How to Make AI Practical in M&A

0:01This episode is brought to you by Firm Room. Searching for a data room that offers simplicity without compromise? Elevate your data room experience with Firm Room, the world's most intuitive virtual data room. Try it free for 14 days, no strings attached. And when you're ready to power up, sign up for unlimited users and 10 gigs of storage at a flat rate of$495 a month. Boost your team with Firm Room, the tool built by dealmakers for dealmakers. Check it out, firmroom.com. Again, that's firmroom.com.

0:41I'm Kisan Patel, and you're listening to M &A Science, where we talk with deal professionals and learn valuable lessons from their experience. This podcast focuses on stories, strategies, and what actually happened during M &A deals.

1:05Hello, M &A scientists. Welcome to the M &A Science podcast, where we learn from the best in M &A to uncover proven techniques for enterprise value creation. If you're interested in learning more about how to optimize your M &A practice or want to get involved with our community of forward-thinking M &A practitioners, visit mascience.com and subscribe to our free weekly newsletter for the latest in industry trends, insightful content, and community events. If you want to keep up with us on the go, head over to LinkedIn and follow M &A Science. I gathered up some of the top AI pros here at Boomi World to break down AI into practical use cases.

1:46It's also going to give you a sneak peek about the AI M &A Teardown series. If you're not familiar, it's a new series that I'm working on to really get past all the hype out there and just show you real use cases of using AI. Hope you enjoy this mini interview. Joining me is Michael Bachman, head of research, architecture, and AI strategy at Boomi, and Chris Capetta, principal solutions architect, also at Boomi. Gentlemen, thanks for taking a little time and chatting more of some nitty gritty about all this emerging stuff, all this chatter about AI. But I want some real answers because it's too much hype and a lot of talking heads.

2:22And some of them, I really don't think they know what they're talking about, but you guys do. You came referred as some credible experts. We hope so. Looking forward to it. I know it's a lot, but we got one mic to share since impromptu interview. If you want to do just a real brief intro on your background. Started Boomy in 2017. I actually came into sales engineering. I was working on the Boomy workflow product, sort of an application development sort of thing, and ventured into the integration area and found myself in product working on Skunkworks AI stuff. So it's been a really enjoyable time to be down the rabbit hole.

2:52Yeah. And I came into Boomy in 2018, just a year after Chris did. But I have a background in memory systems, storage, in storage engineering, pre-sales, as well as linguistics. Great. Let's talk about making AI practical. There is a lot of hype. You were talking about that before. And what we want to do is turn the hype into reality. So we have the pleasure of doing that on a daily basis, especially you, I would say. He calls himself head button pusher because he's constantly on keyboards, trying to make sure that stuff works or tries to break it. Making AI practical really means to understand what goals are and then to also orchestrate how we're going to make LLMs work for customers in a variety of different ways, such as RAG, fine tuning, any type of agent building, that sort of thing.

3:43That's what we do day to day. Cool. Yeah. I'd add to that that the models are pretty impressive as they are, but they can be made more impressive with the right logic and sequencing and steps and extra data. And so all of that sort of stuff is where Boomi has really found itself in a phenomenal position. Sort of contrasted with AI functionality in the platform, we do a lot of AI designing using the platform. So we have a platform that already does data connectivity and orchestration and logic and data transformation, all of those sorts of things that can help an AI model be better. The AI models are really just another point in a potentially 50-step process for us.

4:17So it lets us build out a lot of pretty creative designs quickly that include AI, but that also include connectivity and mapping and all of those sorts of other things. So it's been a really enjoyable thing to explore. You mentioned three areas to really make this dialed in to be useful. RAG, fine tuning agents. Let's break them down. Sure. What does RAG even stand for? So RAG is Retrieve Logmented Generation. It's basically, I like to think of it as a particular use case. If you ask ChatGPT to summarize something in your Salesforce org, it'll say, I have no access to your Salesforce org. But you can take that extra step and you can retrieve the right information or relevant information, even if it's a big, ugly 30-page string of notes, you can feed that to one of these models and then ask your question and it'll respond better.

