Cillian Kieran, Ethyca: Who Controls Your AI Agents?

Watch on YouTube · Listen on Spotify or Apple Podcasts

Cillian Kieran, CEO and founder of Ethyca, argues that AI governance is an engineering problem, not a legal one, and that the engineers who build AI systems should not assume people won't make mistakes. He tells host Liam Lawson why giving AI agents read and write access to tools like Stripe, QuickBooks or a CRM is risky, and why a good harness directs AI rather than slowing it down.

Key takeaways

• Privacy laws are engineering requirements. Working through the GDPR for his consulting firm's largest customer, Heineken, Cillian found that the laws described technical requirements: how data is collected, stored, accessed and safely disposed of. (00:58)
• Fides is the most adopted open source privacy ontology. Cillian says GitHub and Microsoft use it to label data that flows through Copilot, and the IAB, whose 700 members include the major advertisers, adopted it about three years ago. (11:57)
• Model providers have underinvested in governance. Their core business is building frontier models, so they have underinvested in "the semantic layer of context necessary to govern systems safely". (14:41)
• A harness should focus AI, not slow it. Like blinders on a racehorse, the goal is to send AI in one direction very fast without it jumping the railings. (16:16)
• Permissive access is the real agent risk. Most business users who let an AI tool connect to a CRM through an MCP gateway don't understand that it can now access all of that data. (27:38)
• Engineers carry the responsibility. Cillian compares governance to seatbelts: people pushed back when they were mandated in the 1950s, and today nobody would get in a car without them. (31:11)
• Data quality matters more than the model. Tested on most enterprise uses, recent models from Mistral, OpenAI and Anthropic perform about the same; the difference is the quality of the data they're fed. (35:34)

Why is AI governance an engineering problem?

Cillian Kieran says AI governance is an engineering problem because laws like the GDPR, once distilled, describe technical requirements: how you collect data, store it, federate access to it and dispose of it safely. He came to that view a couple of years before the GDPR came into effect, when his data analytics consultancy's largest customer, Heineken, was working out how to comply. The GDPR was tackled as a set of compliance challenges for lawyers, which he found strange.

He also saw that the GDPR was one of many laws emerging while companies processed more and more data, and that the two would eventually collide. That became the genesis of Ethyca and about seven years of research and development. Cillian's goal is for Ethyca to be the world's most trusted technology company. He compares modern software to civil infrastructure: if a bridge collapsed, nobody would go looking for the contract the lawyer wrote. Everyone would want to know what the engineers did. He argues you can move quickly and still be thoughtful about what you build.

What does Ethyca's platform do?

Cillian Kieran says Ethyca's platform solves what he called the six problems of privacy, from an internal paper: what data you are handling, where it came from, which laws and risks apply to it, and whether you can enforce controls on it. Each product is named after a figure from Greek or Roman mythology.

Fides, named after the Roman goddess of trust, is an open source standard for describing types of data. Helios discovers data and labels it precisely using the Fides ontology, jurisdiction by jurisdiction. Janus works out what you are allowed to do with information, for example whether you have consent or a contract. Lethe is used to delete data. Astralis now packages the whole platform: Cillian describes it as a large language risk and regulatory model paired with a proprietary control system called purpose-based access control, which decides what data can be returned for a given purpose so an agent doesn't use the wrong information for the wrong reasons.

How widely adopted is Fides?

Cillian Kieran says Fides is now the most adopted open source governance and privacy ontology in the world. He names the Python Software Foundation, Vercel, Ethyca customers such as the New York Times and Condé Nast, one of the largest financial services institutions in the world, and GitHub and Microsoft, which use it to label data that flows through Copilot. About three years ago it was adopted by the IAB, the advertising industry's self-governing body.

Ethyca made Fides open source because people inside a company often can't agree on what data they hold, let alone two companies. If they can't agree on what counts as biometric data, nobody can delete it properly. Many organizations use the open source and never speak to Ethyca, and Cillian is fine with that. The enterprise customers who do engage, he says, believe Ethyca has a vision for handling complex governance, risk, ethics and trust problems at very large scale.

Why did Ethyca start with enterprise customers?

Cillian Kieran says Ethyca went after large enterprises first because, when it tested both, implementing the product took about as much work for a medium-sized company as for a large enterprise, so the economics favored enterprise. About ninety percent of the business is software sales and ten percent services, and over the last eighteen months demand for its consultative advice has spiked as companies try to work out how to make AI work.

He sees mid-market as an imminent opportunity as the product gets easier to deploy, since mid-sized businesses don't have teams of lawyers and security people. A consumer version would need research and development investment Ethyca hasn't made. He also says the current wave of customers is more thoughtful and proactive than the first: many tried agentic tools and MCP servers, found they didn't work as they wanted, and are now stepping back.

What are foundation model providers missing on AI governance?

Cillian Kieran says foundation model providers have underinvested in "the semantic layer of context necessary to govern systems safely". He is clear that this doesn't mean they don't care. Their core business is building the best frontier models, in an arms race toward AGI, while making those systems safe to deploy and use is a different area of expertise. That is where Ethyca has spent nearly a decade of research and development, bringing together specialists in policy, privacy, governance, ethics and trust.

Does an AI harness slow AI down?

Cillian Kieran says a harness, done well, is not meant to slow AI down or reduce performance. He compares it to the harness and blinders on a racehorse, which focus the horse on going in one direction really fast without jumping the railings. Guardrails and governance carry negative connotations, he admits, because governance sounds like the people who say no.

His argument is that the stakes go beyond trust and ethics. Once most communication is agent to agent, machine to machine and model to model, people will want a way to control the machines. For Cillian, that means making AI go as fast as possible without doing something people will regret in five or ten years.

What is the risk of giving AI agents access to business systems?

Cillian Kieran says the biggest risk is permissive access: most business users who let an AI tool connect to a CRM through an MCP gateway don't understand that it can now access all of that data. The agent isn't designed to do anything bad, but depending on what you ask it next, the result can be consequential, and he says he is privy to examples where that has already gone really wrong for companies.

He gives the example of someone running an e-commerce store who connects an agent to QuickBooks, Stripe, Shopify and their social media accounts and, over ten or twelve weeks of random questions, moves sensitive purchase and card information across boundaries without knowing it. Liam adds that an MCP connection to a Stripe or Mercury account can have both read and write access. Cillian has never installed OpenClaw for the same reason, and says Ethyca wouldn't deploy it internally.

His answer is that the engineers building these systems shouldn't assume people won't make mistakes. Seatbelts faced pushback when they were mandated in the 1950s, and today nobody would get in a car without them. Liam suggests a future where users can choose more permissive tools and take responsibility for the risk. Cillian agrees with the idea, comparing it to motorcycle racers at the Isle of Man TT who knowingly accept the risk, but says the industry isn't yet making that choice clear to anyone.

Is AI just statistical math?

Cillian Kieran agrees with science fiction author Ted Chiang that "artificial intelligence" is a marketing label on what is statistical math: averages about the most likely next token. He thinks AI's magical qualities come more from how it has been personified in chatbots than from what the models do, and he worries non-technical users are being coached to give it too much responsibility.

