Who's Accountable When AI Fails | VC Backed Startups and Fraud | Will You Quit on Me?

One Repeat founder in the piece captures the dynamic bluntly by calling low-risk, high-upside founding opportunities of starting a company and flooding the market, in his own words, as a nice way to get, you know, a job at Anthropic. Has your read on founder commitment changed at all given how accessible and lucrative those outcomes have become?

Even if Anthropic offered me $100 million, what else are you supposed to say as a founder? If you want the money, like, I think some of these fundraisers are advertisements. They're a way to like advertise to the labs that you exist and that you're around.

It's the week of August 3rd. Welcome to The Learning Corner, a Precursor Ventures podcast. Welcome to The Learning Corner by Precursor. I'm Mia Farnam, and I'm joined by our GP and managing partner, Charles Hudson, where we will be talking through the articles and Goodreads that caught our team's attention over the past week. Last episode, you got to meet Alina, who's one of our interns, and this week we've got the other half of our great intern class, Braylon, here on the podcast. Thanks so much for joining us this week. We're excited to have you.

I'm excited. I'm excited to have this conversation.

Awesome.

Well, for our first piece this week, Angela McNeil, author of her blog called The Handshake Protocol, released a piece titled The Architecture of Blame, which centers around the looming question of Who's on the hook when an agent fails? Angela's argument is that blame doesn't become unanswerable because of AI. It becomes unanswerable when systems are designed to lose the answer, and humans have been building those systems through bureaucracy, policy, and org charts for decades. What AI threatens to do is industrialize that pattern, taking what used to require a whole bureaucracy to sustain and running it at machine speed across every workflow at once. The counterintuitive flip she makes is that a properly instrumented AI agent is actually the most auditable actor in your entire enterprise, more accountable than any person on the payroll who can only reconstruct a past decision from memory after the fact. She lands on attribution being a design-time act, not forensics. Identity on every action, scoped authority, and scoped audit trail are the three things you build in advance. So the question of who owned a decision is never a fire drill after something breaks. We've been running, I feel like, the past episodes in themes by accident. And I feel like this week's is on like truth-telling and finger-pointing and blame. Angela opens with a reminder that accountability gaps aren't an AI invention. So they're not new. Long before any model existed, humans built these gaps into their institutions. She gave examples of the insurance claim denied by the policy where everyone agrees it's wrong that it's not covered, but no one can fix it, or the patient record owned by a process that no one has touched or edited since 2019. And the real risk Angela flags isn't that AI creates this pattern, it's that it industrializes it, taking what used to require entire processes and people in place to sustain and running it at matching speed across every workflow at once. So If humans were building these accountability gaps long before AI existed, Charles, what changes occur when you hand that same pattern back to something that never sleeps and scales infinitely in the world of agents?

It's funny, this reminded me a lot of one of the best books I read in a while, which is Unaccountability Machine that James Cham from Bloomberg referred me to a while ago. It's a really good book about basically how systems are designed so that no individual is accountable. It talks a lot about the impact of that on bureaucracy. And I think about this, and as someone who uses a lot of AI and has agents, if agents were really deterministic and followed all the rules, this would be a great framework. They mostly do that, but sometimes agents, like humans in my experience, do things and you go, why did you do that? And it's just like, I'm sorry. I'm like, well, that— I'm not asking you to apologize. I'm asking you to explain why you did this aberrant thing. And sometimes the agents either can't or refuse to tell you why they did the thing that they did. I think one of the fundamental decisions everybody needs to make about AI is like, how much authority and autonomy do you want to give AI? And if you think of like your AI agents as employees, in most cases, you wouldn't give a brand new hire authority to ship to production without somebody looking at the code first. And the same way I have some friends who have AI agents that send emails without human review. I'm like, that sounds dangerous to me. The only way you're going to catch those things is on the way out. I read this as like quality control for agents is a design decision. And it's as much about quality control as it is about blame. And like, how often do you as a human want to have somebody in the loop inspecting the outputs? And I find the challenging thing is you can look at something and it can work perfectly for 2 weeks and then something in the underlying model changes. And a process that was working fine suddenly does fail. And if you're not watching and if you're not monitoring, just like you would any other human employee or literally mechanical process, things drift. So you need to have eyes on these things. And I think this nirvana that you would set something and forget it and it would just run in perpetuity autonomously is nowhere near where we are today, if we ever get there at all.

