What we talked about
Udit Joshi, Product Manager at Google, leads the development of next-generation tools used by Googlers worldwide, integrating Google AI to automate support and sales conversations. He’s a builder at heart. In his past life as a startup founder, he learned how to build products from the ground up:skills he now applies at scale inside Google.
Udit believes we’re in a new era where AI lets anyone create amazing things, with or without technical expertise.
Show notes
Udit Joshi started his career by going door-to-door selling advertisement space on school notebooks to local businesses in India, at a scale that reached millions of notebook copies per month across multiple cities. That same zero-to-one builder instinct now shapes how he thinks about product management inside one of the largest engineering organizations in the world.
What we covered
- Udit’s campus startup, Impressions, printed advertisements on physical school and college notebooks through a local manufacturer, and within two years got published in a national newspaper. He co-founded it in 2013 while still in his second or third year of college, and it has since evolved into a full-fledged UI/UX design studio that continues operating today without him.
- At Google, his team builds the internal support and sales tools used by agents across every major Google product, Ads, YouTube, Pixel phones, Google One, Cloud, and Maps. These are first-party tools that compete with Salesforce, Zendesk, and ServiceNow, built in-house because Google’s scale produces requirements no third-party vendor can match.
- His proudest project was building one of the industry’s first conversational voice bots for Google Ads support in 2019-2020, before AI was mainstream. The bot used NLP to understand why customers were calling and route them to the right specialist team in seconds, cutting average routing time from over 10 minutes to near-instant, and saving the company millions of dollars.
- He now spends significant time designing the handoff between AI agents and human agents. The hardest scenario is a genuinely frustrated customer whose business is on the line, the kind of panic and urgency that current AI cannot read reliably. Getting that customer to the right human, with the right empathy, as fast as possible, is where he focuses product thinking.
- AI has fundamentally changed what he does as a PM: research that previously took days of reading press articles, analyst reports, and watching demos now takes minutes with Gemini’s deep research. He also uses Gemini Canvas and other internal Google tools for document generation, though he always iterates heavily on AI output rather than using it directly.
- His advice for aspiring PMs is to just build things, specifically now, when tools like Google Jules, Gemini Canvas, and vibe coding lower the barrier so far that a product manager can demonstrate a working prototype rather than a written requirements document. He says if he were graduating from college today, he would probably stay on the startup path rather than seeking a traditional job.
About Udit
Udit Joshi is a Product Manager for enterprise platforms at Google, where he has worked for close to six years. Before Google, he co-founded Impressions, a graphics design and unconventional advertising startup in India, and then moved into consulting before joining Google.
- LinkedIn: https://www.linkedin.com/in/uditjoshi92
- Website: https://www.google.com
Episode 36 of the PreVetted Podcast.
Full transcript
Federico Ramallo (00:01.324) Welcome back to the PreVetted podcast where we spotlight extraordinary people and remarkable talent reshaping our world. Our guest today is Udit Joshi. He’s a product manager for enterprise platforms at Google. Hello, Udit, how are you doing today?
Udit (00:19.862) I’m F.E.R. Draco, doing good, how are you? Thanks for having me.
Federico Ramallo (00:24.182) I am doing great. Yeah. Thank you for being here. I appreciate it. We have a lot of interesting things, topics to discuss. There are a lot of developments for the product, product owners, product managers that are very interesting to get your insights about, right?
Udit (00:48.79) Absolutely. And I heard you are visiting the Bay Area, right? I could have invited you in the office and we could have had a face-to-face. So I’d say we can always do that again.
Federico Ramallo (00:58.86) Yes, yes, we can do that another time. I’m unfortunately leaving tomorrow, but happy to do that at another time. that would be nice. Yeah. I was yesterday was at Figma, the Figma offices, because they had a meetup, a meetup. So I was able to visit their office, very nice offices. Yeah.
Udit (01:25.59) Now you should visit the Google campus once as well. There some nice buildings out here.
Federico Ramallo (01:32.908) Yes, that’ll be great.
Awesome. let’s start by, I want to ask about your first role after college.
Udit (01:45.226) Yeah, mean, yeah, it’s been like more than a decade now, but.
Federico Ramallo (01:45.858) Yeah, that’d be good.
Udit (01:49.364) My first role out of college was, in fact, I mean, not a traditional job. It was something that I started during college itself. So this was like early 2010s and stuff back in India. I mean, lot of kids in college were starting things of their own, so like a campus startup. So that was actually my first role, and we did quite a bit. So I co-founded this with my friend. It was called Impressions back then.
Yeah, and after college I did end up joining the consulting role as well, but all of my creative energy I spent early on the startup itself. So it was in the second or the third year of college that we started this. It was a graphics design startup and we also got into the space of unconventional advertisements. So specifically, we worked with the notebooks.
a notebooks manufacturing company, so the physical notebooks that you have. And we provided, and we used that space to sort of print advertisements on it. So we were kind of the middle man going and getting advertisers to do.
print their ads on the school notebooks. So these were notebooks sold across school and college. And you know how densely populated India is, right? So this is a local notebooks company, but even a local company that sells within two or three different cities sells millions of copies per month. So that was the scale at which we were operating.
And yeah, it actually went pretty well. we started that in 2013 or so in a couple of years. We actually got pretty famous and got published in the national newspaper. So that was pretty cool. So was really the first time I felt the thrill of creating something and then getting noticed from it.
Federico Ramallo (03:31.474) wow.
