your-ai-doesnt-have-feature-problem-has-adoption-problem-charanya-kannan-launchpod-logrocket
===

[00:00:00] Sharanya, how you doing? Thanks for joining us today. Of course, happy to be here, Jeff. Thanks for the invite. Definitely. It's funny seeing you now 'cause I feel like the last couple months I have seen Navan everywhere. I get ads on my streaming TV, on my phone, I see them out in the world. Nice to have you here now because, uh, we can kinda make it real.

That's kinda what we're gonna talk about actually, is the fact it is everywhere, and you have this product that is very, very AI-driven, and how do you get people to actually use it? Having ads is great, like whatever. Anyone can pay for ads. But I also hear people talking about Navan, which I feel like wasn't true two years ago necessarily, and it is always good things.

But maybe before that, it's just worth doing the 90-second TLDR. You're VP and GM of Navan Anywhere now. Mm-hmm. But you didn't start out there, so, like, what is that path that led you into product, you know, then here to Navan? So I'm originally from India, where I [00:01:00] started my career with the Tata Group in India.

It's a multi-billion conglomerate there. I was part of their leadership program. So I had the privilege of working in South Korea with the Tata Group in their automotive division. I was the first woman in the management team over there, and it was a phenomenal experience of, like, learning, just being in a different culture.

The automotive space is extremely difficult and complex, and, like, it was a great learning experience. And all of that led me to come to Harvard Business School, which is when I moved to the United States, and then post that, two years at BCG, where you really learn the first principles thinking, which still continues to drive my career and my career growth.

And then post-BCG, I started in tech, PayPal strategy. It was a fun ride, but just doing pure strategy wasn't fulfilling after three or four years because, like, all of the action in tech is really with the product and the engineering teams. And you know, in tech, there's this barrier to entry, especially in product.

It's like if you haven't done [00:02:00] product for the 10 prior years, you are not allowed to enter into product, and it's like a vicious cycle. So it was very difficult to break into product. But then in PayPal, I found this wonderful team in the core checkout product. Once they gave me an opportunity, once I got in there, the growth in product has been pretty fast because your first principles thinking, that sort of like strategic thinking, and then in product, you learn very quickly on how to execute, how to work across, you know, multiple teams, how to drive that execution.

And then Navan happened. When I joined this company, it was called TripActions, and our revenue was probably under fifty million back then. And today, we just had our earnings call recently. We are heading towards really the big billion number shortly. So one thing I wanna dig into, we've all been scarred by old legacy travel management tools, and what makes Navan positive is it just kinda works.

I've talked to hundreds and hundreds of product leaders, and everyone ships some kind of AI product or AI feature. So [00:03:00] much more often than not, the thing I hear is, "We shipped it." And it's amazing, and the people who use it love it, but we get very little adoption. It's not growing as fast and you don't get the percentage of users using it that you think you should, but everyone who uses it loves it.

I do think part of this is just driven by there's just so much new feature scope in every product on Earth right now that, like, it's just hard to cut through the noise. How did you guys look at not just shipping But actually getting people to use it You know, when we talk about AI, I feel all of the conversations in the product and in the engineering circle is on the building of AI part, right?

Yeah. Like the product building of AI part, that's where everybody tends to focus. Building is so easy. Building has become so easy and so cheap these days, at least like building the first prototype. Building a product that lasts is still hard, but building has become so much easier than what it was two or three years ago.

I find that people don't talk about adoption enough, and I think adoption is where the real problem is. Even the large AI companies today, [00:04:00] a year ago, they were like, "Our product, it will just grow." And then now I see so many salespeople and so many marketing people joining these AI companies, product-led growth people joining these AI companies, because even for those massive AI companies, adoption has to be a conscious choice and a muscle that they build to drive user behavior.

Ultimately, like when you think of Navan, yes, all of the big AI companies, the big tech companies are our customers, but then we also have a lot of customers in the Midwest, in Europe, in Asia, and the AI bubble that we see in the Valley, in the tech circles, I don't see that with a lot of those customers who are equally important and critical to us.

