Full episode transcript: AI & PE: The Future of Value Creation

Guest: Ishan Gammampila, Chief Data & Analytics Officer, Apax Partners | Host: Kyle Roemer

Ishan Gammampila: Tools have been there for the last few years. So the first question you should be asking is do I have the right people? If you’re a CEO of a company, do I have the right people?

Kyle Roemer: Welcome back to AI & PE: The Future of Value Creation. We are thrilled to have Ishan from Apax Partners with us today. Ishan is the Chief Data & Analytics Officer in Apax’s Operational Excellence Practice, where he helps portfolio companies turn data analytics and AI into practical operating impact.

Today’s conversation is going to focus on what speed means in an AI-enabled company, how AI changes the pace of insight, automation, process redesign and execution across PE-backed businesses. Welcome, Ishan.

Ishan Gammampila: Thank you. Thank you for having me.

Kyle Roemer: Well let’s jump right in. When you talk about speed in the context of AI – just give the audience a little bit of an overview on what does that actually mean for a portfolio company.

Ishan Gammampila: The game has changed right in the last few years. Like everybody knows, the game has changed and you can see it in your personal lives, right? You’re using ChatGPT and Claude, and you’ve made things run much faster. You’re doing very interesting things. But when it comes to portfolio companies or any company for that matter, it hasn’t translated.

Right. And that’s the part that I’ve been focused on for like the last year or so, is how do you bring that speed into the portfolio companies? And it could be also like a whole slew of things, right. Like getting into insights faster or increasing the speed of delivery of your product or service, or developing things or developing your software faster.

It could be a whole slew of things, but getting that pace that we have seen in our personal lives into the corporate environment, I think is the game right now.

Kyle Roemer: Yeah. I find it such a fascinating dynamic right now because speed. Speed is to your point, in your personal life. Now with AI, it’s in a bunch of different functional areas. And I think with data in particular, the idea of going through the kind of the traditional discovery requirements, gathering, profiling, etc. has – I think that’s it’s a different world today.

And so maybe describe a little bit on like how that’s manifesting – whether it’s at Apax or at a portfolio company of yours.

Ishan Gammampila: Yeah. I mean, there are so many use cases that I can bring to life. The way to think about this and the way I’m seeing this is that any company or person who’s using this technology is not twice or three times faster than everybody else. They’re 100 times faster. And you might hear stories when people say, oh, we are using AI. We are doing a lot of stuff. But at the end of the day, are you really using it in the best way?

And that’s where if you look at companies who are really stepping back and trying to innovate, they’re changing the game. So I can give you an example of one of our portfolio companies in the cybersecurity analytics space. They were trying to solve problems for like over two years.

They were trying to adapt. They were doing things in the normal traditional arc. But then we had a new CPO who joined the company who basically said, we are going to scrap what we have been doing and we are going to reinvent. We’re going to start from scratch. And he has been able to completely pivot the company within a span of about six months.

And now they are like they are getting code delivered. They are doing automated testing. They have completely changed how they deliver their offering. And it’s a completely new revenue stream for them. So that’s just one example. Another one that we should go into detail is the speed to insights.

That’s a big topic, right. Most companies if you look at it – there’s no company. I have heard that say that. Oh, we know how our company operates end to end. Any question you ask, I can answer. This is because of the traditional arc of, how do you get data? How do you get into insights?

We can double click on that topic in a bit. But delivery, development, speed to insights, innovation. Those are like the big themes that I’ve been focusing on.

Kyle Roemer: And it makes I mean look it makes perfect sense. I think the cybersecurity analytics business you describe and, you know, a new CPO coming in and changing how they build product. Are you are you finding. And we’re seeing this with clients as well, that the normal development process, you know, with agents is obviously different today.

But even the roles, how you get to an early proof point on a prototype or design, we’re no longer doing wireframes, we’re no longer doing some of the things that we might have done. Gosh, even a year or a year and a half ago. So how did it – if you look at the product development cycle for that specific business, like what does it look like today and how is that different than maybe what it was?

