The former physicist building an AI assistant for parents | Dilani Kahawala, Anna AI

When we started working on Anna, people were connecting their WhatsApp's and Google accounts to this startup's product. We didn't even have a website. They're happy to let you see inside their very personal world. Parents are pretty sensitive about their personal information. There's like one woman who I talked to and she was like, "I don't care, just take my information, just solve the problem." The world where people navigated by what is defined in a menu is no long with the case. People mostly do their day-to-day work with Anna on messaging. We still have the app and people do use it, but I think it's like people just prefer to work with a product where they normally work in life. Take us a little bit into like where did the first idea for Anna come from? So we experimented through like eight or nine different ideas in like a five-month period, in like a five-month period. I mean that is the beauty of like where we are today. You can build very fast. In 2024, the US Surgeon General made a splash when he issued a rare public health advisory. This one was on parental stress. His report cited statistics that in the US, 41% of parents said on the most days they were so stressed they couldn't function and 48% described their stress as completely overwhelming most days. That's about double the rate of non-parents. Close at a home, the Hilda survey here in Australia found that couples with kids put in around 144 hours a week of paid and unpaid work combined, or about 50% more than non-parents. Anyone listening who is a parent knows that these extra hours are sometimes joyous, sometimes torturous, and actually a lot of them are also just driven by the exhausting mental effort of coordinating to-do lists, activities, and admin. I don't know how many people I know who described the school WhatsApp threads there on as one of the great happiness sappers of their life. My guest today is building something to take all of that off-parents' plates. Delani Karl Waller is the co-founder of Anna, an AI assistant that reads your school emails, calendars, and group chats, and then tells you the one thing that you need to know. Anna might sift through 200 pieces of information to send you just one message. Delani and her co-founder Mario got here by shelving their first company, Braid, and then testing eight or nine ideas in as little as five months. Some of what we get to in today's episode is just what that process of rapid idea generation and experimentation look like for them, and how she knew when Anna was the product that deserved their full attention. She's a PhD physicist, a published YA author, and a product leader from Metta and Atlassian. She now runs a five-person team that's moved to open source models to make the economics work over to Delani. Delani, welcome to Wildtarts. Thank you for having me. It's great to be here. I want to start actually in this conversation, not with the product that we know you for now, which is Anna, but actually for the product, we first invested in, which was known as Braid. Tell the audience a little bit about what Braid is. Braid was a agent team that was going to help software teams handle all the project management. I had been Atlassian, and obviously Atlassian builds a lot of project management software, and I could still see like there was so much as I was like a director of product or something, and I was doing so much just like hands on project management. Where is this one? Is this going to ship? And I'm like, there is a new way to do this with when the AI kind of wave came through. And our goal was just to like, let's just not have to do that anymore. And that was the vision. It's like there'd be agents who handled all the horrible stuff for you. I remember from the early ways in which our team talked about Braid. It was like, it was trying to address a product manager's challenge around what was I think known as a watermelon problem, which was that I think is described. It was like it looks like it's going green on the outside, but it turns out if you open it up, there are all these issues. And yeah, it's like definitely not going to ship on time and all this sort of stuff. It's always green until the last three weeks of the quarter, and then it turns yellow. Why did I not know this six weeks ago when I could have done something about it? And yeah, that happened all the time. Like Lockwook. So you built Braid, but you ultimately shelved it. Tell us a bit about why? Yeah, I mean, when we, you know, we talked to a bunch of customers. We talked to like 50, 75 customers before we went and raised our pre-seed round. There was so much interest from like CTOs, CPOs, because this is a real, real problem. Then we got a small team. We built the MVP, the first prototype of the product. And then we had design partners because it's an enterprise product. It was had to be very secure and carefully deployed. When we started deploying it, and we were trying to get our first few customers online, there are a couple of challenges. One is there wasn't as much enthusiasm when we were starting to deploy as when we were talking to them. And I think this is a classic problem of like, oh, yeah, I absolutely have that problem. I think this is a great idea. But then you, when you actually want to get it live, the enthusiasm kind of when customers have to expend some effort in the implementation. Exactly. It's, and then I think, so part of it was, it's a big problem. But it's like, who owns that problem in an organization? Isn't always clear. Like, is it the CTOs, the CPO? Is it an engineering manager, or is it like higher up? And that was a real, like, you couldn't pin down one person who was like, who's problem it was to solve this or they would like suffer? Everybody was suffering, but nobody was like, but they had the ability to point the finger at something else. Yeah, you could kind of like pass the blame around. And I think we had said, like, the kind of role that we want to like replicate is really like a scrum master or a program managers role. But if you step back and think about it, those people do such different things at different organizations. And I think that was a symptom of like, you couldn't pin down very carefully, like very specifically what they were doing as their job. So we couldn't solve a, like a specific enough pinpoint. Interesting. So you shelved it. And then you went through, I think, in my time at Blackbird, at least one of the most extraordinary periods of what we sometimes refer to as like playing out in the idea maze. You tried something like nine or 10 ideas before you landed on Anna. And I really do want to get to Anna. But I think it'd be great to spend a little bit of time in how you went about the process of saying, okay, it's not braid, got money in the bank, great ideas, a small crack team that we can do interesting things with. Where are we going to go with this? So like, walk me through. How did you start even the process of like filtering ideas? Yeah. I mean, I mean, the first step was just acknowledging that we were going to pivot because I think you can have this like slow death. And I think part of my co-founder is a second time founder because he co-founded Spriggy. Correct. Yep. So Mario was like, don't I don't think we have product market fit. I'm like, okay. And you're right. Because it was like my baby in some sense. Yeah. And just being like, okay, we're going to clean stop. We won't let it have a slow burn. Takes courage to do that. I was super difficult to kind of do it. But actually, like one of the way things like Max was, who is the partner we work with at Blackbird, was like really amazing about it. Like the first time we had the idea, the concept of an idea maze was from him. The process was once we had said, hey, we're going to pause or we're going to stop. It was, you know, the idea is that we're going to have as many turns that having like a bat. The more chances you have, the higher you're likely to have like hitting a homerun. We had some guardrails around like, hey, what kind of things are we not good as a team? What kind of things do we not want to do? But everything else was kind of like game. Just quickly, what were those guardrails? I mean, basically, you know, as a team, we didn't want to like build anything that was like gambling or companion apps or something that was not helping humanity. I guess there were probably like spaces that we weren't going to be strong, like we weren't going to be great at doing hardware, probably. So we were like, what are we good at? What is the team good at? And what do we not want to do as people? And then everything else was like, let's go, let's go try it. And I think one of the things we did was we have an incredible team. So at some point, we had like each engineer, you go try this thing, you go try this thing. I'll help you and I'll help you like get it to market or go post on LinkedIn about it. Like I can be the face of it. But we had each engineer almost like experimenting and running with an idea. One of our engineers, Arif, actually built out one