Use AI, or lose to someone who does [Peteris Erins]

Peteris Erins00:15

There are certainly areas where AI is going to take people's jobs. But in the short term, the biggest problem is that people who use AI are going to take the jobs of people who don't. So become a person who uses AI — and then you won't have that problem.

James Mulligan00:37

Peteris, why don't you introduce yourself?

Peteris Erins00:40

I'm Peteris Erins, founder of Auditless — a crypto boutique consultancy that's been operating since 2018. How I got here: I started with software-engineering internships in California, doing the very old-school version of machine learning — classifiers and that sort of thing, at Twitter. After graduating I moved into management consulting, a couple of years at McKinsey in London. When they acquired QuantumBlack, I moved into a product role there — that's where you and I met and worked together. I quit in 2018 to join crypto and start Auditless. Since then we've done two types of project: strategy work and implementation work, always a mix of a strategic mindset and a deeply technical one.

James Mulligan01:29

Can you double-click on the crypto piece? The generalist understanding is Bitcoin — but that's not specifically what you consult on, is it?

Peteris Erins01:41

At QuantumBlack, the thing I really enjoyed was developer tooling and open-source software. I had the privilege of working on Kedro, which was later open-sourced — and that's actually why I joined crypto. It wasn't the gold-alternative, cryptocurrency angle; it was that I was curious about how smart contracts work. Then I got excited about what they could be used for. Smart contracts power new cryptocurrencies, and more complicated financial instruments — prediction markets, exchanges, lending protocols. They let you create a financial layer where, to build a financial product, you don't need lawyers, you don't need suits — you just write a piece of code and deploy it into that layer. It's taken very different directions over the years, but that's the essence of what got me excited.

James Mulligan02:44

Nice. And recently Auditless has been exploring AI more broadly as a problem space, right?

Peteris Erins02:52

A few moving forces. One is just using AI internally more and more — in development work especially, but also in research. The other is that I write a newsletter every week, and my guiding principle is to write about whatever I'm actually interested in. I learned that back in high school: I had an extended essay to write, spent a whole week trying to pick a topic, and kept procrastinating by studying a data structure for competitive programming. And I thought — why don't I make that the thing? It worked out pretty well; I ended up in some national research competitions. So the newsletter is always what I'm actually reading and thinking about that week. We used to be about 80% crypto, 20% other; now it's flipped, and we're writing about 80% AI. So we're expanding to work with companies outside crypto — private equity is a focus right now. We're talking to a portfolio, getting to know the CEOs, digging into their problems. It's early. We're taking our offerings — the strategy side and the implementation side — and applying them much more broadly. On strategy, that's everything from protocol strategy — what to launch, how to go to market — through to narrative strategy for smaller companies: how they present themselves, how they pitch investors. On implementation, it's developer tooling, security tooling, and smart contracts, which since 2020 have been a major, major thing.

James Mulligan04:57

What problems are you seeing — where do you think Auditless ends up playing a role?

Peteris Erins05:13

It's funny when I contrast crypto and private equity. In crypto you have incredibly talented, incredibly technical people with incredible products, working in a very challenging market — one that's been good at bringing in revenue from the outside through speculation. Somebody creates a synthetic asset; it assumes value through speculation; somebody builds an exchange to trade it, and now they're making revenue — but that revenue is effectively predicated on the speculation. Private equity is the opposite. People are in real markets, with niches that are really interesting and not much competition. But technical capability is a lot lower, and AI adoption is a lot lower. Which makes it a fantastic place to be, because when you're already making money and the company basically works, there are so many opportunities to apply AI. With some CEOs it's taken ten minutes and we already have the first program. It's like: here's what you should do — and by the way, you probably don't need my help. I'll send you a couple of URLs. You've got one developer using Claude? Great — go ahead and do that.

James Mulligan06:46

Can you give me a couple of examples?