Ishan Gammampila: So I’ll tell you a few examples. So you’re absolutely right. It has completely changed, right. Like we don’t write lengthy requirement diagrams or like do wireframes. You build a prototype because this is the point about speed, right? Like the cost of doing an iteration is near zero, right? So you can build something, you can show it, you can scrap it and you can do something else.

So, even in my work, when I do a lot of workshops and sometimes the workshops are related to designing a new system, I would get all the information that I can hoover up before the meeting. I would create a prototype, a clickable prototype, take that. And that’s how I would run the workshop. And that has been a huge game changer in the work that I do, because now people will see the art of the possible. They will see it open up their eyes about, oh, this is the kind of things that we can do. This is the kind of speed that we can run into and they will give feedback. I will take the feedback within half an hour or one hour. I’ll get the next iteration now. Right. So that’s one especially with the cybersecurity company.

They are at a different level. Yesterday I was talking with the CPO and he was showing me how he has created a second brain, a shared brain. It’s not a shared brain for the entire company where anybody who has an idea starts putting that idea into Slack and Slack or email or whatever it is, and they have agents who pull in that information. And surfaces are the new ideas and the new thought processes coming out of that. And now it becomes the pipeline for the new feature development for the product.

Right. So the code development part has been there for the last couple of years, three years. But it’s the two bookends like the testing part. How do you generate new ideas? How do you get it into the funnel and create that end-to-end processes is where things are innovating right now.

Kyle Roemer: I love the idea of crowdsourcing with your domain experts in the business. In addition, obviously your customers are normal things. You might do this. This prototyping dynamic is fascinating. We work with clients now where you’ll have deal teams or finance professionals spend a couple of hours in Claude Code.

And this is kind of what I’m thinking. And then we can take that and then and make it real. Right?

Ishan Gammampila: Yeah. No. Exactly. And even there was another use case that came into my mind. There’s a tech services company in our portfolio, and I’m doing a project with them at a different portfolio company. And they’ve created a harness where they’ve put the UI, the starting point of the UI, and given it back to the users.

So the users can now start adding comments on what they want to change in the UI itself. And the AI takes that, does the next iteration on its own. Right. So this is the whole loop of iterating really fast. And this is the point about speed.

Kyle Roemer: Yeah, yeah, I love that. You know, speed. It’s interesting. Speed has always mattered in the private equity context. Like how much can you get done in your first year of your hold or your first 18 months? How is speed translating maybe to the underwrite and kind of the value creation plan now – because it is it is wildly different than it was a couple of years ago.

So how do you think about that, and how is that translating to new companies that you look at?

Ishan Gammampila: So there are certain things that we used to be scared about that no longer is a problem. Right. And so if we had a legacy software stack three years ago, four years ago, it would have freaked us out. Um, but now it’s a matter of how much does it cost? And is it three months, six months, nine months? What’s the timeline?

Right. You know that you can do it whether you have the people and whether you have the willingness to do it and how much? How much are you trying to underwrite? So that’s one aspect. The second one, at least the way when I look at when I’m being called into due diligence once in a while, that’s the second one that I look at, is the team.

The team matters quite a bit. This goes back to the point about if you find the right person or the right team, there are going to be 100 times more effective and faster than the regular folks. So assessing the team and identifying the gaps up front is going to be a big thing that any due diligence should be looking into.

Kyle Roemer: I want to switch gears. You spend a lot of time in the data, obviously in the AI space, in the portfolio, and I find it really interesting. Coming back into consulting from private equity, I thought that the age of data warehouses and platforms were done. I thought that was a solved thing.

Right. Well, come to find out, there are a lot of poorly built and implemented data platforms over the years, but many of them were in service of reporting, visibility, descriptive analysis, type of use cases today that’s changed with agents and AI. And I’m just curious, how have you seen that evolution in terms of the stack itself?

Ishan Gammampila: Yeah. So this so my background is in data, right? I have been a data scientist, data engineer for almost two decades. And that’s what I’ve been doing. And every time five years ago, if you ask me, how would you go about it? Build a data lake, build a data warehouse. Let’s get your data in order.