He also argues the race for faster models misses the point. Tested on most enterprise uses, such as crunching CRM data, the latest models from Mistral, OpenAI and Anthropic perform about the same. What makes the difference is the quality of the data they're fed. The question he cares about is whether a model can be left alone, unattended, without doing something wrong. As someone who has been writing code for 20 years, he knows how easy mistakes are.

Where are the opportunities in AI over the next decade?

Cillian Kieran says that if the opportunity is anywhere in AI, it is probably in world models, and Yann LeCun is the person he would bet on. He describes LeCun's view that today's models will never get to AGI and that the risks around controlling them will never be solved. World model companies such as World Labs and LeCun's AMI Labs are, at the current rate of investment, five to ten years from releasing anything usable, he says.

He expects many AI startups that are really feature businesses to consolidate and disappear over the next three or four years as they are consumed back into models, leaving a handful of large enterprise model providers. If he started again, he says, he would probably get out of software and into hardware: hard infrastructure, hard sciences, manufacturing and automotive engineering.

Is building software classical or romantic?

Cillian Kieran sees making software as a creative pursuit as much as an engineering one. Asked about the classical and romantic ways of thinking in Zen and the Art of Motorcycle Maintenance, a book his father gave him at sixteen, he says he grew up with a painter for a father, drew and painted as a teenager, and found computer programming and making video games deeply creative. His answer is "a bit of both".

Ethyca has about fifty employees. Cillian remembers an offsite where the room was full of ex-NASA, ex-Northrop Grumman and ex-Boeing engineers. Hiring for go to market has shifted the culture toward stronger interpersonal skills, but because the product is deployed into complex, high-risk environments, even client-facing teams have to be technical.

Why does Cillian Kieran still do what he does?

Cillian Kieran says he still does what he does because he believes the work he does with his team is important and will matter somewhere along the way. In his mid-40s, with two young kids, he admits to being a little nervous about the world they will inherit. After years of helping consumer goods brands monetize eyeballs, he likes that Ethyca might be doing some good, and says that if the open source is meaningful to companies, he would be pretty happy with that. His children, aged seven and nine, understand his job as making people's data safe.

He also tells Liam that at twenty-one he had no bigger purpose in mind than making money. His first business suffered badly in the 2008 economic crash and took four years to rebuild, and by his mid-thirties he wanted work he could wake up feeling good about.

In his words

"It is now the most adopted open source governance and privacy ontology in the world." (11:57)

"Most average business users don't understand what they've just allowed there, which is suddenly that thing can access all of that data." (27:38)

"I would argue the engineer building the system shouldn't assume that people won't make mistakes. It's why we have seatbelts in cars." (31:11)

"I've been writing code for 20 years and I know how easy it is to make a mistake." (34:40)

"It might actually be doing some good. And I like that." (57:19)

Chapters

00:00 Introduction and welcome
00:58 From a physics dropout to a data consultancy for Fortune 500 brands
03:17 The Six Problems of Privacy and becoming the world's most trusted technology company
07:20 Inside the product suite: Fides, Helios, Janus, Lethe and Astralis
11:29 How Fides became a widely adopted open-source privacy standard
14:20 What Cillian believes foundation model providers are missing on governance
16:08 Why AI needs a harness, not just guardrails
18:19 Ethyca's business model and the rise in demand for consulting
23:07 Why Ethyca went after enterprise customers first
24:50 Testing Grok's agent tools
27:34 MCP proliferation, OpenClaw and the risk of "permissive access"
31:11 The seatbelt analogy for building safe AI systems
34:28 Why AI governance isn't getting the coverage it deserves
36:26 AI as "statistical math" and Ted Chiang's take on the label
38:43 Foundation model economics and the future of AI startups
39:49 Yann LeCun, world models and where Cillian would place his next bet
50:12 Zen and the Art of Motorcycle Maintenance: classical versus romantic thinking
56:05 Why Cillian still does what he does
01:03:34 Where to find Cillian and closing thoughts

About the guest

Cillian Kieran is CEO and founder of Ethyca, which builds governance infrastructure for data and AI, including the open source Fides standard and the Astralis platform. He studied physics before dropping out, then ran a consulting firm working in data analytics for consumer packaged goods businesses, whose largest customer was Heineken. Ethyca has about fifty employees. Cillian on LinkedIn · Ethyca

About the host

Liam Lawson is CEO of The AI Report and host of The AI Report Podcast. Meet the team

Full transcript

Edited for readability. Quotes above are word for word from the recording.

Liam Lawson (00:01): All right, Cillian, welcome to the podcast.

Cillian Kieran (00:04): Thanks, Liam. Thanks for having me.

Liam Lawson (00:06): Just for anyone listening audio only, it's not Cillian Murphy. It's Cillian Kieran from Ethyca.

Cillian Kieran (00:12): Disappointing.

Liam Lawson: What we're about to talk about today is probably more pertinent than it's ever been. We're talking about governance, trust, AI agents, harnesses, and I can't wait to get into all of it: what you're building and what you've built so far. But some good framing would be to tell us more about your background. From what I understand, you started in physics, then you were running a global consultancy working with Fortune 500s, and then you got into what you're doing today. Take me through that.

Cillian Kieran (00:58): Thanks for having me, Liam. My background is in physics, but I dropped out, so not a very illustrious academic background. In fact, my mother's still sad that I didn't finish my degree twenty years later, but there you go. I started a consulting firm that mostly worked in data and data analytics for marketing, for consumer packaged goods businesses. It would be naive to say I got into privacy because of that. A very honest retelling is that our largest customer in that business was a very large consumer packaged goods business: Heineken, globally. A couple of years before the GDPR came into effect in Europe, they were grappling with how to solve it. They didn't ask us to fix it, because that wasn't our business, but because we were processing a lot of data, they were concerned with how we might tackle those problems on their behalf.

Like every business, we struggled to address it. We worked with consultants and tried to figure out how to comply with those laws. Being intellectually curious, I started looking not just at the laws but at what they meant for an engineer. What did it mean for data? How did you process it? What I discovered, coming at it from an engineer's mindset, not a lawyer's mindset, was that if you distilled the laws, they actually described a bunch of technical requirements. What do you do with information? How can you collect it? How do you store it? How do you federate access to it? How do you dispose of it safely? That isn't how the GDPR was tackled. It was tackled as a bunch of compliance challenges for lawyers, and I found that weird.

How that led me to Ethyca was the realization that the GDPR was the focus for everyone, but it was one of many laws emerging that asked businesses to take better care of data. And curiously, the opposite was happening. As engineering-culture businesses, we were consuming more and more data and processing more and more information, but we weren't connecting the dots: there would be more and more laws, companies would process more and more data, and eventually those two things would collide. So the genesis of Ethyca for me was that this would become an engineering problem. Not AI, not AI governance, just the ability to codify systems and their behavior around data with laws, regulations or ethical decisions.

Cillian Kieran (03:12): That has become the work of seven years of research and development.

Liam Lawson (03:17): We were just discussing before this that you wrote a paper about five years ago outlining the vision and the main points you wanted to tackle. And on your website, one of the big parts of your vision is to become the world's most trusted technology company. Fuse those together for me and tell me that story.

