The days of making easy money from coding are over, says former Stripe UK CTO turned founder Emma Burrows

0:02Hello and welcome to the Sifted Podcast, the weekly show where we help you get to know the brightest and boldest people and companies in Europe. startup ecosystem. I'm Amy, Sifted's editor and your host. And today I'm joined by Emma Burrows, the co-founder and chief technology officer of Portia AI, a London-based AI agent startup. As I'm sure pod listeners know, AI agents, these apps which use large language models to automate tasks, are incredibly buzzy at the moment, with companies like Sweden's Lovable growing exceptionally fast as a result. But using AI agents throws up a bunch of challenges for businesses, especially around quality and security, which is where Porsche AI comes in.

0:41It offers business customers a library of tools for building and deploying their own AI agents in-house, while promising high security standards to keep humans in the loop at all times. The company raised£4.4 million in a seed round led by General Catalyst in April this year. And before starting Porsche just last year, Emma was the CTO of the payments giant Stripe in the UK, where she met her co-founder, Munir Mouad. Emma was also previously a director of engineering at luxury makeup designer Charlotte Tilbury, where she designed and oversaw the company's e-commerce stack. So a lot to get into today.

1:18Emma, welcome to the show. Thank you. It's great to be here. Thanks for coming into our London studio. So let's roll back the clock to start with. What was so exciting or so necessary, I guess, about the product that has become Portia that you and your co-founder quit your nice big tech jobs at Stripe to launch a startup and get, I'm guessing, a whole bunch of grey hairs. Yes, definitely on the grey hair front. So when Munir and I were kind of brainstorming ideas, we had a pretty different idea originally. And essentially what we wanted to do was build a kind of wealth tech business over the top of a bunch of different existing software.

1:57And we realized that this like intermediate layer of how do you connect software and use LLMs to drive that whole experience was pretty underdeveloped and nascent, particularly when it came to authorization of how do you grant an agent access to those applications in a way that respects people's privacy and security and is, you know, taking into account all of those things. And so that was why we really started Portia was to address exactly that problem. And can you tell us a bit more about what Portia does today and how you're working with some of your customers? Yeah. So Portia itself is built on top of an open source software development kit.

2:36So you can go on GitHub, you can look at all of our code that is open sourced. And the idea is that it's a Python library, so any developer can use it to develop applications. Now, we have kind of two main ways that customers tend to use it. One is we don't really know about them. We never have to know about them because it is open source. And the other is they use our kind of closed source system of evaluations or they want more assistance. And we kind of provide white glove services when it comes to how to deploy agents, monitor their accuracy, reliability, etc. And so you started the company a year and a half ago.

3:11Has the world of AI agents developed as you anticipated or what have been some of the surprises? Because it's just so, so fast moving, isn't it? Yeah, I mean, it's kind of dizzying to keep up with, I would say, you know, particularly being right at the forefront of it. Every day there's kind of a new announcement. So I think the things that have surprised me is both how fast it's moved, but also how far we've really still got to go to make agents production ready. So I think sometimes people feel because you can produce an agent demo so fast that therefore making it reliable is just, you know, as simple as normal software.

3:51But really it's a, the whole system is based on LLMs, which are like probabilistic, right? And so it's a very different way of thinking about and developing software. So it's both been surprising how fast it's moved and also surprising how far we've still got to go. And what are some of the key next steps for the world of AI agents and how, I guess, useful they can be really in the world of work? What will be the next kind of milestones for the sector to reach? So something that we work very closely with is something developed by Anthropic called the Model Context Protocol. And I've referred to it as MCP because that's how everybody refers to it.

4:28And it's really cool because it basically allows a company to say, this is how an AI agent should use my software. And it basically is a translation layer between APIs and agents. But the interesting thing about MCP is, on the one hand, people feel like it's exploded and there's loads of MCP servers. But there's actually not that many that are doing it in a way that uses the part of the protocol that is secure and uses the advanced authentication features. So we launched something last week that makes it super easy to add what is referred to as a remote MCP server. But there's probably only about 20 of them.

5:02And five of those are like random software you've never heard of. So I think in the next six months, you'll see that explode. And then people actually have a way of running agents that is able to use software that's kind of been well developed as opposed to MCP at the moment, which often is essentially running untested code, which is a serious kind of security vulnerability. And what are some of the main challenges right now that using agents throws up for some of the kind of customers, businesses that you work with? Yeah, I think it's partly just understanding the art of the possible and how to design what good looks like.