One thing that she did bring up was the big pros that you get from agents running these processes is potentially having the most accountable paper trail that you've ever had in comparison to people, because she talks about human memory as the one piece that's basically rung true when you have to relook at a process and see what went wrong. I agree with her that there's an opportunity for agents to really step in and really show what the process looked like and what things touch. But to your point, Charles, I've had multiple instances where I've asked how you came— how you as the agent came to a specific Action. And not that it lies, but it gets confused. And then I almost have to be like, that's actually not, that couldn't possibly be where you got the data because that's not where it lives. And it goes, oh, you're right. And so again, we're not there, at least maybe within the build of the agents that I have, like I have to be more robust in terms of how I construct them so that I can catch those little moments along the way. The paper trail piece still exists, even though we're in a different setup of it all.

It's a different paper trail. If an API call to a model returned a result, who's responsible. At least with the human processes, there's like a person you could go talk to. Here in many cases, the agent is some combination of code, API calls, and a model, which all work together in sometimes unpredictable ways.

Carlos, and I'm curious to know too, kind of looking at it as a follow-up, the piece also makes the case that the organizations who can answer who owned this at any moment are actually the ones who can move the fastest, right? And so we talk a little bit about kind of the audit trail that lets you deploy it to kind of those consequential territories without flinch. I'm curious to know from an investing standpoint, you know, do you buy the framing that building this accountability infrastructure is seen as a competitive advantage, or do you look at it more so as not just a, you know, compliance tax?

I think it's really important to understand drift. This is like one thing I've really come to learn in the last 9 months. I have a bunch of models that work for a long time and then they stop. And sometimes they stop because I've learned something new that I haven't shared with the model or the agent. And so the expected result, like the agent's still doing the thing it's supposed to do. It just isn't doing what I want it to do. And that's how, I mean, in other cases, it's something in the plumbing changed. Some of these things will break silently and they won't tell you. And if you don't have good basically sensors looking at the output, things can, can go haywire for a while and you just don't know. If you're going to build these things and use them for important decisions, it's imperative that you have the ability to audit. Audit and analyze. This is what went in and this came out. Did I get what I expected? And what caused the drift? And everybody I know who runs a lot of these agents has a story. Oh yeah, I realized that this thing was broken for 2 weeks and I didn't really realize how broken it was until I gave it kind of a deep inspection and realized, oh, this thing's messed up.

And there are layers of severity. We've talked about this months ago, Charles. Remember when we walked through that article of either was Burger King or McDonald's? 100 drinks were placed with like an AI Honestly, that's just funny. Like, it's just like a funny incident. The range here, or the outcomes become far more drastic when, and even Angela mentions it, if you bring this into healthcare, if you bring this into law, there are just far more drastic outcomes and negative things that can happen if things aren't watched. And so I think that's where the finger pointing and the blame becomes such a higher point of contention because at the end of the day, I think that the piece talks about like the accountability sink. Who's around the drain at the end of the day to take the blame? And if there's almost too robust of a process, no one is. And so there's just no one there. And that's where I think this article brings attention of, is this going to be so archetype that there is no one to blame at the end of the day when things go wrong? And is that actually something that organizations want to design?

I think you're going to end up in a lot of cases where there's going to be somebody whose job it is to use a model they didn't build or use an agent they didn't build. And if it goes wrong, they're gonna say, I didn't build the agent. I was just using the agent the company built. The person who built the model, who built the agent, is gonna say, look, the agent did what it was supposed to do. It's actually Claude's fault, or it's OpenAI's fault. It's the LLM or the underlying model provider's fault. And very few people are gonna be able to fully diagnose who's really at fault. And I think you're gonna end up with a bunch of finger-pointing standoffs. The majority of what happens in corporate America anytime something goes wrong.

Do you think that it will be an assumption in the future that you should, you as the employee should be able to tinker around and improve the agent when it fails? That should be something that's within your wheelhouse?

I don't think most people want that responsibility.

Right.

It'd be like if you were working at an insurance company and you're like, oh, our underwriting model is predicting death rates wrong. I'm gonna go in here and tinker with the weights. Most people would be like, hey, that's like beyond my pay grade.