Peteris Erins06:47

We had one with customer support. A company said it took them four to five days to respond to support tickets. Two people minded the tickets, and some of the delay came from digging into customer-specific software configurations — the customers build their own forms, so to debug an issue you log in as them and work out whether it's their misunderstanding or an actual bug. Sometimes support gets stuck and waits for the developers. The interesting thing was the company was already using Claude Code — the developers were, but the support team wasn't. And I said: support is actually easier than coding, because the same things come up all the time, whereas in coding, when you build one feature you're already looking to the next. So first of all, give your support team Claude Code. They had some reticence — they didn't want to replace the team with a chatbot. But you don't need to go that far. All you need to do is give the team a draft of a great answer. Most of the time they'll press send; sometimes they'll edit it; but either way it saves time. And to make Claude a great support agent, you need a bit of configuration around tone and voice, some ability to learn from past tickets, and some access to those customer configurations — log in as the customer through a read-only interface, a terminal plugged into the API, and a skill explaining how it all works. So that was my recommendation. And I ended up prototyping and building an open-sourceable version of the agent — my last newsletter is about building and testing it. And it does work.

James Mulligan09:11

It's a fascinating space — big players like Sierra are commoditising the support-agent space, rolling out across call centres. What's your reflection on that being a defensible moat? What you described is a custom build that didn't take much time, and suddenly this company has something running. Where's the line?

Peteris Erins09:48

I was honest with them: there's really very little value I'm offering here, and they should just do it — and I explained exactly how. I also built an open-sourceable version, for similar reasons. What we're excited about is two things. One is finding what the future of strategy consulting is. I don't think it's the MBB approach. I think it's a lot more real-time, a lot more aware of data — internal and external — and not structured as two-month projects that take a lot of coordination just to define. Strategy should be an ongoing thing that just happens. The other is that on implementation we actually don't mind how easy it is — the easier it is to recommend a playbook or a tool, the better. The real problem is that people don't know which support tool to install, or how to configure it, or whether to build it or buy it. Those are the things we advise on. So we'll build software on the strategy side, and a lot of playbooks on the implementation side.

James Mulligan11:21

Interesting — so you're productising the service, the consulting, rather than the end-result execution. How about your own personal exploration? You're one of the most tech-leaning product people I've worked with, and an engineer at the same time. How is AI changing the way you work personally?

Peteris Erins11:47

It's an interesting question, because it's a bit like the transition from traditional mining — where people go down into the mines and do the work themselves — to machinery-based mining. Ask the person sitting in the machine how their work has changed, and they'll say, this isn't even the same work. That's the biggest framing change for me: I don't think it's the same work I'm doing anymore. Two things, though. One, it's made me more productive in all the obvious ways — I use a mix of AI across research and coding the most, plus some generative work with images and video. Images are a lot easier; I'm impressed by people who can conquer video. There's still room for taste and good workflow practices. And it's changed working with the team, because they use AI too. Managing people who use AI is a different thing from managing developers who don't. You're figuring out what your context file should look like, arguing about frameworks and best practices, and sometimes you catch yourself reviewing something that hasn't been reviewed by a human first. Because the team does much more, you review a lot more — so you start asking, should I be using agents to add leverage to my review process, otherwise I'm still one person reviewing three to five times the work. That's been a real challenge to adapt to, in a difficult domain: smart contracts have to be right. It's the worst place to be vibe-coding, because a mistake loses a lot of money. We've built contracts from the ground up that manage over $100 million in capital, and that's a scary place to be — you need to find ways to sleep at night. So we've built processes to use AI to speed us up without compromising quality. On a more strategic level, AI has completely changed how I think we should build the company, and what the moats are. One reason we started the newsletter was the 2023 GPT moment. Even before I was using much AI, it was clear the value of software was going to zero — which is a bit of an identity crisis for someone who spent years trying to be good at software. But figuring out the moat — distribution is something that's certainly not becoming less important. There are eight billion people on the planet; if a thousand of them are your true fans, that's worth something, whether or not AI exists. So when in doubt about where to spend time, we build more distribution, because that's always going to be relevant.

James Mulligan15:53

That's really interesting — everyone else can now build the same thing, but they might not have access to the same audience and channels. There are so many threads there. One I find fascinating: you mentioned mitigating the fact that you're reviewing code that might not have been written by a human. What does that look like? How do you solve for that, so the one human reviewer doesn't become a bottleneck?