Udit (03:37.948) And with any startup, it’s really the hustle of doing something yourself. And I you’re also an entrepreneur, so you know what I’m kind of talking about. So doing things from scratch, literally going from door to door to local businesses, pitching them the idea of about, have these notebooks that we’re working with the company on. You can have your advertisements printed on it, so stuff like that. So anyways, fast forward many years now, it’s actually evolved into a
full-fledged UI UX design studio. I’m not involved anymore. I work at Google, as you know. But I think looking back at that experience, creating something from zero to one is something that I really look back on and learn from it. And it’s something that you can’t really learn in a classroom. So I think that was very valuable.
Federico Ramallo (04:30.69) Yes, you have to be there doing and creating to actually have the experience. Yes.
Udit (04:38.109) Right, yeah, yeah. Like we’re very different than like just working in a job, I feel like. Because you’re actually building something, creating something on your own, so.
Federico Ramallo (04:48.994) Right. And you mentioned that you’re now working at Google. I understand your role is product manager, right? So what do you do now as a product manager at Google?
Udit (05:03.283) Yeah, yeah, I think, yeah, so I mean, yeah, I’m a product manager at Google. I’ve been working here for close to six years. And again.
I feel like even back when I was at the startup and now as a product manager, one thing is common. You’re sort of a builder. Even as a product manager, essentially you’re building new things or pitching new ideas and stuff like that. At Google, of course, you know it’s a company that builds a lot of software, a lot of different products that end customers, end users use. I’m specifically on a team that builds software that’s used for Google’s contact centers. So basically for support.
and sales conversations across all of Google products. So think of ads, YouTube, consumer products like Pixel phones, Google One, cloud, maps. So we build support and sales tools for these. So if you’ve ever chatted with or emailed or called Google for support with any product, a lot of Google products are great, but sometimes they don’t work well, so you have to contact support for it. So if you’ve ever contacted support, so that’s where you’ve interacted with the systems that
my team builds, especially on the agent-facing side of the business-facing side of things.
Federico Ramallo (06:19.928) Right. So all the internal tools, right? The internal tools at Google. Yeah.
Udit (06:24.819) Yes, exactly. these are internal customer support and sales tools.
Federico Ramallo (06:31.096) Right, right. I mean, it’s a different company, yesterday at Figma on the Ruby Meetup, one of the engineers showed how they were building internal tools, right? And they build their own frameworks to basically build the complete UI using Ruby without using JavaScript or HTML, right? And it was the same thing. was a…
all the internal tools they use within Figma.
Udit (07:04.597) Right, right. Yeah, mean, especially at Google scale, I mean, there are so many different teams and like hundreds of thousands of obviously employees as well. Yeah, the internal tool landscape is actually very complex and very diverse. Well, like the tools that I’m talking about are like first party build by us itself. So these are like competitors of, I competitors are like tools like Salesforce or a Zendesk or a service now out there in the market.
But for many of our teams, we actually build a first-party tool. Because sometimes the requirements are so unique at Google scale that it makes more sense to build your own tool than work with third-party companies. But of course, teams that use Salesforce and ServiceNow as well. But for bulk of the use cases, bulk of the teams, it’s first-party.
Federico Ramallo (07:45.645) Right.
Federico Ramallo (07:55.68) Interesting.
Udit (07:57.75) And of course, I mean, it’s becoming easier to build things also, right? So, yeah, maybe we can talk a bit more about that, or AI is changing some of that.
Federico Ramallo (08:07.692) Yes, yes, that’s what I wanted to talk about. You know, there’s a lot of hype about AI, right? And they’re hype on both sides, right? Like people are saying that it’s over, you know, over promising. It’s like, this is a great tool, blah, blah, blah, right? And they rely too much. I’ve seen people relying too much on AI, right? And then on the other hand, I was…
on one of the meetups at GitHub and one of the partners at Y Combinator say basically that, I’m paraphrasing here, right? But basically he was saying that there are so many new opportunities of applications that you could not build before, right? That there’s a green field of opportunities now because AI enables or lowers the entry point
to build these type of tools, right? So anyway, that’s, know, trying to give the context of this, right? What is your take on AI and how is AI changing your role?
Udit (09:15.273) Yeah.
Udit (09:25.629) Yeah, absolutely. mean, yeah, obviously there’s a lot of buzz around it and
Yeah, for PMs, feel like the way I was doing work even two years ago has just so drastically changed from how I do it today. So I can’t imagine a time now going back to that two years ago and how I used to do work there. Yeah, I mean, if one given day, like if Gemini goes away, if AI goes away, I don’t know how I’m going to do my work, So maybe to give a few examples. I mean, it’s honestly come to that, right? Because it’s so embedded in everything that you do.
Federico Ramallo (09:52.91) You
Udit (10:01.591) makes you a lot more productive, right? And productive, not in the sense that you, it actually frees up time for more strategic work, right? So it’s not like you have more time now to do things outside work. mean, there’s still always a lot of stuff to do. What you’re expected to do is a lot more now because you can now spend more time on strategy, on visioning.
and things like that. So maybe like to give an example, like a few years ago, if I wanted to understand like what competitors like Salesforce and Zendesk were doing, like it would mean like multiple days of manual research and I have to read through like press articles, analyst reports, watch demos, stuff like that. And today with our internal tools, now Gemini has deep research. And of course, like being at Google, I think one of the big, big, big benefits is you have access to all of the best of AI that Google offers.
There’s deep research, there’s Gemini Canvas, there’s, you must have heard of your Notebook LM, there are other sort of niche tools like Flow for video generation and there are a bunch of those and I could talk a bit about that. Many of which are released outside as well, like on an experimental basis. yeah, externally of course you have to subscribe to a Pro plan, like a Gemini Pro plan. Like here we have access to all of it by default and then dog food capacity as well.
But anyway, guess going back to my example, any sort of research obviously becomes very simple. I can do it in a matter of minutes. It doesn’t take me days. So I think that’s one area. So where you’re thinking about or doing research on your products or on your area. And the second piece of what a PM does is lot of of producing.
