Introduction

Today I want to talk a little bit about how we can use GenAI to do M &A due diligence acceleration.

Speaker Background

So a little bit about me.

From Traditional M&A to GenAI Acceleration

I was doing a lot of M &A work for the past 15 years and what I realized was a lot of

the work that I have been doing could actually be accelerated through GenAI.

So that led me to start this company.

Founding the Company and Early Traction

We started this company earlier this year and very soon we

were able to get to positive revenue.

We have had pretty good traction in the market around

this product.

Problem and Product Overview

So what is exactly the product and what is the problem?

Why M&A Due Diligence Is Slow and Costly

So when you are doing M &A

due diligence you have to really evaluate a particular company and it

takes a lot of time you know you have to screen through a lot of deals you have

to once you get a deal you have to understand what is happening with the

company you have to understand you know how is this company doing versus other

companies how's the market doing is this company worth investing in or not if I

invest in it is there going to be synergies what kind of synergies would I

get and historically when I have done this kind of work have I been successful

and how do I use that knowledge to inform the the the the new deal so all

of this stuff takes a really long amount of time you know met just an example you

know the the investment bankers who are working in this space you know in spend

spend 80, 90 hours a week working on deals, it really takes a lot of time to kind of work

through all that data.

Where GenAI Adds Value

And that's where GenAI actually can come in.

It can actually read through a lot of the documents that are presented to you.

It can understand it.

It can accelerate a lot of that workflow, which is what we have done.

Inputs, Workflow, and Time Savings

so what we have done is you can take a lot of different deal documents whether

it's some sell side document whether it's financials financial statement

industry research and you can you can put it into our system and that would

give you the output that you are looking for and it will it will basically cut

cut off weeks of your work.

Use Cases: Market, Company, and Growth Analysis

So, if you are evaluating a particular market, it can give

you a quick view of the industry.

If you are evaluating a company, it can quickly help

you figure out, you know, how the company is and how it's positioned versus other competitors.

Is this worth investing in or not?

And if you are actually running a company, if you

you are working in a company and you really want to understand how can I grow this company?

What

are the dimensions around which I can grow this company?

This system can give you a growth strategy

report as well.

Financial Analysis Capabilities

If you put in your financial statements, it can give you a detailed analysis

of the financial statement, can help you figure out, you know, how your income statement, balance

sheet etc are going what is your you know financial profile where do you need

to focus on so that's the system that we have built so let me pause here for any

questions around you know what the problem is and what are we trying to

solve for okay if not let's keep moving

How the Agentic System Works

So, we use an agentic AI system.

Agent Teams and Roles

That agentic AI system can really, you know, perform like a person.

We have trained individual agents to perform like investment teams.

So, for example, there is a data

acquisition agent there is a financial statement analyzer agent there is a

partner agent all of these agents are built using different LLMs and they are

trained using different LLMs and once that happens you know they can all work

together to give you the output it definitely requires a lot of different

Data Integrations and Public Sources

data integrations so we have a lot of public public data integration that can

that can work together with our system and so these agents can work on that data to kind of create those outputs.

Benefits: Coverage, Depth, Speed, and Cost

Once we have that, you know, there is a lot of benefit of kind of using this AI based system.

First, it's super hard, you know, to have a full coverage of all the issues.

With AI agents, you can actually go through thousands of sources.

You can have pretty good coverage of the industries.

industries, you can really completely look at all the issues that a particular company

is facing and really kind of do a very quick red flags type of analysis.

The depth of insight is very, very detailed.

You can really go to, at a company level, at a segment level, you know, really kind

kind of look at individual granular segments, business units and start to figure out what

is going on with all of these different segments.

It is obviously very fast, you know, a commercial due diligence work that I used to do used

to take me four weeks, now it can be done in two days.