4:59So all of that can happen programmatically. That's the concept of retrieval augmented generation. You're basically able to help the models respond better by getting information that maybe it doesn't have access to or is maybe outside its training window, or maybe it just didn't even focus on it when the vendor was creating that model. But by getting the right information, even if it's big and ugly, included in the prompt, the LLM can do what it excels at, which is understanding the user's question and parsing through a massive amount of information and using that if relevant to respond to the user's questions.

5:28You're kind of letting the large language model understand what the user wants, and you're giving it enough info to respond better than the model could by itself. I feel like there's a lot of misconceptions about AI where people just think, we do have an A. So the big use case is always the data room. The data room, when you're doing a transaction, has got hundreds, if not thousands of documents, and everybody's like, oh, we'll just slap AI on top of it. And you can, wait a minute, there's some challenges with doing that. You can't just stuff thousands of documents that could be hundreds of pages long and expect it to really digest that information.

5:57Maybe we can sort of hash that out a little bit more of like why that is the limiter. And then I think there's rag in concept versus rag in reality. And it seems like it's been an evolving state. Yeah, definitely. So the context window, the amount of sort of words that you can feed into an LLM. It's growing super fast. And that's really impressive. It used to be even 10 months or so ago, if I were talking to even the earlier versions of GBD4, I would probably ask it 10 or so questions and then it would lose the plot and I would start a new conversation and pick up where we left off. But now you're really at the point, you can have a many hour technical conversation with it and it'll track.

6:31It'll maybe get a bit slower. It'll think a bit harder because it's got to read through a lot of stuff, but it'll still follow and it won't just go off the rails. So it's really handling a lot more context really well. But that's on the scale of one few-hour conversation. If you're looking at a whole documentation set or all of Wikipedia or something like that, you're still way outside what it can do. And also, it seems like some of even these top models that can handle massive amounts of context, they can handle them technically, but they can't handle them well. For example, if you look at like a Google Gemini million token context window or what have you, I've seen some people do tests.

7:04I haven't done these myself, but I've seen people do tests where they say they've hidden an Easter egg somewhere in this massive amount of context. They ask it about it and it says, oh, no, I don't think it's in there. And then they say, no, I'm pretty sure it's in there. Look again. It's like, no, I don't think it's in there. And then the third time, maybe they'll say, I'm certain that it's in there. Find it. And it'll say, oh yeah, it's totally there. I see it now. You can feed it all of this context, but it might not pay attention all that well to all of the context. That's the sort of limitation of the context windows.

7:29I think those will get much bigger and much more reliable in the time to come. But for now, it's better to give a smaller amount, a relatively smaller amount of accurate context. I'd rather give it three pages of really good context versus 100 pages of context where a couple paragraphs are useful. Yeah, when we're talking about a deal room for M &A, you've got multiple parties in there. You've got the lawyers, you've got the printer, you've got the parties that are buying or selling. They're all going to be looking at contracts as an example. One way that RAG could be super useful is summarizing what's in some of those contracts, especially if they're super lengthy, which in M &A they absolutely can be.

8:07If you have a small context window, that's not going to help. Also, if you don't have the details that exist in some of these contracts, you're going to need to augment this linguistic model, which is an LLM or a language model of any size or kind. The idea with the deal room to make that more efficient, just to use your example in M &A, having RAG in conjunction with an LLM is critical. Now you can get into the business of seeing what's relevant in those particular documents, how those documents are going to be validated along the way, and lots of other information that's only going to be isolated to the deal room itself.

8:46If you just let the language model operate in the deal room, first of all, you may not get the outcome that you want, primarily because those models have only been trained on public data at this point, and not on specifics of what's going on in your deal room. Although they'll still give an answer. That answer will sound really good, but it's going to be factually inaccurate to a greater degree for much of probably what a deal inside of that deal desk would require or that deal room would require. It's like the biggest concern is current state is inaccuracy because that's the last thing you want to have in a way is rely on it.

