Check out the conversation on Apple, Spotify, and YouTfrube.
The three titles that are going away (2:31)
Aakash: You manage a team of over 200 people with titles like user experience researcher, designer, product manager, but you told me that you think these titles will all go away in a couple of years. Tell me more.
Srini: I think it’s a matter of time before these titles go away. These titles existed in the past because of a reason. And the reason that these titles existed was because it was a linear handoff that was happening from a product manager to a user experience designer to an engineer building things.
But now, by using AI, each one can do all three things. Which is why I think these three different roles are going to be replaced by what I call a product builder role, where a single person can research the customer, spend more time talking to users, understand their pain points, and then build things and ship things. That, I think, is what is going to happen. It’s probably already happening in some companies, in startups, but I think it’ll be much more widely prevalent in the very near future.
The traditional linear handoff where a product manager writes a product requirements document and then hands it over to a user experience designer, who then takes the product requirements document and builds a prototype in Figma and then creates a wireframe from it, and then hands it off to an engineer. This is a very linear process, and this is completely dead.
Aakash: I love the bold take. So today we are going to walk you through this masterclass of how this product builder role happens in three steps. We’ll walk through the AI PDLC, the EX AI agent that they created, and how it interacts with MCPs in Claude Code. Srini, can you get us started with the AI PDLC?
Data first, not AI first (4:03)
Srini: Yes. Let me show you something instead of talking about it. I’m going to show you how we are building software today.
I’m going to spend maybe about two minutes to talk about the AI first versus the data first approach to building software. There’s a lot of noise out there that everybody is building software by using AI. What you see on the left is the foundation for building the software, which is having an appropriate design system, an appropriate coding system, and a repository that can be used across.
So I call this the data first approach versus the AI first approach. And the difference is mainly building what you see on the right, which is building the foundations that make the AI first approach go faster. Not starting with the AI first approach, but building the foundations, and then building on top of it, which makes things go a lot faster.
The AI PDLC and the Prism hub (4:58)
Srini: So here’s the product development life cycle that we have at Freshworks today. Now this is pretty common across a lot of software companies. Discovery, design, planning, development, QA, deployment, release, and so on. And what we have is an AI agent that’s helping in every step of this process.
That’s what we call the AI PDLC, which is the AI product development life cycle. And we have a governed framework. We have an eval framework for it. And we have built something called Prism, which you see here on the bottom left.
We have our knowledge hub, which has the product knowledge and dependencies. We have the context hub, which has the feature context that’s passed across the different phases. And we have a central skills repository, we call it the AI builder artifacts, which is skills, rules, commands, agents. All these are very specific to Freshworks because we are a billion dollar revenue company, and 75,000 customers use our products, and it touches almost 300 million end users. That’s a lot of users that are using our software. So we need to make sure that whatever we build reaches those people and it’s governed.
So this is our life cycle of how we build software. Today I’m going to go deep on the first two things, which is the discovery aspect and the design aspect of it. I’ll show you the other things briefly, but I’m going to focus mainly on the discovery and the design aspect, where we have a product discovery agent and a design agent, and I’ll show you how those things work.
It all starts and ends in cursor. I was an engineer for 10 years before I went to the non-engineering parts of the job. I did multiple jobs. I was in professional services. I was an investment banker. I was in corporate strategy, and then I became a product leader. And during that course I used to use an IDE to build software. But now, after almost 14 years of not touching code, I’m again spending a lot of time in cursor. So now I’m spending time building software.
PRD Genie and the research burden on PMs (7:07)
Srini: We have built something called PRD Genie, which is used by the PMs to draft the product requirements document. They also use it for spinning up the prototype, and it lives inside cursor. You can also do the pull request from inside cursor to actually build the thing that you have written the requirements for and done the design for.
When you do this, the research burden is on the PMs. They spend a lot of time gathering evidence for functionality, comparative analysis, customer feedback. There are metrics, and metrics come in different forms. How many people are using it? What are the feature requests that are submitted by our community? And then there’s a dependency analysis. This is a lot of burden on the PMs. It’s not just writing the PRD, but they have to get all this information and get it into the product requirements document. And this leaves very little time for strategic thinking. So the PM spends a lot more time on operational work rather than thinking strategically about what they need to build and what customers need.
That’s exactly the problem that PRD Genie has solved. It drafts 80% of the PRD almost instantly. And it’s grounded in real data. It drafts real specs and then it hunts for evidence. It gets evidence on how much users are asking for something, what bugs it solves, what functionality we can do. Baikal is our data lake where we have all the usage metrics from 75,000 customers. And then it looks at the competitive benchmarks to see what problems there are, how the competitors are doing it, and how we can do it better than others.
And then I call this the CPO check, which is an AI agent that reviews the document for strategic alignment, clarity, and edge cases. These are the things I look for when I’m reviewing a PRD, but the AI agent does it automatically now. And finally, it syncs into action. It plugs into the rest of the AI PDLC process across all the teams.
