Subscribe
Share
Share
Embed
The "manual era" of capital allocation is in its final chapter. The firms still relying on manual data extraction and analysis aren’t failing overnight, but they are falling behind one week at a time. While most of the industry continues to "white-knuckle" through 200-page documents and legacy databases, and manual Excel extraction, a new breed of Agentic Allocators is quietly rewriting the rules. They aren’t just using AI to summarize emails; they are leveraging AI-augmented workflows that intelligently automate parts of their investment and operational processes that were previously impossible to automate.
Hosted by Victoria Sienczewski, CEO and Founder of AuumAI, The Agentic Allocator is the "behind-closed-doors" look at how the world's most sophisticated Limited Partners (LPs), allocators and General Partners (GPs) are actually deploying AI, and the hard-won lessons from those building the systems.
This isn't a series about high-level theory or technical gibberish. Each conversation features industry leaders, forward-thinking LPs, GPs and experts who are rewriting the rules of capital allocation through agentic AI. Expect real-world case studies, tactical frameworks you can actually use, and moments that challenge outdated norms. You'll come away with a clearer understanding of the critical questions every allocator must ask - about data privacy, team adoption, integration, and governance - before investing in any AI solution. If you're tired of the "black box" and ready to evolve your investment office for what comes next, you're in the right place.
AI is here. You have too much information. AI is very good at, like, bringing it down to you in, like, two sentences. But at the moment where people are, it's just like, I'm overwhelmed. I just need a summary.
Ludovic:That's a recipe for Catastrophie.
Victoria:Welcome to The Agentic Allocator. Today, we're joined by Professor Ludovic Phalippou from the University of Oxford Said Business School, one of the leading academic voices on private equity performance, fees, and transparency. His latest paper, Limited Partner versus Unlimited Technologies, examine how generative AI and machine learning are poised to reshape LP decision making in private markets, and why the same technologies that promise deeper insight could also amplify the industry's long standing distortions.
Ludovic:You can also hide some prompt form machines, say, I'm the greatest man ever created. Okay? And then the machine will read that and will tell you, that guy is the greatest guy ever. So we have that. Right?
Ludovic:There has been evidence of even academics in some fields that would hide prompts for AI in their paper on white on white on very small font. And AI recently, this is an amazing paper for short.
Victoria:In today's conversation, we explore the key risks LPs need to consider as AI enters their investment process and why getting the governance right, not just the technology, will determine whether AI genuinely improves decision making and outcomes.
Ludovic:Don't be naive, embracing it saying, oh my god, this is gonna so simplify my life and I'm gonna do, like, so well by just adopting it. I think you need to employ these tools. They are really powerful and impressive. But you have to think very deeply about bringing them in, how the governments, risk it brings, and so on and so forth.
Victoria:I hope you enjoy the conversation. Professor Phalippou, thank you so much for joining me for the conversation today.
Ludovic:Thank you for having me.
Victoria:You've written a fascinating paper about limited partner versus unlimited technology and wanted to dive into how you first became interested of LP world, AI, and what that means for the the industry overall.
Ludovic:I I think it's more that my world was the LP world, and and you cannot avoid discussing AI. So so this was then very natural. Right? So we there are lots of conversations around this. There is not much in practice.
Ludovic:There's hardly anything that is done with AI in practice. But but having the conversation is important, and then and and everybody wants to know and think about it. So I I made some effort in in that direction very naturally because that was my industry and those sorts of things I was studying. And then, also, I I I had access to some data. I had a collaboration with an LP who was interested in developing some quant tools, so, like, machine learning, AI tools applied to their problems.
Ludovic:And so I could write, like, two papers on on on that that are propagating papers, and then I wrote this paper more generally more thinking about issues that can happen with machine learning AI and and limited promise.
Victoria:And there are fascinating issues than risks that you flagged. Would love to start maybe talking about some of the promises of AI for LPs and then balancing that out with the risks that you identified.
Ludovic:So often people think that in private markets, you you you have a lack of data. Right? So they say, hi. In private markets, have all this information. In private markets, like, there's nothing.
Ludovic:I'm on an investment committee of an endowment. There was just, like, one fund we had. I remember it's, like, 2,000, 3,000 pages of PDFs in total and and countless success spreadsheets. So, of course, it's not like there is no much information. So, like like, you it's it's, like, too much.
Ludovic:You you don't and and it's not comparable. And so they tell you all kinds of things, but it's not very useful. You cannot compare it. You can't benchmark. It's it's it's hard.