of the ideas we were working on is AEOG for e-commerce, which is one of the ones that was very promising because we had a bunch of small businesses who were interested. And we had built out this experiment where they would get to the point of almost like paying for it. And they'd be like, oh, send us an email. And so we experimented through like eight or nine different ideas in like a five month period in like a five month period. I mean, that is the beauty of like where we are today. That doesn't mean you can build like a proper full on product. But we were thought very carefully about what would what's the experiment that would get the signal that it was worth another two weeks or another month of our time. Okay, so let walk me through that. So you've got these extraordinary engineers who are acting like miniature product. things unto themselves, presumably with a bunch of agents supporting them on the build. And then you consolidate and you say, like, okay, what does it look like getting to the end of a single two week period? What are you measuring to say it gets a yes to keep going? Sometimes you could build it in a way where you could get like a signal where customers, like you could say, well, 30 customers in that AOG01, we had like a fake door experiment where we would do like outreach to a bunch of e-commerce companies and they would come back and like, they would put their website in and we would like build it enough to like give them a flavor of what they would get and then they would put their email into to go move forward. And that was probably the most quantitative one we did and that got us like probably like 30, 40 interested customers. But primarily what we were looking for was the opposite of what the experience we had with people were like saying they were interested. But then when they were like, how do you know their pain is so great that they will then go through some friction or hurdles or talk to a company they've never heard of, probably don't have a good looking website. And that's really, it was customer behavior that we were looking for. So we did one which was called a deputy where we would do like help people work through or complicated for like figure out how to use Google Analytics because like you go there and you're like, oh my god, what do I do. And we had emailed a bunch of companies being like, hey, we built this customer success, customer support people are you interested. And like you would get like lukewarm yeses because no one wants to say no. But then when you're like, hey, let's go to the next step. Like you're looking for really like a yes, I absolutely have this pain point. I'm biting off your hand, yes, exactly. And honestly, we didn't see it for, we saw like glimpses of it. Like the comments one was probably the closest we got. But there were a lot of these that people would definitely say they were interested. But then there wouldn't be enough. And my right that survive AI, one of them got something like 800 uses, yeah, we had a massive number of people who signed up because we built up a wait list. We actually had like this tea or like a tastier product, which was called prompt reactor. Part of it was like a year ago, this just shows how quickly this space is moving. Like we were still thinking about how to structure prompts. And like writing a great prompt, yeah, was like the killer thing to do. And people needed help with that. And this thing got so people still use it today. But that in itself, when you wouldn't be enough for like a wall changing product. So while people were using that, we were like, we were trying to help people get better at using AI, which I think is actually a problem. Yeah. I don't think anyone knows what they actually want to solve that problem. So we built like, we would have like a coach that would like sit in your screen and like help you understand like all the advanced features of like chat gbt or how to build agents. Again, like we launched it and we just watched to see if people would like we announced a LinkedIn, we took it to all of our friends and just see like, you don't do anything. You just see how do does anybody come and use it and try it. And the answer is no. Yeah, it's interesting, isn't it? Because you know, in in pre-sade land, like 800 interested uses a long wait list. Like that's what people are looking for. So many product teams would die for like that form of signal. And what I think is super interesting is that you, with your product background at last in and before that Etsy and meta, like you understood that's not sufficient, I really want to get underneath like this distinction between sort of vanity metrics and like real user behavior metrics. Like what's the thing that for you, you mentioned earlier, like there was no single person feeling enough pain. No one was suffering enough that they were going to buy it. Like what's the kind of signal in the noise for you when you're making those sorts of decisions? Yeah, it's interesting, like it's not like you don't learn this in a big company product team because you already have like they, the founders solve that problem. That's why you're there in some sense. Like we learned it from the braid experience knowing like what didn't work. So the signal, so a couple of things, all of the problems we tackled were pain points because people told us they were pain points, but probably either people didn't know what they wanted. Like it wasn't great enough. It wasn't like the number one problem in their lives. Or we weren't building the right solution or for example in the, in the survive AI case, maybe what people actually wanted was a course or they wanted some accreditation. But the thing that actually the signal that mattered was, I mean, it's like, I feel like I, it's like such a cliche, but it's like people turning up to use a half baked MVP. Like when we started working on Anna, people were connecting their WhatsApp and, and Google accounts to this startups product, then they like, we didn't even have a website. And you're like, wow, they're happy to let you see inside their very personal world. You know, like parents are pretty sensitive about their personal information. There's like one woman who I talked to and she was like, I don't care, just take my information, just solve the problem. So it wasn't like a data, it wasn't like a number. It was more like, how do they walk with their feet? Like they have to vote with their feet to like use your product. I posted about Anna in a Facebook group and it got like 60 responses of like, I'm either building this already, hacking it together the best as I can. Actually, that's another interesting secondary signal that people, like latent demand where people are trying to solve it in some way already, somehow. Yeah. And people like wanted, I would try it. And that group converted into our beta group and they would like sign up to test and try things that were very broken. Well, it's interesting, actually, let's get on to Anna. So Anna is, you know, you might say like a sort of AIEA for parents. Correct. And so your kind of pain point is one shared by many, I think. Take us a little bit into like, where did the first idea for Anna come from out of, you know, you were trying e-commerce, you were trying prompt products, you were trying, you even did a healthcare product, I think for a little while. Yes. And I think I could come from unexpectedly, it was like we were, or everything we were trying was like B2B at that point. Yeah. And Gogo was one of the engineers, he's like a tech lead. He built this little agent, what's that agent for himself? He has a family and he has a lot going on managing like life and work. And he's like, I use this agent and I'm like, I need that. And then, and then I just like posted about it in a Facebook group because that's almost like the first level of like customer validation you can do. And then I got such a response, I was like, hmm, okay. This is different from like I didn't really have to chase customers for this. Yeah. But that signal was so strong, it was worth another two weeks of time. And that's how it started. And it was like two weeks more and two weeks more and two weeks more and here we are. And so how did it feel for you? Because your background is enterprise, Mario's obviously come from consumer. Did it feel scary to be like, okay, I'm now going to be in a potential way to say sell play? Yeah. It was quite unexpected. I was very reticent because until like very recently I would say consumer AI was not a thing. Like everything, especially AI, like everything to date has, that has taken off has been enterprise. Yep. Other than like features that are embedded in other products, right? Yeah. Yeah. Yeah. The person is like, lovable. I would say it's like a consumer product. Through consumer hadn't didn't really exist, other than like chat GPT free version. And so I was like, how are we going to, who's going to pay for this? How are we going to fund it? Because it's quite expensive to build and hence, hence why enterprise AI takes off because there's a lot of money to pay for it. But my co-founder has obviously done this before