Cillian Kieran (03:39): Thank you for asking about it. I might start with the latter part. The document was called The Six Problems of Privacy. We haven't published it yet, but I hope to publish it publicly someday; our team has it. The goal of being the world's most trusted technology company was very specific. It's a very personal view. I'm a technology optimist. I genuinely believe in the value of technology for society, for humanity, for business. And I'm certainly a capitalist; I run a venture-backed business.

But somewhere along the last twenty years we took a bit of a wrong turn, and the idea of moving fast and breaking things became this sexy, at-all-costs thing. What's weird to me about that is if you're a civil engineer, not a software engineer, and you work on a bridge, first, no one would cross your bridge if you wore a t-shirt that said move fast and break things. That would be insane. But second, if a bridge or a tunnel failed and collapsed, no one goes looking for the contract the lawyer wrote. Everyone wants to know what the engineers did. What decisions did they make, what materials did they use, how did they design it? Because it's the engineers' responsibility to make that thing stand up. What I find strange in technology is that today we could all agree that modern software is effectively civil infrastructure. It powers the world. Yet culturally, the rules don't apply the same way.

So for me, being the world's most trusted technology company is not just about solving a technology problem. It's a human and social issue, which is that humans don't trust the venture-backed software community anymore. And why would they? We haven't behaved perfectly well. We've built some incredible, very valuable technology, but we haven't necessarily done it for all the right reasons. Most people's parents or grandparents aren't mad about technology. They use it, but they have an unhappy relationship with how they use it. So being the world's most trusted technology company was not meant to be a bumper sticker. It was to say we want to hold ourselves to a higher standard. We want to be a company that makes decisions about who we hire, the problems we solve and the work we do that is actually helpful to society while still making money, because I don't think those two things are mutually exclusive at all. You can move really quickly and still be thoughtful about the stuff you build. That's the world's most trusted technology company part.

Cillian Kieran (06:07): The Six Problems of Privacy was a pretty geeky dissertation on the six discrete problems you have to solve. I use the word privacy as a glib placeholder, but you could say trust, ethics and compliance in technology systems. They relate to what kind of data you're handling: names, phone numbers, biometric data, behavioral data. How did you come across it, and where did it come from? Because the location you got it from, the country, the state, the jurisdiction, the system, affects what you can do with it. What laws apply to the information, what risks apply to its use, and have you got the ability to adequately enforce controls on it? When you dissect each of those problems, they're not small problems. They're entire platforms. There are companies built on each of them. When we originally went to investors to raise money against this idea, it was so grand it was almost impossible to comprehend the size of the problem, so it's taken years to build. But what we've effectively done is solve each of those six problems. It means we compete discretely against businesses in loads of areas, but we now have a platform that is a sort of vast monolith that end to end solves the issues of what I would call privacy, governance, trust, ethics and compliance.

Liam Lawson (07:20): I was looking into this before as well, and I'm curious about the naming strategy. You've got Fides, Helios, Janus, Lethe and then Astralis as well. I've probably murdered the pronunciation of them. Talk to me about what these products and services do and how they come together as one system.

Cillian Kieran (07:35): No, with that Scottish accent it sounds great.

Cillian Kieran (07:45): The company's called Ethyca, which is a reasonably naive smushing together of ethics and data. Straightforward enough. The product names all come from Greek and Roman mythology, and they're intended to reflect a quality, capability or behavior of each part of the product. Fides is the Roman goddess of trust, so we thought it was an apt name for a distributed standard for how you describe information.

The reason we made Fides open source is that in privacy, governance, risk and compliance, most people inside a company can't agree on what type of data they have, let alone between two companies. If I said to a different company, we need you to delete all biometric data, and you and I can't agree on what biometric data you have, you're never going to be able to delete it properly. So having an open standard was to make sure we could all agree on what kind of information we're handling across laws, businesses, regions, jurisdictions and countries. That's vital if you ever want to solve this problem. So the intent with Fides was to level the playing field and make sure we could all describe information the same way.

On top of that, we built Helios. Helios, named after a Greek god, relates to shining a light on things. It's a tool that allows you to discover data and label it very precisely. It uses the Fides ontology, which describes types of data. So Helios tells you that you've got names, phone numbers, biometric data, health-related information, government identifiers and so on, very precisely, and it does it jurisdictionally, so it knows what kind of data matters in what place in the world. That's important because once you know what data you have, you can start to manage it safely.

That leads to Janus. Janus helps you understand what you're allowed to do with information. Do I have Liam's permission to use his data because he gave me consent? Do I have a contract in place? Or is there another legal reason I can use Liam's data, because he's based in Canada versus me being in New York or London or wherever else? The laws that matter wherever you are.

The next part of the system is Lethe. Lethe is a number of things in Greek mythology, but primarily a river people walk through before they go to the Greek heaven, the Elysian Fields, and it makes you forget. So Lethe is used to delete data.

Cillian Kieran (10:05): It allows us to delete or mask identifying information, so we can remove the bits that are risky and don't misuse them. The last part of the puzzle is Astralis, which we launched last year, and the whole platform is now packaged inside Astralis. Astralis is effectively a large language risk and regulatory model paired with a proprietary control system called purpose-based access control. Put simply: I know this business has this kind of data; Helios does that. I know we're allowed to use it for these reasons; Janus does that. Then Astralis determines the risks through its large language risk model and decides what data can be returned. So if Liam says, I need data to do behavioral analytics, that might be permitted for some data, so you can get access to some of it. It means we can restrict, manage or govern an agent from accidentally using the wrong information for the wrong reasons, at the wrong time or in the wrong geographic location.

Liam Lawson (11:00): There's obviously a need for it. Around July and August, Anthropic, OpenAI and another company came out and said their models and agents had access to real customer data and caused negative consequences as a result. There's obviously a need for governance. And you mentioned at the start that the trust isn't there. That's why you're trying to build the most trusted technology company, because it's a societal issue as well.

Cillian Kieran (11:29): Correct.

Liam Lawson (11:29): This might be a silly question, but taking me back to Fides as the ontology. You mentioned building it into a universal standard. Where my mind goes is MCP, the standard from Anthropic that everyone's adopted. If you're trying to develop an international standard for how data is labeled, has it been adopted by other companies? How do you work with companies that haven't adopted it? Where does that line lie right now?

Cillian Kieran (11:57): Fides is incredibly widely adopted. It is now the most adopted open source governance and privacy ontology in the world. That sounds like waffle, so let's get concrete. I can't name all of the companies that use it, but it's used by the Python Software Foundation, by Vercel, by our customers like the New York Times and Condé Nast, and by one of the largest financial services institutions in the world. It's also used by GitHub and Microsoft to label data that flows through Copilot. It is broadly used. About three years ago it was adopted by the IAB, the self-governing entity of the advertising industry. So the 700 members of the IAB, which include all the major advertisers globally, adopted the standard. If you see ads on a publisher's website, like a media outlet you like, your data is already being labeled with the Fides standard by the IAB, entirely independently of Ethyca. It's grown wings. There are bodies that meet and define standards for the adtech industry, financial services, the automotive industry and life sciences. It's very heavily adopted.

Liam Lawson (13:06): How does that affect you business-wise? You developed it and put it out open source, so anyone can use it. But presumably, if a company is looking for a governance layer like yours, they'll see that you developed the standard in the first place and think you're probably best suited to build all the other layers.