5:35Like, you know, what is a reasonable expectation of a kind of system that's built on a probabilistic LLM? And then the challenge is, you know, when to involve humans in the loop. Basically, like, the way I think about it is if you think about the internet and how it's evolved over the last 20 years, that's evolved for an ecosystem of two bodies, right? It's the machine and the human. And now we're very much evolving into a three-body ecosystem, which is agents, humans, and machines. and agents are like they're neither a one-to-one mapping to humans or a one-to-one mapping to machines and so nothing has really like adapted to take this into account and we're just trying to like bootstrap it on top of a bunch of existing paradigms etc.

6:16And what are some of the main use cases of AI agents and which are some that companies feel perhaps a bit uncomfortable using them for? So in terms of the use cases you know I'm not going to shy away from the fact a A lot of companies are looking to reduce their OPEX through agents, particularly in the more regulated businesses. So a rule of thumb that we often give to companies is if you have a policy that you've had to train a reasonably large kind of group of people on where they have to methodically apply a rule set in natural language and maybe configure a couple of things in different tools, then that's a great example of the kind of thing that you could automate with agents and humans or take some proportion of the mundane bits of that off a human's plate.

7:01So we have an example on our open source repo of a Stripe refund agent. And so what it basically is able to do is read an email, read a company policy, go, does this meet the refund policy? And then after it's finished that, it does all of the work to find the customer in Stripe, find the payment intent, perform the actual refund and then respond to the customer as well. So that's a good example of the kind of things that you can do with agents. And that's working well already today? Yeah, that already is working pretty well. I mean, I think when we talk about working well, the key thing is to be able to run it, you know, hundreds, thousands of times in a kind of mock way and really determine how often is it making the right distinction.

7:42And what are the gray area categories where actually you don't want the agent to make a decision. You want it to go, hey, I'm not confident of my decision. Can a human please review this? And so that's like one of the key kind of control structures that we try and enable. What's the next frontier? I think currently where we're at, if I had to describe it, is single agents are pretty well understood. A single use agent is, for example, the kind of refund example I talked through. So it has a very specific job to be done. Where I think we're getting to, and the technology is only really just getting there is this multi-agent system where you can just feed a whole bunch of different tools workflows etc to the same system and have it produce great results in a whole set of different use cases without having to like do a lot of very specific tuning of that of that specific workflow if you see what I mean for example customer care analysts.

8:38So it can do refunds, it can look things up in your knowledge base, it can go back to you when something about the product isn't working. And that is kind of a multi-use use case where it has multiple things, both actions that it can take, answers it can give, etc. How far off do you think that is? I think that's almost here with the right kind of technologies and frameworks. It's just getting to the level where we can prove and people are comfortable enough to actually like deploy it in that way. So I think that that is still kind of three, six months away, I would say. You mentioned about the impact on the workforce.

9:16How are you seeing both at your company and at people you know, companies you work with, how are product and technical leads thinking about how they grow and use their workforce as a result of all of these tools coming in? I think it's pretty varied. I think a lot of companies are trying to do it in a way that is very responsible and kind of thoughtful from what I've seen, you know. So take the kind of mundane parts of the job and delegate those to agents and look to find new ways within businesses to have those people help the business kind of move forward. But I think the implication of that, even in a responsible company, is over time, you know, maybe headcount doesn't get backfilled, etc.

9:58So of course there will be a workforce implication. And what do you think will be the broader or longer term implications of this? Does this mean there are fewer roles for engineers in the future or there are the same amount of roles because there will be so many more AI companies that need that? How do you kind of feel it's going to all unravel? I think it's difficult to see how it's going to play out completely. I think we already see the problem in the industry in respect to the kind of hiring of junior engineers because the output of a junior engineer versus an LLM applied to coding, actually initially the LLM applied to coding is probably more effective.

10:41And so we need to readjust our ways of kind of thinking about that. And I think there's going to be a short-term period where the industry readjusts its kind of best practices in terms of training engineers because even though an LLM can code at a kind of very junior level, They can't do advanced architecture. I mean, software like Lovable produces great kind of like software that fits in a particular format. But if you want to really push the boundaries of where software is going and how it works, then I do still think you need engineers. I mean, like as an agent company, we use LLMs to produce software as well.