Peteris Erins16:24

I absolutely name and shame people. I'll see something in a spec — the typical fancy language Claude tends to write — and I'll say, that's an interesting use of English from someone who's from Serbia. And it's funny, because most people say, well, actually, Claude writes better English than I do. And I say, no — it actually doesn't. I'd rather have your simplified take, which has the real essence of the problem, than the verbose, sophisticated-sounding Claude version. So that's one thing. The other is just having process for it. Luckily with smart contracts we already had code audits, internal and external. One thing that was unique to how we worked is that we never relied on the external audits. There are always external teams coming in to review the code — that's always going to happen — but it's very risky to do that without doing it internally first. So we'd spend a week or two turning our team into security engineers: they stop writing new code, and they go and try to find bugs. We literally have an 18-step process, just like an auditing firm — same tools, same language, same processes. We don't find everything, but we find the low-hanging fruit, which means the external auditors can focus on the deep stuff, and they have to push harder, because otherwise it looks like they've found nothing. And beyond process, it's setting the standard that you own the work — you're responsible for reviewing your own work.

James Mulligan18:30

The process piece is interesting — what process did the organisation have to check human-written code before AI came along? The places struggling now are probably the ones that had no process in the first place. Your other point, about calling people out, requires a familiarity with what Claude sounds like.

Peteris Erins19:03

To be fair, I really try to avoid the wording "it was written by Claude," because I use Claude too — I'll write things with Claude. So I try to focus on what it is I actually want to change. My teammates know me well enough to know that when I say "I don't understand what this means," or "this is far too complicated a way to explain this," I'm basically saying it reads like Claude. Always try to work out what exactly you're unhappy about — because that's how we add value into the process, by actually having an opinion.

James Mulligan19:48

There was something else — a paradigm shift I think people struggle with, me included. A miner used to go down the shaft and mine it themselves; now a machine does it. Their personal craft hasn't changed — there's a machine doing it, and they become the operator of that machine. How do you help people make that jump, from AI-sceptic to using Claude every day? How do you get them to see the value in becoming the operator, rather than the person who goes down the shaft?

Peteris Erins20:42

"Operator of the machine" is a wonderful way to phrase it. The work we all now have in front of us is finding and eliminating the bottlenecks in the process of spending tokens — a very different type of work from shipping deliverables. The biggest mistake people make is trying Claude for five minutes and going, phew, it's not smarter than me, I'm a better programmer than it is. They forget it's just a tool. The people who hate AI the most are the ones comparing their intelligence against the AI's. And I say — no, no, you're actually right, it's not more human or more smart than you, let's accept that's true. But what if it is just this dumb little tool? Are you smart enough to figure out how to use it and add value from it? You're ten times smarter than Claude? Great — then you should be smart enough to figure out how to use this thing.

James Mulligan22:08

Where do you think that insecurity comes from? There's a general perception that AI is going to take people's jobs. Is that the primary driver?

Peteris Erins22:20

There are certainly areas where AI is going to take jobs. But in the short term, the biggest problem is that people who use AI will take the jobs of people who don't — so become a person who uses it. Long term, even for myself, there's very little of what I do — apart from scribbling essays on the internet and having people feel like they know me — that AI cannot replace. There's a great YouTuber, BPS.space — he wanted to be a space engineer, applied to Tesla, got rejected. So he built model rockets from scratch, learned everything he could, built more and more sophisticated ones, and eventually managed to land them in an automated fashion, just like SpaceX. He ended up getting offers from those companies, and said, actually, no — I enjoy what I do. Now he has a YouTube channel, a big audience, and he sells the parts for the model rockets. So the bear case of my future is that I become a tech influencer basically role-playing being an engineer, doing projects for fun while people follow along. And I'm fine with that. I might eventually become a caricature of myself. But that's the realistic floor case — that the role of humans is to entertain each other. There's also a bull case, though. Aren't you excited by the prospect, James? We're on a podcast, so — so I think you understand.

James Mulligan24:23

I know, I know. It's partly commentary on what we're doing right now. And you promised an unhinged prediction — and you gave me one.

Peteris Erins24:32

So — yes.

James Mulligan24:33

Yeah — maybe we'll all just become entertainers for one another.

Peteris Erins24:37

I think that's certainly an important hedge, and something that'll increase over time. You'll still need yoga instructors in the wild — those people don't need to use AI. It probably takes away some of your credibility if you put Calm or Headspace on in the room. But if you're someone currently using a keyboard to produce all your output, it's possible you find yourself spending much more of your time on distribution than on deliverables.

James Mulligan25:18

Gosh, I've never heard it in such a reductionist frame — anyone who uses a keyboard, that's quite stark. Anyone who uses a keyboard could potentially be replaced by AI.