And so when you think of adoption, there's this thought that like, "Oh my God, you put an AI product out there and it's going to be massive." You know, it requires a conscious muscle building to drive that adoption. And so a big part of the birth of Navan Anywhere, the product that I'm leading right now, our AI initiative, is [00:05:00] around how do you make yourself so seamless and invisible across multiple surfaces that we take care of the distribution where people already exist, so we don't have to pay or drive that user behavior or change.

One part of it is, yes, we have AI inside our product. We've always had, even before the LLM revolution, we've used machine learning inside our product, and our users are used to a very cohesive, strong product with strong user experience. As you said, it just works. It just works for most of our customers.

With that, we've now started embedding conversational booking tools, conversational expense tools, and all of this helps drive a lot of simplicity, time savings, less cognitive effort for our customers. But then there's still the question of driving adoption, and that's when it's like I started using probably December of last year or January of this year, was when I started using an AI tool where I could just connect all of my different apps together in one place.

And that's when I didn't even have to open Jira to submit a [00:06:00] Jira ticket, which was heartbreaking, revolutionary, and I realized I don't have to open Gmail anymore with all of those marketing email clutters. My AI would give me the list of the top critical emails that I had to look at, and so all of the apps were connected in one place.

So when you think of Navan Anywhere, it's such an obvious moat for us to think of distribution in that sense. Where if we can embed the goodness of Navan inside Slack, inside Gemini, or inside Microsoft Teams or Copilot, where our users currently are today, then distribution becomes a no-brainer. So Navan Anywhere was a product that was born thinking about distribution first.

So it is more how do we get it in a place where they already are? What does that workflow look like? Not what's the feature we want to build. Exactly, right. So for the customers, if they have to put in extra effort to do something because AI is superior, then AI is not really superior. The best part is when you just naturally do what [00:07:00] you're doing, and it suddenly feels so much more simpler.

That is the magic and beauty of AI, and that's what we really think about. It's not just building a product that seems seamless, but how do you really embed it in everyday user's life that they don't even have to think about it, that just booking or filing an expense report, all of this business travel and expense, it's like the booking part, people really just wanna get over it and get to the trip as quickly as possible.

And when you're stuck on an airplane, this week, all of the disruptions in London and Europe, like, people just don't want to handle all of the hassle of that. And if Navan can take care of all of that seamlessly for them, where they already are, without any extra cognitive effort on their part, like, that's a winning strategy for us.

Software is not actually the reason we're here, right? We're not here to build software. No one cares about software. But our customers, all they care about is they wanna get from A to B. Their flight got canceled maybe, or they need to deal with expenses. It's just software was the best way to do it, but the best companies look at the outcome, not the feature.

And, and so in your guys' case, how do we put [00:08:00] what makes sense where they already are to just make it work? And then you work back to, what does that feature actually mean? How do we define that? How do we do all that stuff? No one wants another dashboard to look at or another surface to go cover. Thinking about approaching it that way is very, very easy to say, right?

Like, oh, we just want it to magically show up. But, like, the devil lies in every little detail. How did you actually approach that end of it? How do you actually make AI give trustful outputs and do the things that you need, and define that stuff versus just kind of building the feature set? I'm happy to share that we have zero critical hallucinations, right?

And how did we get to that? It's because we didn't treat AI as like one and all, right? AI is one more tool in our hands that is going to make the flow very powerful, and so we always lead with what really makes it work for the user. So the CTO of Navan, he's also the co-founder, especially with this revolution, he's back in the trenches with us, rolling up his sleeves every single day, like coming and [00:09:00] testing out the product.

And he's, "Really, will the user, like, feel this way when they use it? How will the user feel when they get this notification?" That's the obsession we have in Navan, where everybody, leadership included, we're just constantly using our own product and figuring out where we can make it better. So AI is only one component of it.

When you think of a conversational booking experience, it's just so seamless for me. Once I've done it, I don't go back to the classic app anymore because there is a real cognitive tax of using UI, like figuring out, "Hey, where are the filters? Where do I go?" You know, it's just a bunch of things like that.

And then in the conversational experience, you just speak your mind out. You, you can just describe that, talk to the AI in your own words, and then, like, figuring that out and then giving you the options, filtering that. So all of that intelligence and conversational experience comes from AI, and then we have a combination of our UI and our deterministic systems.