Federico Ramallo (11:24.206) Right.
Udit (11:47.431) like documents or slides and things like that, right? So you write a product requirements document or you like a strategy doc or a deck that you present to your executives or stakeholders to get alignment. and stuff like that. like AI helps a lot with that as well.
And I think on that second piece, it won’t do a perfect job where you can just use the output of Gemini and then that’s a ready PRD. But it gives you a great starting point, which you can iterate on. So again, it helps you generate things more quicker as a result. So I think those two places where it’s really changing my role a lot.
Yeah, and like PMs across, think it’s an expectation to start using these because only if you’re using these tools yourself is when you can actually build these into the products that you’re building. Yeah, so you kind of have to act as a user yourself of AI before you can build it into your product. Of course, the use case and the scenarios that you’re building into your product might be very different, but at least you get the gist of how some of this works.
Federico Ramallo (12:58.84) Yes, because you need first to understand what the tool is capable of doing, right? And what are the nuances of using the tool? I think that you don’t need to be a software engineer to use AI. You don’t need to be a mechanical engineer to drive a car, right? Now, if you know the internals, then you can…
understand a little bit better how some things work or take a little bit more advantage of some features, right? But if you can understand functionally how it works, right, you don’t need to be super technical to understand that, right? It’s very easy to grasp those concepts and then use it, right? So I can see how…
as a PM, you can leverage a lot of that technology, right?
Udit (14:05.042) Yeah, no, absolutely. Yeah, I think in our daily work, yes, it’s completely transformed it. Yeah, yeah.
Federico Ramallo (14:16.23) Do you think that AI is going to replace product managers?
Udit (14:22.3) Yeah, mean, I think that’s the natural next question, right? Like, I mean, all of us are kind of fearing our jobs in some ways. And obviously, personally, I hope not. I hope not. But yeah, you never know how smart AI can become. But I think, even realistically, the way people often put it about, like, how AI helps people is, like, it’s…
Federico Ramallo (14:30.456) Hehehe.
Udit (14:44.748) It’s not necessarily replacement. It’s something that’s more of an augmentation to your current role and stuff like that. And that applies to a PM as well.
There’s definitely parts of the job like of a PM like basic documentation, like research and data aggregation and data analysis that I was talking about here, which has become a lot easier. But most of my time now, like I said, is being spent more on strategy, more on influencing people, because now AI is not going to get consensus among like say, 10 engineering leaders and UX designers sitting in a room, right?
AI won’t be able to do that. So as a PM, that’s a very important function, like kind of driving that consensus among different stakeholders, among your executives, pitching ideas, brainstorming, giving suggestions, and stuff like that. there’s a lot of…
work, I would hope, mean, you know, that these systems are just getting better and better. So, I mean, I wouldn’t be surprised if it gets better at that as well. But although I think it might get better at suggesting things of how do you bring consensus to certain things, it might get better at suggesting the right features to build or brainstorm and stuff like that. But it like that the human element would still be required is what I would say.
Federico Ramallo (16:16.77) Yeah, I think that AI is not going to replace PMs, but PMs using AI are going to replace PMs not using AI, right?
Udit (16:28.532) yeah, that’s a good way to put it. Yep, yep. Because, yeah, I think, mean, you wouldn’t need maybe as many PMs or as a matter of any function of, and you’ll be able to do a lot more with the same amount of work that you have in a way. So, yeah, I like that, yeah.
Federico Ramallo (16:49.72) Yeah, so I think that the AI can help you. You mentioned the augment, right? So I think that’s a good way to put it. That gives you more possibilities to build stuff and develop the features and functionality better, right? Because now you have more access to an agent that can support what you’re doing, right?
but you’re driving it, right? You’re still guiding it to the right direction, right?
Udit (17:25.106) Right. Yeah, and it’s really like, yeah, those mundane sort of tasks which it helps you with, but like some of the more creative, the human element tasks, yeah, you still need to do yourself. You still need to go and attend meetings with stakeholders. So you still need to present your ideas to a group of people. So yeah, all of that is there to stay.
Federico Ramallo (17:47.606) Right. And have you done, have you done by coding?
Udit (17:53.595) yeah, Doley. Yep, yep, Yeah, that’s in fact…
Yeah, I I spend most of my time less on writing documents, like I was saying, because now AI helps with a great starting point and stuff like that. So instead of just writing PRDs or product requirement documents, you can actually wipe code your PRD in a way of what you’re recommending so that you can actually show to different stakeholders of what you’re recommending instead of them having to read your strategy doc or your requirements doc.
Federico Ramallo (18:22.928) I see.
Yeah, I think that people talk about by coding, it’s going to replace developers. And I think the same logic applies. I don’t think it’s going to replace developers, but it’s going to give more people access to build prototypes that they couldn’t do before. They needed a developer right next to them just coding.
one typo and everything breaks, right? Or whatever, depending on where you’re building. But now you can build software without knowing or understanding much how everything works. You can declare what you want or what behavior you need and the AI is going to kind of figure out how to build it, right?
Udit (19:19.812) Yeah, and then and like you said like it would be a good prototype or a v0 or a good starting point and of course that requires a lot of effort to actually integrate it with your actual systems and then Productionizing and bring it to life But yeah, like to your point Yeah, building that is a lot more easier, right? So at Google, yeah, we have products like so which is released externally as well So things like say stitch which helps you build like an actual UI
and it builds like the CSS and stuff that can be consumed by Figma directly. So that’s one. And then there’s your Gemini Canvas as well, which kind of here presents the UI of what you want to, of the product or the feature that you’re describing. yeah.