Peteris Erins25:36

Yeah — or your work could evolve.

James Mulligan25:42

What's your more optimistic view of the future? You've been building software for decades — what does the future of software building look like when these models finally do all these things as well as the people currently doing them?

Peteris Erins26:00

The optimistic case is that this is just a wonderful tool — a way to turn the planet's finite resources into knowledge output. But it has inherent limitations. It's really hard to engineer taste; it's really hard to engineer creativity. Talk to a designer and you'll find them a lot less worried than a backend engineer. And there are aspects of roles — product management, for instance — that have always been about understanding the user, the human, figuring out what they're trying to do and how to help them. Those aspects are probably going to stay very relevant, and maybe become more relevant over time. In the past you had companies like Apple, following a very taste-oriented strategy — there were always a couple in a given market, but maybe they weren't dominant. It's possible that becomes a primary axis of competition in the future: a lot less room for the Googles, a lot more room for the Apples. Google never had taste — it was good at shipping a lot of software and solving complicated data problems, not at innovating on its business model. Ironically, Apple is struggling with AI adoption. But the point is about what kind of work gets valued for humans to do — and I think it's compensating for some of the flaws of the AI models.

James Mulligan28:14

I really like that. A lot of the designers I work with feel a bit under threat, with engineers now shipping features and circumventing them via a design system. But I see a similar future — where everything progressively becomes a bit more beige, and the work that stands out is the genuinely creative stuff, with taste applied by a real human. That's what helped Apple win the mobile race: designing something aesthetically pleasing that people wanted to use. Let's talk predictions. You've touched on a few — what do you think the world of work looks like with AI?

Peteris Erins29:06

One big one: trading has always been interesting, because traders were always different people from the rest of us. It's competitive — you're not helping each other, you're competing. Your performance boils down to your P&L, in a cutthroat environment, and often you don't even have access to great tools. I think a part of our jobs may become a lot more like a trader's job. And this is a scary thing. If AI basically means every job becomes "how much compute do I get, and how do I use it effectively," then everyone with a Claude plan or a Codex plan has, as part of their job, managing their tokens — because they're going to have a budget. And that's exactly like trading. How are traders ranked on the floor? Each has a P&L, some way of observing how much they're making, and as your P&L improves you get access to more funds to trade with — aka more tokens. So a lot more jobs reorient around having some observability of how performant individuals are with their tokens, and then you get bigger and bigger token quotas. Becoming a VP won't mean you now manage seven people instead of three — it'll mean you now manage a budget of half a million dollars worth of tokens a year, instead of a smaller one.

James Mulligan31:02

A couple of interesting things there. One — AI is less like a tool and more like a resource now, almost the equivalent of headcount. And I agree organisations will start thinking in quarterly token budgets: how much does a team get, how much does a leader get. The trader analogy is fascinating, but it's predicated on each individual having a measurable outcome to determine performance — and that's not how most organisations are currently set up. Do you see a pivot there too?

Peteris Erins31:42

I do. CFOs are going to be asking the question: why is this person getting half a million dollars' worth of tokens a year — what are they spending it on? We'll get better and better at putting metrics around it. It's similar to product-market fit: the exciting moment for a company is when the lifetime value of a customer exceeds the customer acquisition cost. Before and after that, you're usually running a completely different company. The same is going to be true of tokens — the companies that figure out how to profitably spend tokens, and know that they're profitably spending them, are going to have a huge advantage. We don't have the language for it yet, the same way we once didn't have LTV versus CAC, even though it seems really simple now. We'll build those frameworks over time. In industries with complex creative projects — making a movie, working out its budget, seeing how it does — some AI challenges are going to feel like that: let's spend two million tokens on this and see how it goes. Other domains will be more defined.

James Mulligan33:02

I think you're bang on. Any other predictions?