So these are the 12 steps. I call it the 12 phase specialized intelligence, from idea briefing to requirements to knowledge, competitive analysis, and so on. These are some of the things. Competitor analysis, customer feedback analysis, quantitative metrics. So it’s not just the qualitative. Customer analysis and competitive analysis is qualitative, and this is quantitative metrics showing what the usage data shows and what we should be doing. And finally, internal dependency analysis. You can imagine, when it’s a large company there are dependencies on platform, there are dependencies on systems, all the other things. So it also does the dependency mapping and analysis. And finally it does the automated visual prototyping, so it goes from having a spec to getting a final screen. And finally it does the quality assurance.
Why cursor is the environment (9:54)
Aakash: Why cursor?
Srini: Very good question. Cursor is an integrated development environment, and I like cursor for three reasons.
Number one is I can get started, and it’s a no code IDE. A completely non-technical person can use it, as opposed to the other IDEs that are out there in the market, which mostly are geared towards an engineer. A completely non-technical person can use it and they can get good at it. That’s number one.
Number two is you can pick and choose the models. Not everything needs the latest LLM model. You can pick and choose the models that you want, and it’ll perform based on what needs to be done.
Number three, it has connections to Figma and other third party sources. My favorite thing is to use the Figma MCP plugin, whereby it can read the Figma mockups that we create and it can create code from it. So these are the three reasons why I primarily like cursor as the development environment.
Initializing the project (13:04)
Srini: So this is the cursor environment, and what you see here is a Freshworks demo. Which means we have created a project here. We have the artifacts. I’m going to initiate this.
Now I have a problem. Let me describe the problem I’m trying to solve. We launched an AI agent studio for our customer experience product called Freshdesk back in November of 2025. And we recently launched our EX AI agent studio, which is the same as CX but for employee experience. And we did that within like four months. When we launched the product, one of the pieces of functionality that was missing is the performance analysis.
So think about it as, you know, a lot of people working in different companies. Imagine an AI agent is answering questions about, hey, my Zoom is not working, what do I do, or I need to get access to cursor, or I need to request my time off. All these things an AI agent is answering questions about. But as it’s answering questions, it’ll answer some of it, some of it won’t answer. Imagine you are the IT admin and you want to see how the AI agent is performing. That’s what is called performance analysis, how the AI agent is performing. That functionality we need to build. So we built it for CX, we are going to have a fast follow for EX. And now I need to build it. So that’s the project that I’m going to show you.
I’m going to start with initializing this project. Okay. So I initiated it and it’s a 12 step process. Remember what I just told you earlier. It’s a 12 step process in terms of building things. What it’s asking for to initialize the AI PDLC, it’s asking for the business unit, epic ID, feature team, all this other information that it needs. And I’m going to give the information that it’s asking for. So the business unit here is EX, which is employee experience. I’m giving it the epic ID and the team.
So once I give this information, it’s going to run, it’s going to plan its next moves, and it’s now setting up an artifact demo so that it can initiate building this project and the relevant files that it needs.
Okay. So this is going to take some time. In the interest of time, I’m going to just show you how this looks. So instead of doing all 12 steps, which will probably take another 30 minutes, I’m going to shortcut it. I did this earlier in the day today. So I’m going to say that I’m staging what I was able to do in cursor earlier. So for one and two, it can ask me questions and the rest of it, otherwise it’ll just keep asking questions. So I’m just going to say, look, don’t keep asking me questions. Just proceed to the next step.
Okay. Now it’s going to ask. Let’s give it this. This should not take too long. It’ll probably take like 30 seconds. What it’s doing now, it’s like a normal developer. Think of it like a developer. It’s creating the directories and it’s creating the files that it needs. It’s initiating everything. It’s creating the scaffolding that’s needed to build it. And now it’s completed it. So that took like 15 seconds.
Now I’m actually going to describe the problem statement that I said earlier. So it’s a long one, but essentially what it is, as you can see, the problem here is EX AI agent studio, they don’t have any unified view of how their AI agent performs across the employee channels. So we already have that for CX agent studio, and that’s what we want to build. And the users are IT and HR admins, IT managers or HR business partners, like think of a CIO or a CHRO or a service desk lead.
Why now is because we just launched EX AI agents and customers are starting to use it and they want this functionality. And now I’ve given it, and I said, look, you don’t have to reinvent the wheel. There’s always something like this that exists for CX. So you can pretty much take what it is and customize it, because employee experience people experience it through Microsoft Teams, Slack, and so on, for custom experiences data. So there are some nuances that have been added here. So I’ve given it this, how it needs to ideate.
Now it’s going to go ahead and start looking at the problem and come back with some questions that it has. Think about it as a product manager. When somebody tells you, hey, go build a performance dashboard for an agent studio, as a PM you would want to ask some clarifying questions on what exactly is needed. Which is essentially what the AI agent is thinking and saying, okay, what are the things that I need to understand in order for me to create a very comprehensive product requirements document and a design mockup that can be used. So it’s now planning for those things, and all these commands have been given as part of the init process that I told you about earlier.
So these commands have been done and now it has come back with the questions. Look at some of these questions. It’s asking what logical questions would be. How deep should the analytics drill down go? Where do ticket level details live? What is the primary persona’s core workflow? What do success metrics look like in six months? These are all domain specific questions. ITSM process framing, service catalog, all these are things that an IT admin would understand. So it’s asking all the relevant questions so it understands the domain. Which is what I said earlier when I started, which is having the appropriate data is the basis for building faster.