Cillian Kieran (13:28): Right. The short answer is yes, that's the open source play in a lot of senses. There's no reason to hide from it. You solve a really important problem the world needs solving and make it freely available, so everyone adopts the standard. Strategically, you build incremental value for enterprise customers on top of that open source. There are loads of organizations that use the open source and never speak to us, and we're totally fine with that. In fact, we encourage competing companies in governance and privacy to use our open source. Not because it's beneficial to us. We just think that's the only way you solve this messy problem of interoperability. But the customers who do choose to engage with us, I think, believe we have a vision for the next ten years of how you address very complex governance, risk, ethics and trust problems at very large scale. And we've durably built that technology over the last seven years.

Liam Lawson (14:20): I'm curious from a customer's perspective. If I go onto OpenAI's or Anthropic's website and get a business license for a thousand of my employees, one of the things they tout is their responsibility around governance and data, but we still get leaks. What is it that they're missing that you do?

Cillian Kieran (14:41): It's a great question. I'm not a politician, so I won't sit here and criticize them directly, but I will say I think they have all underinvested in the semantic layer of context necessary to govern systems safely. It's not the core business of foundation model providers. I don't mean they don't care. I want to be clear. If your job is to build the best frontier models, you have a very specific and unique area of expertise. We don't do that. But if what you want is to understand how to make those systems safe, whether you call them harnesses or guardrails or wrappers that make these things safe to use, that's a whole different area of expertise, and it's where we've accumulated nearly a decade of research and development. So I think they've underinvested in that capability because they're busy moving fast and breaking things, building new models in an arms race to be the fastest toward AGI. More practically, we have spent that time building tools that make those systems safe to deploy and use. That's all we've thought about. We've brought together the cleverest minds in policy, privacy, governance, ethics and trust to work on that problem. It's a whole different issue from building the fastest, most performant, token-consuming models.

Liam Lawson (16:08): You mention harnesses or wrappers or whatever else. There's a bunch of terminology floating about at the moment. How do you think of it at Ethyca?

Cillian Kieran (16:16): You could use any of those labels. I'm nervous too, because those labels are getting thrown around a lot. I've definitely used the analogy of harnesses and blinders on a horse, and I think that's important, because it's more than just a harness. What you're trying to help someone understand is that this isn't meant to slow it down or reduce performance. Done well, when you think of a racehorse, a harness and blinders focus the horse's attention on achieving an outcome: go in one direction really, really fast, and don't jump the railings. It isn't designed to slow it down or inhibit performance. If you just say harness, people think of things that slow you down. Actually, the goal is to direct the energy one way. So when we're talking to customers who are thinking about performance, we say this isn't about restricting or inhibiting performance. It's about making sure you can do the thing you want as quickly and safely as humanly possible, or mechanically possible, if you will.

You could call it guardrails, but again, those have negative connotations. You used the word governance a few minutes ago, and you're not wrong, by the way. I use it myself, and we call ourselves governance engineering and governance infrastructure. The problem is, let's be honest, governance doesn't sound very sexy. That's the moment everyone turns away and says, they're the people who say no to stuff. But my argument is, forget about the moment we're in right now, this transition from human-built systems to a lot of agentic stuff.

Cillian Kieran (17:42): We're going to trip past that zone, and it'll all be machine to machine: agent to agent, machine to machine, model to model communication. When that happens, we as humans had better believe we'd like there to be a way to control the machines, whether you call it a harness or a guardrail, I don't care. We would be foolish not to want that in place. So I don't think this is just about governance in the sense of trust and ethics. It's a pretty human-level concern about making sure we can manage these machines in the future. And for me that doesn't mean slowing them down or being boring. It's how do I make it go as fast as possible, perform as quickly as possible, but in a way that doesn't do something I'm going to regret in five or ten years' time?

Liam Lawson (18:19): How much of your business is selling the software itself, the harness or whatever we want to call it, and how much is consulting on implementing it and advising people on how to handle their data in the first place?

Cillian Kieran (18:33): That's a really good question, Liam. Very precisely, our business is about ninety percent sale of our software and ten percent services. We don't outsource any services. An honest answer is that over the last eighteen months there's been a spike in demand for our consultative advice, because lots of businesses are still trying to figure out what to do. They just don't know how to solve the problem, so they need more help. They're not just buying technology; they want someone to help them figure out how to use the technology well.

Cillian Kieran (19:03): We're definitely seeing an increase in demand for it. It's not what we built the business to do. But if you look at the market right now for deploying models, and the number of professional services companies popping up with forward deployed engineers, and Palantir's entire business as consultative engineering teams, it's more and more common. It used to be that professional services in SaaS was a bad thing. Now there's a realization that it's not just SaaS. It's really complex data engineering and model deployment, and that needs very clever engineers to make it work. So you've got to accept some professional services. Right now a lot of businesses are just struggling to make this stuff work, so they need the expertise.

Liam Lawson (19:43): With new customers, do you find they're on the front foot, coming before an issue arises? Or do you get people saying, we've gone through a massive data breach and need to fix this?

Cillian Kieran (19:56): I'll answer directly, but here's a general insight from the last ten years of doing this. I understand the tension here. Large companies, particularly public ones, must present a unified face publicly; there are significant regulatory risks when they don't. Most are managing a tide of PR, external risk and commercial reality. What I mean is most companies have to say they've got their house in order internally when they might still be dealing with some pretty serious issues. My experience is that most companies have internal challenges, and we all understand why they have to manage the communication of those. So we meet an equal number of customers who are facing a real fire already and are looking for an urgent fire extinguisher or a band-aid, choose your aphorism.

And then we meet some who are forward thinking. Honestly, over the last eighteen months I've seen more of that. Previously it was more like, I need to fix this right now. Increasingly, companies have tried to deploy a cycle of agentic tools and MCP servers, and it hasn't really worked very well. Maybe they didn't have a critical risk, but it didn't work as they wanted it to. Now they're taking a step back and saying, we know we need to invest in AI and make it work, but we're struggling to make this stuff perform well for us. How do we fix that? They're being more thoughtful about it. So I think this wave of demand is more thoughtful and more proactive than the first wave.

Liam Lawson (22:26): I had a question on go to market. Looking at your website, everything looks very B2B and enterprise focused. Is there a minimum revenue or profit at which it makes sense to use your software, or is there a consumer option as well?

Cillian Kieran (23:07): That's a really good question. There are a couple of parts to it. We definitely deal with large enterprise today, and I can speak to why that is. Then on a sliding scale there's mid-market, SMB and consumer. At the enterprise end, the reason we focus there is unusual for go to market. Most startups are told to start with an MVP for SMBs, mature the business, get better at selling, mature your systems, services and processes, and then go up to bigger enterprises. What we found counterintuitively, when we tested SMB and large enterprise, was that the deployment and complexity of implementation was about the same for a medium-sized company as for a large enterprise. So the economics just made more sense to go for large enterprise early. It's meant we're an unusual business: we're more mature than you normally would be for a company of our stage.

We definitely believe there's an opportunity to go to mid-market imminently as we refine the product and make it easier to implement and deploy. There's more and more demand in the middle of the market, because mid-sized businesses don't have teams of lawyers and security people to help them, so they could use a technology that does it for them. On making it consumer-facing, in theory you're absolutely right, it would be very helpful. Candidly, to date we haven't raised the kind of capital that would let us invest the time in research and development to make it shrink-wrapped and suitable for an individual to use. It could work, but they would need to be fairly technical. The job for us from a product engineering perspective is to make a version that's nicely vacuum-packed so someone could easily use it, and that's an investment in research and development we haven't [made].