At least it gives you a quick answer and you know, you can accelerate it to the same level

less than a week.

so it it does reduce it you know pretty dramatically the cost is pretty low and

so you know depth of insight coverage speed and cost yeah yeah yeah that's a

Current Performance and Human-in-the-Loop

that's a good question so as I said you know we have actually 15 plus clients at

this point who are evaluating our system and subscribing to our system after

analyzing that whatever they have been doing so I wouldn't say it's 100 % there

but you know it will give you a first 50 60 percent of the answer there that you

can then kind of focus on the insight and and the human layer on top to get to

the right answer so I wouldn't say out of the box it gets you to the hundred

percent answer but it takes but it does a lot of the work the the hard work that

needed to do to kind of look at all these different documents, kind of put it all together.

It would take all of that, it would put all of that together, it would give you an answer and now the expert can

actually make an expert judgment.

But the expert's time is spent in analyzing

the insights and creating insights, not doing the detailed work.

Yeah.

Commercial Model and Data Onboarding

Subscriptions and Project-Based Work

So, it's a subscription service plus it's also project -based service.

So sometimes, you know, before they subscribe, sometimes they actually ask me to do projects.

And once they evaluate the project, then they say, OK, well, this is good.

We might want to subscribe to you.

You actually.

Public Data Connectivity and Private Data Uploads

So if you are looking at like a company analysis, you can actually not upload anything and it will go out and do a lot of public analysis.

it's already connected to SEC it's already connected to Edgar it's already

connected to public filings it's already connected to transcripts so all of those

things are already pulled in so you just have but if it's a private company

obviously there isn't that much data but it will still pull in a lot of data and

you can build these reports without uploading anything but if you upload

some things then it gets better so for example a private company information

you can get a lot of market competitor commercial data but you wouldn't get

financial data so you would upload financial data and then the whole report

Quality, Hallucination Control, and Robustness

would be complete yeah how does the model fight disinformation so we take a

lot of time to figure out the right answer and there are many ways to doing

that first is a hallucination is a problem so for hallucination you know we

Prompt Decomposition to Reduce Hallucinations

We have worked on trying to figure out how do you minimize hallucination.

So we kind of break the prompts down into thousands of prompts.

If you actually ask an LLM model like a very specific question, the chances of hallucination

are lower.

Versus if you ask a generic question, then it has sort of agency to talk about a lot

of things and hallucinate.

So we have tried to kind of first figure out how to really break our questions into very

specific question so that's one thing then we have a generative model and then

Generator–Checker Dual-Agent Workflow

we have a checker model so the generative so there is agentic team that

is generating the answer and agentic team that is checking the answer so that

checking the answer team actually their whole task is to say whatever this this

team has generated is wrong and identify five different reasons why that is wrong

so it basically would do that and it would it would take its own context to

basically say this is wrong and so both the teams now need to work together to

come up with an answer that is acceptable to both the teams so these

are the two agentic teams that are working together to generate these

answers that's a second approach the third is and as you might have realized

if you prompt chat GPT one one question and if you ask the same question again

It might give you a different answer.

So that's another sort of approach that we had to work through to get a robust answer.

Consistency via Multi-Angle Questioning

So for that, we we basically use kind of an FBI interrogation system type of thing where you ask like different questions,

the same question in many different ways to try to get to the right answer.

And we kind of get to the answer by asking questions many different ways and arrive at a more robust answer.

So there are many sort of systems in place to make sure we are producing the right quality,

because that right quality is essentially what is needed for the clients that we have.

Why Quality Matters for Financial Clients

The clients that we have are private equity companies, are consulting companies,

our investment banks and our portfolio companies so if there is any anything

wrong then then you know then the trust is broken so we had to really work on

getting the quality high okay with that that was that was it on my side I want

to give you a little bit of demo of the system also sorry did you have a

Technical Stack and Model Selection

a question yeah they they are used we're using kind of all these standard sort of

Agent Frameworks and LLM Choices

agent building method langchen etc to sort of build these agents we are using

underlying different agents have different underlying sort of LLMs for

example data acquisition agents you know they need to be aware of they need to be

aware of like the recent news.

Selecting Models by Task (News, Checking, Sizing)

So what we found was Gemini or Perplexity might actually have a

much better understanding of the recent news than let's say OpenAI.

So this agent that is

like tasked to get the right sort of latest news answer will be a Gemini or a Perplexity.