Now I’m going to give it answers. Here are the answers. I said these are the things that you need. Analytics drill down should be at this level, agent, topic, subtopic. The owner is IT or HR admin, and the landing view should be performance overview. It should create Sankey charts for better visualization. So I’m going to give it the instructions and it comes back with questions.
What do you think, Aakash, so far?
Aakash: This is the first time anybody has used the Grok family of models, and I think the performance is really good and really fast. Is that why you chose it?
Srini: That’s right. I used to use Claude Opus and I’ll still use it. But this seems to be working really well. It’s really, really fast. Nothing took more than like 10 or 15 seconds, as you could see.
Aakash: Yeah, I mainly use one model and a little bit of another. So watching this go 10 times faster is blowing my mind.
Walking through the 12 stages (19:46)
Srini: Let me actually show you while this is running. This is going to take a few seconds. Let me show you something that I already have created.
So these are the 12 steps. Let me just walk you through the steps that happened here. It’s a lot. First, when I answered, it started gathering the knowledge. It started gathering knowledge of epics, of how the analytics should be, and it gathered how the analytics screens are laid out right now. Then it cross-checked with the competitor analysis. How are competitors doing it, like Atlassian, ServiceNow? Just looking at how they are doing things. And it flagged some things for review just to confirm.
And then it did the market and competitor research. That’s stage four. Stage five is voice of the customer. What are the customers saying? What is the feedback we heard from customers? Then metrics. These metrics are coming from our Databricks data lake, which has all the usage data.
So let me just show you how it looks. As you can see, it literally wrote SQL queries to query the data lake. You can see here it wrote a bunch of these SQL queries and it defined what metrics are needed and it extracted the metrics. And finally it gave the data, what the data gaps are. It identified what data gaps are needed and it came up with what data we need.
So you can see a lot of stages that have been passed. When I said everything is grounded in data, what I meant is different companies have different ways of writing software, like the PRD, what the steps in the PRD are and what the amount of qualitative data versus quantitative data is that you use. Every company’s different and every project is different. This particular one, we have this 12 step process as you can see on the right side. These are your markdown files. You see the idea brief, requirements, knowledge gathering. All this is the different stages that the AI agent just went through to create. And this is a final checklist.
They call it the CPO review. Somebody really liked me in the product management team, which is why I think they called it CPO review. So the agent studio performance dashboard, they run through the CPO review, step one checker, self-administered, final approval. These are all the things that a normal product manager would go through. They would have to write all these things. Imagine it would take days and days to do this. The agent did this in a matter of seconds, not even minutes.
So you can see here, going back to this, it’s completed all the steps. Steps 3 to 12 are already completed, and that’s what I just showed you.
What the finished PRD looks like (22:38)
Srini: Now that it has completed these steps, the next thing is to go look at how the PRD looks, because ultimately that’s the outcome.
So you can see here it’s created a PRD. This is the markdown file. Let me make this bigger. So let’s look at what this PRD looks like. This is the PRD for agent studio performance dashboard.
It’s figuring out that employee experience has something called a service catalog which is different from what the CX product had. So it has considered that and it has accounted for it. It gives you the problem statement, what the goals are, what’s in scope, out of scope, and then it shows you who the users are. Service desk lead, IT manager, and so on. And then it lays out what the solution is.
Then it lays out the market and customer evidence. What did our customers want when they were using the product, and how does the market look. And then the important part, quantitative evidence. This is what I meant by our data lake built on Databricks. It shows you how many catalog requests we have, how many ITSM accounts we have, and by region. It said, okay, what are the things that quantitatively are needed in order for us to build the capability.
And then it built the user stories, user stories which are positive scenarios and negative scenarios. It built a bunch of user stories. You can see almost eight user stories. And then it built the requirements. Here are the functional requirements that it built, non-functional requirements. These are things that we have been doing for the last seven or eight years when we were building software, but it was done in a very manual way. Now everything is automated. It also has the recommendations on pricing and packaging, what’s the rollout plan. It came out with all these things.
Now this is the PRD that we can take. So now the PRD is done. What do you do now? What you do now is build the actual mockup.
How do you trust an AI generated PRD (24:36)
Aakash: The worry I always have with these AI generated PRDs is, for metrics just as an example, we’re looking at the quantitative information. There’s a million definitions of a metric. Did it really pull the right definition? Did it look at the right thing? We looked at how it was detangling experience you had in one product versus the other. Does it really have the latest version of what’s live on the internet, or does it have the latest based on Jira, which might not match the product? Some of those details always get me nervous when it comes to AI generated PRDs. How do we have confidence that it resolved those correctly?
Srini: This is where the initialization process is very handy. When you use AI for building completely new things, like think of Lovable or Replit or any of your favorite tools, it’s very easy to use AI, or even Figma Make for that matter. It’s very easy to build new things from scratch because there’s no precedent, there’s no 75,000 customers that are using the product.