in this space and the thing that convinced me I think was, you can just move a lot faster in consumer, just way, way, way faster, just because the feedback loops are faster. The feedback loops are faster. You can just like iterate quickly. You can get product out quickly because you don't have to like go to five people to get a sale, like a contract signed. Yep. And I understood the problem because I was the parent who was experiencing it. And I think the end of the day that probably had a large part to do with how excited I was. Yeah. Indeed. I mean, you were kind of personal connection to it. Yeah. And it's interesting actually because I think there can be strength and also weakness in having a personal connection to the problem, right? Like the strength is you know it yourself. You can use yourself as a proxy for other users. The challenges are you in fact a good proxy for the kinds of users you need to scale. Talk to me a little bit about that. So you've got a family of your own three kids, a partner with another busy job, intake, you're a founder yourself, no end of like the amount of sort of cognitive strain on you. Talk to me about how you've thought about building out the user profiles for Anna with that knowledge in mind. So the strength is that I'm like. I call myself the chief data labeling officer. I just have honestly, the first 80% of the product, it was my data, I connected my WhatsApp and my Gmail. And I was like, just take it all, figure out how we build the first version off this data. And that was great, 'cause I think I'm pretty representative of the very first chunk of parents that we wanna go after, which are two busy parents with round the clock jobs that have multiple kids who are school age. And so you're managing your life, your work, and you're the PA for your three kids. And so you're just constantly dying under this burden of non-stop admin. So I think it was more a strength early on than it was very helpful. The point where I think it runs into limitations is, the first users, like the first 100 users we got were like barrier tech moms. So they know how to use Cloud Code or Cloud Code Work or what an agent is or that you talk and you ask and you shall receive. That's kinda how you work with AI. What's interesting is when you start going to the next layer of parents that you want trying your product, there's like a education, level education you have to do. First of all, most people wouldn't consider hiring an assistant because in the real world, that's like hyper expensive, but you have this problem. So you kinda have to help people understand, well now you can have an assistant if you want. But the idea of working with an agent through SMS, to through I message or WhatsApp, and you just ask for stuff like you were the human and you get it is not a thing that people do yet. And that came so naturally to me because I spend all day just like driving cord. (laughs) So that's where the limitations are. How do you understand this customer who lives in a very different world from you? - Yeah. I'm gathering from that that some of that's familiarity with AI and its capabilities. Talk to me also a bit about the trust element there because you're dealing in people's very private worlds. But also if you get it wrong, it's like is there a kid left somewhere in a soccer field never being picked up? Talk to me a bit about how you've thought about that 'cause it's still a team of about five. - Yes, it's five of us. It's three engineers and I guess we're all building in some ways to me and Mario and three engineers. And it's a very extensive product because, so there's two parts of trust. One is data privacy, right? Like who can see my most? Like my medical information is in my email. Like what's up conversations? It's like privacy and trust are like really, really critical. And a lot of parents suck the number one question we get. I am too, right? I'm putting my kids information in this app. So actually one of the things that really helped was because we built Braid. Braid is like a project manager for software teams and is a project manager for families. It's like closer than you think. - Full circle my friends. - Yeah, exactly. But Braid had to be sock to a compliant enterprise. We had sock to before we even raised money. So we were set up. Like we have a team that knew how to build the infra that meant we weren't like putting your personal information in a MD file somewhere. It is very properly like enterprise grade. And that is just by virtue of having done it before once. - Yeah. - And so we were very confidently able to tell people, hey, this is how we store a data. It's encrypted. We don't train any models on your data, which is pretty important to people. - Yeah. - And you can delete it. Like if you're like, hey, this is enough for me, it's gone. It's not like we sell it to anybody. So we make some pretty strong guarantees. And it's very important to earn people's trust. The second part of trust is like, does the thing, does Anna do what she says? She will do. And as you know with even like, Claude, sometimes not. - Yeah. - Sometimes not. It's not overtly wrong. But like humans, it's gonna misinterpret what you say. And when we use Claude code or co-work, you're constantly correcting giving your feedback. But the parent, the average parent doesn't have time and does not want to do that. Which is why, you know, when we stand out from a solution like open call, where you configure everything and you maintain it, you build it, people give up after two weeks. 'Cause it's just so much work to like coach and maintain this agent, which is like flaky. So we have a very high bar for where we have to be with getting things right. You can't get the sucker of time wrong. 'Cause there will be a sad child somewhere. Have we got it right? 100% no. We're always getting there. But the bar is, I would say, much higher than because it's a consumer product where it's your family's like schedule. We just put a very high level of high requirements for what we need to get to. - So break that down. What does that look like underneath the hood? At what level of sort of intentional rigidity around the models you use and the way that they're deployed and what's the trade-off that you've had to accept as a result of that? - Yeah, I think the, so if you, if you tell called raw, just like manage my life, it's not gonna get a lot of it right. It'll probably do like a 60% job. And then you'll have to do a lot of training and coaching and just perpetually live with some level of imperfection. The way we've structured Anna is that we do many, many, many levels of, we take in, obviously, in just all of your emails and your WhatsApps and your school app information to understand what's going on. And then we have many, many levels of, it's this noise, it's this signal, based on what we know about you, is this about your kid? 'Cause like, I get an email from my primary school and it's got 17 dates. (laughs) Only two of them are relevant for me 'cause my kids are in year two and prep. So like, is this relevant for you? Like, do you care about this? Do your kids actually do soccer? And then extract the right information. And we have a lot of guardrails around extracting things correctly, making sure like you check that something's not already on the calendar. So there's a lot of guardrails, but we also have a massive evil suite that is, and you built that all yourself. Built that all ourselves because one challenge is like, people start using Anna for all sorts of things. Like, we didn't expect. It's not just, yeah, tell me a bit about that. What are you seeing in use of behavior? The main thing that people use it for is to pick up two dudes that they would otherwise miss and manage their calendars. So like, the soccer match gets moving and I'll be like, hey, this moved you on mid-opted your calendar. But then you see people using it for researching their trips or figuring out what meals to cook or to research a bird they gift or to set reminders. Setting reminders is like a big thing as a parent. It's like, don't forget, I need to pack the Bluey costume tomorrow morning. - Always, just always have a good costume. - Always, just have it there. And add milk to my shopping cart. Find me like holiday camps for the kids. And this is just, it's kind of like because it's an agent, it can do a lot, but at the same time, we wanna make sure that the level of accuracy at which, and I can perform is very high. So we then have to like, we have evals around, do you get dates correct? Do you put white things on the calendar? Do you check what time zone you're in? Is it near you? Did we check that we got the child assignment correct? And then people, most of our pain customers actually have two people in an account 'cause it's two parent household. - That's a collaborative. - So then you have this collaboration dynamic of like, well, this email came in about your school, but my husband's the one that actually handles the school stuff and I'm the extra curricula's parent. (laughs) Anna has to like learn that. And so these are the kinds of like complexities. - Yeah. - So we have