11:22But the software still has a way to go on that front, I would say. What are you or what are other companies thinking about in terms of that junior engineer challenge? I guess this has been in some senses the kind of training of engineers has been a problem for a while, right? Because senior engineers have always been an expensive resource. And so I know companies have kind of not wanted necessarily to get their senior engineers to spend a whole bunch of time training junior engineers. And now I guess you've got this problem where software can do parts of their job for them. But, you know, eventually the senior engineers will retire and we will need new senior engineers.

11:59So what are some of the sort of training pathways that people are thinking of? I don't think people are thinking that carefully about it yet. That's why I feel like there is a period of kind of short term pain that we'll have to go through to kind of readjust the ways of thinking about this, I would say. Would you still recommend becoming an engineer as a career path for someone young today? I still would, partly just because I think if you're the kind of person that gets joy out of it, it's just the most fun thing to do. So why not align your job with something that you're passionate about? But I also think, you know, to your point about senior engineers, will it be as lucrative a career in the next 10 years?

12:42I don't think it will in general. You know, it's not going to be the easy path, I would say, that it has been. You know, if you're an average performing engineer, I think those salaries are going to get driven down as a consequence of the value of software getting driven down. And I actually think in some ways that's probably a good thing for the industry overall. You know, we've kind of ended up in this situation where big tech almost ends up on a bit of a pedestal. So I would still recommend it, but more for the do it for the joy, not because you expect necessarily the incredibly lucrative career out of it.

13:13Are you seeing it have an impact on salaries already? I would say not yet from our perspective, but we also take the philosophy at Porsche of, you know, we have a pretty lean team. We want a team that is very much able to think and drive the company forward. And so, you know, we're willing to pay for that small lean team. where I do think it will kind of have an impact is on those kind of junior salaries or folks who like they don't love the pure art of software engineering they kind of went through it as a kind of degree to find themselves type thing um we'll see how that that evolves first yeah we published um recently some data from the salary platform ravio which showed that roles for entry level engineers dropped 72 percent over the past year so yeah we're seeing that already what about Well, for people who are engineers, they love the job, they love the art and the skill of it.

14:09What do you think they need to do to future proof? Yeah, so the bit of this equation that we haven't talked about so far is the kind of disproportionate salaries that you see go to kind of AI engineers. And on the one hand, I think some of that is ridiculous. And on the other hand, I kind of get it because it is a very different way of engineering. You know, like we haven't hired many people who come from just like a very research heavy AI background. But there is definitely quite a steep learning curve of learning to deal with like a stochastic system under the hood. And sometimes it's like a bit exhausting, you know, like trying to work out in 3 % of the cases why something has gone wrong when you're used to being able to write code and have it execute exactly how you wanted it to.

14:50So I would definitely say like any engineer out there should be pretty familiar with how to use the large language models within software, be playing around. This stuff is so easy to tinker with now that there's really no excuse not to kind of try it and make sure you understand the art of the possible. How's it changed the way that you work? I mean, personally, I hadn't coded for a little while prior to doing this startup. And so it was quite fun to come back into being hands-on for a period. And I would say like the ramp up speed, you know, we use a lot of the engineers kind of have free reign of what they want to use from an AI perspective.

15:29I personally use Cursor quite a bit and ChatGPT quite a bit. And Cursor was very good, particularly as I was kind of getting back up to speed with like writing software every day. And now I probably have shifted more to ChatGPT as a consequence of like my issues are less about kind of day to day syntax and more about kind of how do I think about this? And I want to have a conversation with ChatGPT more as a like co-worker than because I need advice on the low level code, which is what Cursor is particularly good at. How have you found the transition from big tech to tiny, tiny tech and, you know, at least what half of the responsibility for the success of the company on your shoulders?

16:07I mean, it's been pretty wonderful, I would say, from that perspective. You know, like, at my heart, I love kind of building stuff. And so being able to do that in this particular time of technological innovation is just like the most incredible opportunity. but it's also like it's pretty exhausting as well you know trying to keep up with all the trends that are going on is exhausting trying to work out like what does this mean for business models what will stay and be sticky when everything is so easy to copy etc so it's it's generally been great but it's it's more of an emotional roller coaster than I than I predicted I would say.