Federico Ramallo (20:05.132) Yeah, way that I’ve been trying to describe vibe coding is, vibe coding can, so the LLMs can intuitively figure out how to put code together to make it work. But they don’t understand, but the LLMs don’t understand how to design it.
so it’s scalable, so it works in production. I think that’s where the limitation currently is. It’s going to become better in the future, but people are trying to use Vibe coding to build from idea prototype to product. I think that that’s the wrong approach because, yeah, it could work, but…
It’s basically a literal software engineer. It’s a literal junior engineer, right? So if you give literal requirements, it’s going to fulfill them. The application could work like 80 % of what you expect, but the underlying design, the architecture, all of that, it doesn’t understand, right? It simply doesn’t. So whatever is built is not going to be scalable. It’s not going to be easy to modify in the future, right?
I think that we are at that point where we still need software engineers that can understand that, right? So they can build that code, right? But for the scenario of prototyping, I think it’s great because it gives so many more people the possibility to, hey, I built this, right? And it works and you can click on it, right?
Udit (21:51.956) Yeah, no that that’s so true right and especially like in established products in companies right now, so there’s one Category where yeah, you can wipe code and build a standalone new app like like a simple thing like maybe
that you want for your specific use case. Even Google, we have a tool called Jules where you can build an app yourself, itself. And you can actually productionize that as well. But that’s a standalone app that doesn’t integrate with lot of existing products or your ecosystem or your system.
Yeah, yeah. So maybe as a new.
Federico Ramallo (22:25.858) Yeah, for those type scenarios, think that the, you know, vibe coding is great. Yeah.
Udit (22:32.327) Yeah.
Federico Ramallo (22:34.254) And the other part is when you kind of know what you’re looking for, but you don’t know the specific word or you don’t know the specific, you know, where it exactly is. So you can ask the AI, hey, I’m looking for whatever it is. It’s somewhere here. Please go and look it. And it’s going to figure out, you know, the, you know, I found it, right? So.
So I use a lot the copilot to do those type of things. I don’t know the specific instructions, so please help me, give me a subset of functionality, behavior, or you know, I just need a little bit of JavaScript code and then I can review it and say, it works. So we, by we, mean like with other software engineers, we talk about that is an eager,
naive junior software engineer. He wants to help. He doesn’t know what, but he wants to help. He doesn’t know the underlines, but he wants to help. He’s eager to help.
Udit (23:47.664) Right, yeah, and he’s probably quite intelligent, but sometimes can be unpredictable. You don’t know how it may behave. That’s how AI inherently is, right? It’s not deterministic. I think the analogy that they sometimes use is traditional software is like a calculator, like one plus one is always two, but with AI you never know what answer it will give.
Federico Ramallo (24:12.494) You
Udit (24:13.819) So yeah, kind of, yeah, like it’s, I think what they say, like it’s a very intelligent puppy who’s very good at learning and like if you train it with a lot of good data, it will give you really good responses and it’s super smart at understanding a lot. But yeah, it can go wrong sometimes. So, you know, you need to know how to kind of, how to deal with it. Like if you’re including or building AI into your products.
Federico Ramallo (24:38.264) Right. People talk about prompt engineering, right? But I remember who I have to look up who, but one person talked about prompt alchemy, right? Because we really don’t know exactly how this thing works, right? And it’s so unpredictable or it’s not deterministic.
We’re kind of trying to figure out how to give better prompts, but it’s a best-guessed scenario. There’s not a lot of engineering behind it because at the end of the day, we don’t understand how it works internally. There are a lot of we understand, but with AI, when you ask to parse a document and put it on a vector database,
gives you like 1500 dimensions that we don’t have any idea what they mean, right? The AI knows, right? But we don’t, right? So that’s what I’m saying that at the end of the day, we don’t know exactly how it works internally, how it thinks internally, right?
Udit (25:51.527) Yeah, absolutely.
Federico Ramallo (25:55.042) Yeah. So what other ways have you found to use AI that it’s interesting that you can share?
Udit (26:07.281) Yeah, guess one aspect is, I mean, as a PM in my role, how I use AI and I of talked about that. So like, mean, like for productivity, like web coding for producing prototypes and stuff like that. And then of course, like how do we use AI in the products we build, right? So like I talked about the space that I’m working in. So it’s a customer support and sales tools. essentially in the business of making teams and business operations more efficient, like more smarter.
And really with AI, elevating their role from an executor of tasks to a supervisor of AI agents. So I think, yeah, taking the example of a typical customer service call. So imagine you’re a small business owner calling Google Ads because your ads are not working.
So an AI agent can obviously automate a lot of that, can understand your issue, can chart out a plan of how to solve it. And if it’s simple enough, it’ll go ahead and solve it itself. It might give you some knowledge, recommendations of how you can solve it, or it might actually execute in a few tasks, or it may create tasks for certain, at times you may need a human involved.
And at times it could be a scenario where, so what we’re really thinking hard about is, so where AI may not kind of like, how does that handover from an AI agent to a human take place, right? So, say if it’s a really frustrated customer because their ads are not working and like the business is on the line. So they want like a solution right now. So, and AI may not be able to understand that level of panic or that frustration as easily.
So a huge part of my job is designing the system to recognize the limits of AI, what is it good at, what is it not good at. So ensuring that at the end of the day, we to keep our customers happy, right? So we have to get that customer, that frustrated customer to human, maybe the right empathy and expertise as fast as possible.
Udit (28:12.955) Yeah, so I think that’s a lot of like, where we’re spending time thinking about like how to kind of handle those exceptional scenarios and how should the handover from AI to human take place in this realm of like customers and sales.