I think it has switched now, right. You don’t need to create that big investment in that foundation up front. But the real frustration is every time I got a vendor to go and do a data warehouse or a data platform or just a reporting like get some dashboards up. It takes three months.

For the business to see the first report coming out. And every time the first report on the dashboard is wrong, the KPIs are defined wrong.

So the question that I was posing to some of my vendors and our internal team was like, it has to be a better way to do this. And how can we take best practices from the software development lifecycle and try to reinvent how we do data warehousing? And I don’t think you need to get your data in order to get started.

You will have to get to that place to do interesting things and more sustainable things. But the model has switched.

Kyle Roemer: Well, let’s talk more about that model. So you wrote the article. You’re alluding to the Substack piece, like what happens when you show stakeholders their data on day three? Like it’s fascinating for those listening. Please read it. We’ll include it in the podcast brief. But you ran an experiment around data.

Let me just tell folks about it.

Ishan Gammampila: Yeah. So we ran the experiment internally as well. But the place where we ran the experiment, we ran it at the portfolio company as well. So this was back in, I would say January, February. So there’s a portfolio company. We had been helping get their reporting organized for the like the last few years.

And there they ended up with having two separate data warehouses, that was built over the years by two different teams, a CRM system, marketing system, so on and so forth. So we spent about one week just saying, let’s not worry about the structure. Let’s connect all of these through MCP to Claude. Right. This is at the time and Claude 4.6 came in and everybody is like, oh, this is great. So we were like, let’s push the boundary. So we connected it. We went there, and we spent two days sitting with the business executives saying, okay, let’s just ask questions, don’t worry about the data structures. Let’s just ask questions. And the results that we started getting and the insights that we started getting from that experience was insane – and few things happen, right. The business users started asking questions that we didn’t even think about, right? And they started going down the rabbit hole and like trying to find like, okay, what happened?

What happened? Like, why is this happening? And it was surprising how good Claude was stitching things together to get an answer, and was answer right all the time. Absolutely not. Right. I mean, it had gaps, but getting started, getting that thought process was it was a big unlock. And that was a that’s when I was like, okay, this can be the model, right?

You can start with connecting your data sources through MCP into something like Claude. Start asking questions, understand what you want to build and then go and start building it properly.

Kyle Roemer: Having worked in the data space a long time as well. Like folks are frightened a little bit of giving stakeholders access to data. Yeah. Like is it clean, is it transformed? Does it meet all the different requirements? I think the reality today is like, look, it doesn’t need to be perfect, but gosh, wouldn’t it be nice if stakeholders started to reveal to the data team the quality issues, process issues where there’s nonconformity, and taking that because those are real things versus going through the traditional let’s profile data, let’s go through requirements gathering, etc..

Ishan Gammampila: Yeah, no 100% agreed. And this is the part the feedback loop coming from this, right? So the business users start using it. They identify gaps, they identify definition issues. And we even create a feedback loop, which is so easy that you tell Claude to create a summary of your chat and send it to the central team. And the central team uses that to now do the next round of the data warehouse or the data like the data cleaner.

So that loop was very important and it expedited, I would say like years of trying to get the data warehouse to where it should be. So that that was a very eye opening. Interesting use case.

And I have another example, if you’re interested.

Kyle Roemer: Yeah. Please. I would love it.

Ishan Gammampila: So we have a portfolio company which is in the CPG space and they have to manage inventory. There are a lot of retailers. The way it works is that the retailers would send Excel files, Excel reports to say this is how much we have sold and this is how much inventory we have, and we want to order this much in the next two weeks.

But given these are in Excels, there’s a team of people who would be sitting down and just trying to make sense of it and manage the manufacturing pipeline, and if we don’t supply them with the amount of inventory that they need, we start paying penalties, right? So this problem came to us and my boss was a partner in the company. He basically said, I’m going to vibe code this. I’m going to build a system. So he created like 50 parsers to parse these Excels and got the demand data from the system, created an environment where you can forecast the demand and how much of manufacturing capacity we need. And he did it in a way that it’s a self-healing platform where if the format of the Excel changes the AI figures out it has changed and it adapts.