Our deterministic systems for things like payment, which is extremely critical and crucial, we're not going to let [00:10:00] AI just haphazardly decide, like, how to do payments, right? For things like payments, we still fall back to our deterministic systems. So by knowing exactly where the decision fatigue is low- And where the risks are high, just fall back to deterministic systems.

You don't need AI for the sake of AI. And then where it really adds value, that's where you bring the user experience inside. Another aspect of building the seamless AI product is, six months ago, I was having this banter with my design and my front-end engineers. I was like, "You know, you guys better learn how to do evals and how to do vine coding and all of that, because design and front-end is going to be obsolete because everything is going to be conversational."

I couldn't have been more wrong. I was wrong. Because the more we put out the AI product in front of users, there's also conversational cognitive load. Mm-hmm. How many times do you just wanna type like long answers or think of our AI returning five hotel [00:11:00] options for you? How do you describe that without images?

How do you potentially pick a hotel without ever looking at the images? And so this is where we think a combination of AI plus UI plus the deterministic systems, that is the future of retail. So the real judgment or discernment is, where do you bring in AI? Where does it make sense to bring in your agentic systems versus where do you fall back to your deterministic systems?

And sometimes just agentic solutions for all is just not the answer. And that's how we've gotten to a zero-hallucination system that users love, because a lot of our work that the product and the design teams do is just truly understanding the cognitive load for the users. When does a UI simplify things for them versus when does conversational make things easier for them?

The question you asked, how do you make it a good experience to select a hotel if you don't have pictures? I would wager who's probably gonna have the best insight into that is a product person who understands or someone who understands the customer deeply and kind of a UX designer- [00:12:00] Mm-hmm ... who can actually think about, like, here's the constraints, here's how we would approach it, and here's how people, you know, work through these things typically.

But when it comes to the human taste, bring in people, and that's kinda how we've looked at it here as well as where do we focus our tool set and where do we focus what we're trying to do has very much centered around what are humans really, really special at, and what's the superpower there, and how do we enable them to do more of that, and what's the stuff they hate?

Maybe I'm wrong, but, like, I would wager your favorite part of product is probably not Let's go into Jira and try and prioritize a backlog This is a controversial take, but we never do that. In fact- Yeah ... our engineers write a lot of the Jira stories. Yeah. We have a very, very lean product team. Even before the AI era, all of travel, which today is over three billion in GBV as of last year reported numbers, and even much bigger this year.

It's a massive product. A couple of billion dollars in GBV. Yeah. And we've been able to achieve this with fewer than five product managers. Because product's role is really like when there's ambiguity, figuring out what the trade-off is [00:13:00] resolving that. Yeah. And then working across sales and marketing and commercial to really drive that adoption, and then really tracking the metrics to understand.

You know, we don't track, like, success of features. What we truly track is the success of the overall product. Is it driving revenue? What does our margin look like? And so that has really been how we designed our product team. Our product team was never about Scrum or Jira or anything like that. And so that's just a necessary tax we have to pay just because we're a growing company, we need records and all of that.

And so simplifying Jira with AI and never having to open, at least in Jira, is probably one of the best gifts that AI has given us. Yeah. To your point, you just want travel decisions to be easy wherever you kind of are so you can- Mm ... you, I don't wanna log into some kind of platform to do all the whole thing.

I just kinda want a couple options presented to me, and I wanna figure it out from there. The engineers writing tickets, I'm sure also Jira is, is their least favorite part too. So, like, that's where we've looked at how do we enter at spots where people naturally, [00:14:00] it's not, not even just not their superpower, it's just not what they want to do.

It's not what brings them fulfillment. How do you kind of bring the parts that are good to the human level? But speaking of the human level, one piece I'm curious how you guys approach, because I've seen this differently everywhere, but there's, like, a common theme. Writing code is one thing. It's deterministic.

For the most part, it's going to do the thing you program it to do every single time. In a year, it's gonna still be doing that. AI is not that way, to be as simple as, as possible to say it, right? Like, it's almost, you know... I think at one point it was, I described it as training my five-year-old how to do something.