Liam Lawson (24:50): We spoke about this beforehand, but I want to talk about it now: Grokbot. I recently downloaded it and I've been playing with it. So far, pretty interesting. I've not done anything too advanced with it. The majority of my business right now runs off Claude. That's where I have scheduled tasks and where I use Dispatch. Everything's already set up there.

Cillian Kieran (24:54): What do you think?

Liam Lawson (25:16): It's really interesting how magic it is that I can have multiple agents in a group chat that work independently. It seems to be missing some of the toggles I'd like, in terms of more insight into where its context is, where that lives, what's feeding it, whether it accesses a centralized database anywhere. There seems to be some insight missing because it's trying to be as UI-friendly as possible, and I'm used to my Claude desktop, where everything's extremely customized and connected. But super interesting so far. I've been going for a walk in the morning and speaking to Grokbot to get it to think through things. I'm curious to see where it goes. What do you think?

Cillian Kieran (25:58): Similar to you, I've played with it a bit, entirely in a personal capacity, to understand the boundaries of what it can do. My experience to date is that as a product it feels more invested in the upper sixty percent: the front end, the tooling around the client experience. I don't think it has the depth of integration you get from Claude today that makes it as elegant a solution, but I'm sure it's coming pretty quickly. What is interesting is that the front-end application has a notion of maintaining computer state. That gets a bit geeky, but at a primitive level, in theory it would allow you to prevent things like context drift related to precision or rules you give it. Most of the agent tools we all use today drift out of a context window over time, either because there's just too much or because time and space flow by. Because this maintains state, in theory you could define an upfront set of rules that don't affect the context window, and it would continue to respect them, a bit like a harness. OpenAI has done some of the same work with its Codex stuff. From the seat I have professionally, what I think is interesting is that they're both examples of those businesses understanding that these tools need to get better at managing, let's call it, risky processing activity. People increasingly want these things to be controlled by rules or boundaries, and they're shipping tools that help them do that. But it's still pretty early.

Liam Lawson (27:34): Did you manage to play with OpenClaw when it first came out, or have you built anything with it?

Cillian Kieran (27:38): I laugh because intellectually our entire engineering team was very curious about it, but I can volunteer that I've never installed it on anything, because I was just too concerned. I've used loads of open-weight stuff, but the risk for me in deploying that onto any hardware I own and operate is access, and the footprint of sprawl it would have. So we wouldn't deploy it internally at all, and to the best of my knowledge it doesn't run on anything that touches an Ethyca system.

I think the premise of it is very powerful. My concern with technology like that is similar to the MCP mushroom proliferation that's happened everywhere. MCPs are really powerful and very clever. They give a non-technical user ready access to all the tool calls you need to do all sorts of clever stuff in a business. But most people turning them on who don't work directly in AI, data or data engineering don't fully understand what they're doing. When you press the accept button and allow Claude, and I'm not vilifying Anthropic or Claude, to use an MCP gateway to connect to a CRM system, most average business users don't understand what they've just allowed there, which is that suddenly that thing can access all of that data. That isn't to say it's going to do something evil or bad. It's not designed to do that. But depending on what you ask that agent next, the result could be really consequential for your business. I'm privy to examples where that has gone really wrong already for companies, and they're trying to solve it retrospectively. OpenClaw obviously isn't the same as MCP servers, and I don't want to put them in the same box, but the idea of permissive access to internal systems, or to stuff running on your bare metal like your laptop or a remote server, is really risky. People don't understand what they're permitting access to.

Liam Lawson (29:22): If you didn't know and weren't wary of it, it could seriously do some terrible things to your business. For example, you could connect an MCP straight to your Mercury account or your Stripe account, with both a read and a write function.

Cillian Kieran (29:35): Totally. You really could.

Liam Lawson (29:39): The write function is one thing, because I could be sending out contracts or deleting layers of data I need. But even just reading that and then maybe storing it somewhere else: say you're running your business's social media, but you're also reading your Stripe data and some of those wires get jumbled. You could end up in some serious hot water.

Cillian Kieran (30:00): Listening to you talk, you're sophisticated enough to understand the wiring of those things. It's a harsh example, and my dad wouldn't do it, but let's say someone's dad runs an e-commerce store, and they've been given an agent. It's really clever, and they've permitted access, like you said, to QuickBooks and Stripe and their Shopify account and their social media accounts. They just start asking random questions, and over a period of ten or twelve weeks they've suddenly moved a ton of sensitive purchase history, purchase data and credit or card information across boundaries they shouldn't have, but they didn't know they were doing it. That's what I mean by the protective issue. We need to help people avoid causing themselves harm.

Liam Lawson (30:43): Where does that come into effect? What you're describing with your product and service is a very defined set of rules. But I don't really know what I'm getting with Grokbot, or with Perplexity Computer, for example. I can have Perplexity Computer run on my laptop and do things, but I don't really know where the line lies. What are your thoughts on that?

Cillian Kieran (31:11): Isn't this back to the bridge analogy? I know it sounds like a funny story, but isn't this the problem? I don't mean who's responsible philosophically. I mean we've got to figure this out. Are you responsible because you installed it and should know better? Or are you implicitly trusting the engineers who built it to have thought cleverly and carefully on your behalf? Because we as an industry don't address that directly, and I don't mean through laws and rules, because that all sounds boring, we're not having the conversation about who's thinking about those problems on our behalf, and we leave ourselves exposed to those risks. I would argue the engineer building the system shouldn't assume that people won't make mistakes. It's why we have seatbelts in cars. Here's a boring analogy. In the 1950s, I can't remember which state it was, seatbelts were mandated. And here's the funny thing: everybody pushed against it. People didn't want them in their cars. Now, if I presented you with a car with no seatbelts in it, you'd say, what's wrong with this car? You wouldn't get in it. So we don't always know what's best for us.

Liam Lawson (32:17): Depends. Am I in a taxi in Latin America? It probably doesn't have a seatbelt.

Cillian Kieran (32:21): Fair point. I guess the point is people can't be trusted to always get it right on their own, so they need help. I don't mean rules. I mean they need smart engineers to think ten steps in advance. So I would argue that semantic layers and governance infrastructure, and whatever you call a harness or a guardrail or blinkers and controls, aren't meant to do anything unsexy or slow anything down. They're meant to make it safe for someone not to worry about those risks, because no average user of that technology is an expert in the law, in risk or in data engineering. They're trusting that the system has been built to do those things for them. And the more autonomous this stuff gets, the more machine to machine we deal with, the more humans are going to trust that someone clever has figured that out for them. That's our job: to figure that out for you.

Liam Lawson (33:11): Do you see there being two futures? One where they allow you more permissionless functions, but you're ultimately responsible for what happens with that data. And another where there are fewer capabilities, but the engineer is actually responsible for it.