Peteris Erins33:16

A spicy one. McKinsey's been around for what, a hundred years — seen as a fairly enduring organisation, with structural reasons why it could keep growing. But I think this is the most dangerous moment traditional consulting has faced — potentially the biggest existential risk it's ever had. It could actually fail to become relevant in this world. A couple of reasons. One: the transformation space was always something no other company wanted to play in. Whenever something new came along — lean manufacturing, say — there was a massive pool of revenue that just went to the consultancies, because nobody else was going to teach organisations how to do it. They extracted a major amount of value from every big technological or process innovation. The same thing started happening with AI: consultancies have been very relevant in AI transformations historically. But now OpenAI and Anthropic are creating transformation organisations, and there are SaaS tools where agents run around doing the implementation and chasing things up. Some of those companies are pretty well funded. Two: if you see consulting as a chain of knowledge work — a set of playbooks and experts, where you gather data and then synthesise it — you start asking, why isn't it real-time? Why does an analysis take a quarter? Why isn't intelligence part of how the parts are built? At some point it just becomes too slow, and the knowledge output of a fixed number of humans is bounded, versus the potentially unbounded output of a bunch of well-coordinated agents. Those two shifts are going to be really hard for consulting firms to get around.

James Mulligan36:17

Could any of these consultancies proactively build a proprietary set of knowledge that gives a better answer in real time?

Peteris Erins36:28

There's certainly a point where it's very easy to outperform what a consulting team can do, even with expert networks, because that knowledge is inherently limited to the people they have. I'm in the middle of a strategy project right now. This week I was curious about funding in the AI space by category, so I ran a bunch of workflows, found 400 companies, had an agent scan through all their funding events, and built a picture of the whole thing. That's a piece of work I never would have attempted as a consultant to do in one week myself — and it wouldn't have been easily outsourced either, because a lot of the time you either do it yourself or try to find a report where someone's done it, and the report is never exactly the thing you need. In consulting you often end up working with very limited data. You have an ideal picture — I want the 20 customer segments, the 200 competitors, the revenue each competitor makes in each segment — and you end up with a 10-by-10 table full of assumptions, and you make decisions from there. In a world with agents that can go and search, you get a lot closer to that ideal table a lot quicker. And that compounds the value of the process — you get there quicker, you make decisions faster. So these firms are going to start competing on very different axes.

James Mulligan38:17

Fascinating — so we're all going to become traders, and consulting's getting disrupted whether it wants to be or not. If you were to train all over again tomorrow — just finishing school, or just going to uni — what would you train to do today?

Peteris Erins38:43

That's a tough one, because I think about it a lot in the context of raising my son. My own training was a maths degree, and the way I explain why: I wasn't necessarily going into academia, or into a job directly using mathematics, but it felt like something where, if I could figure out something that complicated, I'd become quite good at learning things — and then I could learn whatever I wanted. That turned out pretty well, and I've always been excited to learn new things quickly. That feels like a core skill in this world, because it's really hard to predict — this isn't a linear transformation. When railroads came along, maybe the right person could have figured out the impacts. But with this ever-increasing scale of intelligence, it's really hard to predict the second-order effects. So being much more adaptive and learning a lot is great. Second, on distribution: communication becomes a lot more important —

James Mulligan40:16

Storytelling.

Peteris Erins40:17

Yeah, storytelling. Third: taste, having an opinion, creativity. And it's interesting what role maths, science and engineering play. You could take the stance that you don't need to learn to code, because Claude will do it for you. Or the opposite — that learning code and algorithms is just a really good way of thinking. A lot of consulting boils down to planning, and from doing algorithms I found planning quite easy. So there are aspects where maths and science are still very useful, both because they make your brain think in a way that applies to other domains, and because the intuition you build by doing something helps you use the tool better. You probably won't learn assembly language — that may not be as relevant — but you'll learn a powerful higher-order language and do some things with it, and that's useful for having intuition about what's happening under the hood when you're vibe-coding.

James Mulligan41:46

Fascinating — or we all just end up as YouTube influencers, doing weekly shows on our favourite hobbies.

Peteris Erins41:53

Hence the storytelling. It's funny — we mock teenagers who say "I want to be a YouTube influencer," but they were strategically right the whole time. What they meant was they wanted to do silly TikTok stuff, but what they should be doing is a podcast like this — and they were directionally very right. The younger generation always intuitively figures out what's going to matter next. Previous generations figured out that maths and science would really matter. The current generation figuring out that distribution and social media really matter — they're not wrong about it.

James Mulligan42:47

That's an amazing hot take, and it's starting to sound very true. Peteris, thank you so much — that was fascinating and hilarious at the same time.

Peteris Erins43:03

It's a fun world to be in, for sure.

James Mulligan43:06

Next time, then.