evals around basics like dates and times and getting details correct, but we also have like, did we read an image correctly? Did the assignment go to the right parent? Did a parent have to correct Anna? - I think the way we've landed with, and probably the most powerful part of this is when someone gets frustrated with Anna, Anna just funnels that right back into a ticket. I feel like hey, this person, you got something wrong. We're gonna figure out what went wrong. We'll create a bunch of tests that then perpetually live in our eval suite so that, and then there's an automated fix. That is what I mean about like, I think consumer that is really wonderful. You just get so much data, and you don't have to ask them. - No, they're using it. - And it's like legitimate data. - Well, and I guess with a product like Anna, relative even to a bunch of other consumer products, if they are choosing to use it, there are multiple touch points in a day, let alone a week. - Yes, yes. - So I imagine they're just amassing of that user behavior. Can you tell me a little bit about something you like, 100% believed was gonna be like a critical part of the product that isn't, and maybe something that like radically surprised you. So we always had this bad. When we first started, Mario was like, always should have is a phone number, and you call Anna, and you're like, talk to Anna. And it was just like the most futuristic version, right? Like you literally like call this person, and I'm like, no, no, no, no, like parents need an app. We're still used to like buttons. I think the truth is somewhere in the middle, I'm still like the app girl. and Mario was like, "Hey, we gotta live in the future." - Which is funny, right? The idea that the future is going back to a telephone call is kind of quaint. - It's all like SMS, I message and voice now. So what's surprising is people just do, mostly do their day-to-day work with Anna on messaging. We still have the app and people do use it, but I think it's like the world where people navigated by what is defined in a menu is no longer the case. And that's certainly for early use. - Changing very, very, very fast. And people just prefer to work with a product where they normally work in life. - Even if you would think messaging is pretty limited from a UI perspective. At some point you do run into the limit because if you're managing many, many different things in your day, you only have one thread. - You only have one thread and you're like, what is the way that thing goes? So at some point you need to go and have a place where you can find it again, but I've been surprised by how much of it is messaging. - Yeah. - So we are now doing a lot of work around Anna, how Anna responds in a message to be like concise and fun and interesting. We just didn't appreciate how important that was. - Well, talk to me a bit about that. So I'm always interested, I know there's always been a debate about like, why are all assistants female. Talk to me, that was obviously an intentional decision by you and I'd love to learn a little bit about tone and how you train a tool to be like helpful but not clippy, like kind of annoyingly helpful. - Yes, it's a hard balance. Actually, the name was funny because as we were on this journey of like iterating with as many ideas as possible and we were like, this is like the next idea and we were sure we were gonna like, we want to something else as we had not. At that point everything was like, you get a month of our life and then it doesn't work on moving on. So Anna was honestly like, we needed a name for the product and it wasn't my idea, it was actually GoGo's idea. He's like, how about Anna? Everybody, pretty much anyone on Earth can say it. - Yeah, exist in lots of languages and it's like easy to spell, easy to remember. And that was kind of it. I think we were at a point where we knew there were things that really mattered which was like, are you solving a big point point? And it's easy to get stuck on like, are we naming it correctly? Is there like a bad name, a good name? How does the branding look? That's all secondary to a solving a problem. So we just picked quickly what we thought was gonna, he's like, I need to ship this thing. What are you gonna call it? I'm like, pick one. When I woke up, he had like three names and I was like, I recommend so I'm like, go for it. And that's where we are. But you have like hobbies of the world and so on. So we didn't honestly like dwell on it too much. But the tone I think is interesting. We're only now beginning to appreciate how important it is that there is a personality there and it feels like someone was initially as like, really worthy, like oh my god, just stop already. So we've like tried to get her to be more concise and more to the point and we're still working on it actually. And how personalized is that to the user? Like if you get used to users that like give it to me and bullet points, can you stipulate? You know, we did lots of tests. We're like, every time you like respond to me and make sure you get at least four emojis. And you know, like she will do that. Yeah. So you can customize it, but there are some things when you're talking to someone on messaging, the personality matters a lot. Like just like there are people that you really like working with and there are people you don't. And it's not like anything. I wanna work with that agent more than the other agent. Yeah, yeah, and they, like some days like I just wanna work with Codex more than Claude. And I'm like, is this probably personality actually? Yeah, interesting. But like there are some things like you should always be concise. There should be, it's a fairly linear kind of medium. So you need to do things that are like, how do you make it more interesting for people? So it's like something we're actively working on. It's like always a work in progress, but it matters quite a bit. Yeah, that's so interesting. Do you hear people like talk back to Anna in a particular way? Like, that was really annoying. Yes, so this is great. This is, I think one of the best things about our product is like you never have to do user research. Ever. 'Cause Anna knows when someone's annoyed at them. Like you just, oh my God, like how could you get that wrong? Did you check my calendar? Why did you check my calendar? It's already on there. I already told you three times. Kate's super snarky today. Like we're gonna need to work a bit harder with her. That is the best feedback you could get when you're not watching and you're not interviewing someone. You just know, like Anna knows when they're frustrated and we'll go back in like, okay, we need to like get these things, we got these things wrong, we need to correct it. Yeah, yeah. And you do, you definitely see people getting frustrated. Super interesting. So we've talked a bit about like why also doing all of this IDMA's exploration and building Anna in this new phase, you're doing this all while the technology is moving so quickly under our feet. We've had folks on the program, including the Anthropic, folks Caitlin and Angela, who I interviewed and they were sort of their advice to founders was build just slightly beyond what the models are capable of. If you're gonna be building an AI native product that's reliant on a foundation model or even frankly any model, like build just beyond what they're capable of now, because you'll probably like by the time you start shipping and getting product in the hands of users, it will then be possible to do. What's your sort of take on that and how's that, I mean, has that informed the way you think about building Anna? Yes, we kind of experimented and learned this by feel, because frankly, part of when we were doing braid, the vision we wanted to land, the models weren't there, like agent agents weren't really, yeah, 18 months ago, yeah. So I would say we probably build a little bit more than a few months out, because for example, like one of the early bets we made was voice, that we will have the way that people will talk to Anna, primarily today's eye message or WhatsApp, but it will be voice. And early on when we were building voice models were terrible, they just were not good and people would try our voice and it would just be like, she would respond to you like 20 seconds later, you can't have a conversation. Yeah. But like, for four months after we had our initial version, Gemini Live came out and we had to do a lot of work still, but like the big jump in quality. So I would say don't build for now, build maybe six months to eight months out. And is that with any knowledge of what the, because it's not always easy to know what these, these labs are working on to be like, yeah, they've prioritized voice, so I can foresee we're gonna have something useful to go on. You're still taking bets, right? Yeah, you're still taking bets and you never really know, but you kind of look at the trajectory of what's happened in those sort of the last two years and you guess. But what happens is if you build too far out, consumer doesn't understand it and it's not a good experience. Right. Because you're marrying technical capability