That’s why it’s important to have the appropriate references to what it needs to use. That’s what the initialization process does. When it created these markdown files, you will see that in the markdown files, it specifically tells you which version it referenced and why it referenced that. So there are specific instructions that are given as part of the AI PDLC process.
So it’s not completely automated. The way I would describe it is AI is not running completely on autopilot. It’s actually a copilot. There is still value for the product manager. If the product manager is doing it, for the product manager to do it. If the engineer is doing it, the engineer should be doing it. So the value shifts from doing operational work to judgment. Okay, which version should I give, and how should I prompt the AI to do things faster?
So going faster doesn’t mean you sacrifice quality. And the way to do this, of not sacrificing quality, is to ground it on things that it needs to reference so that it doesn’t hallucinate. You can’t afford AI to hallucinate in building software. It’s just not acceptable, because then 75,000 customers are going to get affected. Of course, there are quality checks in place. There are checks and balances that we put in place. But the important thing is to ground AI on the things that it needs to know.
From document to prototype in Figma Make (27:02)
Aakash: Got it. And now let’s see that prototype. How do we go from document to visual?
Srini: Yes. Let me show you the prototype. Let me start with Figma Make. Same as how we started with the Freshworks init process. You see here that we have the Fresh Service starter, and we have Dew which is our design system. And Fresh Service, think of it like a scaffolding.
We have two main products, Fresh Service and Freshdesk. Fresh Service is the scaffolding in which all the functionality that the users see sits. So I’m going to start with this scaffolding of Fresh Service, because the functionality that we are trying to build right now, the performance dashboard, is inside the Fresh Service scaffolding.
So I’m preloading the Fresh Service scaffolding so that when I create the prototype using the PRD that I created earlier, we don’t start with a blank canvas. We start with a canvas that’s already used by 75,000 customers.
And this is the scaffolding. When I say scaffolding, what I mean is this is literally the screen that half a million of our user base sees when they log in. This is the ticket screen and this is the settings screen. These are all the buttons, and it has the onboarding screen. This is the template that all of our users see today.
And what I’m going to do with this is I’m going to say, using this Fresh Service starter kit, build the analyze section from the PRD. So I’m going to go back to cursor, take the PRD that I had, and paste it into Figma Make here.
Okay, it’s pasted it as a link, as you can see. And I’m going to tell it to plan. Let me actually go here where I’ve already started, because it’s going to take some time. So what I did here is I gave it the PRD and I said, using this Fresh Service starter as the shell, build the analyze section. And it started reasoning it and it has started building it.
How much of the analyze section should I build? It’s asking now. All eight modules, P0 modules, overview. I’m going to say all eight modules. So now it’s going to start building all the eight modules. Remember the eight use cases that I just showed you in cursor. Those are the same use cases. It’s calling each of those modules and it’s going to start building all eight modules in the screen, which is what you’re going to see.
Aakash: So is Figma Make your preferred prototyping tool because it has access to your design system and you can easily bring it into Figma files, or is there any other reason?
Srini: Two reasons. One is what you just said, it has our design system. But we have built our design system into Claude as well. Many of your listeners might know Claude has Claude Design, where you can also feed in your design system.
Number two, a lot of our user experience designers, a lot of our product managers are very familiar with Figma, and Figma also lets you connect with other external data sources like cursor and other places. So it makes it easier from an adoption standpoint. When you’re changing the ethos of how software is built, you don’t want to change too much of the tools that they are using. And the tool usage can evolve over time. Now I see a lot of our PMs and user experience people actually using Claude Design and Figma, and this evolves. We’ll be sitting here 12 months from now and maybe everybody will be inside Claude. It’s just the current best.
Where judgment still matters (34:36)
Srini: So this is going to take some time. I’m going to show you how this actually looks. This is a fully built out design. It’s going to take at least four or five minutes. So I already did this. I’m going to show it to you, and I’m going to show you that there is actually a bunch of back and forth that I did on this. So it’s not straightforward that you just give it the design.
So here it said do design components. Then it asked me some questions. I responded. And it actually didn’t get it right the first time. It had a couple of places where it was not using the design system. So I had to tell it that there are some specific components that it has to use, and I had to give it references to it. Then it changed everything.
Finally it came back, and then I realized that this design does not work on narrow monitors. So I said, you know, people use, real users, some of them have wide monitors like the one I’m using now, and some have narrow monitors, some have small screens. So I have to prompt it to say that. And this is where judgment is important.
So coding and business knowledge and experience of writing product requirements was important in the past, but now judgment is very important. A really experienced designer would know that, hey, it needs to work on different screens, and that’s by understanding the end users. So that’s why I said, okay, it has to work on narrow monitors. Then it responded to that.
The Sankey chart is this cool thing that you see here. For those of you that don’t know this, this thing that you see here is called a Sankey chart. A Sankey chart shows you what percentage of topics are based on feature requests, what is software application errors, like that. So this is the Sankey chart, and the Sankey chart was not showing up. So I had to fix that.
These are all the things that I did, and finally it came up with this. So this is actually the working product. You can see that this is a prototype of what the analyze and performance dashboard should look like, and this is now a completely built out mockup that Figma Make, essentially the agent running within Figma Make, has built for us.