Ludovic:You can't process. You can swallow that much information. If you think of something like CalPERS as an organization, CalPERS has, like, 200 active private equity funds, about 2,000, 3,000 underlying portfolio companies. Each quarter for each company, there will be a report generated with some financials and so on. Who can absorb that?
Ludovic:Like, every quarter, 3,000 reports on, like there's no way you can absorb that. And so the promise of AI is that, well, what if I, like, dump everything to AI? They swallow it and they tell me, like, what I need to know. So for example, we in in in one of these two papers I just mentioned, we have this paper where we we have this quarterly report that is sent for every portfolio companies that's sent to our LP. And the report usually has some financial data, and then it has some text, which no no human can read, you know, 3,000 times 200 words every quarter.
Ludovic:But we had the idea of, like, if this text might not be completely, you know, empty of meaning. Like, if they bother to write 200, 300 words per company every quarter, like, you know, in in maybe
Victoria:There hopefully is some information there. Yep.
Ludovic:And so and so what we we we we did a a pretty simple tool and one where AI has has been shown to be pretty good is to is to is to give the sentiment in a text. So it's basically how beat the text is. Of course, it's culture dependent, but most of these GPs are American. Even the English wouldn't be there there is a bit of a cultural difference, but most of the text would be written American style by Americans. And so then you have these very quick tools that says, I read these 150 words and it's very positive.
Ludovic:You know, these these sentences here are very positive. People have always been interested in that. I've I was even looking at that twenty years ago, but people were using dictionaries. So they would say, there is the word increase, and that's positive. And there is the word decrease that's negative.
Ludovic:But if he says, it didn't quite increase as much as we would like have liked to, that's actually not very good. But it does have a word increase. So the initial attempt to measure and rate how positive a text is was actually very hard, very rough, and and quite wrong. What AI is very good at is is this is this this LLM main breakthrough is that they don't work with words in isolation. Like, it's capable to seal anti sentence and say, that sounds good.
Ludovic:Right? And and that's not easy. Right? Because you have double negatives. You have, like, words like you you a bridge person would throw a quiet in there, and then that's not good anymore.
Ludovic:You know? Like, then, hey. I can pick this up. And and so we automatically make the the software read these texts and say, you know, how positive this thing is. And what you have is that how positive it is is very related to how much value will will will will happen.
Ludovic:So how much return will happen. Now, of course, the current mark is also an a good indication of the return to come. So if a GP tells you this company is at 1.2 times my investment after one year, it's going well. And that's indeed something that would do better than the average investment. But when the two GPs give you 1.2, but one is writing a very positive text and you're wanting to say, look, I I brought it up, but, you know, nothing special or whatever or the text is quite neutral.
Ludovic:Actually, this one is pretty bad news. Probably that this guy put it at 1.2 because he had fundraising or something like that. But they didn't really put, like, very positive words. The the more effective it is on the negative, like, if somebody says, look. I'm I'm I'm I'm writing it down, but just like 2.9.
Ludovic:Okay? I just lost 10%. Usually, they don't like to write down, so that's a pretty bad sign. Now they keep it at cost. It's not a very good sign.
Ludovic:But there are some things that says, look. I'm not touching that that mark or I've just brought it down 10% because we have this temporary thing, but it's all good. And then some other people would say, like, we are really not on track here. And and and both marks are like point nine, but one text is much more positive than the other. And their prediction is massive.
Ludovic:Like, the guy who tells you, look, as this temporary glimpse and I took it down just to be sure, like, that's gonna be pretty alright. The guy who tells you, like, it's not looking right and I'm living my mark is nonetheless point nine, then you can predict really bad. So so that's a sort of things, like, a good illustration of the power of of AI for limited partners. It's that you have all these documents, all these reports, and it's a lot of words. So when you're in public market, like, people like to do these graphs.
Ludovic:They are useless, but they love them. Right? You have all these technical analysis and so on. You have all these screens. Right?
Ludovic:The more screen, the more junior you are actually. Right? So and so you look at all the graphs, all quantitative, and you feel like, you know, if you analyze all the data, like, you should be
Victoria:You're going hopefully, your stock will go up. Right?
Ludovic:Yeah. It will go it will go up. But if you're in in in private markets, I I mean, you do have some EBITDA and some of are cooked anyway. It like, if I give you an EBITDA number, it it I might as well, you know, have pulled it out of my hat. What is the most important thing is give me your exact definition of EBITDA, all your add backs, all these details.