Our next The next piece is a TechCrunch article covering two new academic papers, one from Imperial College and Emlyon Business School and one from the University of Toronto that map fraud patterns across VC-backed startups from 2000 to 2023. The headline finding is that companies with venture funding were more likely to face fraud charges than those that didn't take it. And startups launched during overheated markets with weak investor due diligence are 19% more likely to later commit fraud. As a follow-on to those statistics, researchers introduced this concept of facading and gives the example where founders have moved surface lies about performance to reinforce facading, where fake customers, contracts, and invoices are generated to back up those lies and going as deep as to where entire parallel realities are constructed around capabilities that simply don't exist. What makes this especially poignant in, in the callout that investors are not just passive victims here and in some cases co-create fraud by continuing to back founders' previous previously accused of misconduct. The researchers call for more accountability on the investor side, including SEC audits after startups hit large investment thresholds and investors being held liable for corporate governance failures. Reframing fraud is really a solo act. So obviously, you know, a lot there, definitely timing, you know, in terms of kind of what we're seeing in the market. But then also, as we continue the conversation around compliance and accountability, I'm curious, Charles, you know, Given that, you know, initial thoughts, you know, at a high level.

It matches my experience, which is that we've had a bunch of episodes about revenue. Like, what is— we spent more time talking about what is revenue than I thought we'd ever spend in the modern era. I think part of that is because founders and investors exist in this ecosystem. It's a feedback loop. With investors, you're like, I want to see more growth. Some businesses can really achieve that growth through honest means, and other people say, look, they rationalize, if I can just get this round done and if we fudge the growth, we'll go back to doing it the right way on the other side. And I think we generally know how that happens. One thing I've learned is that there is far more fraud, in my opinion, in the venture startup ecosystem than ever gets reported. And part of the challenge is no one has the incentive to report it. There are cases I know of where both a founder and a VC know fraud has happened, and there's usually not always a great reason to report. If you think about like the Charlie Javits thing, that was when a big public company has egg on its face over an acquisition, your options are to prosecute the fraud or just say, hey, we didn't do a good job on our due diligence, like my bad. And, uh, that's not going to happen. If a VC and a founder know, if the founder and VC both come out about this fraud, they, everyone has to take accountability for what happened. The investor has to say, hey, this thing happened under my watch. I should have been paying more attention. I didn't ask enough questions. The founders to say, I was like a criminal or I did something wrong. And in many cases, there's fraud at companies that are not consequential. Companies like go to zero or they don't matter or they wind down and people bury the fraud or the bad action with the company, which is a shame because oftentimes those people go on to do the same thing in their next company. And investors may or may not learn the lesson that they learn when this fraud is more public. Most things are not Theranos.

Right.

They're just— they're not. Oh, this company lied about a couple of accounts and that enabled them to get financing done. But it wasn't a lot of money for the person who wrote the check and the company failed anyway. So you're like, well, what's— what's the point? I do think the pressure right now that companies are under to show these crazy growth rates. And the irony is sometimes you learn after the fact that the crazy growth rate was never real or the crazy growth rate itself was a product of dodgy processes. And suddenly everybody has engaged in collective bad behavior chasing a target that was never real. I'm 100% certain there will be companies whose ARR, revenue reported growth ramps, we will look back on them and say, oh, that wasn't what we thought it was. And maybe we shouldn't have all been chasing that.

They break down levels of facading into 3 categories, and they start with surface level, which is what I see all the time. I actually don't even know if I can really call out and slap the hands of founders or investors. It's just, I mean, it's media. It's being able to put your numbers out for people to see. And since they're private companies and they're so early, I think that exists throughout our ecosystem. That's just like something that you have to sniff out as an investor. When somebody gives you an update about their company, something that's a company that you haven't invested in, like you have to then do the diligence to see what am I missing here? If it's all positive and rosy, the deep facading. That's what you're referencing, Charles, of like, that's what Time's bestselling novels are built off of. Like, there are no story— those happen, those are quite rare. And so that's why when they, when they come out, like, it is such a blockbuster event. I know the piece talks about different remedies that we can make to solve this. And one of them was broader SEC audits. I find the friction in that, knowing that the way that I know venture works, especially in hypey bubbly markets, new rounds with new valuations just happen so quickly. There's no time to do that. And then it kind of ties into the process of why founders enjoy staying private longer. Yeah, now more than ever, because there's just so much more at stake when you go public. So again, all themes we've talked about, but it's unclear to me how you break the cycle, especially in a market that we're in today.

Yeah, I mean, you could— I think what you said is really true. In hot markets, most people make the calculus, hey, if I ask a lot of questions and do a lot of diligence, I might not win this round. And if I'm going to do the round, make the investment anyway, what's the point of doing all this due diligence if I know, if I already made up my mind?

Yeah.