Aakash: And the PM applies their judgment really in finalizing the PRD, if there are any missed areas that it’s wrong on, in iterating here on the design. And then I’m keen to see what happens on this final step.
Inside agent studio (37:01)
Srini: That’s exactly right. What I just showed you was how we use the AI PDLC to build the core functionality of the performance dashboard in agent studio. Now I’m going to show you how the actual agent studio itself looks.
This is the agent studio that we released maybe two months back. Think of this agent studio as Claude. Claude is used by millions of people around the globe to get a lot of things done. People summarize things, people brainstorm on things. But what the agent studio is meant for is two specific things. When an employee joins a company and they need help with their IT issues or HR related issues, or they want to provision new software, how a company can build very domain specific agents in IT and HR to serve their internal employees. Most employees are either in Slack or Teams, or they are in a support portal to access their needs. And that’s what the agent studio does.
So this is our agent studio. What we have here is AI agents. What I have logged into right now is Fresh Service, which is used by millions of users. And here is the agent studio, which is inbuilt into the interface that they are already used to. And we have AI agents here. What the AI agents do, you can see here, IT agents, HR agents. Some companies actually prefer to have IT and HR together. These are the pre-built AI agents. And you’ll see the libraries here. When I go to AI agents, these are the pre-built IT and HR agents.
Now I’ll show you the workflows. Think of workflows as building blocks. For example, there is a workflow for creating an incident inside PagerDuty. Imagine your service goes down and you want to create an incident inside PagerDuty. That’s the incident one. And here there is a password reset. If you want to reset your password, most companies have either Okta or Azure. Here’s the password reset in Okta. Here’s the password reset in Azure AD.
These are pre-built workflows that come out of the box. The way we have come up with this is we looked at 80% of our customers, what the most often used workflows are, and we built those workflows. Despite that, a lot of our customers actually have built their own workflows. Look at lock computer. If some computer is locked, Jamf is used for unlocking. And there’s a bunch of other workflows that a lot of our customers have built.
Knowledge, the service catalog, and guardrails (39:44)
Srini: Then the next one that I want to show you is the knowledge. If I go to one of the existing AI agents, let me look at this IT and HR agent. Think of knowledge as what feeds the AI agent.
These AI agents cannot hallucinate, which means they need to have guardrails and they need to be governed. And one of the ways in which we make sure that the agents don’t hallucinate and they’re governed inside the enterprise data is by giving it data to train on. That’s what we call knowledge.
Knowledge can come in the form of URLs. In this case we have fed the Microsoft 365 help and support, and it ingested all the things related to Microsoft Office 365 so that people don’t have to waste time searching for things. They can just get answers.
The second one is policy. For example, how to do the withholding form for W4. Employees, when they join, they want the W4 form and they want to say, am I single or married, or how my tax withholding should be. We have actually documented that in what we call a standard operating procedure, and that’s the document that exists here. There’s a bunch of solution articles that people create, in different categories, whether related to IT, hardware, HR.
And then finally the apps. Not all the knowledge has to live within the agent studio. There is knowledge that exists most often in third party apps. Enterprises tend to store their data in Google, so we have a connector to Google. Or SharePoint, we have a search connector. Or Confluence. A lot of the knowledge can reside there. So you can connect your apps in your companies to fetch the knowledge from third party sources outside of Fresh Service. So Google, SharePoint, Confluence, and so on. That’s knowledge.
Then I’m going to show you something very specific, which is a service catalog. Think of a service catalog as a list of things that you can have to get information. Here is an example of Fresh Service inside our company. If I want to request a service, so think of a service as, hey, I need access to Slack, or I need access to employment verification, it’s in here. And there are some collaboration tools, if I want to have access to Zoom or Box.
This is somewhat of an IT term, where the IT department, the CIO’s department, maintains a list of all the services that are offered inside. For example, HR applications, they have a bunch of HR applications here, and software installers, so they have Adobe. Think of it as the approved software that an employee can get. So they cannot randomly ask for things. They have to ask within the service catalog. That’s what this is.
And similarly there are some instructions that you can give to the AI agent. What is the business context, what do we do. Freshworks as a company, we have 5,000 people and we all use this. So we can give what Freshworks is, what this business is, what we do, and we can give some custom instructions as well if we want to. Some of the financial services companies probably don’t want PII data to be shared, that kind of instruction you can give.
And then there are configurations. There’s multilingual support. People are spread over different countries, and sometimes the AI agent is asking questions in English or Spanish or German. So you can support multiple different languages here, and how it needs to act. That’s conversation behavior. So bottom line, you can customize the AI agent to behave and act in the way that you prefer.
And finally you can test the agent here once you build the agent. So primarily all these screens are used by an IT admin to build an AI agent, because they know their users really well. And before they launch it they can test it, and here are some tests that they can do. And finally they can deploy it. They can deploy it within a support portal, or mostly what we find is our customers deploy it within Slack or Microsoft Teams, because that’s where most of the users are.
