31 design team. What that has resulted in 77:33 is like engineers who are now doing far 77:34 more like staying in the zone for far 77:37 greater using like whether it's cloud 77:38 code or codeex or any of those or 77:40 augment or any of any of those tools to 77:42 actually just you know be far more 77:43 productive. And there's some real wins 77:45 here like one reason what we had was um 77:46 like Finn works in Slack. But when we 77:49 were building that it was built very 77:50 firmly from like how do we use AI here. 77:53 So it was like let's build one perfect 77:55 Slack solution. Let's document all of 77:57 our principles and then let's have cloud 77:59 code right the Microsoft teams the 78:01 Discord the WhatsApp it's every other 78:03 solution. So we went live to production 78:05 with Slack but I think we have like 78:06 everything else now in like public beta 78:08 like and it'll all go live. So we're 78:10 finding all these like it's a lot of 78:11 like 1.2x wins but then every now and 78:13 then there's like a 50x productivity 78:15 boost that we're finding. So I think 78:16 that's probably the biggest way in which 78:18 like the product has been is being built 78:19 from an AI native way. on the go to 78:21 market side of the house, we've been 78:22 slower, but like I think we're like 78:23 we're looking at like we've trained a a 78:26 a GPT if you like on like all of our 78:28 marketing copy, our principles, our 78:30 content, our visuals, etc. And we've 78:31 been using that to produce a lot of like 78:33 um you know stuff like event invites and 78:35 things like that. 78:36 >> Yeah. 78:36 >> Well, Dez, thank you. 78:38 >> Cheers. Thanks very much. 78:39 >> Yeah.