Keeping Humans Accountable in an AI First Workplace With Nansen

[00:00:00] The leading issue of agentic AI in businesses right now is ensuring agents act with compliance guidelines. And Denodo applies guardrails across your entire data estate. By aligning your company's data infrastructure under one system, these guardrails perform consistently across your platform. So start scaling your business and start with Denodo. Simply visit denodo.com to learn more.

[00:00:34] What happens when every employee gets an AI agent and the CEO uses an assistant to plan meetings, build products, and occasionally negotiate domestic peace? Yep, it sounds like a glimpse of tomorrow. But my guest today, he's already living this dream. So in this episode of AI at Work, I speak with the co-founder and CEO of Nansen.

[00:01:01] And we're going to talk about what an AI First Workplace looks like in practice, how Nansen has given each of its roughly eight employees an agent, and one, my guest, has used one to create meeting agendas, and another to build a product through Telegram, all while going for a walk with his daughter. So we'll discuss why people remain responsible for every AI output though, and when work should be sent back to the kitchen,

[00:01:28] and how machine-readable documents make agents useful, and why evaluation matters before an agent even thinks about handling anything financial. So this is a conversation about experimentation without surrendering judgment. And hopefully you will leave this episode today with a few practical ideas for using AI boldly, while keeping the humans accountable for the result. But enough from me.

[00:01:58] Let me introduce you to my guest right now. So thank you for joining me on the podcast today, Alex. Okay, so everyone listening a little about who you are and what you do. Yeah, thanks for having me, Neil. I run a company called Nansen. A lot of people know us for blockchain analytics. People see what's happening on a blockchain in real time. And right now we're focused on building the product where you can trade everything on-chain with AI agents.

[00:02:26] So we are really big on agentic trading. We think that's the future of how both institutional and retail investors are going to be interfacing with the markets across the board. So that's what we're building. Fantastic. And there's so much I want to talk with you around. I especially got a little bit geeky about agentic trading, AI. But before we talk about the tech here, can you tell everyone listening a little more about Nansen,

[00:02:56] the problems that you're solving and why understanding what's happening on blockchain networks has become so important for, indeed, investors and businesses alike?

[00:03:34] Yeah. And overall, we're going to spend the whole podcast talking about blockchains. But that's kind of the background when you think about what blockchains are and why they're useful. And increasingly, people are trading and transferring other types of assets on blockchain. So historically, people know about Bitcoin, they know about crypto.

[00:03:59] But now you're seeing tokenized stocks, you're seeing tokenized commodities, stable coins, real world assets being traded and transferred on chain. So the interesting thing that we're pursuing is that because blockchains are transparent, you can see what everyone's doing in real time on a blockchain. That's different from traditional markets.

[00:04:23] And we think this is a really good environment for AI agents to be unleashed because they can see everything that's happening. And we're basically trying to make it easier for retail investors to invest in any type of asset that is on a blockchain. And you can see what the smart money is doing when you think about both individual investors and funds that are doing well trading and investing.

[00:04:49] We track all of the profits and losses that people make down to the individual addresses on a blockchain. And the thinking is that retail investors are going to benefit the most from AI because we obviously don't have a team of, you know, PhDs running our hedge funds, you know, you and me as individual retail investors. But AI agents kind of give you that promise where you can actually have them do a ton of research.

[00:05:16] They can scan data. They're much more disciplined. They're less affected by emotions than we are as humans. So the idea is that we can kind of supercharge the retail investor and give you incredible tools to do better when you invest in trade. And this is not just a job for you because I was reading before you came on the podcast that you're using AI, for example, to automate almost every aspect of your life, not just your work.

[00:05:44] So for people listening that may be a little bit nervous of connecting inboxes and calendars to AI, what does your typical day look like for you? And what tasks have you genuinely handed over to AI and had benefits from? I'm curious what you've experienced here. Yeah, there's so many things I could choose from here. But maybe one thing I can say is that at Nansen, at our company, we're about 80 people.

[00:06:09] And we have basically provisioned one AI agent to each individual employee. And so I have my own OpenClaw. That's the tool we use or the framework we use. My OpenClaw is called Winnie. And so Winnie basically prepares a draft for all my agendas, for all the meetings I run. And this is not just kind of randomly created, obviously.

[00:06:39] It looks at all the meetings that we have run in the past. It has full context on our objectives and key results, our strategy, the culture of the company, and so on and so forth. So let's say on a Monday, I wake up and Winnie has prepared. Here's the agenda for the product engineering and design leadership team that we're going to do today. And then I'll basically chat with Winnie on Slack. And then we'll kind of iterate towards the right agenda.