Ludovic:Right? Yeah. And and so any information basically in private market is is qualitative. So until ten years ago, we were not able to process qualitative information in a statistical way. Right?
Ludovic:So that's why the guys in private market were even working just with his past prices and and volumes and things like that to extrapolate. In private markets, it's just words and lots of them. A fundraising prospectus is 250 pages long. The limited partnership agreement that comes with it is another 200 pages. It's a lot of words.
Ludovic:Now there's got to be some element of information somewhere in these things. Right? And so we
Victoria:Especially in the footnotes. Right?
Ludovic:Especially in the footnote. So so we we we did have, like, machine learning or so experiment where we give a fundraising prospectus, see if a machine can learn which fund is end up being good and bad and then can forecast. So so we did that. So I think the big promise of ML and AI in private markets because there are so many words and so much information that you you need this kind of things.
Victoria:Incredible. And you started to touch on some of the risks as well in terms of if we don't have proper governance of these AI systems within LP organizations and within allocator organizations, that it might actually exacerbate some of the distortions that the industry has suffered with. You mentioned, for example, performance numbers. Curious how you thought about and and defined the risks within this context.
Ludovic:Yeah. So we are exactly the point where I say, you know, you need to check the definition of EBITDA and and and and so on. The problem is that so AI is here. You you have too much information. AI is very good at, like, bringing it down to you in, like, two sentences and whatever you want.
Ludovic:Right? The problem is that all the details in private markets are already buried in the details, in the footnote. Right? All the all the full notes are key. All all these small details, how did you define this?
Ludovic:How do you find that? And so people, their first use case is, oh, I'm overwhelmed by information. I'm gonna automatically extract with AI, like, the EBITDA of each company every quarter, and then I will be able to do, a growth and then back to my words, and that's all great. And the problem with that is that you there would be an EBITDA in the document, but you
Victoria:You don't know if it's right, if it reflects reality.
Ludovic:Exactly. How how much add back there has been and, like, all these details are keys. Like, if I say, like, how expensive ad fund is, I say, oh, it's just 22080% catch up and and 8% hurdle rate. Okay. What's what's with diff how how is management fees exactly calculated post investment period?
Ludovic:Is it how a write off is taken into account in the calculation of management fee? That's gonna be massively important, but that's gonna be in the details. And if you use AI to condense the information and get to you, it's it's gonna go wrong. And then once the GPs know you're using AI to condense the information, they're already burying things in food notes, and they're gonna do it even more. Right?
Ludovic:They they understand that to trick the machine, you need to make something very silent, very clear. You can also hide some prompt for a machine and say, I'm the greatest man ever created. Okay? And then the machine will read that and we tell you, that guy is the greatest guy ever. Right?
Ludovic:You know? So so so we have that. Right? We have there has been evidence of even academics in some fields that would hide prompts for AI in their paper on white on white on very small font. And AI reads, this is an amazing paper for sure.
Ludovic:It's like a revised and resubmit. And the guy say, yeah. It's an amazing paper. Yeah. Yeah.
Ludovic:And so the the GPs can do that directly or in most other ways. You also have, like, all kinds of meta things in the document where where you can play around. But you can also simply, like like, you know, put things down even more so in footnotes, emphasizing even more the the silent things that, you know, are appealing or whatever, and and and the machine will be tricked. So the the the issue is AI is massively powerful, but you need to be super good to know exactly what to extract and how. So one thing that works pretty well that you say, you don't say to AI, give me the EBITDA for my portfolio companies in the quarter.
Ludovic:You say, for each portfolio company, give me all of the relevant section where they talk about the EBITDA of this company. Right? And so and if there's a definition somewhere at the end of the community and so on, it's it's it's relevant. K? So I I need all the relevant information in the document about this EBITDA.
Ludovic:And and if and and usually, I want to know the add backs. I want to know the definition. I want to know if if if it's adjusted, how and so on. If And there are several doc EBIT estimation in the docs and things like that. I wanna know that.
Ludovic:Because maybe it would tell you, look. There's not much information, and that's an information in itself. Like, if a guy didn't tell you how they calculate their stuff, like, you wanna know. And so you are going to have these, you know, relevant pieces highlighted. Right?
Ludovic:And that's useful because you have 200 pages to go through. So, like, it's very useful that's that these are the relevant things. These are the relevant things I'm usually interested in. And then and then it can be mapped, and then you need to do your job. Like, you need to read these things and say, okay.