The Slack demo, from W4 forms to mortgage letters (43:52)
Srini: What I’m going to show you is how, once an agent is configured, how does it respond to questions in Slack. So we use Slack. I’m going to show you a demo of Slack.
What we have done now is created an IT and HR agent. We call it hire to retire. Which means from the time somebody comes in to when they retire, these are the things that they can use the AI agent for, how the AI agent can help them in the different phases where they need help.
So imagine this is Srini. I just joined Freshworks and I want to say I need access to Slack. I assume that I have requested access to Slack, and Slack is here and it has already installed Slack. It uses Jamf. Now I have Slack.
Then what I’m going to do is, I have some questions regarding my health insurance policy. Previously what used to happen was I had to email HR and somebody in HR will say, hey, contact this person or that person. Now I don’t have to do that. Here is what I call the concierge, the live IT HR. This is my private channel and I can use this to ask questions privately.
So I can say here, what are the mental health benefits offered as part of my insurance policy. If I’m the kind of person that wants a private answer, then I can ask this here. Or in some instances HR might actually say, hey, this question might help others, so I can ask it in a public channel. So this is HR help. HR help is something where people can go to ask questions about HR related things.
So here it’s saying your insurance policy includes several mental health benefits, and it gave me a bunch of benefits that I have. And it’s also showing the source. The source is actually a document. So if you remember, I showed you earlier how the HR team can actually put in a solutions article. So it’s coming from one of the solution articles. The same thing happened in my private conversation as well, it gave me the same response. So it really depends. It’s all customer preferences. Some customers want a central channel, some customers want private. So you can achieve both things from here.
Then I can say, well, what about forms. It used to be that when I needed a form I had to either go to some archaic website which nobody remembers, or again email HR. So here I’m just asking a question to my AI agent saying, hey, can I get a blank W4 tax form. And it comes back and boom, there’s the W4 tax form. You can see here it says here’s the thing, and I can access the form directly from Slack. So that’s the tax form.
I’ll give you a couple more examples to show you the power of this. I want to generate, let’s say, an employment verification letter. I just joined, my W4 form is taken care of, and I want to apply for a loan. I’m buying a house, I’m applying for a mortgage. So I’m going to ask it for an employment verification letter.
Remember how it used to be when you needed an employment verification letter. You had to either email your HR or you had to jump through hoops to get this. Now I’m just going to ask the question here. I need an official employment verification that I generate for my bank. Let’s see what it does.
Now remember, this is very personal to me and it needs to know exactly. So it’s asking for details. It’s asking why do I need this, what’s the address, what’s the purpose of the letter. These are all very valid questions and I’m going to give it answers. This is the organization name, Acme company. Here’s the address. And the purpose of the letter is that I’m applying for a home mortgage. So I’m giving it the information.
Remember, this used to take multiple emails or potentially even a phone call conversation, or going to an HR department and sitting down with someone and spending 30 minutes explaining to them why it is. Now it’s verifying the information. I’m saying yes, all this looks good. Now it’ll go to work and it has to generate an employment verification letter. There you go. One moment. Okay, there’s my employment verification letter.
Aakash: Wow. You see this in Slack?
Srini: There’s my employment verification letter. I wish I could be making this money. I don’t. But that’s where it is.
Aakash: I thought you just doxxed your salary.
Srini: I joined way before 2025. So there you go.
But you see how it is, right? It’s the AI agent helping. Imagine, as an employee, all of you must have joined companies where it took hoops. You need to know someone that knows someone to get work done. You know how it is. But now you can just literally go to this AI agent and the AI agent is going to help you with day to day tasks.
So what we have heard from 20,000 odd customers that we have for Fresh Service is it significantly improves experience. People used to have SLAs, which is service level agreements. Now there is experience level agreements, which is how satisfied am I as an employee. Customer satisfaction matters. People talk about customer satisfaction, but employee satisfaction is also very, very important. I would actually argue that it’s way more important, because if you have happy employees, that will ultimately result in happy customers.
So that’s the power of AI agent studio for you. Look, this took five minutes. I was able to show you how to set up an agent and how easy it is to set up an agent in our agent studio, and how easy it is to use from inside Slack.
Fresh Service MCP inside Claude (49:41)
Aakash: Amazing. Now, how do MCPs and using this with Claude come into the picture?
Srini: That’s the last bit. Let me show you how a lot of our customers actually use Claude to connect to Fresh Service. For Fresh Service, we launched a Claude connector.
Let me log into Claude right now. I’m going to open a new chat and this is already connected. I have connected this Fresh Service MCP. So think of this as a company that uses Claude to query data from Fresh Service and get information and resolve things on the fly from Claude.
I’m going to ask it a question. I’m going to say, hey, fetch all the tickets with Windows 11 related issues reported in the last 60 days and generate a visual report with root cause analysis. If people were to do this today, they actually jump through three steps and there are like three people involved, because somebody has to fetch the data, somebody has to create a report, and then somebody has to look at the report and say, okay, what caused it, root cause analysis and all of that. Now it’s just a single prompt. I’m going to give it a single prompt. It’s going to figure out. This is going to take a couple of minutes.
It says this is a Freshworks instance. By the way, talking of governance and security, it is governed, which means none of this data leaves outside our environment.