with like user adoption. And the voice example was a really good one. Like it was unusable and people just get frustrated because the tech wasn't there. And there are a few things like that today that are challenging. For example, like we're on a, when you're on Claude, when you're using Claude, you're on a laptop. You have access to your local browser, you're all your passwords are in there. When you're on a phone, we actually, it's very difficult to do all of that, like passwords, like an iPhone is very restrictive on like what you can get from other apps. And people are trying to solve it or we're trying to solve this problem, but it hasn't been solved very elegantly yet. There's a little bit of friction there that I would say today that we think will probably be gone in a year. But you still, so that's one where it's not the best user experience today because we're really on the edge of what is capable. Yeah. And you have to do a lot of work around tests and the evals and like putting in guard rails to make sure that that works well to compensate for the underlying inputs. So often like our evals suite is compensating, but the fact that these models are built for coding. They don't know that a soccer match moving is very important for you. So they don't inherently have that get things right in the same way they would ship code. Yeah, interesting. Yeah, that's a fascinating insight into also the subtle values that go into these models as well, right? Like, you know, if you're built for code, you have an understanding of the importance of like getting things right from a security standpoint because begin to price hates when, you know, there's an issue with security, but they're probably less likely to have that sense of like the interaction with the real world and the person and their experience. Like a lot of our testing and evals are trying to build a harness around coding models to make them work for humans in life. So I'm interested, you're like deep in the world of agent AI, which is so nascent. I'm really interested in how people are responding to this idea of coaching an agent to be the version that they need it to be. As you said earlier, some parents are all over at the Bay Area, parents know what they're doing, many others don't. What does it look like for you from a product roadmap? Like how much of that reality do you expose? And how much do you choose to keep underneath the hood? 'Cause you don't want new user caring about what's actually happening. Or like maybe a good example would be like, What is it? a user onboarding experience look like where what you're doing is extracting value for like which of the parents does the admin for school versus the extracurriculars or whatever while not getting in their way. Yeah, it's it's a big part of what we do and you have the open-claw parent who like they just love tinkering and they just want to do that but what we realistically see is they will probably do it for a few weeks and then kind of taper off. It's kind of a spiking interest. Our dream is that Anna just knows you like it takes about a minute to onboard and then like 10 minutes later Anna understands you enough to know what's important for you. Yeah, through context. Through context and she understands pretty much everything she needs to know about your family like 80% of it from the get go. And so she knows what tasks matter to you. She knows what dates matter. She knows when to move like which calendar to put things in. She knows like what allergies your kids might have and like what schools they go to and therefore can reason through that. I think we're in this phase of like having to coach agents and correct them. The dream is that that is not the case and like when you're getting close to hiring a human and a human does need a little bit of onboarding but they have a very fast ramp up I would say. So they need to know a little bit about your family but they know inherently like I shouldn't bother them about the discount they got from like their Pilates studio because yeah that's not relevant and this is just not that's noise. So our vision is to get to the point where Anna doesn't need much coaching and we do a lot of the testing is around making that happen but parents still do have to nudge her be like this is my preference don't do that don't do this at least if like probably for like a week or two. Yeah and then I imagine you run into the issue of like one parent has one preference the other parent has the other and it's like okay how am I trading off these against these uses is my best bet just to expose if why are you have different views on this thing exactly yeah fascinating. So I'm then interested in kind of what you're observing about where the real value lies for Anna. It seems obvious that like task management is like helpful and removes cognitive load which everyone is you know feeling burdened by. We have some products in our portfolio and some companies where actually the impact is like one level out is extraordinary so to use Holter as an example there an AgTech platform collars for cows you know AI for a farm discovered with their dairy farmers that being able to save them time in the mornings by bringing the cows automatically to be milked that's like 45 minutes more sleep at four o'clock in the morning can count for a lot they like picked up on these signals around the sort of improvements in well-being for farmers which have traditionally had this real issue around mental health challenges because the job's super demanding and relentless. Even lower rates of divorce I mean these are anecdata but lower rates of divorce amongst farmers and that's not making them a multi billion dollar company alone but that's the sort of fodder of people's real life experience with a product designed to solve a technical challenge that they have which is about farm management but ends up having this like quite personal effect. How do you think about you're already in the land of someone's personal life? How are you measuring or thinking about the medium to longer term value of what Anna could be to its users? If we could give a family like two parents or one single parent four hours back in their week but they were not constantly just thinking about did I do that permission slip do I have like what I would have done it tomorrow that's the dream and you know the the way we think about like timeline is like maybe today we save them like two hours but can we save them like 10 hours a year from now and what that looks like and this happens to me is I'm sitting at a dinner table with my three kids and you know we have like 20 minutes to kind of just like get everything eaten but like I'm always like okay well what what's going to go in their lunch box now and then like what's coming up tomorrow working life is so busy you only get so many hours with your kids and that time is kind of spent on this like mental load stuff that's what a lot of parents experience like so if you could just give them really 20 minutes that are just like talk to your kids that's the dream I think the other benefit that people should experience is like if you had a human and they were really good they they were lovely to just work with you would feel supported and I think that's where like the personality and the way she responds in the way you know she gets something doesn't understand exactly what you want but it's like clarifying feeling supported and knowing that somebody has your back I think is like the second thing so it's actually great to wake up in the morning but this is what I get to do it's like give people free time yeah that's really that's it yeah especially because you know the value of it yourself yeah it's like time is all parents want yep just like 20 minutes to go for a run indeed so I want to move to how you're actually thinking about building the company itself you're in this interesting position where I mean as you said earlier you've got a five person team three engineers out of Indonesia if I understand correctly how are you thinking about what your needs will be you you've got this meta experience where you've got agents that orchestrated to deliver anner as a product to your users you're obviously there for a pretty on the bleeding edge of understanding what agentic building of the company could look like and could deliver for you we've had dda from culture amp on the podcast this season talking about how his prediction is that you know companies there won't be as many thousand two thousand person organizations into the future they'll get smaller and smaller because of agentic AI what's your expectation about how you're planning to grow your own team the types of skills you're likely to need and how might that have changed even from when you started braid yeah I mean it's changed so much I we have a five person team and I think the product that we've built is five years ago would have been like I know like a hundred person team and you know we're a preceded company so we have to be careful with capital we leverage agents up the wazoo basically you just have like we basically run out of how many called code accounts we can