Yes. I think it's a little older than a five-year-old now. Yeah. But, like, how do you kind of approach that kind of training the AI and the fact that it is kind of as much of a human as it is code? Honestly, this has been the best of times and the worst of times, right? For for all of us. So this has been the most fun product management has ever been as far as I've experienced it.

But this has also been the most difficult period in product management because the technology is [00:15:00] changing every few weeks, and typically when your sales team come to you and say, "Hey, this is broken," you go back to them and say, "Hey, now it's fixed. That problem will never happen again." And now, for the first time ever, I'm finding myself in a situation where I can't say that.

I'm like, "We've made an attempt to fix it." Our evals show that there is an eighty, eighty-five percent probability that it's fixed, but then is it really fixed? Like, we don't know. It's AI, and it's non-deterministic, and so there's a little bit of that. So if you think of it as a pyramid, right? The base of the pyramid is absolutely what I first said, the most critical flows, extremely high stakes, which is literally the payment transactions.

That part, there is no need for, like, AI to go and do that. The API calls are deterministic world works perfectly fine for that. It's fast. It's quick. There is no judgment involved over there. As soon as the user hits book, you go ahead and book. So having that sort of, like, deterministic systems for the financial thing is, like, the solid base.

That's where we get a lot of our zero critical hallucinations from. [00:16:00] But then the second part of the pyramid is just really having our own infrastructure. So instead of using large language models, our phenomenal engineering team, they've sort of taken an open source model, Qwen, and distilled it, and we find that that sort of distilling and deploying that model, which is fine-tuned for our own use cases, that has improved our accuracy rate by at least five or six percentage points compared to just using the large language models that are available for plug-and-play in the market.

And then the third aspect of the pyramid is truly how do you as an organization do evals? So today, all of us come together, product, design, engineering. We all read these conversations. So we have a weekly review meeting where the CTO, the SVP, they all join, and we're all sharing chats and conversations that we've read through.

And so there's this deep, deep culture of all of us are reading these conversations. And so we have systems set up with [00:17:00] agentic reasoning for all of this. It flags to me every single day a couple of chats that if things have gone wrong, and I should probably read that. There's just this culture of constantly reading those conversations to understand how we can fine-tune that.

And then we go ahead and, like, we break a very large prompt into smaller, more auditable prompts that lend itself to more easier auditability. So there's just this whole culture part around how you do evals. That combined with how we've built our systems, which is distilling open source models. We don't use external third party.

The CTO's organization built this tool called Cognition. It's our own AI infrastructure that we run. And then ultimately, we still fall back to deterministic systems for extremely critical high volume payments. So a combination of these is what I think is leading us towards building that very robust AI product.

Is it still perfect? Exactly what you said. It's, you know, at this point, it's probably not a five-year-old, it's more like a fifteen-year-old. So nine out of ten times, it's great. And that one time, that's where legal comes in. We have a [00:18:00] disclaimer, "Hey, this is AI," right? "Use your judgment." I wanna take that topic and extrapolate and branch off it slightly because there's one other thing that I talked about.

One of the fun things I get to do aside from hosting this show is we do this whole kind of series of dinners for product leaders across the country where we bring people together. And a thing that comes up really, really often is what is the future of the product role in AI? I think we've all seen the LinkedIn post about, like It's the end of product.

I don't think it is. But you brought up something, I think you wrote about it, or we talked once before a while back, and you brought this up, and you put it just really, really well, so I wanna kinda dig in. But, like, I do think at some level there is going to be a flavor of product people, of PMs, that no longer exist.

I think you nailed it when you said, like, a lot of PMs just optimize for the wrong thing. And what does wrong versus right look like here? Yeah. So some of this comes from having interviewed or mentored close [00:19:00] to 100 PMs. Yeah. And having worked with at least 30 PMs over the span of my career. And in all of these, it's a very broad classification, but I think I put PMs under two buckets.

This is too simplistic, but bear with me. One bucket is product owners, right? Mm-hmm. And so they think the product role is very much maintaining Scrum like Jira, and they're responsible for the delivery of the engineering team. Guess what? The engineering team doesn't need babysitting. Especially now in the AI era, engineers have become extremely strong, right?