Right? So this is another point about speed, right? It was done in one week by not an engineer, but a slightly technical business user. This is the part of like don’t wait for the data to be perfect. Start, iterate, learn and adapt.

Kyle Roemer: These examples are really powerful. I think, you know, we get asked a lot by sponsors and portfolio companies for their portcos. Like what does AI-ready data actually look like? And I think that definition has changed based on what you’re describing. But there’s, I suspect, a spectrum of when data is ready enough for certain use cases. And maybe just give an example of when data needs to be more accurate, more integrated, and more governed. To be able to take on a specific use case versus what you’re describing in some of these, like quick hit examples.

Ishan Gammampila: Yeah. I mean, it’s a good question. The short answer is I think data is good enough to get started regardless of your use case. Right. But there are two different aspects to look at. One is are the people ready?

Because one important thing that we realized while doing these exercises with our portfolio companies is that if you are in the mindset to say to prove that AI is wrong,

You will succeed 100% of the time you will find something that the AI is doing wrong. But if you come with the mindset to say, okay, this will get me to like the 80–90%, but I still have to do some checking.

That’s where – and you keep on improving – that’s where you get the best. The takeaway there is that you have to find the right people who understand that, right? Because if you give it to the

wrong people, there will be some people who will say, oh, you know what? I’m going to prove this wrong. I’m going to find an issue. So that’s one category that you should be careful about. And the second category is they will be like, oh, I trust this completely. And they would make wrong insights and wrong decisions.

So that’s one bucket of things that where you have to worry about. And then wherever you need accuracy 100% accuracy. Like say for example, if you’re doing cash reconciliation or something like that and you’re trying to use AI to do some automation,

then you have to be careful. You need to have checks and balances. But the good thing is you can have AI to do the checks and balances to right? So you need to figure out how much human in the loop you give a system.

Kyle Roemer: The people side is is super important. And I think it’s all about I think going into these things expectations like are you expecting to find issues? You’ll find them. Of course you’ll find them. But can you get to a good enough starting point with this whether it’s a model or a tool, etc.. And then iterate from that point?

I think that’s really important. The people side I think is really interesting. And maybe there’s two angles just to explore for a minute. One is, in all of these portfolio companies, there are folks that are more or less AI-enabled and AI-native and all of those things. But there’s a lot of folks that are just in the early innings of using these tools. So I’m curious, what have you found to be successful in enabling or training folks to get more up to speed on using, whether it’s models or some of the products that exist?

Ishan Gammampila: There are a few things that will make it successful. One is you have to make the use cases relevant to what they do, right? Because if you if you come from a data background, you think you will connect the dots. If you give an accounting person a CRM use case, but it’s usually not the case. So if you’re talking to an accounting person, you have to give them something that relates to them.

The other thing is, what I personally found very useful is short-form videos. Rather than getting them to do tutorials or whatever, like short-form videos to show like somebody in their team doing something and showing them this is what we did, right? This is how we went about it. This is how I used Claude in Excel, so on and so forth.

Kyle Roemer: I do want to explore like how have data teams evolved as well, and what are the types of whether it’s skills or domain expertise or pod structures you’re seeing. Move the fastest, having the most impact.

Ishan Gammampila: Honestly, I

They still haven’t stepped up. Yeah. At least in my view. So I work with a bunch of vendors. There’s one vendor who has really stepped up and embraced this. And I’ll tell you the patterns that I’m seeing there. So they’ve used, not specific tools per se, but they’re using Claude.

But the way they go about it is kind of question and answer kind of methods to discover what the data model should be. They’re using Claude to understand the domain a much better. And they go back to the client with questions, very pointed questions and answers to get the feedback. That is a very interesting pattern we have seen, because if you really think about it, business users are really great editors, but they are not creators.