[00:07:08] So that's one kind of basic example. Obviously, I've done a ton of other things. I actually created the first version of one of our new products through Telegram with one of my other OpenClaws. So I was out walking with my daughter and the stroller. And I think she was having a nap in the stroller. So I just was on Telegram and talking to my OpenClaw.

[00:07:37] And when I got back to my house, I could see the product that the OpenClaw had built for me. And it actually ended up becoming our command line interface. Cool that now thousands of people actually use. So there's a bunch of different things. Obviously, in my personal life, both me and my wife use AI a lot. Sometimes we broker disputes between us. I'll have my agent talk to your agent.

[00:08:06] And then that's a way to de-escalate a situation. And yeah, there's a bunch of different ways I use it. Oh, I absolutely love that. I'm going to have to think of that next time I have an argument with my wife. I'll get my agent to talk to your agent. I've not got time for this. But seriously, a lot of companies in the workplace, they're racing to deploy AI with an almost deploy first and then review later mentality. But many are now slowing down. What do you think is most misunderstood about what AI is actually good at and what it isn't?

[00:08:36] You probably are someone that is experimenting and using so much on a day-to-day basis. Where have you noticed its strengths and weaknesses? Yeah, I mean, I think a lot of people get caught up in sort of early issues with AI. And maybe they will have tried something a year ago. And they sort of assume that that's still a hurdle for them today. When in reality, these models have become incredibly good.

[00:09:04] So I think like hallucinations is obviously something that people talk about. That's relatively rare, I would say. And, you know, the newest models, even some of the newest open weights models like Kimi and GLM 5.2 and so on. I wouldn't say it's like a fully solved problem.

[00:09:20] But I would say, generally, I think it's better to assume that agents and these LLMs can do more than what you think they can do to push them. And then see where you hit the limits and boundaries, as opposed to assuming that something you kind of tried a year ago is not going to work today either. So that's kind of one way I would think about that.

[00:09:50] And then people often weirdly don't think about the risks of not using AI, right? So if you think about it in a security context, which is very common, people are concerned about the security implications of using AI, which is understandable and it makes sense. But they maybe don't think about the fact that the attackers are using AI. So if you don't use AI, you're kind of, you know, bringing a knife to a gunfight in some ways.

[00:10:17] So I think you kind of need to arm yourself with AI in, you know, safe and responsible ways. But also your competitors are probably using AI, right? So I think like the risk of not using AI may be greater than the risk of using it in many cases. And when I asked our head of security to roll out these open claws at the company, he was very, of course, nervous about it.

[00:10:43] And he said something like, look, we're basically giving knives to children here. But if we don't give them these knives, they're going to probably get even sharper knives themselves. And so, you know, you might as well try to give people like a safe version of AI that they can use rather than some janky thing they put together, which maybe puts them at even greater risk. Such a good analogy there.

[00:11:08] And also when I was doing a little research on you, I was reading how you say AI should accelerate judgment, but not replace it. And that is a massive talking point right now. The importance of critical thinking. It's almost a paradox because to get the most out of AI, you need critical thinking. But the more you use it, or a lot of people that use it don't use their critical thinking. So which is it? Almost outsourcing their critical thinking.

[00:11:31] So where do you draw the line between delegating work to AI and then keeping that human responsibility for that final decision? Yeah, at least at the company, what we say is we expect you to use AI in pretty much everything you do. But we also expect you as the human to own the quality and the output of the process or of the AI. And so this is kind of feedback we give people.

[00:11:59] It's great that you're using AI, right? We see the end dashes and the text and all that stuff. And in some companies, it might be sort of frowned upon to use AI when you write things. We kind of expect that people use AI, but we also expect you to not output slop, right? And so I think the basic principle is that the human has to own the output and the outcome. And so that means you as the human are responsible for the quality.

[00:12:29] And another phrase I like to use is you often have to send things back to the kitchen with AI, right? You kind of get a meal. It's like, actually, this doesn't taste right. There's something wrong. Let's send it back. And then you do a couple of iterations on it. And so I actually think quality is probably one of the primary ways that you can differentiate as a company in 2026.

[00:12:54] Because it's so easy to ship something that's kind of 80-20 or kind of okay, but not amazing. But to do the work of iterating towards really great quality, it takes, I think, a human in the loop and someone with great taste and judgment. Yeah, I love that line about sending the meal back to the kitchen. Because even the best AI models, they can hallucinate, lose context, or confidently produce the wrong answer. I think we've all seen examples of that.

[00:13:23] So when you're designing workflows, how do you design them that take advantage of AI speed without allowing some of those mistakes to become business problems? Yeah, I think the main thing, which is a huge unlock in how you think about using AI at work, is to make sure that the artifacts you're working on, whether it's a document or code or a design or an architecture diagram, is all machine-readable.