create but there are different modes I would say you know like go go who is like a pioneer he was just experiment with so many things like he he has like the largest number of called code accounts going in parallel at any one time I think he once told me he's got like 11 it's like I think he has like running simultaneously at some point I think anthropic was like no you can't have anymore but his mode of operation is he's doing a lot of exploration all at once and then like on the other end you have like Mario who is like very carefully like he does like all of our payments and onboarding which has to be like taking all of that exploration and like refining it and making sure that the customer experience of the end of the day is like reliable and solid yeah so stepping back I think company our goal is to hire as few people as possible which I think is for me that was never like that's it that's an anathema as a goal exactly it's like we just want to hire as few people as possible yeah probably a good metric is like revenue per employee like how could you max out the revenue per employee I think we are not there yet at a point where in a year from now for example agents can do everything we're not there yet we still need really talented engineers but those engineers are more like architects who can coordinate orchestrate a bunch of agents yeah and then have the agents escalate to them if we see a customer problem recurring you're like okay I fixed it and sometimes agents will say they fixed it and and they'll be like here's the proof I fixed it but it's still not the best solution so we will look at when there are recurrences of problems and we'll escalate that to human or there might be an architectural decision that we need to make where I wouldn't be able to do that as a vibe coder so you still need really talented engineers but I do think companies software companies in particular will have dramatically fewer people because even things like go to market I can't run sixteen agents in parallel because those agents are not optimized to do the things that I want to do yeah and I and I think is a lot of attention split that humans are not good at yeah managing but I think we will get to a point where there are very few people needed to run an effective company and the more people you add even back in the day you don't go sometimes you slow down just because the communication overhead is so significant so you've run teams of like 150 right yeah at Atlassian so you know what it looks like to also run teams of humans trying to achieve a software outcome I'm curious like is 150 agents easier than 150 humans harder how are they different like you're still tasking both with work you're still checking to your point you're still checking the quality of that work yeah they still need to communicate unto themselves in some optimal way what's your what's your guess the agents are are easier in that like they're like a known quantity. They're all kind of the same. Like they all behave in a very similar way. You know where they're gonna fail and you know what they can do. And you can kind of micromanage them to the extreme. In a way that you can't micromanage humans and you shouldn't. - It's demolishing. - It's demolishing. What you can't replace is like 150 people who all have high agency and are driven and will like autonomously go do things. The problem is like in most large organizations you don't have 150 people who are like that and there's conflict and there's disagreement. Like a lot of the time in a large organization you spend just trying to figure out what to build and agree on it at different layers. That I don't think there is a, at some point you have to kind of agree on that and you still have that conversation. But I think what's become so much faster is we would like write a strategy and then like debate the strategy. I think today you just like build it. - The market will tell you. - Like build it and take it and see if like it's usable to a conversation rather than like debate on the theoretical. - Mm. - So I think 150 person organization can be, not today can be run with a five person team. - Maybe not to the first. - And maybe not even in a year but I think it's going that way. - Yeah, fascinating. - I want to come back to, you mentioned earlier, you're still pre-seed. So you're having to make a bunch of trade off decisions about where you spend your money. You may have phenomenal access to relatively cheap labor in the form of agents but actually then it becomes a question of where are you spending tokens. - Yes. - Talk to me a bit about what you've learned in the last six months and some decisions that you've taken at Anna to optimize for that. - This came so early, like having to think about costs that was super surprising. So our product, Anna specifically analyzes basically every piece of information that comes into your life, analyzes it carefully multiple times and checks against many, many different things. So our work is very token intensive. You might not see it but Anna might have to analyze 200 pieces of information before she tells you about one thing that matters to you. And so behind the scenes, you might have gotten five messages from her in a day but there's like non-stop work happening. So like most people, we started with like frontier models because they were the best. And they still are, I would say. But we're very quickly we're realizing that we were spending way more, at the time we hadn't even started monetizing. We just was like, oh my God, run away. - Yeah, the loss per customer. You were like talking revenue per customer. - Yeah, it was insane. The amount we were spending and because we're making like really careful judgments about what matters to someone's life, we were using quite a lot of like very strong thinking models. So very quickly we had to be like, this is not gonna work. One, we have to start charging people. Like five years ago in consumer, I think the play is that you get as many users as possible. You don't monetize, you don't worry about monetization as early. We had to start monetizing pretty quickly. I think we're very reasonably priced considering how much it costs to serve. But the reality is like, unless you're like Elon and you have this money you can throw at it, you're gonna be burning too much. And then the other thing we did very quickly is we had to migrate to open source models. - Yeah, so you came off the frontier. - Yeah, we came with the frontier. But hopefully around the same time that open source is kind of catching up. - Yes, and honestly it was really only possible the last five, six months, where the deep sea, the GLMs of the world are now, they're not exactly like even we can see that they are not exactly as good as Opus or Soul. But they're close and they're close enough and with an Eval suite, you can do well. - So what's the differential there between the sort of model where you were using the frontiers and the model where you're using the open source? - What we see, like we ran a test yesterday because we were measuring like how accurately there's Anna, do with certain types of complex tasks. And if like a frontier model can do like 85%, an open source model is probably like 83, 82%. - So it's close. - And cost wise, what's the differential? - It's like varies, but it can be like one-fifth of the cost to 120th of the cost. - Yeah, well. - And it's dramatically different. And I guess what is quite clear in a lot of products at the moment is the early days of AI exploration was like, obviously we're using the best frontier model to solve every task. Then it becomes apparent like, oh, the economics are starting to bite. How do I understand is this a real task that deserves, you know, does it deserve opus or does it deserve the lower tier model, even if you're gonna stay in the kind of frontier space? And I guess you are in a really interesting position to start taking a point of view on like, what's the task? Like how much of am I gonna wear myself through evals? And then how much just risk and imperfection am I going to wear on behalf of my customer for a task that I think can cope with like the coffee is of 4.7 here and a 4.6 there and it's probably okay. - We were pretty clear that we weren't gonna take much of a hit for the customer because we're already like, like we said, like consumer doesn't have a lot of patience for you getting it wrong. So that's why I think for us the test suite is so important because if you're like 10% off, that's not, that's still not good enough, I would say. So you are, we are now beginning to like think about different models of the same class, like thinking models and trading like an open source model for a frontier model but also you know, like speed matters a lot. So Anna and I was doing like, there was a time when Anna wouldn't get back to you for 30 seconds no matter what you asked, minimum. And like that's not okay, like people, that's just like not a conversation. So we've had to build this agent orchestration harness basically that has like, is able to have a conversation with you quickly and then like route things to the background for like