So if you give them something to edit, they will tell you 78 things that is wrong with that. But if you ask them to design what they want, they won’t. So having that quick feedback, quick iterative loops, so the challenge I gave to this vendor is to say you guys used to take three months to get the first dashboard. I want it in two weeks. I don’t care how you do it. It could be you directly connecting to MCPs, but put something in front of the end user and then iterate.

Kyle Roemer: Yeah, I mean, I love that. I have a strong bias for our teams. Like you need to start with the models first to some level of fidelity on whether it’s requirements, data model transformations, etc.. But, you know, we’re naturally getting pushed by a number of clients, like how do we move away from the traditional, even the traditional sprint model into what can we run in parallel from a prototyping perspective and a production perspective?

Get these things moving so that people are seeing things that can react to things. Provide feedback while you’re doing all the hardening, stitching, everything else you might need to do to get things in production. I think that that just feels like the model of the future right now.

Ishan Gammampila: Exactly. Couldn’t agree more on that point.

Kyle Roemer: I think when you look at a big topic and not just data, but generally AI and workflows is there’s a lot of companies that are putting AI on top of preexisting workflows that, frankly, that humans have done for a number of years. Right. We have a belief that for certain workflows, not every workflow, but certain workflows, you need to reimagine them with agents versus just apply to them.

Where are you seeing success in that, or where are you seeing also some pitfalls in companies trying to do this?

Ishan Gammampila: Yeah, this is a great question. I’m seeing this all over the place. The risk you’re seeing right now is you’re trying to automate without reducing complexity. You’re taking an existing workflow and trying to say, okay, I had 78 steps. I take Excel from here, put it there. So, for example, we were looking at a portfolio company and the office of the CFO automation. And we’re looking at the bank reconciliation process. And on average, it takes about 3 to 6 hours to do a bank reconciliation.

And we are like, what is going on? And then you see that they go and download a file from the bank portal, do a pivot, copy it into another place, and then do a bunch of work, then upload it into NetSuite. There’s so much of manual work and like kind of swivel chair work, but the moment we try to introduce AI, the initial instinct is, can I do that? Pivot faster, right? Can I do the copying from this Excel to the other faster. So that’s the problem. That many people fall into the trap that they fall into.

So the way we’ve gone about this is a very old concept. Probably 30, 35 years old at this point. It’s the jobs to be done framework.

So let’s step back and let’s look at what’s the jobs to be done and what’s the best way to go from point A to point B? If it is a bank reconciliation thing, you have cash coming into the bank. You need to make sure that you tie that back into your ledger. That’s it. Right. And how you get about get about that. Let’s try to see what’s the easiest way, the minimum number of steps to get there.

Kyle Roemer: Maybe just for folks like what’s the what’s the best example right now that you’ve seen of a workflow just being reimagined with AI?

Ishan Gammampila: Let’s look at like maybe two right. One in the back office and one on the product itself. So on the back office side, I’m spending a lot of time in the office of the CFO trying to automate the workflows there. And again, we’ve gone through the entire office of the CFO workflows and identified like 54 jobs to be done.

And the whole point is to say, where’s the data? How can I get this, um, in the shortest path possible? So that’s the kind of like thinking end to end rather than piecemeal. So clear cost cutting, right? Clear efficiency improvement workflows.

On the other hand, I can give you another example from one of our portfolio companies. This company has a very complicated pricing engine, and the pricing engine sits between your CRM, your ERP, and many other places. Uh, many other software systems. Uh, think of a pricing engine for, like, a CPG company which computes your rebates, what kind of discounts you should give, so on and so forth.

So as you can imagine, these pricing engines are very complicated, very complicated to integrate. It takes long sales cycles. Their sales cycles were about 6 to 9 months or more.

Because of that they were only going after the large companies. But what they did about honestly, by this time, it’s about a year and a half. Two years ago, they said, let us look at the jobs to be done. How many jobs to be done? Jobs are there from end to end, from trying to get a sale up to delivering how many jobs to be done.