[00:13:50] It's something that an agent can easily read, right? And I often kind of, or not so much anymore, but I used to run into this where I would see people worked in, let's say, like a FigJam or a Miro or something like that, because they want to have something visual, which is great. But the problem with that is that it's often not that accessible to an agent.

[00:14:13] So I think probably step number one is to make sure that you have artifacts that both humans and agents can read. It's not enough that only agents can read it, because then you're really delegating everything to the agent. So we've actually moved towards a lot of our documents internally just being documents in GitHub. So it's kind of odd. Like the finance team, it looks like an engineering team because they have their own GitHub repositories, right?

[00:14:42] Just like you would have code in GitHub. And increasingly, we just have markdown documents instead of Google Drive documents. We have CSVs and markdowns to replace spreadsheets or HTML documents that are more interactive. So I think that's kind of the biggest unlock to think about. Is the artifact I'm working on actually machine-readable? Because then I can make use of the agents.

[00:15:12] And then I think it goes back to what I said earlier. But I think probably the second part is you obviously can't have it only be machine-readable, because then you're fully delegating to the agent. So finding something that kind of works for both the human and the agent, so that you as the human can have your sign-off on it, and you can review the output and make sure that it's high enough quality. I think that's probably the most practical piece of advice I could give people.

[00:15:40] And I also read that you highlighted that AI isn't physical, because it cannot load a washing machine, even though we would like it to, or observe the real world without any reliable human input. So what does that limitation tell us about where AI is today, and also why high-quality data and good record-keeping still matter so much? I mean, to be fair, it is kind of changing, right? We are getting humanoids that increasingly we will probably see,

[00:16:10] especially here, I'm in Singapore, and in East Asia, it seems like there's kind of a high appetite for having robots around us in the real world. Like if you go to Heidi Lau, which is a Chinese restaurant chain, and you see robots bringing out the food. And I saw a video the other day where, actually it's funny, the laundry example, because there are going to be condos where you have a shared robot that comes to your door, picks up the laundry,

[00:16:38] and actually puts it in the washing machine. So maybe that example is very quickly getting dated. But so I think, you know, I think it's just that the world of bits has less friction for AI. Obviously, in the short term, the world of atoms is still slower, harder to iterate on. It costs more to experiment with. But I also think the world of atoms is maybe the next frontier

[00:17:06] that a lot of entrepreneurs and a lot of people in tech are getting excited about. So I'm actually super, super curious on how the next few years are going to be. And I think maybe if I would make a prediction, the southern part of China, specifically the greater Bay area, Guangdong, Hong Kong, Macau, that area will probably be the first like metropolitan area

[00:17:33] where you see a lot of robots out in the streets and in the condos and, you know, around. So, yeah, I'm pretty excited about that, actually. I think it's going to be an incredibly, incredibly novel experience for us as humans to like coexist with robots on the streets. Incredibly cool. And if we were to look ahead to that future that we're talking about there, what do you think needs to change before you can trust AI on even greater autonomy,

[00:18:02] whether that is in your personal life, your business or financial decision-making, what technical governance hurdles still need to be overcome? So I think this phrase or this term evals is very helpful. So for evaluations and when you build agentic systems, you kind of have this like religious focus on evals. You need to make sure that the agent is doing the thing that you're optimizing it for. And so in our case, we run trading agent evals.

[00:18:32] We actually run a bunch of different evals, like how well can our agent read data? How well can it understand significance of a news item or, you know, a piece of data that it discovers, whether that's like a big piece of news or a small piece of news and maybe not that important. And so I think running evals is kind of the most practical way to think about it, that if you're building a product or building a piece of software,

[00:19:01] building a vertical agent that tries to solve something, you need to have an incredibly good eval loop where you're evaluating whether the agent is actually good at doing the thing you're optimizing for. And so in our case, the most foundational eval is basically, do these trading agents actually manage to make money? If we put them in a backtesting scenario, which is the only thing that's feasible to do if you want to iterate really fast

[00:19:30] and also not spend a huge amount of capital, you put them in a backtesting environment and you run evals to see if they actually make money or do they lose money. And we actually have another layer above that, which we call the optimizer or the coach, which looks at these individual trading agents and then tries to advise them on how they could do better, how they can adjust their investment strategy, their trading strategy. So to kind of zoom out a bit

[00:19:59] and apply that same lens to other domains, I think like if you have an incredibly tight feedback loop where you are running evals and you're understanding how you can get the number to go up effectively, that's basically how you do it. So it's the same with, you know, think of self-driving cars and other things that are very high stakes. You need to make sure that it actually works. You won't sit in the backseat if you aren't very sure that this self-driving agent

[00:20:27] has performed incredibly well, both in an experimental scenario and also in the real world. So I think it's analogous to trading. People will want, they shouldn't trust something that is completely new where you have no information about, you know, how well it works. But over time, I think we'll be able to build up the trust by publishing more of our results of the evals and so on and so forth. So yeah, I think at the end of the day, people should definitely not just trust, they should also verify.