longer. Talk the audience through a little bit about how you designed and architected that. - Yeah, so I mean, I guess Anna has two sides. One is we're passively ingesting all the information and then there's a set of agents or you could call it like pipeline that is thinking through all the information that comes through but then there's also the way that the user engages with Anna, which is a conversation and you can ask Anna at task calendars as well. And then that conversation layer early on, we were, we just had like the called code harness honestly and that's what was doing our MVP work but it was slow. It was accurate but very slow and people will be like, what like Anna doesn't respond to me for 30 seconds and it's not like when you're using called code on your laptop where you will go to often do something else while it's working. Consumer doesn't have that patience and on SMS or I message or what's up, it really makes a difference. So we've gone and we are now in like the third iteration of our agent harness, which is our custom harness and ideally we buy, we don't build but with so much of this we're so on the edge that there wasn't anything we could buy. So we have like a three layer agent harness where there is a conversation agent that's just fast and it's like it's the mouth and there's an orchestrator that is like, she wants research and a task and she wants me to like go check the calendar and she like what brain dumped all of that to me. So I'm gonna like queue these up and route it to the right specialized agents who will go to the work and some of those agents take time 'cause research takes time where it was like adding a calendar is faster but all of that is kind of handled so that the user continues to have a relatively fast conversation and we're always pushing to get even more snappy with it. - That's the sort of frontier for you is the sort of speed lesson sequestion. - Yes. - Yeah, so that's really interesting. I think we've sort of covered like even in the short, even in the short life of Anna so far. - Yes. - You've already run up against the limitations of AI technology as it's rolling out from all of the big players. We talked about token economics or tokenomics and we talked about the kind of harnessing layer and like what you're optimizing for. What's the next thing that you anticipate you're gonna need to build or the next buy build decision because it doesn't exist in the market? - Right now one of the, and this is not like a long term, this is like next week kind of thing. We've tried many iterations of how do you get an agent that lives on the cloud, not on your laptop, not on like a local machine and is operated on your phone to do work for you on the web. - Maybe explain parts of the audience that don't understand what, like they might imagine isn't that already possible? - Yes, that's a good, so let me give you an example. So if you want Anna to go find us a question like hey, when did the Roman Empire end? That is simple, that's like a web search that we can do anywhere. But if you want Anna to go, So add milk to your grocery cart. That is an app that you just doesn't have an API. Yeah, that is behind the logs or whatever. Yeah, that behind a login. And if you do it on your laptop, code code, the way that people use code do this would be like it's like a little browser extension and like code will go open it up and like go like click around and add something to the cart. But you can't do that on a phone. It's very difficult because phone is much more secure for good reasons and will not let you like do things in other apps that you don't own. And so when you want to do something on the web, which is where we get a lot of admin done like pay bills, book a doctor's appointment, get a reservation, you know, see what my subscription status is. It requires you to navigate to the right place, which absolutely does not have an API of any kind to log in, give credentials securely. Yep. And do you remember your password? I don't remember password. Anything. Yep. And so how do we store that and then how do we reuse that without asking you again and again? And then how do we like find possibly a janky website and like navigate around and get things down and get back to you? Yeah. That I think is a lot of people trying to solve this, but it hasn't been solved that elegantly yet. Interesting. And you know, it just reminds me like the surface area of Anna is gigantic because you've basically taken on the surface area of a human. Yes. And their utilization of tools to help them get through life. So it's extraordinary to me what you what you're doing with a tiny team. It's it's wild. I think like Mario is telling me like, I cannot like this product is so enormous. Like how do we get our hands around quality in all the areas that matter? And that's kind of the real real thing that we struggle with and focus on. We haven't touched on it yet, but I have to spend a little bit of time on delineating the person before all of this. So you have a PhD from Harvard in physics when and worked on a niche part of the large hydrogen collider. Talk to us a little bit about what all of that looked like with the delinear then have anticipated the delinear of now and what translates like what what about being a PhD level physicist has helped you become a product designer and a founder. It's funny because my I met my husband on the first day of U.D. so I was like 17. And I all I wanted to do was do theoretical physics and be a professor. And then he's like, what happened to you? False advertising. Yes, exactly. I think it's honestly like I'm probably an extreme example of like the reality of careers today, which is like you just have so many opportunities and you don't even know about when you're 17. I gentick. I was definitely not to age you, but I think you know unless you, and this was just like two years ago, which it wasn't didn't exist didn't exist. I think the common I don't know what the common theme is other than like I like to work on very difficult problems and like I don't know, I just like thrive off or feel happy when I'm like constantly challenged and probably feel a little bit stressed. I don't know if that's healthy, but like my husband says you just like to play life on hard mode all the time. I think that's a compliment. I think that's a compliment, but I think it stresses him out watching me sometimes. But I think it's just like I like to solve hard problems and I like to think about problems that haven't been solved before, which is what physics is like very first principle thinking. And that's somehow I think how I ended up after even after in product management for so long wanting to start up just like I, when I'm thinking about something I have not done before, it's just very exciting to me. I think that's the theme honestly. And the everything in between is when I was doing uni here, I wanted to be physicist and I didn't feel like there were that many other options. And when I went and did, when I went to Cambridge during PhD, I was like, my god, all these people like building stuff and doing stuff. What is the start up? And that's when I really discovered tech and I was like, this moves way faster than physics. Like it took me, it takes like a year and a half to write a paper, write one paper. Yeah. And like five people read it. And I'm like, it's not where I, and sometimes those five people are not even on the like Nobel Prize laureate list. So it's not even like you can, yeah, a lot of phenomenal researchers out there, expanding enormous effort, pursuing this sort of intellectual frontier in a way that none of the rest of us appreciate. Exactly. It never makes it. And like physics is like one of the oldest. So it's like so narrowed down and moved so slowly. And I probably learned something about myself in my early 20s of like, I like things that move fast. And so like when I went to med, I was like, okay, this is the right pace for me. This is like things move fast. That was just like a learning, I think about myself. Yeah. And so applying like what I'm naturally good at to that pace landed me here, I think. So you did McKinsey etsy meta Atlassian before founding really quickly, like maybe let's take those, you know, those are each storied institutions with their own cultures and their own ways of operating in the world. Most people have heard their names. What's the thing you've taken from each of those workplaces that informs how you work now even in a counter intuitive way? Yeah. I mean, McKinsey is the most interesting one because it was very challenging. And I probably wasn't cut out from like the way I work and my personality to be a consultant long term, but oh my God, everything I have learned about communicating, which is in my opinion, most of what a good product manager does is communicating well, sharply and asking the right questions. I learned from McKinsey. And that skill is still like, hey, I'm so bad at writing and like structuring. That skill is still probably one of the most valuable skills I have. It probably got me accelerated the rest of my career from Etsy, I think like customer like user