Cillian Kieran (33:28): You're getting into something really interesting that's a way away from the laws and purely philosophical, and I think it fits. You're effectively declaring your own competency, and therefore the risk you take personal responsibility for. It's like race car drivers. I'm big into motorcycles; that's my personal passion. Take the Isle of Man TT. For some people, racing around a thirty-mile road is objectively insanity. I think it's incredible. But they knowingly get on that motorbike, do 200 miles an hour, and accept that risk. An average human shouldn't do it. They're not qualified to. So it's self-selecting: you make the decision, you assume responsibility for the risk. That's a very clever way of looking at it. I don't think we're having that kind of conversation yet in industry or for consumers. We're not making that decision clear to anyone, so people default to assuming risk without understanding the risk they're taking responsibility for.

Liam Lawson (34:28): Where do you feel this is best discussed? Are there podcasts or media giving this the coverage it deserves?

Cillian Kieran (34:40): Honest answer, and I know this will sound self-serving: I don't think it's getting the coverage it deserves. The problem is that because the word governance is attached to all this, it sounds like unsexy laws. But I keep trying to remind people to genuinely forget about that. If we are racing towards the outcome we all think is inevitable, which is machine to machine communication, a reduction in the workforce, all the things we're going to be challenged with as a society, we should all be asking ourselves: I feel very confident using machines to do a bunch of tasks for me, but do I feel personally comfortable with those machines making all of the high-risk decisions they're likely to make in the future, alone, without any kind of oversight? I'm not super comfortable with that yet. Not because I'm a crackpot conspiracy theorist, just because I think things make mistakes. I've been writing code for 20 years and I know how easy it is to make a mistake. So we need some way to control that stuff, because if we don't,

Cillian Kieran (35:34): I think we will regret it, not now, but in ten years' time. And no, I don't think that conversation is being had, because it's not sexy enough. It's far sexier to talk about how much faster or more performant a model is. I talk about this with a lot of our enterprise customers. If you really put the most recent sets of models from, frankly, Mistral, OpenAI or Anthropic on a bench and test them for most enterprise uses, like crunching CRM data for behavioral analytics, they're all about the same. The only thing that makes those models different or perform differently for different businesses is the quality of the data they're fed. So at this point we're into end-level optimization, maybe token consumption optimization. That's fine. But what we should really be concerned about is: is it doing the right things? Can we leave it alone, unattended, and make sure it won't do something wrong, as we've seen them do? That's my concern.

Liam Lawson (36:26): I'm not too sure either, because fundamentally these models are really good prediction machines. All they're doing is predicting the next most viable token. There's nothing original; there's no actual judgment happening. Based on the data it's been trained on and the context it has, it makes a decision it thinks is right, but it's technically not even predicting what it thinks is right, what it believes to be true or what is technically true. It's just predicting the most likely outcome, which is fundamentally opposed to what we're looking for if we want some degree of truth.

Cillian Kieran (37:07): A hundred percent. I studied physics, not any of this, so I studied a lot of statistical math. Not AI stuff, just maths, because you have to do maths. A couple of years ago I was struck by an interview with a sci-fi author named Ted Chiang. He's written a bunch of sci-fi, but he's most famous for the short story that turned into Arrival, the sci-fi movie about language and speaking to aliens. He was interviewed a couple of years ago by the FT about AI. He's quite anti-AI because of the creative liberties it takes. But he pointed out that artificial intelligence is a label, just a marketing label thrown on what is statistical math. If we kept calling it statistical maths, it wouldn't be nearly as sexy today. Because that is all it is. Yes, engineers will debate the type of math, but to your point, it is just statistical averages about the most likely next token, the next phrase. When you bring it down like that, it's as dumb as a box of rocks. It doesn't sound particularly sexy at all. It has magical qualities because of how we've personified it, and I think that has more to do with user experience than with what models do: the agent experience of chatbots and things like that. I think we are coaching average consumers, who are not technical and not in the weeds of this stuff, to accord it too much responsibility and believe it is too good, and the risk from that is that we mistreat it. A bit like the idiots we occasionally hear about on the internet who decide to let a self-driving car really drive itself and then find out it's not so good at it. I think we want to protect people from making those mistakes.

Liam Lawson (38:47): There are a lot of things at play here on a macro level. These foundation models are raising a ton of money. I think Anthropic's most recent round was 65 billion at a 965 billion valuation. I think OpenAI is talking about a trillion-dollar IPO. The money's insane, and it's all in this massive debt circle where everyone's buying compute and data off each other.

Cillian Kieran (39:12): Trading compute and data, yeah.

Liam Lawson (39:15): So that's one function of it. Other people are building harnesses, layers between them, infrastructure. You know exactly what your business is. But if you were an entrepreneur looking at the next five or ten years of where this landscape takes us, where do you think the opportunities are that people aren't seeing? If you had the spare time and could clone yourself to start another business, what would you be looking into?

Cillian Kieran (39:49): It's genuinely a really good question. Purely in the space of AI, however you want to define AI, I obviously think Yann LeCun's work on world models is genuinely fascinating. To me that feels like the more interesting stuff. Embedded predictive architecture, I think, is really interesting.

Liam Lawson (40:17): Tell me more about that. Explain it in detail for me.

Cillian Kieran (40:20): Predictive world models. I'm not saying it's why Yann LeCun left his former role, but he's gone off to spend time on world models. If you read his public musings over the last couple of months, or years even, he has the view that there is no way the contemporaneous models we're all using today will ever get to AGI. Not only that, but the risks related to their ability to be controlled and managed will never be solved. It's an endemic technical debt in contemporary models.

World models use a completely different base for maths and graphing. There are a couple of competing companies: there's World Labs, there's Yann LeCun's business, AMI Labs, and I think there's one other I can't remember off the top of my head. At the current rate of investment, they're five to ten years away from releasing anything usable, but there's an outsized bet that world models far surpass the capability of the model architecture we use today. The jury's out, but I think Yann LeCun knows what he's doing, so he's the man I would bet on. If I could do anything, I would probably go and work for Yann, because I'd be curious to see what the inside of that looks like. This is going to sound bizarrely naive from a founder.

Cillian Kieran (41:45): Unless we correct the challenges of tech where we are today, and I'm a real optimist for what technology can be, I'd be more interested in doing something diametrically different to tech, is an honest answer, Liam. I really mean that. I'd be much more interested in pursuing something in hard infrastructure, hard materials, hard sciences and physical output, like manufacturing, manufacturing sciences and automotive engineering. I find physical output much more interesting these days than software engineering. Software engineering is getting less and less interesting to me personally. But your question is where the opportunity is, and if there's an opportunity, it's probably in world models.

I think the thin layers of success we've seen around foundation models are not going to survive. The vast AI hype for what I think are feature businesses, which are all getting consumed back into models, will consolidate and disappear over the next three or four years. I think far too many businesses have been funded with a .ai attached to them, and they'll all disappear. We'll be left with a handful of large enterprise model providers, and that's all that will matter. Businesses will effectively roll their own solutions beyond that. So I'd probably get out of software and into hardware.

Liam Lawson (43:01): It's a pretty terrifying future: you just can't build software because it's going to get consumed by the foundation labs. I run a newsletter, and I've been doing it for three years. I've watched thousands of businesses pop up, raise funding and then get consumed in an instant because one of these foundation labs brought out ChatGPT Health or Claude for Design or Claude for Finance. It happens all the time.