Federico Ramallo (28:31.48) Yeah, I think that there are lot of opportunities for workflow automation, knowledge database, so having an agent that… And again, the agent is right now, it’s a very fluid term, so it hasn’t been quite defined yet. It’s like anything LLM that it’s connected to within a box, right? Anyway, the…
having context of your data, your application, then that allows you to get information and then eventually with agents, can also run actions. So the LLM can figure out, you need to do this. Do you agree? Yes, no. And then move forward.
Udit (29:23.923) Yeah, totally. But yeah, as you were talking, I mean, like, how do I use AI in my personal life, maybe a bit more, yeah, I could talk about and…
Federico Ramallo (29:34.582) All right. Yeah.
Udit (29:34.674) I think at Google, there’s a new AI sort of tool that they keep launching almost every week or every day, if that matter. And we get access to it even earlier on in a dog food capacity. And lot of it gets launched in the Google Labs. And I think there’s an external-facing labs.google.com as well, where anyone can experiment with some of these tools. So recently, I’ve used this tool called Flow, which helps you.
Federico Ramallo (29:40.792) Hehehe.
Udit (30:04.147) like generate videos and a new thing that they released was they can generate a 8 second video based on like an image clip that you can upload. So basically uploaded my dog’s photo, it was just a regular photo, it wasn’t a video of him and it created a very natural video of the dog running around on the beach. It was a photo of the dog on the beach and then they created a video of that.
of him running around and then kind of playing there and stuff. I thought that was pretty cool. Yeah, yeah. And with just one single photo, and it actually made a video out of it, which was just mind-blowing. And I think another one, I don’t know if you’ve heard of, sorry, are you going to say it?
Federico Ramallo (30:32.366) That’s nice, that’s amazing.
Federico Ramallo (30:43.758) Right.
Federico Ramallo (30:48.878) go ahead, go ahead.
Udit (30:51.491) There’s a project, Mariner, which is also another thing that kind of blows your mind. It was released during the Google I.O., the last one that happened, where you can actually ask the AI to do something on your behalf. So for example, if you want to search for the top five restaurants in San Jose, and it actually goes to a… It’ll open a new tab, a Google Chrome tab.
And then we’ll search for top restaurants. And you can see all of that happening on your screen. So it’ll search for it. And then it’ll scroll through the results. It’ll open some of those results. If you say that I’m looking for a restaurant which is within whatever, like five miles away where I’m living, and with a restaurant that has this specific kind of cuisine or this specific dish that you get. So you can actually open each of those links and search for that. And then.
kind of give you the result at the end of it. So it’s all running in the background, but you can actually see it running as well. So, which I thought was pretty cool. So you can actually see what it’s doing on your Google Chrome tab. Yeah, so which was also pretty amazing. You can let it run in your background, so you can come back in five minutes and then just see the results, or you can actually see it live in action. yeah, those are a couple of things I’ve started. you.
Federico Ramallo (31:55.31) Great.
Federico Ramallo (32:08.642) Yeah, that’s very interesting. I mean, because we’re getting to a point where we can automate more and more complex tasks.
Udit (32:16.627) Yeah, yeah. I like in your personal life, right? Like, I mean, I mean, of course, like searching for restaurants, I mean, it’s simple enough to do. So I may do it, but yeah, if it’s some complex with a lot of different criteria that I have, then, or even like say, yeah, on Instacart, like, yeah, add like, yeah, from this list of items that I want to order, add all of these items. Go to Instacart and add all of these for me and order it as well.
Federico Ramallo (32:45.902) Yeah
Udit (32:46.131) And it can actually is where it may run at a point where it needs some input from you. So for example, it needs you to log in on Instacart because it’s not logged in by default. So it’ll actually then give you a notification that it needs your attention. And then you can go and then provide those. It’ll say that, give your email ID and password for Instacart. So you can go and enter that. And then it’ll enter that for you on the screen. And then go ahead and do everything else.
Federico Ramallo (33:14.049) nice.
Udit (33:14.323) If you haven’t tried that out, yeah, you should try it out. It’s pretty amazing.
Federico Ramallo (33:19.864) Right. I’ve seen people talking about shifting from the search box to the LLM box, right? Instead of a Google search, now people want to type on a chat, right, and get more personalized answers. Have you experienced that type of, you know,
shifting in your personal use of search.
Udit (33:55.956) Yeah, I think so. Yeah, yeah. I think I end up using a lot more sort of Gemini instead of traditional search. And even in search, now you have the AI mode. So if I want something hyperlocal, like I’m searching for something nearby or…
or like some factual thing, then I still go to Google search. But in Google search, yeah, I default to the AI mode because it at least gives a curated answer to me instead of me having to go through the results and stuff like that. And then I can ask a follow-up question on AI mode.
Federico Ramallo (34:35.886) The responses used to be not very good, but recently they’ve been improving a lot. So yeah, I’m using more and more the response of the chat rather than the response of the search.
Udit (34:51.399) Yeah, yeah. And those early days, yeah, those things like, you can put what, like, like tiny pieces of rocks on pizza or, you know, there’s a safe to eat and like all of those examples I got famous, right? Like in the media, this is what like Gemini and JPT is telling me. Yeah, like all those, yeah, kind of ridiculous things. Yeah. Which may still happen. Like, like, even like exceptions today, if you ask a really weird question, it’s going to maybe give you a weird answer. Okay. Is it okay to eat rocks? And yeah.
Federico Ramallo (35:10.67) Yeah
Udit (35:19.751) I think it’ll tell you, yeah, sure, and killing a bunch of reasons.