experience, Etsy is a very beautiful product. Yeah, user experience and just watching people who are really good at that, how they think. Matt, I was probably like really my like where I learned what kind of product manager I am. And it is a, it's a place where you are given some ridiculous goal and you like go figure it out. Like I landed at Metta, I was in the ads team and my goal that year was a billion dollars in revenue. No pressure. No pressure. Did you get that? We did get there. And Gogo, who is on my team, was the engineer on my team. So that's the connection. Well, that's the obvious takeaway is just take Gogo with you. Exactly. Yes. It's one of those places where like the pace, they don't put too much emphasis on like product management principles. I mean, those things are there, but it's like, here's your goal, go figure it out. And I think like not everybody likes that, I, I appreciate the clarity. Yeah. And it was, it was fun. And then you did a lasian. And you did a lasian where I joined a lasian in the barrier and then moved back to Australia. And I think that lasian was a lot about me moving for an IC to a manager and figuring out, okay, how do I actually like learn to motivate people and motivate different people who have different, care about different things and learned a lot about, I think, the enterprise world. And that, which is why like when I wanted to find something, I was like, there's so much that needs to improve in the way we work. Yeah. So the other thread of Delani that may not be obvious, though, will be if you Google you because you're also famous for this, is you're a YA young adult author published. Like square that for me, theoretical physics, building a career in tech, and then you decide you know what I'm going to do, I'm going to go write novels for young adults. This was definitely a side hustle, which succeeded. Well, I mean, I, I hope one day to go back and write the sequel for it. It was like basically an escape valve for me, like working a matter of stressful and working last year was stressful too. If you like, in any place where you want to do well, and this was a way for me to just like, it was actually very closely related to physics in some ways, like because it's a very pure, yeah, like you build your own world and you like figure out what laws and rules apply there. And except this one doesn't have to pass any experimental verification. And actually, probably the publishing timelines are fairly similar, but maybe more than five people read it. Yes, slightly more, yeah, but probably not as many as I, you know, I aim for JK rolling, but yes, don't we all, don't we all. And anything that you think you've taken, because obviously the, the building of a world, like all of that is deeply creative. In this world of sort of agent, AI, you talked about earlier, like actually you can have all of these agents, you kind of know what the pitfalls are because they are sort of replicas of one another, at least at this stage, like what have you taken if anything from world building and the creation of an entire universe for readers that might inform the way you think about product, if anything? Yeah, it's all about does it make sense to the person using it at the end of the day? Right? Like they're actually pretty related in how you communicate with someone who is about to to get a varied. deep and complex experience, but they don't know what you're talking about on day one. So you have all these rules and all these structures in your brain about how the world's going to work or how Anna's going to work and what her like, agentex system is behind the scenes and she has an orchestrator and domain specific, like they used to think of any of that. They just need to experience and be never actually see any of it or understand any of it in depth. And so it's like this idea of how do you reveal just the right amount as you go? There are some fantasy books that are like all about explaining all the rules on all of that. I'm more about entertainment. Can you keep or use like engagement and entertain without them having to understand? Interesting. The machinery behind it. Delaney, I'm really interested. You've, because you're building in this particular moment in time and you are running right up against all the constraints of the existing tooling that you would use to build Anna, that's forcing you to build things yourselves in house that are that ideally you will buy and probably even in 12 to 18 months maybe purchasable. But right now you've got a problem to solve, so you've got to build it. Meanwhile, you've also got this huge data universe. As you said earlier, like actually, the scale of people's pain point is they give you access to a lot of information about themselves and that allows you to do this really customized and beautiful experience for your users. You also mentioned earlier that there are these like open-club parents. And you mentioned that part of the reason why you chose to build Anna was like a positive signal is some people are trying to build agents for themselves, sufficient of a pain point. So in the universe of all of those features, how are you thinking about long-term defensibility for Anna as a product? Yeah, I think the biggest thing here is the fact that we are going after parents and we really care about building something that works for parents. And getting it right, there are obviously a lot of agent platforms left right and send us open call, there's hermys, there's GrocBod, there's a bunch. But like life is complicated. So if you imagine hiring an assistant that's going to knock it out of the park for you, they have do a lot of different things, they have to understand you, they have to look at all your data sources, three reliable, they have to communicate in a way that makes sense to you, they have to understand when to talk to you versus your husband. There's a lot of nuance that you need to package into. I think this is like vertical AI for consumer in the sense that like Harvey is still going to like win I think in legal because you know, I've tried to use court to write contracts, but it's, it's never going to do, do what a lawyer needs it to do specifically. And I think the way that we need to package the way we analyze the data, the way we figure out what's noise and when to talk to you, the channels that we talk to you through, the agent architecture that like makes this happen behind the scenes, all the pieces that need to come together. I think make the experience seamless will mean that taking an off the shelf like or like a general agentic platform and making that suit your needs is always going to be more of a barrier. And parents are just very time poor. And they just want to think that works because this is not their priority to tinker. So I think that us just getting better and better and better and better at doing this really well. And parents like a large population of the world, right? So for us, if we just pretend we're set like we, that's what we just need, we need to get right. And I think I, I worry a lot less about whether like open air is going to come and release a parent product. Just because I know how hard it is to make it work so well for that category. So Delaney, our final question is always what's something you've changed your mind about? Oh, so many things. I think that's the theme of the last two years. One thing that I have changed my mind about with Anna, I think is voice and user interaction, like shifting from UI to a conversation or whether that's messaging or voice. And I was pretty on the UI side of it for a while. Let me build it a look beautiful. We'll put the button here. Exactly. And the thing that began to convince me was just like I started using this for flow. And I was like, I mean, it's changing very quickly to, it's unstoppable how efficient it is, like talk and type and have a conversation. And I think that was pretty dramatic change. It's actually interesting to think about how that will affect workplaces. Yeah. Because we're used to the relative silence of the click of people on there. And then they go to a meeting room to go have physical conversation with other humans. It's actually curious to think about how, you know, how we will work in a world with more like people talking to agents. Yeah. Yeah. Yeah. I don't know because we are remote mostly. So I no longer have not experienced an office where people are like talking their whisper flow or like talking to their court agents. That's kind of really good noise cancelling indeed. Delaney, thank you so much for joining us for another episode of Wild Hearts. If you want to learn more from other ambitious people building, designing and creating the world that we all want to live in, then please hit the subscribe and follow button. It would mean the world to us, the founders, the operators and the investors who join us on Wild Hearts. This podcast is a labour of love from Blackbird and our production team. The show is produced by Joel Connolly of Blackbird. Our marketing genius is Laura Coford and our editor is Andy Jones. Thank you all so much for listening and we'll see you all next week.