Like, coding takes less than fifty percent of the time. They have the other fifty percent of the time to reason with AI, to think through things, to, like, think through what it means for the broader idea of the product. So then the PM's role becomes much broader than that, and that's the second bucket of PMs, who constantly think about, "Hey, why are we doing this?"

Am I building this feature because one customer pinged me saying they want to have this, and am I just blindly, like, going and building that? Am I just playing this [00:20:00] whisperer role where what the sales tells me, I just go relay that to engineering? When engineering builds it, I go back and say that to sales and marketing.

Is that my role? Definitely not. My role is something bigger, which is Think about how will this product outlast and live five years, 10 years, and what needs to happen for that to be true? A, your customers need to absolutely love this product. And when you think about that, your competitors become obsolete.

Like, why do you even care about your competitors? Because if you are constantly ahead in delivering value for your customers, that's all you need to think about. The second thing is: How do you do that in a way that is sustainable for the company? How do you think about the cash flow of the company, the revenue?

Do you know what makes GBV? What drives revenue? What drives the cost levers? And how do you manage all of this optimally with your product design and build? Now, this is what AI cannot do, meaning somebody has to frame the problem and give it to AI. So knowing this why, knowing how it ties to the broader commercials of the company, knowing the value it brings to the [00:21:00] customer, and just all of this being able to navigate ambiguity and bring clarity to the team and sort of driving with that conviction during that ambiguity, this is what I mean by what a real product person can bring value.

And this is something that AI cannot disrupt, not in the next two or three years, because that abstract reasoning, knowing what problem to solve, knowing how to frame the problem, right? Like framing the problem. Everybody knows there's a problem, but how do you frame that problem? Mm-hmm. Give the problem eyes and ears and a figure that people can go after and solve it.

That is where I think you still define the problem to your AI, and then your AI can solve it, mostly. But then you are the one who needs to know what problem to go after and what to solve. And you can use AI to reason and think through that, but a lot of that has to come from you, and this is where good product managers truly shine.

And to me, that's very different from, I call them product owners, but, like, generally they're very busy. They're in meetings all day. [00:22:00] They write Jira tickets. They even write, like, pretty strong PRDs, but it doesn't add to the end outcome, and it doesn't drive value for your users or your organization. So then, like, that's how I would differentiate this.

The first bucket will definitely go away, and we're already seeing that happening as people are upskilling, and the second bucket is much more valuable, and it will always be relevant even in the AI era. I think we all think back, there's very few people you ever thought, like, "God, that person is vital to success here."

Mm-hmm. "And they are incredible because they write a mean Jira ticket." Or like- ... they're really good at process and alignment. But you do, when you think of this person always has the ability to look at a problem and crystallize it in just the right way where we know what to do next. Yeah. Those are the people you look to whenever you run into a problem, who you go like, "We can't live without that.

We need that." And it, it just becomes more and more clear when the barrier to process and fast and all that stuff starts to float away, what's left is, like, that [00:23:00] person brings a great insight that made this thing that we were building better than it ever would have been. Definitely. Insight over process, and to me, that's the biggest mistake most product managers make.

A lot of product managers we've hired and who weren't a fit, they come in and they start to talk about the process, right? Yeah. They start with the process and the process, and that's fine. You know, 20% of it is still processes. You need that for the team to scale efficiently, but then if you're just thinking about processes 80% of your time, that's not what we want.

It's truly the insight and the clarity from ambiguity, what problems to go after. Like, that is what is really valuable. Exactly. Okay, there's one last thing I am super, super curious about, and I think there's a lot of people building even great AI products, 'cause this is the thing that starts to come out now, is think about how we've moved through the phases of AI, right?

For a little while it was like, "That's cute. It's a chat bot that once in a while even sounds a little human, but it's wrong all the time, and who trusts it?" And then if that was your only experience with it, you were not very happy. But as it grew, right [00:24:00] now, I mean, what, just the other day, one of the millennium problems in math was solved.

I think we can all agree coding it does a great job on, but these are all the places where I think we all agree, like, it's really, really good at are deterministic fixed outcome- Bounded problems, yeah ... that's... Okay, I'm stealing that. They're bounded problems, and they have an answer. Where I don't think even, like, as you watch people criticize, say, 5.1, Astra from OpenAI, everyone I hear says, like, "It doesn't think very well.