So now what it’s doing is it’s looking at the data. I already have a pre-baked one here. You see it’s the same thing. I gave it the same command and it created a beautiful report. You see here, this is the report. What you see on the right is a Windows 11 update regression.
What it said is there were 12 tickets across 60 days. There are two separate rollouts that happened. So a Windows 11 patch happened and that resulted in the tickets getting created. And you can see the shape of the incident, July 10th and 13th, more so 13th is when a lot of the issues happened. Six incidents happened on the 13th. So clearly some Windows 11 patch was applied. And it shows you the root cause analysis. The two big things are kernel driver regression, so some driver was applied which made blue screens appear, and there’s a patch that was applied and the install hangs.
And it also gives you the reference of all the tickets. Here’s a ticket. Here’s the cluster it belongs to, when it was created, what’s the status of the ticket, has anybody responded to it or not, and what are the actions to take. So it actually gave me more than what I asked for. I just asked for what is the root cause analysis. What it showed me is actually the recommended actions. Reply to tickets at the pilot ring before the wider rollout. This is Claude being way more proactive in doing things than just being a passive help desk person that has to wait to take a response.
Right now what I’m going to do is, I think the live demo that I had should have come back now. Here’s what it is. Well, it’s still working. But I already did this.
So what I asked next is help me answer these tickets with a relevant response by searching the knowledge base articles available. So it went back and used the Fresh Service MCP, and the MCP went back into Fresh Service and got the KB articles. And that’s how it said what do I do when the system freeze happens. And this is like ticket number 158, and it says random freezing after a Windows update is usually caused by background reindexing and driver optimization, as is the case with most problems with Windows. It said restart the system.
Aakash: How many times have you heard that?
Srini: Right. And it gave solutions for each one of these things. System freezing, Windows lagging, and so on.
Then I said, okay, there are a bunch of these 12 tickets. There’s no way that I want to do it one at a time. It actually suggested, you want me to send the ticket replies the way I did for July? Then I said, okay, great, go ahead and send these replies to the tickets. And it actually sent the replies to all these tickets. I’ll show you how the ticket responses were sent, and it’s actually logged in Fresh Service as well that these tickets have been now answered.
This is the power of MCP. Remember how I started it. If I’m an IT agent, I’m spending all my day just responding to tickets. Now I don’t really have to do that. What I can do is I can come in and just ask, okay, what happened in the last 10 days or 60 days. Then I can get all the information. Then in one shot I say, okay, help me answer these tickets, because I may be a new agent, I just joined so I don’t know much about any of these things. Usually the IT agents are either Windows experts or server experts. Now you can be an expert in the overall IT environment.
It’s not only answering questions. It’s getting the answers and it’s phrasing the answers, and it can even send the responses right from here. So this usually used to take, imagine each ticket usually takes about one or two hours to resolve. Now 12 tickets, think of it like 12 to 24 hours. We just did this in like five minutes.
And that’s where a combination of the Fresh Service MCP plus Claude is helping the users, in this case the users are IT agents, to be way more effective in catering to their end employees. So remember what I told you earlier, by using the AI agents it was making the employees way happier, the employee satisfaction increased. In this case it’s making the IT agents way happier, so the IT agents are very happy because they don’t have to spend as much time digging through information and chasing things, and it’s making them way more productive. And as a result, even the end users who are submitting these tickets, they get immediate responses instead of having to wait. So that’s the power of MCP.
What PMs need to learn now (55:34)
Aakash: So we just walked through the AI PDLC, what this looks like on an example product where you’re having your users build agents, and then how they can connect that with MCP into your agents. What do PMs need to know who haven’t built AI agents and MCP features, in order to build this next generation of product features?
Srini: I think be curious. The world is evolving so fast that the entire employee experience and the customer experience has literally turned on its head in the last 12 months, and it’ll continue to evolve and it’ll drastically change.
I would say think about every single workflow. A bunch of the people listening are working in SaaS companies. You’ll have to unlearn a lot of things that you’ve done and relearn things. The entire software was built for humans. What we just saw with MCP was not a system that was built for humans. It was built for an AI agent, it was built for Claude to be more effective.
And you have to think about CLI. Everything that we did in cursor was from a command line interface. I was literally chatting away with cursor. Everything that I was doing in Slack, I was just chatting with an AI agent. So there is no concept of I have to build a thing for a person. You have to build a thing for a human being and for AI agents. So you need to completely unlearn how you are building things and relearn how you build things for not just human beings but also for AI agents.
The PM to engineer ratio shift (57:01)
Aakash: One of the craziest things you shared with me is that Freshworks is changing its ratio of PMs to engineers from when you began to now. Can you walk through that evolution and what’s driving it?
Srini: Yeah. So when I started, which was like 18 months back, we had almost 1 to 20 in terms of PM to engineer ratio in some teams. In some teams 1 to 20, in some teams it was 1 to 10. So somewhere between 10 to 20. And we used to have one UX person for every two PMs, and then 20 engineers for every PM.