We found that the checker agent needs to be like a Claude because Claude is very good at sort of

poking holes it seems like then for identifying market size we found that some other LLMs are

Ensembles, Guardrails, and Scopes

much better we kind of use an ensemble of agents to work together and you have to define guidelines

you have to define guardrails etc for every agent so that they can operate within sort of their

defined Uniscope.

Demo: Industry Report Workflow

So let me show you what a report looks like.

Defining Scope and Geographies

So for example, if

you are evaluating a particular industry and you wanted to say, hey, you know, I

don't know, I'm evaluating, you can say any industry, you

know let's say toy you know or you know toy manufacturing or something like that

and here are all the countries that we support so we are our our customers are

primarily in the US but we are getting trialed in Europe in in Australia and in

Japan so you would that's all you needed to do then what what happens is then if

Document Uploads and Context Building

you have any documents you can upload those documents so for example for

private company you might have financial financial statements or something like

that for industry report it doesn't matter it these might be other industry

reports that you you might want to upload and then what it does is it it

It spends a little bit of time to understand

what toy manufacturing means, who the key players are.

And then you generate a report.

Report Generation Time and Output

So that takes about 30 minutes to generate.

And that report will look like the following.

And this is about a 100 -page report

that comes out that can really help you get up

to speed on any particular industry.

Flexible Industry Definitions

The interesting thing about this report is it doesn't need to be like a Nix code based

or you can basically come up with any combination of words and it will figure out which industry

it is.

What the Report Covers

So it will give you a historical overview of the industry, why does the industry

exist, what is current state, future state, what kind of

industry segments are there, what are the products, what are the key players,

shares, industry trends, who the customers are, how do they buy, what is the market size

of the industry and it evaluates the market size by looking at, you know, existing reports

that are out there or it would look at, it would do its own sort of triangulation, it

would do its own math to figure out, you know, what is the size of the industry.

Market Sizing Logic and Triangulation

So, for example, it will basically say, okay, let me add, let me say how many children,

so this is a preschool education, we want to find out what the market size is.

It will think very logically, it will say, in order to figure out the market size for

preschool education, I'm going to actually see how many children are there between zero

to five years of age.

Of them, I would need to figure out how many of them are enrolled in a preschool and what

is the average monthly fees.

For each of these, it will figure out some sort of assumption and will provide you the

source for that assumption.

Then it will give you a sense of what is the size of the market.

So it actually like when you ask it a question, it actually goes back and understands, tries to understand that question, breaks it down into smaller questions, tries to answer the smaller question and then kind of answers the bigger question.

Operating Models and Expert-Level Reasoning

So very sort of logical agents that we have built here.

It can highlight, you know, market growth accelerators and inhibitors, how the industry operates.

So, this one is interesting because, you know, there is always a question about experts versus non -experts.

Can you really kind of go into a particular industry without being an expert and create outputs that are as good as an expert?

So, I think the answer to that is yes.

I think the answer is we are getting to a level where AI can actually kind of get to

a level which is equal to, you know, many years of expertise because if you start to

walk it through the right approach, then we'll figure out, you know, how the industry works.

Once it knows how the industry works, you can build on top of that and, you know, kind

of build that expertise very fast.

So this is an example, you know, like for this particular industry, it kind of really

understands how does the revenue get created, what are the cost structure, what

are the operating metrics and so now it's already gotten up to speed on that

industry.

Deals, Competitors, and Positioning Dimensions

You know what is the deal timeline, you know how many deals are

happening in this space, who are the competitors, who are the large players in

in this space?

Where do they operate in?

How are they all competitively positioned?

So it identifies what are the dimensions of positioning,

so like scale and coverage footprint, program quality,

child outcomes, customer proposition, brand experience.

So it will figure out exactly how any particular company is

positioned, which is super important because when

when you talk about growth strategy, you really need to understand how a company is positioned,

how the industry is doing, and use that to figure out how it can actually grow.

So, all

of these details that we are building, the AI is actually learning how this company is

positioned.