Audited financials would catch some of this stuff, but in general, I find that orchestrated fraud usually starts at the top at most startups. And so I do think if there were more corrective consequences for board members management and investors, there'd be less of this. I don't think any of those people want that to happen.

Yeah.

And so I, I think I've just sort of concluded that a certain amount of fraud and misbehavior is an occupational hazard of being a venture capitalist who invests in private companies. And my job as a VC is to try to sniff it out before I make an investment. And if I find it after I make an investment, to try to fix it. But most VCs are non-control investors. And so in some cases, the only remedy you really have is to try to get rid of the person in charge of the company, which most VC funds aren't willing to do. It's also like most of the things I found out in our portfolio companies, in the rare case of something bad has happened, it has been a lower-level person telling me, hey, this CEO is like up to something, or like, I'm resigning because I'm really uncomfortable with the way that we're presenting financials to the board, or the way that we're talking about our pipeline, or this upcoming product. It's usually like an internal junior person on their way out the door who tells you like, look here, this is where the bodies are buried.

Yeah. And the behavior starting at the top, it doesn't really take that Hollywood form of people gathering into a room usually. And this is how we're going to map it out. It's usually just pressure, pressure coming from every type of angle that instills this new energy in your organization of we need to move quicker. And I don't think we can do that without fabricating. A bit of this in some way, shape, or form.

We go to great lengths in Ventureland to not call fraud fraud. Facading and fudging, and sometimes people are just straight up lying. They're just lying. Plain language makes these things easier to understand.

My last piece on what my line is for surface level facading. Yesterday on LinkedIn, I think I saw one founder post like a growth map, obviously in like the normal XY axis. That shows how much they've grown month over month. And it literally was just a line. There were no numbers and it just showed a line that increased over time with like a lot of interactions and comments. And I was sort of like, yeah, I mean, it's not false, but it could just be the y-axis is literally like a dollar to like $10. And like, that's just like the beginning parts of how people kind of get into the swirl of how can I put out signals that things are going well, but also fall in the realm of what's passable, which that's what's scary. And but I think it's part of the job to what you just said.

Yeah. And I think about accounting just kind of having that background all the time. We kind of lean into the trust but verify. That's kind of just always been the written rule of, you know, being able to scope out those opportunities and what those may look like. And so definitely spot on in the sense of, you know, how you got to sometimes just pull— let's say pull the rug up to see what's under there. So our last piece is from The Wall Street Journal titled VCs Are Asking Founders, Will You Quit on Me? By Yulia Chernova covering how the AI talent race is fueling a wave of founder poaching and acquihires that is leaving early-stage investors feeling shortchanged. Backed by that stat, the venture-backed startup acquisitions in the AI sector jumped from 100 in 2014 to 322 last year and are already on pace to exceed that in 2026 with 245 deals through early June alone. The dynamic that AI labs and big tech companies are dangling hard to is offers that founders is happening. And what we see after VCs are already invested at high valuations is that founders are now walking away with large equity packages while investors often just get their money back. And so the article highlights VCs now explicitly asking in diligence, will you sell to a lab in a year? Which is a question that they say never had kind of asked before in an attempt to re-found a commitment through long dinners, conversations rather than pitch decks. One repeat founder in the piece captures the dynamic bluntly by calling low-risk, high-upside founding opportunities of starting a company and flooding the market in his own words as a nice way to get a job at Anthropic. And so Charles, I wanna pause for your thoughts on this piece and specifically has your read on founder commitment changed at all given how accessible and lucrative those outcomes have become?

I try very hard not to ask people questions where like there's only one right answer. And if you say, are you going to quit? Everyone's like, no, I'm going to, I'm in it for the long haul. I'm like totally going to do this. Even if Anthropic offered me $100 million. What else are you supposed to say? As a founder, if you want the money. And one of my friends this weekend, we were talking about this very, this very theme. And he's like, I think some of these fundraisers are advertisements. They're a way to like advertise to the labs that you exist and that you're around and to give them some sense of what it would take to get you. We have one of our existing portfolio companies, one of the co-founders is probably gonna go join one of the big labs, and it's a lot of money. And his co-founder was like, this is what they offered him. I was like, he's probably gonna take it. It's a lot of money. I didn't feel like wronged. I was like, look, my job is to find people who want to build great companies and who believe that like the best bet they can make is on themselves. As a founder. I think if you're running a company and somebody offers you $20 million to join a lab company in like compensation over a couple of years, if you don't think you can make that in your own startup, like risk adjusted, you're going to take it. And so like the ways I can— the only ironically, some of the ways to solve this are A, find people who are just maniacally committed, but every founder tells you that. Two, find companies that raise so much money that like they can afford to remain independent and that the lure of like a big payday, it doesn't matter because they already have the resources. Or 3, just accept— this has been an issue in every cycle since mobile, because in my mind mobile is really when aqua hires became a thing. And I think the model providers are doing exactly what they should do, which is if they see talent that they think they want that could be effective in their organization, they should go get it. And if they are able to pay more. And those people think, hey, I mean, the truth is what these people are really telling you is, I think I am better off working at Anthropic. I will make more money for myself and my family than I am being the founder of this company. And that's as much an assessment about Anthropic as it is about the thing they're working on.