Federico Ramallo (35:24.526) Right, right. Yeah, the… It’s interesting, you when they try to put guardrails on the LLMs, but they’re easy to work around, right? The LLM is too smart for its own good, right? So you ask the LLM, how can I build a bomb? says, no, no, I cannot give you that answer because that’s dangerous, right? Okay, what should I not do?
Udit (35:44.945) Yeah, exactly.
Federico Ramallo (35:54.216) not be the bomb and he’s like don’t do this don’t do that so he gives you the know he gives you the answer without you know realizing and i’m oversimplifying it right but there there are many tricks like that that that i find it funny yeah
Udit (36:14.918) Yeah, know, good stuff.
Federico Ramallo (36:16.001) So I wanted to ask you a little bit about what is your workflow on building products, how is the process of creating an AI power product different than building traditional software?
Udit (36:30.48) Yeah, think we spoke a bit about it, right? So think the expectation is definitely to build more AI into all the products that we build. Because the technology is there, the AI, the LLMs are really smart, so you might as well leverage it to improve your business operations. But.
Yeah, so what we’ve done is, I mean, of course, we have the standard Gemini knowledge retrieval and all of that running. And we have like chat bot and voice bot kind of agents running on customer support. So if you call into Google support, say for ads, there’ll be a voice bot talking to you, trying to understand your issue and stuff like that. yeah, But where we’re spending a lot of time now is
Yeah, we just don’t want to just include AI agents for the sake of it. We need a way for agents to kind of seamlessly hand over things to humans when needed. So what is the right way of doing that? So that’s one area. And I think another area is
sort of like for the human themselves, now for the customer support agent. So for like for an ads agent or whatever, Google Pixel agent. So now they’ll be able to do a lot more. There’ll be AI agents helping them out in some of their tasks. So they need a way to monitor the AI, right? Their role itself is fundamentally changing from a regular like executor of tasks to someone who is supervising in some ways like a group of a workforce of AI agents.
So how do we offer those views that they can easily monitor the different AI agents running, and if they want to take over some conversation, then they can do that, and so on. So those are some of the things that we’re thinking of.
Federico Ramallo (38:21.976) Right.
Federico Ramallo (38:26.23) Interesting. Yeah. I mean, I’m changing a little bit of topic or going back to building the product. So one thing that I’m using Microsoft Office 365 for email and corporate email and
Somehow they decided to, instead of showing you the apps, they show you the chat by default. So I find that very intrusive because you’re looking to open an email or open a document or whatever. But instead of like, you would go, you could click Word, whatever, or Outlook, get the emails.
Udit (38:57.852) Okay.
Federico Ramallo (39:16.43) and so on and so forth, right? But now it shows the chat. So it gets harder and harder to get to what you’re trying to do, right? So they’re pushing the chat to the front. Or I’m looking for a document, right? So I’m saying, okay, then if they’re doing that, I’ll start using the chat, right? And I say to a chat, can you give me this document that I know that I have on my OneDrive storage, whatever, right? And the response is generic, it’s, oh,
There are many documents like that on the internet. It’s like, I don’t care. I want my document, right? Search on my files, right? I say that and it doesn’t understand. So the way I see that is a bad feature implementation, right? Because they don’t understand what the users are looking for to do, right? And there is no obvious way to actually change that, right?
to change that behavior. So they change the flow of the homepage, right? And basically affecting every day, the workflow for every day, every user, So I find that very intrusive and useless because it actually, yeah. Sorry, it’s a little bit of a rant.
Udit (40:31.375) Mm.
Udit (40:35.57) Right.
Udit (40:40.082) Yeah, it’s actually, mean, kind of, it’s actually making a job harder than what it actually tried to make it simpler and so on. Right. So yeah, I think, yeah, that the user experience aspect is so important, right? So you just don’t want to shove it into like whatever you’re building. It’s kind of, yeah, you need to do it in a sensible way, in a very kind of thought through way for the user. Yeah. So it’s a lot of, I think, I mean, at every like sort of tech company, whenever you’re building products, there’s a lot of user research that goes on.
Federico Ramallo (40:46.691) Right.
Udit (41:09.456) I like real user research and not just using, say, the AI mode of the deep research in Gemini to get user research, actually talking to users about which of these modalities would you prefer. Would you prefer a side panel of a chat interface, or would you like to recommend it at only certain points of your workflows and stuff like that? I think that becomes a lot more important because it started to offer a lot more capabilities and new experiences, but we’ve got to be.
kind of cognizant of doing it the right way.
Federico Ramallo (41:44.514) Yeah, I think we’re going to see more and more use of AI, but I think that this particular implementation was not the best one. And basically they miss understanding what the user wants. As you said, you need to talk to real people to understand what they actually need and want.
Udit (41:54.652) Thank you.
Udit (42:05.85) Right.
Thank
Federico Ramallo (42:08.96) Awesome. you know, another thing I wanted to ask is about which project are you most proud of and why?
Udit (42:19.953) Yeah, I like that. And that’s maybe taking me back a couple of years in the early days of Google. So at Google, you must have heard Sundar sometimes say, we used to do AI, like in ML, before AI was cool. So even back in 2019, 2020, was actually real, in a way, AI stuff that we used to build. So one of the first.
projects, or my first role at Google was as a product manager for Google Ads, Google Ads support tools, specifically working for the Google Ads domain. So we had a challenge where, so there of course are millions of customers calling phone support specifically, and they’re calling for different issues, right? So some of them would have like a billing issue or a policy issue.
or something about their performance not being up to par, stuff like that. So there was a large team of agents on the Google Ads support team, which is staffed. And their only job was to understand what the customer’s issue is and then transfer the customer to another team, which is a specialist team who can actually solve their issue. So it was very simple job. All they had to do was to understand the issue and then transfer the customer. So my goal was to create a way where we could do it in a more scalable way.
