Reimagining the Future of Work with AI Employees

Intro 00:01
Sarah, welcome to Startup Success, the podcast for startup founders and investors. Here, you’ll find stories of success from others in the trenches as they work to scale some of the fastest growing startups in the world, stories that will help you in your own journey. Startup Success starts now.

Kate 00:17
Welcome to startup success today. I have Surojit Chatterjee in studio, who is the CEO and founder of Ema, which is a universal AI assistant. Welcome.

Surojit 00:32
Thank you, Kate.

Kate 00:34
I’m excited to talk to you. I want to get into all things Ema, because obviously this is of great interest to me and our listeners, but before we do that, if you could just set the stage and give us, you know, an overview of your background and what led you to found Ema.

Surojit 00:51
I spent a lot of time in companies like Google and then Coinbase. I actually started the mobile ads group at Google. Did that for 10 years or so, and made mobile ads one of the largest businesses for Google and in general. Today, mobile ads is more than half of Google’s revenue per year. I also led Google’s product search business, product ads, Google Shopping, and grew that from $7 billion to $25 billion in revenue per year.

Kate 01:26
Wow. So you played some integral roles at Google.

Surojit 01:31
Yeah, I happen to be at the right time, right place, most of the time in ads team and running very large parts of Google’s overall business. And then I went to Coinbase, and I was running all of product and also responsible for revenue at Coinbase. I was still Product Officer at Coinbase for more than three years, and this is the time when Coinbase went public, so I did a lot of work expanding the regular base, from 200 millions to multiple billions of dollars. Expanding 10 x’ing user base, diversifying the product portfolio and helping out in the IPO process. One of the most successful IPOs.

Kate 02:15
What an exciting time that must have been.

Surojit 02:17
It was, yes, a lot of learning. And post that I really wanted to do something on my own and stepped out of Coinbase, thought a little bit with different things, wrote some code myself, got together with a few of my ex-colleagues, and founded this company. We started working on it late 2022 and formally incorporated in late Feb 2023. That’s the story. And we are building a universal AI employee, not just an assistant.

Kate 02:58
Oh, okay, all right.

Surojit 03:01
Employee for any enterprise to help them automate business processes.

Kate 03:07
Okay, so is this something you kind of were thinking was needed back in the earlier stages of your career? Is that how you ended up in this universal AI employee kind of path?

Surojit 03:22
Yeah. Look, I ran very large businesses and companies in the Valley, and we always have the best people. Google Coinbase – very talented folks. But I saw what ends up happening is somewhere between 30% to almost 50-60% of the time, these very talented people are doing repetitive, tedious jobs. Tasks that they don’t want to do that’s probably not very value additive, but has to be done to keep the lights on. And we went through some as any startup goes through rough patches. We went through some tough times at Coinbase. And those times, and in general, you’re always thinking, how can you make your teams more productive, more efficient? How can you free up your most talented people to do the most creative and value added work? And that was the inspiration. And if you look beyond just tech, the amount of inefficiencies are even larger. If you look at like beyond pure tech companies,

Kate 04:28
Absolutely, I need an Ema.

Surojit 04:33
That was the inspiration. What I saw was LLMs are great in generating words, predicting the next words. That’s what LLMs do in simplistic terms. But in an enterprise, you need to predict the next action in order to create the action plan. That’s what humans do. Create AI or AI agents that mimic human behavior and can take on more and more of the tedious parts of somebody’s job. So we call this then AI employees, because these are actually like a mesh of AI agents, multiple AI agents, working together to perform a complex task, which may require multiple steps in a sequence, sequence of actions. This task is kind of adaptive, so based on the context of the situation in the enterprise, they can adapt. And that’s why it was very hard to automate them previously. But now I think there’s technology to do this, and we have seen great success with many, many, many customers in the enterprise.