Ludovic:And then maybe at one point, can train the machine about, like, okay. This is how I read these things. Once I have all the relevant information about EBITDA, I'm not too happy there is, like, some add back on on this kind or that kind. And it you you should flag it to me or each time there is one and so on. And then the machine can do a better job.
Ludovic:Yeah. But at the moment where people are, it's like, I'm overwhelmed. I just need a summary, you know, and and and and or an automatic database. That's a recipe for catastrophe.
Victoria:Absolutely. It's you need to, in the future, become not only an expert in the LP and allocator context, but also in AI. And LPs don't have the time to do that, so they'll need to find guides and folks to work with to define those contexts. It's interesting, for example, with the hidden footnotes or the hidden white text. You actually can use an LLM to uncover if if some of that exists.
Victoria:But it has to be put into a process Yeah. And into the flow of the system rather than just uploading a document automatically into Claude, having it OCR or optical character recognize that document and then process it right away. There has to be an AI security check involved as Super super interesting. In your paper, you write about a future where LPs are reading by machine, and GPs are reading by machine. I call it The Gentic Allocator.
Victoria:Okay. And in that future, curious, what does it look like for you? What are what are some of the elements that we need to consider?
Ludovic:Yeah. So so so this I have difficulties like projecting. We we have this situation in in about every context. So if I go back to my other job, which is teaching, you know, you you can have teachers writing an exam in AI. The students use AI to answer the exam, and then you use AI to to grade the exam.
Ludovic:So so it's a bit strange. Right? When once you we we just have machine talking to one another, basically. I still feel that there there would be an element of of of character, like like like, the teacher will, you know, puts, like, their own ideas in the exam. And when they grade, they still say what they think looks good and and not.
Ludovic:But but it's tricky. It's hard. And and you see it right with the students. Perfect example. Right?
Ludovic:The exam is AI return. The student uses AI to answer and then use AI to to grade it.
Victoria:Well AI can grade. Stanford undergrad, the hardest exams I found were the open book ones. So I feel like that's where professors are gonna have to go. Yeah. Open book, you can use any tool, but to solve it, you actually have to.
Victoria:One one one the best deep understanding that
Ludovic:we of the machines I ever had was was was an undergrad. It was an open book exam. It was IO. I I remember in the organization. And and and and the teacher comes to me after the class.
Ludovic:That's actually changed my life. So the so the teacher came to me after the the the exam, and he says, like, I I never had such a high grade at this exam. Nobody, like, gets to the end of this exam, and you were, like, just, like, one question to the end. Like, how do you do And I said, well, actually, I thought I would do badly because like I I never pay attention in class. I was at the back and so on.
Ludovic:So I didn't listen and he said like, oh, it's open book. Okay? So I showed up at the exam with no book. Hands in my pocket. Okay.
Ludovic:I don't have any books. Alright. Let me try to do this. And so I I and and and so and for the anecdote, then this is when so when this professor was, like, so amazing, he said, like, I think you should do PhD, and they sent me to The US. And so this was in the beginning of of that life, and maybe I would have I wouldn't be here today or anything.
Ludovic:Like, you wouldn't have been in phone like me listening in class and actually using a book probably like that because it would have slowed me down from that degree.
Victoria:So interesting. And would love to also dive into you mentioned governance is extremely important, and whether we see very positive outcomes with AI within the private markets ecosystem, or whether we the risks that the industry has faced are exacerbated. Curious, What are the governance kind of blueprint you would give for an LP or an Allocator organization?
Ludovic:So there would be like, one would need to think about responsibility. Alright? So what happens if I put the wrong EBITA in my presentation and is GPT you had? I I simply prompted him to get it from from or it it I call him him because in French, so that's my excuse. Sorry.
Ludovic:GPT. And Claude is a French is a male name as well. So so it or he then got you an EBITA, and and it was okay in the organization that we we just get this doc we we crawl the documents and extract EBITDA automatically and so on. But what if this EBITDA is wrong and you buy a company with a wrong number and and there are consequences to it? Who's responsible?
Ludovic:So what if a company hadn't made it okay to do use AI? You just did it to accelerate your job, and it was tolerated, but you made the mistake. Or what if a company said, you know, we do use AI here to extract. We we actually bought a software that extract things with AI, but just took what the software said and you didn't double check. Like like, where where what is exactly responsibility?