And scalable, not just to save costs for the company, right? To even make it better for the customer who’s actually calling in, to being able to get them to the right agent, to the right specialist team as quickly as possible. So we built essentially a conversational voice part back then, which leveraged NLP, so natural language understanding, and then you.
So which would understand what the customer is talking about and like the board would ask you a couple of follow-up questions as well if it doesn’t quite understand like what your issue is. If you’re just going to say that I have an ads issue it’s going to ask you like a follow-up question. Yeah, please could you tell me a bit more or like contextually understand based on what your response is and then you’re transfer you to the right team. So if it’s billing policy, your YouTube ads or whatever that is. So yeah, I mean that ended up saving like a lot of your millions of dollars for the company like
Udit (44:30.253) like by doing that specialist routing automatically. But also, like it’s all like a problem for the customers, right? So where customers had to, it would be like at least 10 minutes into the call before they would be routed to the right agent who could actually solve their issues. So now instead of 10 minutes, it’s just taking a few seconds to get you to the right agent in a way.
So think that was unique in the sense that it was one of the first times where we were leveraging something like this. it was the first time we were doing this in the industry as well. It wasn’t as common in the industry. You traditionally had like simple IVRs like press one for this or press two for this. And then you could have like a long list of press 10 for this, press 100 for this. I it could just go on and on. So we actually had a conversational voice pod. It’s one of the first ones in the industry out there.
Yeah, it was super interesting. Internally, there was a lot of cross-functional alignment that we had to get, like with our legal teams, policy teams, with our UX conversational designers about what are some of the right things to tell to the customer, how should the voice part sound very conversational, natural sounding. We don’t want to make it sound very robotic. At the same time, we don’t want to create the perception that this is an actual human. the caller should know that it’s actually a…
It’s a bot talking to them and not a human. So like, stuff there as well. So yeah, I think that’s one of the projects that I’m most proud of.
Federico Ramallo (46:04.27) That is very interesting, yeah. Because with the volume of calls, you can save the company money, but also improve the user experience. Because every time you call for support, if that process gets dragged and dragged and then that’s usually the contact between humans and Google, right? Because…
usually it’s all the software, right? So that experience is critical to keep the customers happy, right? And when they’re calling is because they’re having an issue and they need, they’re in pain and they need assistance as quickly as possible, right?
Udit (46:49.42) Yeah, precisely.
Federico Ramallo (46:53.142) interesting. What is your one principle you never compromise on when building products?
Udit (47:07.535) Yeah, I think, yeah, guess, yeah, we talked a bit about this. And there’s always, like, I mean, general advice as a PM. Of course, you say that you should always think about the user and so on. But I think, yeah, I’ll talk about something that’s maybe unique about me. So I talked about how my background spans. traditionally, I had an engineering degree. And then I started the campus startup, so which was that scrappy.
user first kind of mindset there as well. then, so my consulting work, understanding high level business needs. So like, I think all of that culminates to my PM role here. So where you sit at the intersection of all of these worlds and act as a translator. So I think you have one, so coming back to your question, like, so one principle that I don’t compromise on is like, it’s very important to translate the customer’s pain point into
into what they really need, right? So even in the AI world, you have to translate a complex AI feature into clear business benefits that you can bring to the organization that you’re building for. You have to translate a high-level corporate strategy into an inspiring vision for the team. And essentially, the best PMs are the best translators in a way. So I think not forgetting that.
So where irrespective of, yeah, you will have all of these tools at your disposal, there’s still some level of translation that you have to do, some level of kind of understanding and then kind of producing the right recommendations out of it, if that makes sense. So I mean, you just don’t use the outputs of an AI just as is. You have to still make sense of that and still apply your thinking to it and your translation to it to come up with the right recommendation.
Federico Ramallo (49:07.756) Awesome, so one last question that I have for you is what advice would you give someone who wants to move into product at Big Tech or product at any company?
Udit (49:20.079) Yeah, yeah. think, yeah, I think, yeah, like with product and like we’ve kind of, talked about this, this kind of been a theme of our chat, has been to like just build things, right? Like now more than ever, it’s easier, like it’s very easy to build stuff. So you can just, yeah, go to like Jules and like build an app of your own or like use Gemini Canvas to build a prototype or something or.
Stuff like that. So it’s the best time to start building something like as a fresher out of college if I were to graduate today I think it was probably a better idea to continue on my startup journey Then then like trying to like work towards getting a job Because like of all the tools that it’s available there like you can actually easily build something and that’s especially important for getting into product and like and and Big tech and it’s going to become like more and more important. So
Because those are the kind of skills that be required. So kind of thinking about what to build, actually going ahead and building that stuff, going from zero to one, having that mindset. So instead of actually learning some of those hard skills for an entry level job, which might become obsolete soon, it’s some of those kind of
soft skills or like that mindset of building more, building something new is what I would recommend.
Federico Ramallo (50:50.414) Basically keep moving forward, building stuff and learning from that experience. Interesting.
Udit (50:56.025) Yeah, yeah. mean, if you have an idea of something simple in your daily life, which you have a pain point, then yeah, just build an app for it maybe. Yeah. If it helps you, like a reminder for, I don’t know, like making sure that you have to, like give your pet dog this medicine once a month and sort of like, you need to maybe go to a website and order it every month. So you can actually build a tool for that or something like that.
Yeah, build it for daily experience, build it for something more than that as well. Publish it out there. Yeah, and you can see where it goes. So at least you start building the mindset. Like it may not turn out to be successful because, I mean, no one except you might actually use it, but that’s okay. You still build something. I think that’s what matters.