It doesn't do this, it doesn't do that," which is tough. But the other problem people run into, and this is probably gonna be the harder to solve one, is software used to be a great place when it came to selling a lot of it Because you made it once, you sold it, and almost every user was like marginal increase in, you know, margin for you was great because you just basically...

Like, the cost was making it. You know, compute was minor, unless you were doing some crazy data crunching or something. But now every instance we sell, you pay for inference [00:25:00] for every request and everything. What I hear in a lot of places is we all have to deal with margin compression, which I think has generally been true, except Navan is public, so this can be seen.

Your margin has actually gotten better. Yeah. What the hell, Trona? What's going on? 'Cause that's, that's bonkers. So there is a very big component of travel. Travel is still, you know, when you're stuck in a plane for five hours on the tarmac where the plane cannot take off, right? No AI can come and solve the problem for you Those are the stickiest situations where a human with all of their travel experience can come, like problem solve, put you on a different flight.

And maybe AI will get there in two years, three years, I don't know. But for now, it's a problem that humans solve. Mm-hmm. In Naavan, we have a lot of like phenomenal travel counselors who help out our customers all the time in these extremely sticky problems. And so what we've seen with AI for post-booking and support is, um, by the way, this is true even before the LLM revolution, [00:26:00] where we were using basic machine learning to help guide our customer support and chat with customers.

So our deflection back then was sort of in the thirties, like one in three conversation, like AI bot would effectively deflect. But right now we see with LLM and just the much more higher levels of inference and reasoning, we are able to see that one in two conversations are very effectively handled by AI, where it helps find an alternate flight for you, cancel your hotel, rebook you into another hotel.

Like it can do a bunch of these different things for you. It's just phenomenal. And then the extremely sticky, like very gnarly, difficult use cases where you're four travelers with a minor and you have five stops in different cities, and like now you're missing your connection flight. Like all of that, AI knows when to transfer that to a human with judgment who can understand, empathize, and figure out with their infinite stay compute power.

I still think human brain is unrivaled in its ability to do things. So that's when it transfers to our human specialists who are pretty phenomenal in [00:27:00] solving these complex gnarly problems. So the reason where we've been able to drive a margin outcome, and we're still getting better at this, is being able to use AI to automate all of those regular but still difficult post-booking conversations and assistance.

Mm-hmm. And making sure only fewer conversations reach our human counselors who are extremely important to solve those very, very sticky travel issues. The second aspect of the margin is also the fact that we've gotten much better at building our own infrastructure, maintaining our own infrastructure.

We're not paying for third parties. And the fact that now that we are distilling models which are almost like at one, you know, I'm ballparking here, but at least like twenty percent of the cost of the bigger AI models, like this is all contributing to margin expansion, and it'll only continue to get better.

I love to kind of hear that because we all have seen the Miro acquisition, Airtable, and the SaaS world is panicking about the multiples there. It's good to see kind of the other end is companies like Navan, [00:28:00] where you're figuring out these problems, and you figure out in the post-AI, with AI world, how do you continue to add value and how do you continue to grow and do things like, no, you don't need to have margin compression and AI.

You can find ways to grow, add that excellent experience, and also find ways to do it in a better economic way than you did five years ago. It's a really innovative story, and I kind of love across the board here how you're looking at this, right? How do you figure out what to build? Mm-hmm. How do you figure out where to focus it?

And it's not feature first. It's what's the benefit we're trying to do for the customer? So it's really interesting to see how you all are doing it, is build that all in, and then how do you do it in a way that is still economically viable? 'Cause at the end, we all wanna make great products, but we have to return some money to our investors, too.

Yep. Well, Charanya, I appreciate you coming on. This has been really, really insightful. Great info. I appreciate how deep you went across all these things and how Navan's doing it. I hope to, uh, stay in touch and love to hear how this kind of continues, goes, and look forward to the future here of what y'all are doing.

Great. I enjoyed this conversation. [00:29:00] Thank you for the very thought-provoking questions, Jeff. Appreciate it. Thank you very much. Have a good one. Thank you.