9:18And then it's giving you inaccurate information or missing some of those points you may be trying to dig into. 100%. Two ways that I'm thinking about it. One is summarization that's going to help all of the parties understand what's in the documents. Two is mapping. So the idea of understanding how one document or one process could map to another, even in a deal room itself, that could be critical. I'm no expert on deal rooms and what goes on there. I have dealt with some of that in the past. What I can say is that language models are really good at reading in text and responding to it. Rags just make that whole process a lot more efficient.

9:55How do you define this sweet spot? The reference I've heard before is taking unstructured data and structuring it. I don't know if you have other references to sort of give a good perspective on how LLMs are most useful. Yeah, so they're useful in a lot of ways. The nice thing about LLMs, to your point, is they are dealing with unstructured data. And the unstructured data represents 80 to 90 % of all data that's created. Yet it also represents a problem that's been largely unsolved heretofore. If I can rely on an LLM to provide some sort of structured feedback based on unstructured data, that's really cool.

10:34And so what I often say is, the nice thing about LLM technology is this. Before LLMs and with structured data, traditional machine learning or discriminative machine learning could be applied, but it forced the human to sort of think like a logical computation system. For instance, a SQL statement is not a natural language thing that we would do in communication. You don't say select star from this table where this condition is met, do these things and return these things. That is very much a structured data return. Instead, I want to ask in plain human language of any kind to the LLM to go and provide that retrieval for me.

11:16What the LLM is doing is it's offloading the thinking, in air quotes, to the actual model itself. Now, I as the human don't need to think like a computer. The computation system is emulating human-like thinking. For me, and I think yours as well said, LLMs really excel at handling the grunt work. I think you still generally want a human in the loop. But if an LLM can make a human's process 80 % faster and better, that's a huge win. So I think LLMs can handle the grunt work. And I think that can be both from the perspective of summarization and aggregation. It can also be sort of a companion, a conversation to help you work through an idea.

11:55So often at this point, if I want to learn something, I no longer go to Google. I fundamentally have changed the way I interact with information. I'll start a conversation with an LLM. I'll get some info from it. Then we can have back and forth dialogue. I can get to the idea or the end result that I was looking for more effectively than just throwing Google searches up against the wall and hoping someone has both done what I'm trying to understand and happen to write it somewhere that I'll stub my toe on it. It's a different way to engage with information. And in that way, I'm able to have some, I find pretty productive dialogues with LLens to work through different ideas and concepts and designs and all of that sort of stuff.

12:32So true. I got a flower to meet our engineers next week and they want me to do a workshop on just teaching them sourcing through value realization M &A. I'm like, whoa that's a lot i threw it in chat got an outline gave it some feedback like we expand this area and it's just that's all i needed was just to have that little conversation totally agree now we referenced discriminative versus generative and it sounds like discriminative was the prior iteration of ai that's still pretty quantitative in nature that you're getting some technical components you know some rules that you got to really define for it to get value versus generative is more qualitative we reference the conversation and it's i think that's what's bringing it to being more of a breakthrough thing that everybody's referencing.

13:13You're definitely on it. No, discriminative is probably the better way to call classical or traditional machine learning. A lot of our colleagues, especially in this very incipient phase of generative AI being launched wholesale through chat GPT, coming into the zeitgeist, all of that stuff. It's like we referred to traditional ML as what I now call discriminative or what the industry or what data scientists have called discriminative forever. So the idea is one is basically very quantitative and focused on structured or semi-structured types of data. And the other one is generative, which is not only filtering out information, but it's also generating new information as well.

13:55So mathematically, one is just taking from a set of data a subset of potential responses. The other one is doing that and also generating new responses as a result. And that is great for unstructured data. So you're right on the money there. So I'm actually going to say some things. And I'm interested in your take on whether I'm understanding discriminative AI correctly. I would think of discriminative AI as often used in classification. So you would train a model specifically to do something like recognize the sentiment of an Amazon review or something like that. I would think of that as saying this is a positive or it's a negative, it's an either or, it falls into these buckets.