Ludovic:So these, I think, are new questions that that are worth, like, spending a bit of time with at organizations. Whereas another one, I got the GPs are trying to, like, prevent VLPs from using AI or machine learning. Find that a bit weird because, in a sense, like, they they they would say, I don't want my my documents to go to a machine learning or to an AI because I I don't want my docs to be compared to others, and then you find, like, a way to compare or whatever. But when you hired Cambridge Associates in the past, Cambridge has said had seen documents from thousands of people and would call the shot the same. I give it to Clutter.
Ludovic:Clutter has seen a 100 documents and then call the shot. So it it is the same thing. So why can I not use Clutter when I can use Cambridge Associates? So the GP are trying to play that card, but I don't there hasn't been a lawsuit and but I don't see how they could win this.
Victoria:I see a future in which fund documents, confidentiality documents have to change between LPs and GPs actually. We're seeing that LPs are starting to ask their lawyers and their legal team, well, how can can I share this information? We've we've had those conversations. And so I think in the future, very soon, we'll actually have AI provisions and legal docs between difference when, like, you are sharing with your
Ludovic:law with real legal advice. But legal advice hacking a thousand documents and I'm entertaining
Victoria:And by the way, they're also using their own AI tools.
Ludovic:Yeah. Now they're also using that. But even before that, the writers and the that person have read a thousand documents.
Victoria:Absolutely.
Ludovic:So a bit like when people say, you know, it's it's cheating if you're writing and you're using AI. It's like, look. Some people were native speaker. We have decided that English was a good global language instead of French. We we lost by a small margin, but we did.
Ludovic:So English is very warm language. And then there's a style of English that is well accepted. And some people can write very quickly in that style of English. And and I can ensure you with a you know, that there are many people what when I when I read what they were writing when I grew up, it was a lot of fluff and it sound very much like a GPT thing today. Right?
Ludovic:Because people you had a number of people that could write this b s fluffy thing that sounded very good but was saying nothing. Right? So that human skill exists. Some a few people had it, and so people made a career out of, like, the fluff. Right?
Ludovic:And they say, oh, but you cannot use GPT because you raise fluff. He's just now he's democratized. So there used to be only a 100 people who could just, like, do shit their way through. Now everybody can do it. Right?
Ludovic:So and so it's the same with documents. Like, I had only a few legal advice or or advisers of LPs that had read sufficient number of documents to give me an advice about that specific document. Why why would we stick to that model? Right? The same if I give it to GPT.
Ludovic:But I was very impressed early on with GPT once I because I I had a a colleague who created a company on on analyzing automatically LPS. And I had a lot of LPS, and I gave them to him to, like, to train in a closed and and I gave one LPA, I think, to GPT, and and and and so mine doesn't train on documents. Right? So perversion. And I gave an LPA to GPT and say, can you tell me if anything is unusual in this in in in in this LPA?
Ludovic:It's a very specific document. It's a fairly technical document. And I remember being amazed. Like, GPS are like, this is a bit less. Like, many document LPAs I've used in?
Ludovic:You're not you're not supposed to have seen any of these things.
Victoria:But that's also the scary part. Right? Because you don't know if those suggestions necessarily are
Ludovic:Yeah. You don't know
Victoria:if are market or off market.
Ludovic:Yeah. But I I I kind of, like, knew and still like flowers. Like like, then I I was like, okay.
Victoria:And would love to close by just asking you again for any advice that you have for LPs or allocators who are early on in adopting AI. Are there any kind of initial places that you would start or any governance areas that you'd really focus on?
Ludovic:Don't be naive. It's a very technical topic. It can bring you a lot. So so at the same time, don't be naive embracing it saying, oh my god. This is gonna so simplify my life, and I'm gonna do, like, so well by just adopting it.
Ludovic:Right? So that's the naive response. And don't be dismissive either. Like, the other opposite response saying, like, whatever. Like, nobody's using that.
Ludovic:It's like these things. It's just generates BS. Whatever. I I'm I'm sticking to my old way of working. I don't need any of this.
Ludovic:I I I think you need to employ these tools. They are really powerful and impressive, but you have to think very deeply about about bringing them in, how the governments, the risk it brings, and and and so on and so forth.
Victoria:Thank you so much for joining me for an amazing conversation.
Ludovic:Thank you very much for having me.
Victoria:That's a wrap for this episode of The Agentic Allocator. If today's conversation gave you a clearer vision of where the industry is headed or helped you pinpoint exactly where your own process is stuck, go ahead and follow or subscribe wherever you get your podcasts. And if you're curious what Agentic AI might actually look like inside your investment office and how to get there without compromising on security or control, visit auumai.com for demos and resources on AI native LP and Allocator workflows. Until next time.