Yeah.

The irony of it all is that the companies that this happens to the most frequently, I would go out on a limb and say, probably raised at valuations that are at the higher end.

Of the bubble.

And so the tension is tough of investors being like, this is worth paying up for, and then feeling like, oh, like this actually came out in a direction that I probably didn't want. But what can you do at that moment?

I think if you invest in a company where the founders are AI researchers doing AI researchy things, part of your memo has to say there is a chance these people will get poached.

Yeah.

And that they will end up at the model provider and I will maybe get back my money or my money and then some, but this won't be a home run. And like that, ironically, if I were making those investments, that would be like my principal fear is not the companies go to zero, it's that they go to 1x and you get your money back in some weird acquihire, acquilicense deal. They go make a bunch of money and you're left with this shell of a company that you have to figure out what to do with. I wish I had better news.

There's something really interesting sitting at the intersection of this piece and the fraud article that we just discussed. On one side, investors are setting expectations pretty extreme that founders feel pressure to misrepresent what they built. And then on the other end, they're now asking founders to prove that they have conviction to stay the course no matter what. Gizmo, how do you like hold both of those dynamics at the same time? Right? I think we talked a lot about kind of like some of the struggles from both ends, but those, you know, pressure moments, yet, you know, moments where you're also trying to build conviction. How do you kind of balance those dynamics?

We're in this world where For investors, the only companies that really matter are the big scaled outcomes. That's it. That's kind of the whole game right now. It's the whole business. Kind of stinks, but it is what it is. Which means investors are trying to figure out very quickly, is this company SpaceX? Is this company Databricks? Is this company going to make me famous and rich? My friend Reese, like, Reese sent like one of these like memes around with like the parent in a pool with two kids. And one kid's on the ledge. It says like, MegaFun VC. It has a kid on the ledge where it's holding its hands and it says current portfolio founder. And the kid's like drowning, the parents like not paying attention to them. Then at the bottom of the ocean, it has like a skeleton in a chair and it says portfolio companies that won't even help the fund get to 1x. And it, I do think like we're in this dynamic where investors want to win the deal and founders want to get the money. And then as soon as that's over, investors are like, great, is this company going to matter? If it's going to matter, I'm going to put pressure on it to continue to grow. And if it doesn't matter, I don't care. And founders are like, I want to be in the bucket of companies that matter. I know the bucket of companies that matter are the high performers. I need to make sure I'm a high performer.

I think the difference, or something that feels meaningfully different here, is that the term of high performer, I think, now is tied heavily to signal, like outward signal. That can come in the terms of high valuation, large round, large social media presence, building in public, all of that, um, which I don't know, like we've seen over cycles, there, there isn't all too much data that that is the recipe for creating strong outcomes. But we've, we as venture in the ecosystem, to your point, I think we are just in this cloud of figuring out what to make sense of the entire investment stack from LPs to GPs all the way down to founders.

Yeah.

And so people are trying to figure out where their priorities lie. And so I, to your point, like if I'm a founder and I'm presented with a $20 million cash offer to go join a AI lab company and I just raised $6 to $10 million, like I'd debate that pretty heavily. And I actually would say, I would actually appreciate a founder who's like, Yeah, if that came along, like, I would look at it because I think that's a little bit more realistic than you telling me like, I'm here until the day I die. Like, okay, like, well, I've seen companies shut down all the time, so it's okay that you can have like a different answer to this question. But I don't know if that level of transparency will actually ever come out candidly.

And I think honestly, most people don't know. It's one thing to be like, oh yeah, if it showed up, I wouldn't take it. How do you know that?

Yeah, you don't. Yeah. Well, we are all out of time. Thank you, Charles, for joining us. Thank you, Braylon, for joining us as well. To all of our subscribers, please Like and share with friends, and we will talk with you all next week. Thank you. Bye.