Once it understands that, that's when it can make decisions about where it

can go grow.

Yeah.

Yes.

Yeah.

Yes.

It will identify the research reports on its own.

If there is a researcher.

Yeah.

So that's from the Web.

So basically, like the way that

it works is we have, you know, thousands of questions.

We go out to the to Google, for

for example, and we input those thousands of questions.

We get to the reports that are

generated or the search results.

We figure out which search results have the most citations.

So then we take those tens of thousands of reports, we combine it together.

Then that

becomes our knowledge base.

And from that knowledge base, we can actually ask additional

questions around each of these dimensions and then so we don't use like

the training of LLMs we actually build these knowledge base and work from that

so any of these things you can kind of click on it and it will open you open

where the source is coming from so it's very transparent how it's getting to an

answer so in that way you know you can understand the competitive positioning

Public Data Tie-Ins and M&A Screens

And then I mentioned that we are already connected to public data systems so that public data systems, you know, you can identify, you know, in this particular industry, who the companies are, you know, how are they growing, etc.

And then it would it would also do some sort of M &A screen.

So sometimes, you know, people want to say, hey, this is an industry that I want to get into, which are the companies that I can go by?

So it can highlight some of these screens.

Domain Chatbot with Deep Context

We have a chat bot here that is actually now going to be much, much stronger than a chat GPT because it has gone through the tens of thousands of reports.

It has read through the reports.

And so now it has so much more context than chat GPT on on that particular industry.

So if you ask any questions, you know, it will be it will be much better.

Automatic Competitor Discovery

and so yeah so this is an example of an industry report yes no we will identify it the AI agents

will go out and say how many companies exist which are the right competitors it will actually

figure out for this particular industry who would be the right so you don't have to do anything like

Like, as I was showing, you just have to provide the name of the industry and for that industry

it will go out and identify exactly who the competitors might be and so it will bring

it together.

And those data are given to you?

Yeah, yeah.

It's not just publicly available, it's also, yeah, it is all publicly available but it's

not like there is a list somewhere.

It will go out and say, okay, for this company, you know, what is the association of this

company with other companies.

so who could be competitors so it will bring all of that together yeah yeah we

Agents Deployed in Client Environments

are actually thinking about having these agents because the thing is many of our

clients don't want to upload the data on the and so they want us to give them

there they want us to kind of give them their our agents so that they can be

just like you know agents for hire basically and they operate within their

environment.

So, we are thinking about that, yes, yes, yes, yeah, yeah.

So, I think the

Source Selection and Language Coverage

Citations and Source Reliability

answer to the, to the first question is, it is largely citations that we are using to

make sure we are picking up the right sources.

So, we try to make sure that the knowledge

that we build uses sources that have the most citation.

We're leaning on Google and other search engines to kind of

get us to the most cited sources.

Multilingual Ingestion and Output

Then the other thing about AI like second question around English is that

AI is super interesting.

I didn't even know that like if you put in documents in Japanese, it will read up the documents

in Japanese and spit it out in English if you wanted.

So basically that's not something we have done, but based on the LLMs, it's already fluent in all languages.

So you can put in documents in Japanese and get the answers in English.

And so you can put, so you can run, so I have run a bunch, like I just did a project where there were 40 different competitors that I had to evaluate across Indonesia, Japan, China, Thailand.

And the languages were all different.

And so the local languages were different.

And the, all I had to do was just run that.

it will go out and look at their local news, local reports in the local language, will interpret it and spit it out in English.

So, it is, it is, it is sort of like all the, it could read all the languages and can spit it out in all languages.

Yeah.

Positioning vs. General Research Tools

I am asking this question, because I am not sure I fully understand, I am just trying

to think, it seems like in the agents that you built and in the data that you crunched

and all these information that you have made it more than chat GPT, but how is that like

a research tool like notebook LM, which has spectacular ability to do research, I don't

know if you know that one.

Yeah, I definitely know that.

Verticalized Agents vs. Generic RAG

I think it's about the agent the way you interpret the data notebook LLM you can kind of put in data

and and then you can ask questions around it for our system it is already pre -programmed to

solve for that particular vertical and so we have spent a lot of time making sure that everything is

consistent, everything is ticking and tying and it's particularly focused on that vertical

use case.