16:45Where do you stand on or I guess what's your approach to how much do you try to keep abreast of what other companies are doing, what the big tech companies are doing, what the latest breakthroughs in AI all over the world are versus just focusing on what you're doing? I think we try and stay fairly abreast of it. I think that that is a natural consequence of being quite embedded in it as well. We want to make sure that for the content that we produce, it's really up to date with the latest trends of what people are talking about, thinking about. And so as a consequence, that just means you sit in a lot of forums where people are talking about this stuff.

17:24And so, yeah, it actually feels more natural than I expected to stay on top of those trends once you have a very good kind of base level of understanding of what is going on in the industry. And then we run like company hackathons every two weeks. Sometimes we dog food our own stuff. Sometimes we try and build something new. and sometimes you know we play around with something that's happening in the industry to understand what's going on there. What does dog food mean? Dog food is it's a kind of terrible google term where like you have to eat your own stuff that's like not really ready for human consumption.

18:00Perfect yeah and when you were at Stripe thinking about starting your own thing who did you speak to or what kind of conversations did you have what kind of thought processes did you go through yourself to be like do I actually want to do this will I enjoy doing this am I I don't know brave enough to do this does this make sense for all this other stuff I've got going on in my life what was that process yeah so um I've worked with startups for quite a while and I've wanted to do this for a while um and really the kind of time of life that I was in right I've got two two young kids was something that held me back and kind of financial state stability etc And like I was always pushing to do that kind of sooner.

18:40And my husband is always the like voice of being incredibly rational about decisions, which is useful. You need someone like that sometimes in your life. But at the point at which all of this change happened, it just felt like this is the time to go and kind of discover what you can do and build a business around it. And so it just was a whole bunch of different factors kind of coming together. You know, my daughter was nine months at the time and like I knew we weren't going to have any more kids. And I'd done four years at Stripe, which was a great experience, but felt like I was kind of ready for a new challenge anyway.

19:14As a female founder of an AI startup, you are quite a rare creature. Why do you think that is? I think I was lucky in some ways. I think my family in some way is just quite contrarian. You know, it's just always taught me never to listen to the background noise of what you're supposed to do. And I think that that is useful as well. But when you're trying to simultaneously parent and run a company, you know, there's a lot of discussion about things like mum guilt, for example. I don't really I don't feel it that often, to be honest. Right. I know I'm doing what I love. I know that's the right choice for me.

19:52And I'm not going to pay attention to societal pressure around that. And so I really think that has driven me to follow what I love without caring about whether society tells me that that is the right normal thing to do. And that's probably the most important thing because like society does tell women, I think, a lot that this is a male dominated industry. And there's various factors around that that are easy to let it get you down, I would say. Yeah, and I guess I feel like there's that whole thing of you can't be what you can't see. And unfortunately, it does feel like this huge rise of the AI industry has come hand in hand with having just very, very, very few female role models at the top.

20:38And so I guess I personally am quite worried that a lot of the kind of good efforts that were made to, you know, increase diversity in the tech industry in this new sector, which is where so much money and so much attention is heading, is not really being the fruits of that work is not really being seen, which just seems like a huge shame. Yeah, no, I totally agree. You know, like if I was to think about where I felt like the workforce was heading from a tech perspective five years ago, I felt like we were on a pretty good course. You know, there was definitely a lot more kind of roots in than there were maybe when I was starting out.

21:16And now I agree, unfortunately, like this is a bit of a repeat of the trend of the like 90s, whereby it was really the kind of most hardcore people that were able to make a career out of this to start with. And I think the same thing is happening again. And it's happening on top of a political landscape that is difficult from that perspective. So it's been pretty disappointing to see some of the trends around, you know, how much people care about this or how popular it is for people to care about this. So I think it's really unfortunate on that front. Yeah, and it feels like the kind of narrative, at least in spaces like LinkedIn, around this 996 hustle culture.

22:01I think I shared on Slack with a vomit face the other day, a founder saying, I went into the office at 4.45am the other day and my co-founder had been there 15 minutes before me. And I personally, again, I just don't really see who that is helpful for or who that kind of, I don't know, that showing, is it showing off? I don't know what it is, but I wonder, surely nobody can work that hard for that long. But it's creating this feeling of there being this gang that are working in this way. And if you're not, then you're not part of that. Maybe you don't belong is my kind of concern around it. Yeah, totally.