And they realize that they have about 100 plus. And they said, let’s put an agent on all of them. And some of these agents could be deterministic agents. Some of them could be just rules based agents. Some of them will be AI agents. And that change their entire motion on how they do work, their sales cycle, their sales cycles went from 6 to 9 months to about 1 to 2 weeks, where they literally go and plug it into the system and say, you guys play around with it for two weeks. If you don’t like it, pull it back, because it’s so easy now. And because of that, they managed to go down market like the small and medium sized businesses.

But this is taking having the guts to step back and say, let’s look at our offering end to end and let’s try to reinvent how you do it.

Kyle Roemer: Yeah, I love that. And it’s a perfect it’s a perfect example of using jobs to be done in a, in a go to market BD process. And to your point, which I really appreciate, not everything needs to be an AI agent also like some of these things are just, you know, maybe it’s deterministic, maybe it’s machine learning. Maybe it’s just rules based to do an activity or push data somewhere or all of that. And I think sometimes we can overcomplicate this with thinking an LLM needs to take care of each step, and that’s just not the case. Yep. 100%.

Ishan Gammampila: 100%.

Kyle Roemer: A lot of pilots happening. They’ve been happening for certainly over a year. Not always. Are we seeing things get to production? So what are you what are you finding that’s preventing pilots, you know, actually getting production lines and into core processes?

Ishan Gammampila: Yeah. So again, even on this multiple angles to look at this one, I think we touched upon it is not having the right people and the right mindset. I think this is this is probably the biggest one because the tools are available. Tools have been available for the last couple of years, but if you don’t have the mindset, good example is like seeking perfection from day one.

That’s not going to happen, right? Like you have to get to that 80% and from The difference with this technology is getting to 80%. Takes 1% of the time, right? Not 20%. It takes less than 1% of the time, but the last 20% is still a grind. But trying to find use cases that are good with 80% and then increasing it to get to the 90–95%. That kind of mindset is very important.

So you need to have the right people. Next is you think going from vibe coding into production-level builds is going to be just pushing your code? What do you vibe-code into production that doesn’t work right? So this is where the whole new topic of harness engineering comes in. Once you create the right harnesses that’s where you can start. It’ll take some time to build the harnesses, but once you build it, it’s going to be much easier to do more deployment. Do the right thing, have this same standardized data model. One standardized data model in the back end. So, so if I had to summarize it, those two points, people and try to find the right guardrails.

Tools have been there for the last few years. The first question you should be asking is do I have the right people? If you’re a CEO of a company, do I have the right people? Uh, and it’s always a very tough question, but the only way to do that is just look outside on what your competition is doing. What other portfolio companies in your group are doing, how fast they are running.

Just try to understand. And if you see them running at a different pace, that means you don’t have the right people who are thinking about the problem in the right way. That’s I think the biggest point because take out of all of this and going back to the tool point, your competition has the same tools. So if they have the right mindset and the right people, they will start running much ahead of you, right? So how do you use the same tools that are available for everyone, but start building things in a much faster, much more sustainable way?

But how? Yeah. How do you increase the velocity of what you’re doing and mistakes? Again, the last point is like it’s easy to redo things. So don’t overthink it. Just jump onto it. Just iterate. Make mistakes. That’s the best way you will learn.

Kyle Roemer: Yeah I mean I think that’s so well said. And like in the age of AI, people still matter. I think folks should understand that people still matter and the types of people and how they’re enabled. But the process and how you kind of rethink how things get delivered, I think is as important.

So yeah, this was awesome. I’m like, love the conversation. And there’s just so much goodness in how you’ve described the impact in the portfolio, I think from changing the idea of quickly prototyping and getting data in the arms of stakeholders to rethinking the people side, the use cases you tackle, the processes and the workflows, and in particular, this whole the concept, as you alluded to, it’s been around a while, like jobs to be done.

Like have that mindset as you go in and look at processes and workflows and apply, whether it’s AI agents or models or rules-based engines to those things to really get to a better place. I just love the idea of like, look, it’s okay to fail a little bit. Absorb a little bit of pain, but it doesn’t need to be perfect.

This was an awesome conversation, I appreciate it.

Ishan Gammampila: Yeah. This was fun. Thank you for having me.