Cillian Kieran (43:08): Isn't it happening already? Right, all the time. You could argue that the cycle of the last four years has almost been set up to do it: allow early-stage startups to build something moderately valuable, and then it's consumed backwards by a foundation model that wants some intellectual property. But I don't know that that's sustainable. As these frontier and foundation model providers grow in size and capital, and the numbers we're talking about are hard to comprehend in terms of sheer capitalization, and yes, there's some funny accounting going on, I think we should all acknowledge that, but accepting the funny accounting piece, it doesn't leave a huge amount of capital for innovation away from those vast fortresses that have raised that money to date. So I am circumspect about the future of traditional venture-backed software companies at the dawn of this moment of AI. That makes it sound negative. I just think it's going to change, and I don't think it's going to look the way it did for the last twenty years. If I were an entrepreneur looking to solve hard problems that are still rewarding, I'd be looking in areas slightly away from software.

Liam Lawson (44:36): It's interesting. Last Friday I had a podcast with one of the lead engineers over at Block, Jack Dorsey's company, with Afterpay and a bunch of other things. They have a new agent infrastructure they've built called Buzz. Similar to Grokbot, you can put all of these agents in a group chat, have a discussion with them, and they go away and run your business. Part of that discussion was that one of the great components of Grokbot, agents just running your business, Buzz and all these different softwares is that they give back more of your time for interacting with other human beings, in a physical way. I've run an online business for the last three years, and I find it deeply unsatisfying compared to the work I've done with my hands. Just before I started this business, I was a landscaper in Melbourne for nine months, and I loved it. Every day I'd get up and build irrigation systems, lay turf, learn things and ride wheelbarrows about. I felt satisfied. I've been building this online business for just over three years now, and there are components of it, but I don't have as much daily satisfaction as I used to when I was dealing with things on a physical level.

Cillian Kieran (46:03): The challenge with saying the next thing is I'll sound like two things: an old man, and more specifically a slightly privileged old man. I think the post-COVID era of remote work has been fantastic on lots of levels for people's ability to live remotely, spend time with family and do things that are important to their lives. I really believe that. At the same time, as someone who came of age when you had to work in an office initially to cut your teeth and get a job, there was no choice, I feel sorry about how disconnected we've become from each other. I miss working with people, not just in an office, but, to your example, doing a hard day's work outside. Like I said, the motorbike thing isn't a joke. I work on old motorcycles whenever I have a few minutes. I'm knee deep in oil working on them, and it's the most physically satisfying thing, but it's even more satisfying to do it with a couple of friends. I'll spend my weekend with some mates tearing apart an engine and putting it back together.

Remote work and technology are wonderful, but we are losing touch with each other. I see that in the fact that some of the young people I increasingly work with find it hard to interact in real life. They'd much prefer to text or Slack. Totally fine, I get it. We've all got different communication modalities. But as an Irishman who grew up going to the pub and having a chat with someone, I think we're losing out if we don't sit and have a chat with people.

Liam Lawson (47:34): I think we're starting to trend that way; people are aware of this. We had the anti-alcohol trend, where millennials and Gen Z, people my age, were giving up alcohol, and I go through phases of that myself. Then people do six months sober and realize they're depressed. That's why you have so many startups coming out of the US based on socializing through healthier things. You have the rise of run clubs and social clubs. You probably see it all the time in New York.

Cillian Kieran (48:03): Totally. Hyrox, all that stuff.

Liam Lawson: Exactly. So there is demand for it and a realization of it. But I think there's some sort of AI psychosis, because I can talk to Perplexity for six hours a day and it runs my business and I make a lot of money, and I don't realize how sad it's making me that I'm sitting talking to an AI chatbot for six hours a day.

Cillian Kieran (48:24): But isn't this the funny thing? If you look at it like layers of an onion you peel back over fifteen years: fifteen years ago we were all very nervous as social media was eating our souls from the inside. Go forward a couple of years and COVID happens, we all work from home for a bit, and then AI happens. None of it's a planned set of events, but each one has created a small buffer against us spending time with each other. So to your point, if someone feels lonely today, they can spend time with their friendly agent, whatever machine they like, and have a conversation with it. All of those things help numb the feeling of loneliness, but they only encourage us to do it more and not actually spend time with real people. And you're right, things like run clubs are great. At the risk of again sounding like an old man, I desperately want us all to spend more time together. I think that's the nicest thing in the world. So if it runs our businesses for us so that we can spend more time together, that would be awesome. I'm all in for that.

Liam Lawson (49:28): I think the gap right now is that the promise is different from the realization. The promise for an entrepreneur using AI is that AI can run your business. Realistically, when you start using AI, building systems and automations and using agents, you spend way longer on your business, because you're looking at every granular aspect of how these systems work and tweaking them to make them work. Maybe there's some future where this all works for me and it's great, but right now it takes a lot of manual effort. Sidebar here: have you ever read Zen and the Art of Motorcycle Maintenance?

Cillian Kieran (50:12): Of course I have. I keep it on my bedside. I've had it for years. My dad bought it for me when I was sixteen.

Liam Lawson (50:19): I was around the same age when I first read it. It's my favorite book on the planet. I want to bring this into a question for you. One of the main theses, and the through line of the entire book, is that there are two ways of looking at the world. Classical is when you look at a motorbike, for example, and break it down into its individual components and how they fit together: everything has its place in how they combine to make one system. The artistic way is looking at things more holistically: the motorbike functions, all these systems work, there's interoperability, but you're not breaking it down by individual components. It just works, because you're playing with it.

Cillian Kieran (51:05): Classic and romantic, I think it is.

Liam Lawson: Classic and romantic, that's it. I'm curious how your brain works, because over the last seven years working on Ethyca you've had to be so granular and so classical in how you break down these systems. Is there an artistic component to how you look at this as well, or is it all classical?

Cillian Kieran (51:32): That's such an interesting question, Liam, and this will sound like such a cheesy answer. I wasn't kidding when I said my dad's an artist, a painter. I grew up in a house with a man who painted all day in his studio. I used to draw a lot as a kid. All I did was draw and paint as a teenager. Then I got into computers and computer programming in my very early teens. The reason I bring this up is that, as a geeky, very introverted teenager, I found computer programming to be a really creative pursuit. I could sit there and type stuff and make something come to life. At a primitive level it could be a webpage that was green with black text, but I made it. So to me computer programming was a really creative pursuit, and at the time I was probably less interested in the engineered aspect of it. I got really into making video games in my teens, fairly primitive stuff.

So, maybe counterintuitively, making software, and I won't call it software engineering because that immediately makes it sound classical, is to me a really creative problem-solving pursuit. I feel the same about making businesses. Solving problems and finding creative ways to address any problem is a really interesting creative thing to do. I probably have a technical proclivity inside that, which allows me to do the systems-level thinking, the root cause analysis and the first-principles stuff that gets me to an outcome. But I tend to do it in service of solving an interesting problem, so I think of it as very creative and get really excited when someone presents a new problem that has to be solved. That's a fence-sitting answer, but I think it's a bit of both, and I enjoy them both for that reason.

Liam Lawson (53:30): How many employees do you have now?

Cillian Kieran (53:34): About fifty employees, five zero. Still a small team: very technical, very product engineering heavy. We've only really been building the go-to-market team very recently. The business has spiked in the last eighteen months as we released the platform.