14:34And I would think of that as how most traditional machine learning models functioned. The large language models are doing a lot more. They're trained on so much information that they can, in really interesting ways, do a lot of those things quite well without sort of purpose-specific training for that. So I can, with the large language model, give really good instructions to say, I don't want you to answer this question. I want you to classify the sentiment of this. So in that way, the large language models are kind of a fascinating leap up in that they can do a lot of the things that a discriminative model could have done, but they didn't require purpose-built training for that.

15:11They just know so much that they're able to handle it. And again, like I said before, it's like you filter out data and then you generate new data. So classification techniques would definitely use that and be useful even in generative models. I'll also say you talked about classifiers or classification in general. Clustering, classification, regression, prediction, prescription. Those are the elements that I would say are really in the domain of discriminative AI. Those can still be used by generative models and indeed are because you are predicting the next token that's going to appear. So there are these embodiments of discriminative modeling inside of generative models.

15:50But really, generative models are using different types of technology, like transformers are completely different than even classifiers were and that sort of thing. But the idea between these things used in tandem is really where I think the value is going to be. You use quantitative and qualitative methods. You use structured and unstructured data together. And you bring these models together in such a way that they can be used as ensembles. And with that, you create pipelines of better and better outcomes. And this is why it's important to fine-tune, to have RAG, to have agents that are purpose-built to fetch certain data or perform certain functions, and then feed that back into what these models will ultimately do to lead to an outcome.

16:32But I agree with exactly what you said, too. Let's talk about fine-tuning. What the hell is fine-tuning? Yeah. So fine-tuning, it could be differentiated from pre-training. Pre-training is when you give it basically the internet and you tell it, here's how you predict the next word. Fine-tuning is when, in the sense of a big model, they're making it so that it responds to questions instead of thinking it needs to generate a follow-on question, that sort of thing. From the perspective of me as a builder, I find that the fine-tuning I'm able to do generally happens... I don't have the compute to do pre-training myself.

17:03So I'm able to basically take a model and I could do fine-tuning to turn it into a discriminative piece of functionality. I could do some fine-tuning to say, I want to have the Llama 3 model do some classification or some routing or something like that. But for the most part, in my experience, fine-tuning is best today at modifying the tone of how a model responds, not overriding the facts of the underlying model. So I would also actually add prompting into this. I would think of using good prompting and instructions to have the model do what I want it to do. I would think of good retrieval as having the model know what I want it to know.

17:41And I would think of fine-tuning as having the model sound how I want it to sound. And I kind of use those in different ways. I would say fine-tuning for me today, not actually that useful. There are definitely really impressive fine-tuning things that are going on on the scale of like an open AI or an Anthropic or something like that. They're doing probably, I would imagine, hundreds of sequences of different types of fine-tuning for security and for answering and for reasoning and for all of those sorts of things. But on the scale of what I can do to just take an open model, fine-tune it myself, yeah, I think fine-tuning can modify the tone.

18:12But I think if I want the model to know different things, I'm going to retrieval 100 % of the time. And I'm just having the vanilla model be given much more information and context in the prompt rather than trying to override the underlying facts of the model by fine-tuning it. I love that answer on so many levels. The only thing I would add to that is I think there's a trade-off at some point where you're going to do retrieval and then fine-tuning. And one of those areas could be very domain-specific responses. Like you were saying before, you take a fine-tuned model and that's going to be, what did you say, on the tone?

18:50If you want your tone to sound a lot like an auditor, then maybe you have a model that you're fine-tuning on auditing type of language and nomenclature. Basically giving it a corpus of data that is commensurate with the type of actor who would do the performance of an auditor, as an example. You may come out with a way where you use both. Maybe you don't want to use the LLM, the big linguistic model to answer routine questions that only an auditor would repeatedly ask under certain circumstances. Maybe that's a good time for a fine-tuned model to come in and retrieve the right data in the process.