Versus more platform type things that appear to be strong, but when you start

to really stress test it, there's lots of inconsistencies start to emerge and things

may not look right.

Consistency Across Large Reports

Like for example, just to build a large report like this, if you

have individual sections, how do you make sure that what is being said in one section

matches the other section and so if you use notebook you know you upload some

data and you ask it one prompt and then you ask like six six prompts later and

you as the same same question it would not be the same so it is it is again

about consistency and trying to make sure that the agents are specifically

trained for this vertical use case notebook notebook is primarily like a

a rag -based, so it's just like it just takes the document, it sort of like creates its

own sort of cloud around the data that it has and it sort of works through that, it's

just a rag -based system, it doesn't have an agent.

Best-of-Breed LLM Ensemble

No we use the, we figured out which LLM models work for which use case and we took them and

we combined it and that's what we are doing.

Yes.

Data Usage and Licensing Perspective

Use of Publicly Available Information

So, if it is available publicly, we are

using it.

We are being sued publicly available data.

Right.

I think that is their fight right now.

To us, it is like if it is available on Google,

I can see it and so I can read it.

Transparency and Source Linking

So, this tool is more for organizing the data that

I can see.

And you saw that we can click on a particular tool, it will take you to the

article.

Human vs. AI-Generated Research Quality

Questionable, like I don't know if the market research from IBIS world

that is human generated is any good than any better than let's say what chat GPT

would produce.

I would say that you know some of like the human who human

generated articles weren't masterpieces in any case right they had they were all

flawed also.

So, I don't think, I mean, I think what AI is doing is like, it is definitely

generating more, but that doesn't mean it's wrong.

Yes.

No, I agree.

I agree.

And I think

Customer Segments and Use Cases

where we are right now, we are where, you know, we primarily serve private equity clients,

consulting clients they have established businesses and for those established

businesses you know they the history sort of is like five years ten years

back and all of that is used to build a report and five ten years back it was

mostly human generated content I think so yes we're using we are working with

Serving Large and Small Private Equity

all sizes thus the some of the larger private equity companies are building

their own AI systems to do this but they are not they don't have growth strategy

type of capabilities so we are building we are we are kind of supporting them on

growth strategy for smaller private equities they don't have just the same

evaluation type of capability so we are supporting where the use cases are

different but we are working across the range we are also working with

Consulting and Investment Banking Use Cases

consulting companies banking company investment banks who are doing sell

sell side work so they need to kind of evaluate a particular industry they need to evaluate

a company so so they they can do that they can they can use our system yes yeah so there are

Deliverables: Concise vs. Detailed, PPT Export

like two versions of the reports there is a concise and this i'm just like highlighting here

there is a concise version there is a detailed version what you are seeing is the more detailed

version but there is a shorter version also that is more summarized so and you

can download it in PowerPoint and basically slot it into your your the

Freshness, Dashboards, and Triggers

Real-Time Updates and Monitoring

deck that you are working on so then the first question was how updated is our

information our information is updated to the day because again we are kind of

of running these queries and as long as it's on Google and so on, it is real time.

So some of the use cases that we are seeing is, you know, a large corporate has a number

of targets that it wants to acquire.

So now it basically like says, okay, create dashboards for me about information about

each of the targets.

And so when you refresh it, it gets kind of updated every time it refreshes.

And so you can bubble up if there is any sort of event that makes it more likely to kind of get a deal done.

Sales Enablement and Event Triggers

And so you can use that for sales also.

So like if you have an enterprise sales group

and the enterprise sales group wants to kind of talk to the specific enterprise

and is just waiting for the right event to happen

to talk to this particular customer,

then you can kind of create dashboard for that.

And so it can create like sales triggers.

Conclusion and Next Steps

Yeah, so I think that's basically what I had.

Yeah, if there are any questions,

I would love to stay in touch.

This is my LinkedIn QR code.

And thanks for the engagement, it was really fun.