22:43It is difficult, even as someone that considers themselves a bit of a contrarian, to completely drown out that narrative. And unfortunately, I think it is true for a certain kind of, you know, you're just out of university and what you love is programming. and maybe you don't have a lot of other hobbies and this is the thing that drives you. So it doesn't feel like work to you. I work pretty much as hard as I possibly can with the combination of two kids and that doesn't equate to 996 because my husband would kill me if I worked every single Saturday during the day or whatever. But it does pretty much equate to I do family and I do work and there's not much else besides that And that's kind of an OK compromise for me.

23:28But I also think this whole thing around hustle culture, if you're working that hard every day, I'm not sure how you can continually make the right decisions. Because the other thing about running a startup is it's exhausting because you constantly self-question, right? And that's part of what makes you come out with the right answers is a certain amount of deep criticism of your own product. now trying to do that on four hours of sleep personally I find is not where I am then you know productive to work out okay this is the thing I stick at this is the thing the team needs to twist on etc so I think there's a lot of just like you grind because you have to but also just grinding for the sake of it and burning yourself out isn't isn't sustainable either you mentioned you've worked with a bunch of startups in the past and you've kind of I don't know if dabble is the right word but dabbled in angel investing you were part of blossom capitals angel program and the sequoia scout program which you're not part of either of those anymore i don't think the blossom one is still going but are you still an active angel um what did you learn from that experience seeing seeing the startup universe from that that perspective yeah i i mean i think i am still an active angel but not that active because like to the point about time and focus right i just have to kind of focus on the business right now.

24:46And so opportunistically, if something comes across that I think is super interesting, then maybe, but mostly I'm just pretty focused and don't allocate time to that right now. But, you know, programs like the Sequoia Scout program, it's just like a great way to learn. I think my better experience in terms of experience with startups was more on the tech advising side, which is either formal or informal, because then you actually really get a sense of what are the problems that startups have to grapple with but nothing can really prepare you for the difference between being in a full-time job and running your own your own company so did you think you were more prepared or more aware of what was coming for you than you in reality were you know what probably but also there's also kind of a joy in realizing again that you can build and that in some ways it's not that different as well from the day job right ultimately you're building a product you're trying to listen to users and so I have both been surprised by the emotional resilience it requires but also how similar to some jobs it feels um so it's like a weird kind of dichotomy I would say how did you meet the team at general catalyst your lead investor yeah so um general catalyst did a kind of dinner with with stripe kind of a few years ago now and so I met our partner there and then you know would kind of periodically go for coffee.

26:08And I eventually got to the point where I was like, I'm thinking about this thing. And actually, like, she was great, I think, for helping us work through a process. You know, they were the first ones to give us a term sheet. And I think that that was actually a brilliant kind of forcing function for us to be like, I think we could have just sat on our hands for another five months and done user discovery. But actually, you know, I think Juliet really knew what she was doing in terms of like the conviction that she saw even if like we would have gone through more user discovery to get there. And Juliette's actually announced hasn't she she's leaving General Catalyst to start her own company.

26:45She has yeah. How does that feel as a as a founder to kind of I guess lose in that sense your your original believer or investor? It does feel sad like no doubt about it we picked General Catalyst for a whole bunch of reasons but Juliette was a big part of that so I'm not gonna not gonna deny that but there's also something good about it as well because now we have someone I mean Judith's going to stay as kind of more of an informal advisor and we value her opinion and that also means I've got someone that understands the VC world and understands our business well who I can go to for advice without feeling like it's also an investor and so there's kind of a niceness to that as well.

27:24Makes sense. So we're recording this podcast in the second week of July and this is the week after a whole bunch of founders published an open letter in Sifted calling on the EU to pause the rollout of its AI Act. They say that the Act as it stands would hurt Europe's competitiveness and stifle innovation. Where do you stand? That's a great question. I'm probably pretty in agreement. I think we are still at the stage of seeing how legislation can carefully help AI. And I think it's important that Europe doesn't fall behind. So like if Europe was the only set of countries in the world, then I might feel different about the AI Act, but we're not.