[00:20:57] And I think that's intuitively the approach you would take if you jump into a self-driving car, for example. And I suspect you will have set off a few light bulb moments with people listening today. And for those people that are listening, maybe they want to follow in your footsteps. They want to get more value from AI without falling into the trap of over-automation. Any practical principles or habits that you can share that have served you best? And also where people should start if they want AI to become

[00:21:27] a genuine productivity partner rather than just another tool that they chat with like they do on Google or something. Yeah. I mean, I think it's the best time ever to get your hands dirty and just play around with some of these tools. The recommendation I've given to a lot of friends, family members, etc. is if you can kind of set up an OpenClaw or a Hermes agent, people might have heard of these, in some simple environment,

[00:21:56] whether it's on an old laptop that you have or an old computer, or if you're a bit more technical and you can spin up like a cloud instance, that's probably what I would do. And then make sure that you don't have too many sensitive pieces of information on it because you might, you know, you don't put your credit card details into that entity and so on. But you kind of just go a little bit nuts and experiment a lot. That's kind of the overall piece of advice I would give.

[00:22:27] I would say it's the best time in history for people who are curious and have high agency. And so maybe what I'll end on is people who I think have four traits will do extremely well in the age of AI. If you have high agency, if you have a great sense of problem selection, if you're working on the right thing, because you could work on so many different things these days and create so many different types of products,

[00:22:57] if you have great judgment and taste so that we get that quality assurance that we talked about earlier. And then if you are able to communicate clearly, both with other people, but especially with the agents so they understand what you want. Those are probably the four traits I would sort of nurture in yourself. And then you combine that with an appetite for experimentation and not taking too much risk. Then I think a lot of good things will come out of that. Fantastic advice. And before I let you go,

[00:23:27] if we have any traders listening, Nansen turns on-chain data into actionable intelligence for traders and allows them to see what smart money sees and act before the market. And for anyone listening, interested in learning more about that before you give any details, are you any big announcements this year? Any teasers you can leave us with and what we can expect? And also where listeners can find out more information. Yeah. So you can go to nansen.ai. Right now, we are actually running a really cool raffle

[00:23:56] where if you buy out our trading features, you get to take part in our raffle, which we're giving out $100,000 weekly. So that's definitely a good reason for people to try it out. We also have a big announcement that's coming in a little bit, I think probably within two months, where we will be giving out early access to some of our most loyal users to a new product that we're launching where you can basically trade

[00:24:25] fully autonomously with our product. So that's kind of the holy grail we've been working towards for a long time. And we've been investing a lot of the different parts of the stack. And now it's all going to come together in this new product that we're going to ship later this year. So stay tuned for that. Oh, wow. You've left us on a teaser, a cliffhanger there. We're going to have to get you back on in a few months and learn more about how that has evolved and what you're going to be working towards next year as well. So we'll get you back on the show.

[00:24:55] For everybody listening, though, I'll put links to everything you mentioned there in the show notes so you do go check out, join the community, learn more information about that. But thank you for talking about all this stuff in a language everyone can understand today. Really appreciate you, Tom. Thanks so much, Neil. I think Alex's advice there to send AI work back to the kitchen is such a useful reminder that speed should never excuse poor quality. Yes, use the draft. Test the idea. Ask the agent to try again,

[00:25:24] but keep a person responsible for the outcome. And I also think his practical starting point is also refreshingly simple. Experiment in a contained environment. Avoid sensitive information and then learn where the agent performs well before granting it wider access. So make documents readable by both people and machine and then begin to build those evaluation loops around the result. And if an agent will influence financial decisions,

[00:25:54] trust has got to come from repeated evidence rather than just blind faith or enthusiasm. And I think also we should shout out those four traits that Alex recommends, which were high agency, strong problem selection, judgment and taste, and clear communication. Yeah, this is a AI podcast, but it's these human abilities that become so valuable as producing an average first draft could quickly become so much easier.

[00:26:25] But over to you, which part of your work could an AI agent accelerate without taking that ownership away from you? I'm interested in what you're doing here, what's working, what isn't. So please pop by techtalksnetwork.com, share your stories, and we'll keep this conversation going. But that's it for today. Time for me to go now. I'll be back in your podcast feed real soon. Remember, drop by techtalksnetwork.com. If not, I'll see you here again very soon. Bye for now.