Federico Ramallo (51:47.286) Interesting. Great. Great. So we’re running out of time, but I appreciate, Udit, having you here, sharing your experience. Any closing thoughts that you want to share?
Udit (52:02.795) This is great, Rodrigo. Thanks for inviting me. Yeah, had a great conversation. And I’m sure if I come back to this in six months from now, things would have changed so much. So whatever we talked about today would seem like a distant past because the rate at which things are changing with AI, feel like it’s so fast that, yeah, it’s really exciting. It’s a really exciting time to be in tech and to just be, yeah, be out there.
Federico Ramallo (52:31.512) Yes, yes, I agree. mean, the amount of new things that are new tools that are coming up with AI power tools, it’s amazing. And I think that this is going to keep increasing and the quality of the tools are going to be better and better.
Udit (52:50.917) Yeah, totally. Great. Thank you.
Federico Ramallo (52:52.814) Right. Yeah. So thank you very much, Uri for joining us today.
Udit (52:57.531) Thank you. Have a nice day.
Federico Ramallo (53:00.078) Have a nice day. I cannot stop. Let’s see if I can stop. Yeah, I cannot stop recording. My UI says reconnecting and I’ll edit it later. Yeah. But anyway, thank you, Judith. I think it was a great episode. I mean, it’s still recording, but I’ll cut it. Yeah. Yeah, it was great. mean, how do you feel about the episode?
Udit (53:11.715) I’ll be there.
Udit (53:20.281) Yeah.
Udit (53:27.215) Yeah, yeah, I I liked it. I enjoyed the conversation. Thanks again for doing this.
Federico Ramallo (53:31.576) Great. Awesome. So I’ll send you the video once I edit it. I’m trying to edit the next, probably in a week. I’ll send you the video so you can review it. Let me know if you have any feedback on to change anything. By the way, is there anything you want to cut out, edit out, anything like that?
Udit (53:51.282) Yeah, nothing top of my mind. guess I mean maybe naturally I don’t know if you have tools to remove some arms and hands and stuff like that. Yeah, which I think some of these tools have.
Federico Ramallo (53:59.663) Yes, yes. I remove silences and then it has two cameras. So, you know, when you talk, shows you, when I talk, it shows me, you know, those type of things. know, we filter the audio. I edit the beginning and the end. And then what we do once we have that long video and we have a schedule for publishing, we also do short videos. Because, you said
very quotable moments basically, you know, where we can do little short clips that we can share on the networks, so you know, I think that’s interesting because on the flow of the conversation you have great ideas that come naturally, so that’s kind of what we want to highlight and you know, people having short attention span so they can see that and say wow that sounds great, right?
and then they can have access to the full video if they want. So that’s kind of the idea.
Udit (55:00.401) Yeah, exactly like in a lot of podcasts like Lenny’s and all of that he kind of shows like the major like quotes or clips in the beginning and then he starts the entire podcast and stuff so it kind of it gets you into it and then But there’s really cool stuff like how many of these do you record like in a week or so?
Federico Ramallo (55:20.738) So right now I’m nine this week. Yes, it’s a lot. I mean, I’m hearing different experiences from different people, very interesting. So yeah, it’s great. I think it’s a great way to connect at a deeper level than, you know, on a conference or, you know, so many other places,
Udit (55:25.478) wow, okay, that’s amazing.
Federico Ramallo (55:47.894) or a meetup where there’s 30 people so it’s too much overload.
Udit (55:55.926) Yeah, and I’m seeing, think you get people from different industries, right? It’s not just tech and stuff, yeah. It’s just nice, it’s a good diverse group of people and perspectives.
Federico Ramallo (56:08.238) Right, right, yeah. And if there’s anything I do for you in the future, let me know, right? I’m around. If you’re planning to go to Guadalajara at one point, let me know. Right next to Guadalajara, there is a town called Tequila, where all the tequila comes from. know, yeah, Agave, yeah, yeah, And there is a train.
Udit (56:28.603) really, there’s actually a town named, yeah, it’s like a gawe basically that’s built, like, drawn there.
Federico Ramallo (56:37.166) I’m telling this to people because I never went to because it’s, know, when you’re local, you don’t go to the touristic things. So they take you through the agave plantations. It’s a train. It’s a luxury train from the fifties. They give you some samplings of tequila throughout the road. Then they take you to the factories. show you the factories. So it’s a very interesting and unique experience. And, you know, I
Udit (56:43.393) yeah. Always.
Federico Ramallo (57:05.574) I need a friend to do it. Yeah, I’ll let you know when I’m coming back. I’m going to be back in August. So yeah, hope maybe we can meet in person at one point.
Udit (57:09.015) Actually, that sounds pretty amazing. Yeah, yeah, That’s a great tip. Thanks, yeah. You should definitely stay connected.
Udit (57:23.375) Yeah, absolutely. Yeah, yeah, let’s meet over coffee. If you can, yeah, visit the campus or something. Yeah, let’s keep in touch. It’s exciting. I like what you’re doing. Even at some point, I want to start something of my own as well, like start doing something, like building something. Yeah.
Federico Ramallo (57:40.748) Yeah, yeah, I think it’ll be interesting because I build stuff like I’m building privated AI and the feedback that I’m getting is I’m building it but I’m not following link startup. So I’m building without validating my assumptions of if it’s what the user needs, right? So it will be interesting to build something together at one point, yeah.
Udit (58:04.079) Yeah, Yeah, yeah, let me know when you’re back in the bay.
Federico Ramallo (58:09.272) Well done. Thank you very much, Udit, for joining. I’m looking forward to see you.
Udit (58:14.735) Yeah, same here. Thanks, Rodrigo. Have a nice day. Bye.
Federico Ramallo (58:18.498) Awesome. Thank you. Bye-bye.