19:28So it sounds like you're like pre-prompting. I'll use an example using Dealroom here again because it's the one I'm most familiar with. But we have our AI platform that we build on top of whatever learning large language manager they're using. but we'll have instructions in there that basically says only provide responses to information that you can reference in a document. I mean, you got to be able to reference where the source was. Don't give me anything outside of that document. It's just give me responses of things that you can actually reference. That example of fine tuning. I would say that's an example of probably prompting and retrieval.

19:58The instructions you're giving it are good prompting. You're telling it what you want it to do and the documents you're showing it are the retrieval. That's where you've just fetched the relevant documents. I would say that's probably not fine tuning. I think of fine tuning as modifying a copy of the model. So instead of calling the Llama 7 billion model, I'm calling the Chris Capetta May 8th copy of the Llama 7 billion model. So it's actually fundamentally changing and getting a different copy of the large language model. And I don't know to what degree this conversation is like a Boomi sort of relevant thing versus just general technology thing.

20:29But I will say, I think fine tuning will become stronger in the years to come. And from what I've seen, using Boomi to pipe the data in and format it and prepare it from wherever it may sit in whatever format it might sit. As fine-tuning becomes more useful, Boomi will continue to be a good sort of pipeline for that data. We definitely use Boomi example because we are working on doing, if some of you listening aren't familiar, but we have a new series called AI M &A Teardown. And so we're going to have probably both of you as guests to actually do some deeper dive and use like actual Boomi examples for demo.

21:00So yeah, happy to talk through those. Anything else on fine-tuning? Because I do want to touch on agents. I think we've covered a lot of ground on fine-tuning. I do. There's only one thing I would like to say. I often think of fine-tuned. I think you just said it. There's the Lama 7B variant or some smaller model. And then there's the Chris Capetta variant on that as well. I could see fine-tuning as we multiply ourselves and our agency. I can see fine-tuned models for each one of us as humans to interact with the world as if it were in our voice in various contexts. Again, it's probably going to involve a combination between retrieval and fine-tuning.

21:37But that fine-tuning component with that linguistic model, again, that's the key. Language models are essentially taking syntax and converting that to semantics. The idea is, if you're going to construct a sentence and you're going to derive meaning out of it, language models are really good for that sort of thing. And if you can put it in your voice, that I think is going to be a really good frontier for where fine-tuning is going to be helpful. Very cool. Let's talk about agents because you talked a lot about agents earlier today. Let's define what those things are. Yeah, it's sort of a nebulous world right now.

22:10I think you might have a couple of slightly different answers here, but let's see where we go. I think of agents as functions. In Amazon, AWS, they're Lambda functions. They're task-based types of fetchers and putters and getters and those sorts of things. They're there. These functions are ephemeral, and they perform a particular purpose and then either go away or are somehow latent. Fundamentally, as an infrastructure person, I think of agents as just that. They're basically task executioners. It's a term that has really spanned and grown in the last year or so. I would say probably the technical definition of an agent I could think of as anything with generative AI using tools.

22:54And to that degree, I think there are a lot of agents out there that are doing really cool things. For me, I think the most exciting element of agents is when you can start to get into higher level goals and have it make decisions as to how to achieve those goals. And then even beyond that, you could imagine a scenario where you could give an agent or an agent swarm, a bunch of agents interacting, some goal or some sort of high level, just general directive. Say, hey, we would like you to work on food shortage in this nation state or something like that. And then have it be able to have both the autonomy, the access, the decision making, and the tool set to go out and decide what goals go into that initiative or objective, and then decide what tools go into those goals.

23:36There's a varying degree of agents at this point. And I think they can span from low-level decision-making AIs that have access to external tools of some degree, all the way up to HAL 9000, but a nice version of HAL 9000, hopefully, is what we would go for. So basically, have it actually do things. So for me to prompt it to go change my flight, that would be an example. that you've got tasks that can actually do. And tools to go actually make the change and send the emails and call the APIs and all those things. So yeah, 100 % I'd call that an agent. Yeah. And the other thing too, we talk about this all the time, but making sure that results are presented in the right sequence or if it's unnecessary to do that, if it doesn't matter if you do tasks or perform tasks in parallel, that's fine too.