In the last three to six months, what I’ve seen in teams that are adopting it is we have evolved into one PM to maybe one engineer, and sometimes there’s not even a user experience designer. And it’s gone to the drastic level where teams are operating so fast. And there are not a lot of layers as well. You don’t need to have these sprint planning sessions, six month planning sessions. You plan in two weeks. You release things in two weeks.
Now we are seeing that with our agent studio that we just launched. We’re able to see that although the ratios have changed, the velocity of how we release things and deliver things to our end customers has significantly improved. So imagine coming down from six months to two weeks. That’s the pace at which it has changed. And I would be surprised if in the next six months we go down from two weeks to maybe releasing things every two days, every week. It’s almost unheard of.
And in SaaS there’s also now room to do alpha beta testing. You can release things for a control group. You can even personalize software for every persona. It used to be really hard and you had to go through a bunch of bells and whistles to customize the SaaS software for different personas and for different industry verticals. And that is going to significantly change in the future, where you’ll have personalized customized software, personalized software for personas, customized software for industries.
The roadmap for other CPOs (59:15)
Aakash: For your fellow CPOs out there who are running product teams that aren’t where you’re at now, what is the roadmap to transform your team into this new way of working?
Srini: The number one thing to do is not just have an edict across the organization but actually do it. I was an engineer, but I was an engineer like 15 years back. So you have to set an example. Show it, not just say it. Because a bunch of PowerPoint slides and a ton of documents, people read these things and they’re like, okay, whatever, this guy sitting in an ivory tower preaching. You don’t have to preach, you have to actually show.
So I picked up software development again in the last six months and I’m really enjoying it, and I’ve probably written more code than maybe when I was an engineer. And that’s saying a lot, because I was an engineer for like 10 years. What I just showed you, the demonstration that I showed you, is an example of things that I’m doing. So at every level people have to show it. And once they see it, PMs will be like, well, if my CPO is doing it, why can’t I do it. And I’m not saying that I’m doing it, there are pockets of people. So set an example. Show it rather than saying it.
And there are pockets of people in every organization that are leading edge, that are bleeding edge, and they really want to change the trajectory. Identify those people, identify those champions. So we do a monthly product team meeting where literally it’s showcasing. It’s demos. It’s people showing demos of things that they’re doing.
PRD Genie that I showed you earlier was something that one of my team members came up with, and he said, hey, I built this PRD Genie, I want to show it to everybody. We showed it to everybody. Then we standardized it. The whole cursor thing that I just showed you, one of the guys in our design team said, hey, I’ve completely reimagined the product development life cycle, can I show you? Those are the people that you need to identify and champion.
Shipping more without dropping quality (1:01:16)
Aakash: So the worry with embracing this new way of working is that you are shipping more, you’re shipping more frequently, but your metrics aren’t going up. Kind of the age old product problem, like 90% of what you ship doesn’t actually move the metrics in a positive direction. How do you not only ship more, but ship at a higher quality?
Srini: I think it’s not about shipping more, but even when you ship more, you learn faster and you can fail fast. These feedback loops that we have right now.
So I almost think of the product’s job as, a third of it is launching the product. The other third is actually making sure the users adopt it. The last third is monetizing it. Because once you ship things you need to make sure that people are using it, and then you need to make sure that if people are loving what you ship, then they should be prepared to pay for it.
So I would say what this does is you fail and fail fast, or you succeed and you succeed way more effectively than what you did before.
How Freshworks hires AI PMs (1:02:15)
Aakash: So circling back to where we started. These titles might dissolve, PMs might have to become PM plus engineer, this new product builder role. How are you personally hiring people, hiring AI PMs at Freshworks? How are you assessing whether they have this level of capability?
Srini: Look, this is fairly new for everybody. I think we look for curiosity. Are you the kind of person that is open to change? And it’s not just people that we hire from outside. Even for people that are there, we’re making sure that we have a systematic training program to train people, and we do these innovations in pockets to make sure that it works.
And when we hire people, we look for, because not a lot of people might have done this before, but you should be open. You should be curious. And I really want to see, when I interview a PM, I want to see their git repository. How much of the things have they done on their own? Because you can’t teach passion to people. You can teach skills, you can teach technical things to people, but teaching passion is not something that is teachable. So people should inherently have it.
So what I look for is, have people done things on their own? What have they built? So the interviews are not more of, hey, tell me about this, tell me about that, but why don’t you open cursor and just show me what you did.
Aakash: And do you guys do structured case interviews, or is it mainly like show us what you’ve done in the past?
Srini: So we do a combination of both, depending on the level of the role and what we are hiring for. We do case interviews and we also do, okay, show me what you’ve done.
Where to find Srini (1:03:47)
Aakash: Freshworks is hiring. I keep seeing PM and AI PM roles. If people want to get in touch with you or hear more, where can they do that?
Srini: Please contact me on LinkedIn. My LinkedIn is https://www.linkedin.com/in/srinivasan28/. Yes, we are hiring, so please reach out.
Aakash: All right, for my money, seems like a very exciting product job. Find the link to his LinkedIn below in the description. Thank you for tuning in. And Srini, thanks for dropping so much knowledge.
Srini: Thank you, Aakash. It was a pleasure.