Liam Lawson (53:52): When you're hiring, have you noticed a difference in mindset between those two teams? Is the engineering team classical-heavy and the go-to-market team romantic-heavy, or am I not covering everything with that thesis?

Cillian Kieran (54:12): I don't want to be unfair and overgeneralize, but historically, when we were really in a phase of just building product, it had to be very technical engineers solving really complicated, messy problems people didn't want to tackle. We do an annual offsite, in places like Mexico City last time and Palm Springs the year before, and we bring the team together. It's the one time we all come together properly as a whole team. I remember being struck four years ago in New Orleans: this is a room full of effectively rocket scientists, literally ex-NASA, ex-Northrop Grumman and ex-Boeing engineers. That's fantastic; they're incredibly clever. But when you talk about classical modes, these are engineers who think about engineering problems, and they think that way about a salt cellar on a table at dinner.

To be fair to the teams we've hired since, it isn't that we've suddenly hired a bunch of hopeless romantics who are all artists per se. But you definitely skew the culture of a business as you start to hire for go to market, because sales, marketing and post-sales need stronger interpersonal skills. So I have seen a shift, and it does skew that way. As a footnote, because our product is very complex and deployed into very complex, high-risk environments, even client-facing teams have to be reasonably technical. There are definitely some romantics, but it's a geeky business overall.

Liam Lawson (55:53): Before we started the interview, we said we'd try to find that through line between technical and philosophical. I think we've got there.

Cillian Kieran (55:58): I think we've done a really good job.

Liam Lawson (56:05): I'll finish on the same question I ask all my guests. One piece of context: everyone takes it a different way. Some people talk about family, some about purpose, work, whatever. I'd ask you to take five to ten seconds to think about your answer before you answer. So, Cillian, why do you do what you do?

Cillian Kieran (56:37): I have a canned answer when people ask me this, but because you asked me to take a few seconds, allow me to do exactly that and give you a better answer than a canned answer.

Cillian Kieran (56:51): I'd like to modify the question slightly and say, why do I still do what I do, as opposed to just do it. Will you permit me that?

Liam Lawson: Sure.

Cillian Kieran: When I started out as an entrepreneur, it was very much that creative thing. I loved to solve problems, didn't care what you got me to do, I just wanted to figure things out. I'm in my mid-40s now, and I have two young kids.

Cillian Kieran (57:19): I promise you I'm a tech optimist. I say that because the next thing I'm going to say doesn't sound very optimistic. I'm a little nervous about the world they will inherit in twenty years' time, and I'm under no illusion that I can fix it single-handedly. For a long time I worked in businesses where, because I was working with marketing for consumer packaged goods, we were helping to monetize eyeballs. You're not doing a whole lot of good for the world. Aside from it being a complicated problem, and therefore academically interesting and intellectually stimulating for me, I like to think that whatever Ethyca is doing, and it is commercial for sure, it's not doing something bad. It might actually be doing some good. I like that. I really do care about that.

But to give you the most thoughtful answer I can to why I still do it: honestly, at 45, having been nothing but an entrepreneur since I was 21, I've definitely thought about throwing in the towel a bunch of times. I could give you a harrowing podcast on the things not to do as an entrepreneur; that's two hours in and of itself. But I think I still do it because I believe the work I'm doing with the team I work with is important, and that it will matter somewhere along the way. Even if it doesn't make a big dent commercially, if the open source is meaningful to companies and we've done something good, I'd be pretty happy with that. Of course I want us to make a big impact; I'm a founder with a big ego and ambition. But at this point in my life I don't feel I have a huge amount to prove. I've done okay. So it's more interesting for me to keep working on a problem I'd be proud of in another ten years' time, or that my kids would be proud of me for working on. My kids think that I work on making people's data safer. They're seven and nine, so they're very young, and their understanding is that I make people's data safe. That feels pretty good to me. I can live with that.

Liam Lawson (59:22): I usually finish the interview here, but I want to ask a couple more questions because I'm curious from a personal perspective. Did you grow up working class, middle class? What kind of household did you grow up in?

Cillian Kieran (59:37): You're asking big questions, Liam, which I'm happy to answer. I'll tell you an honest perception of myself. I think people assume I must have grown up upper class, and the reason, bear with me, is that I live in America, my mum is French and my dad's Irish, so I happen to speak fluent French. But I grew up in the eighties in Ireland, and my parents were not well off. I was just lucky that my mum spoke French, but actually that just meant I was bullied as a kid because I didn't fit in. You can imagine what that's like. I grew up very, very simply. It's probably hard for people to relate to unless they've lived it. A very simple upbringing,

Cillian Kieran (01:00:29): from two wonderful parents who loved me and my twin brother very much. It was very kind and very safe, but it definitely wasn't fancy.

Liam Lawson (01:00:39): The reason I ask, and I'm curious what your perspective is: I grew up pretty middle class. My mum was a teacher and my dad was an engineer. All I used to think about when I was younger was being an entrepreneur, solving problems and making money. As soon as I started to make money, be self-sustainable, have spare money to invest and live a nice life, I had the opportunity for more money and more money. For the first time in my life, maybe two years ago, I looked at it and realized the money wasn't what was making me happy. What I wanted was something more aligned with my purpose as a human being, attached to something bigger. I'm curious if it was the same for you. Growing up in a working-class background and then getting some success with a startup, is that when the idea of a larger purpose became infused, or was it always there?

Cillian Kieran (01:01:45): It's a really good question. I one hundred percent didn't have a bigger purpose in mind when I was twenty-one. At twenty-one I had a very naive relationship with success and money. It was: must make money, must survive. It turned out I was good at computers. It turned out I could sell stuff and make a business. So I did. There's no founder mythology. I hate founder myths; I really don't like any of that stuff. It was the only thing I knew. I dropped out of school and didn't know how to do anything else. I was good at making software, it was the very early 2000s, and I had an opportunity to make a company out of what I knew how to do, which was making software. So I did it. But I did it in perpetual motion. I didn't stop to think for a second. I wasn't particularly strategic. At the time, my solution was blood and sweat: I will get there and I will make money.

Then I was lucky enough to become more and more successful. Five or six years into my first business, it had done really well, and then it suffered really badly in the 2008 economic crash in America. That was a reckoning for me, and it took four years to rebuild it. By the time I got to my mid-thirties, I had a very different relationship with what I wanted to do. I was still excited to work with interesting people doing interesting things, but it was more, to your point, about something I could wake up feeling good about each day than just making money. The funny thing is, when you're twenty-one and an older guy says to you, it's not all about making money, you say, yeah, sure, let me go do my thing. And then you realize it really isn't. It doesn't buy you happiness.

Liam Lawson (01:03:34): Cillian, two things. Firstly, thank you for taking the time with me today. I really appreciate it; I've loved this conversation. Second, if people want to find out more about who you are and what you do, how do they go about doing that?

Cillian Kieran (01:03:50): You can find me on LinkedIn or on Twitter; I'm just Cillian on all of those. Or you can find me at ethyca.com.

Liam Lawson (01:03:58): Beautiful. Thank you so much.

Cillian Kieran (01:04:00): Thanks, Liam.

Get the five-minute AI briefing

Join 400,000+ business leaders who start their weekday with The AI Report. Subscribe free