Proof of Work: AI Value Creation
00:22:37 · Stuart Willson
[00:22:37] S0: Yeah. Uh you spoke to reliability and you spoke to a robust data model that is critical to ensure reliability. What does a robust data model look like for a fund that is exploring portfolio monitoring for the first time? What does that look like? Is it a folder with a bunch of files in it? I know it's not. But uh what is it what does that data model look like? And what are the steps that one needs to take to go from Excel sheets to something more robust? Jeff? Jordan, do you want to get started?
00:23:14 · Joe Zein
[00:23:14] S2: Yeah. Uh I can get started with this one. Uh a robust data model is definitely not an Excel sheet. The thing is you need auditability, you need, you know, uh version version, you know, you need to be able to see the versions. You need to be able to know who edited what, who did what, you need to have rules on top of it, and Excel obviously does not allow you to do any of that, right? Excel is amazing because it's a canvas, it's a free canvas, right? Open canvas, or you can do anything you want, right? But there's no way of checking rules. There is no way of limiting someone to add to add a row. You see an Excel sheet, if you have added access, you can do whatever you want with it. And when we say a governed and proper and a good data model, I think one, there is those key criteria that we have to have, which is auditability, traceability, version history, all those different things. The second thing is this is on the call metadata layer or on the envelope layer. The other thing is, well, what do you use a data model for, right? You use a data model for to query it and to write to it. So there is years of computer science experience of like people who have played around and toyed around and tested different different data structures to see what is optimal for retrieval and what is optimal for writing data. And what is optimal today, and why this is becoming more of a topic of interest today, is because you also want to feed something to DLLM that is structured enough, right? There's this analogy that I like like to give. You go into a library and you see the books all over the place, and someone is asking you to give them a summary about a specific topic. Today, the way DLLM is gonna work is if you just give them a file system, right, of a bunch of Excel files and PDFs, is gonna go and try to read all the books based on text work on text similarity and try to come up with a summary. And you see why this could be problematic. On the opposite side, if you have everything in clean tables that are kind of self-documented and that are organized in a way and that have constraints, both at the database level but also at the metadata level, that that help you understand that you know each row is an individual transaction.
00:25:22 · Joe Zein, Stuart Willson
Each row, so so now you have uh that library is all of a sudden organized and it has some rules and assumptions that you can trust to go on and build your analysis. So, for example, if I tell you that in this library, in that section, that's the health section, on that row, you're only gonna see books, books about the heart, then I can, you know, I know that I can only go that I only go need to read the books that are on that shelf and not read all of the books because there's not gonna be one misclassified book that is not in the right section. [00:25:55] S0: And so there's a real difference between AI doing search over a pile of documents versus querying governed structured data. How do you actually move a firm from one to the other? And how long does it take?
00:26:15 · Joe Zein
[00:26:15] S2: So tools like what Yann's building, like CRMs, those are uh a big step in one of those directions, right? And in that direction, where people have adopted at the end of the day, if you're using specific SaaS tools, those are already governed models, right? Like they they're using databases under the hood, they've gone through the job of abstracting that world, trying to design it in tables that are referencing each other. So those are you know a big, you know, help us bootstrap that implementation. But then you you will always have other than your CRM, your portfolio monitoring, the different SaaS tools that you're using that you can pull tables from using the API, you still have a bunch of files living in the file system. And there's a lot of techniques that we use that we go and pipelines that we run where we're you know reducing the set of all of the data into a canonical set of data, and then you know, removing all of the duplicates, and then removing all of the files that were that we think are not useful. And then we're using AI on top of that to be able to classify files. Is this a SIM? Is this a um you know, what kind of document is that? Is it a presentation about a portfolio company or is it an IC deck? And once we were able to process all of these, we're able to look into those documents, do extractions of things that would have not gotten extracted by the portfolio monitoring system, right? Because at the end of the day, the portfolio monitoring and valuation system is for anything post-closed. But then you also have all this material about all the deals that were looked at. Those never made it into something like 73 strings, right? They may have made it to some extent into the CRM, but not fully, right? And so capturing, looking back into those files, first selecting that subset and two looking through those files in a cost-efficient way and extracting those records and putting them in that database where we would have already pulled the records from the databases from 73 strings and from Salesforce and from different places. And then on top of all of that, deduplicating entities, because the last thing you want is to have a reference to Goldman Sachs that is GS somewhere and have it in you know 73 strings as Goldman Sachs Inc., you want to know that these two are kind of the same entity.
00:28:29 · Joe Zein
And it's this multi-step process that we usually take to move all of the unstructured data of a company into that structured database that is yet that is then used, that is very well documented to be used by AI.
00:28:43 · Yann Magnan
[00:28:43] S1: Don't have uh much to uh to add to uh to what Joe Joe said. I mean, I think you covered it all at the end of the day. Uh what I would add is uh that um it's a lot of a change management uh type of process. Um you whether you start from a blank sheet of paper, like there's a lot of uh change management at that point in time, or whether you start from uh um uh another system, legacy system, and you want to move to something that is more um AI native and that will help you uh uh generate more value out of the uh LLMs uh algos that are uh out there. Uh at the end of the day, you have to change a lot of things. Uh you have to change uh how people think about uh data, how they think about processes. You have to spend time understanding what's going to be the correct data model that's going to uh that's going to help you monitor your assets. But I think the opportunity is much bigger than that, is actually to be able to extract intelligence, insights across the portfolio. So a lot of a lot of thought has has to be put into that to make sure that you put in place the solution that really makes sense and that will really take you to the next level. So change management is something critical in terms of timing. It's not fast, it's not something you're going to do in a couple of weeks. You know, most often would be a quarter, two quarters, three quarters, depending on the size of the client and on the size of the portfolio and whether portfolio covers multiple strategies, multiple geographies, multiple sub-businesses. As again, you want to be able to uh you want to make sure that you are able at the end of the day, once you've done that work, that you extract intelligence and value not just at the portfolio company level, but across the portfolio. So that requires a lot of uh a lot of time. Yeah. Some time.