Kate 05:42
Wow. I mean, it sounds impressive, if you can get beyond, you know, into that stage you were talking about that action stage, predicting action. And you feel like you you have,

Surojit 05:54
Yeah, I think we have, you know, this is of course, a very long term vision. yes, yes. I think most of the future enterprises will be highly, highly automated. So more number of humans will create businesses that are, like $10 billion or $100 billion worth, but have maybe 1000s of people, not hundreds of 1000s of people. And I think there’ll be a lot more, a lot more enterprises, a lot more companies in the future, because it’ll just become easier and easier to start a company. Because today, you know, to build a company and to scale a company, you need a lot of people with different types of skills. Hiring is often the number one challenge for most employers. Hiring the right person at the right role. So imagine you still have to hire, you still need humans, but if you can automate a lot of the tasks, you can just move faster, so that’s the longer term. I think in the short to medium term, we have seen a lot of success in a number of areas in enterprise where you can automate to a large degree. Customer support with a bunch of customers using us in automating resolution of complex tickets. This is not just deflecting like in a chat bot. It’s things that become real tickets that require humans to take some actions. We can do that. We have seen a lot of progress and a lot of good results in sales and marketing automation, like automatically generating leads, helping sales people prepare for a meeting, helping them prepare a proposal. You know, these are the things we have to do every day that can be done through AI agents, through these AI employees. And a variety of other use cases where we have deployed. Of course, it’s going to be a journey. We are still very, very early in this technology. Lot to be figured out. But I think the future is going to be very interesting.

Kate 08:05
For sure. I mean, those use cases you mentioned, that’s a big deal. You can take that off someone’s plate. (Absolutely.) I mean, it really saves time. And it’s the repetitive task that, like, for example, the sales use case you shared, right? Like building a proposal or preparing for a call. I mean, all the information’s out there. So if you can help streamline it, I see what you’re saying is, in that regard, it’s more like an employee versus, you know, just an AI assistant.

Surojit 08:38
Absolutely, absolutely. So sometimes the challenge in an enterprise or a fast growing company, like we’ve been working with FinTech companies like TrueLayer in Europe or Moneyview in India, their problem is their business is growing really fast. But can they grow their customer support, the face of the business? And sometimes they are not able to grow. They can’t hire fast enough.

Kate 09:04
Hire and train fast enough. Yes, I’ve been there. Yes, I’ve been in startups where that was a real issue around customer support. Wow, that’s fascinating. So that’s a big differentiator, then for Ema.

Surojit 09:19
Yeah, I think the big thing for us is, I’ll say a couple of things. One, we are a horizontal layer. (Okay.) Instead of doing like one thing and specific things. I have a hypothesis and I can go deeper into it, but building as a horizontal layer actually makes us stronger in any vertical. I know this is a contradictory statement, and I could go deeper into it, but we are a horizontal layer. The idea is we are building a factory of AI employees. (Okay.) So you can monitor them, govern them, from one dashboard. And I think increasingly that’s becoming a challenge for Chief AI officers like CIOs, CTOs, CIOs. If you are deploying generative AI inside your enterprise, you want to know what data this particular agent, this generative AI program is accessing, because it’s accessing sensitive data in AI for only when it has access to data. So it’s accessing your most private, most sensitive data. Is it protecting the data? How is it using that data? Imagine doing that check with hundreds of different tools across the enterprise. That’s a nightmare.

Kate 10:34
It sounds. It sounds like a nightmare.

Surojit 10:39
So our value prop, first value prop is we are a horizontal layer, so you don’t have to do that only once, not hundreds of times. And from one surface, one dashboard, you can basically get a view of, Oh, there, here are all my employees working along with humans. You can control kind of access, like which humans can work with each AI employee, you can also control which AI employee has access to which data. And we protect all PII and PHI. We automatically redact any PII. We automatically do things like copyright check for any document we generate, right, so to make sure you know AI didn’t copy something. So lots of safeguards. It’s fully compliant. SOC 2, GDPR, HIPAA, everything you can think of we went ahead and got the compliance done. So very safe, very secure, enterprise friendly. The second big thing is it’s highly, highly accurate. So we have created our own fusion model. It’s a mixture of experts model that combines multiple public models so you don’t have to worry about Oh, is Gemini better today? Is Quad getting better? Should I move to GPT-4o? We have a model that actually does that periodically, automatically and figures out how to combine, not just route with particular model, but how to combine multiple models together to A remove bias, so think of it as talking to multiple experts. (Right.) So that kind of has the side effect of removing bias that any particular expert may have. And B improving accuracy and at the same time lowering cost. Because not everything has to go to the most expensive model or models. (Right.) So highest accuracy at the lowest possible cost for enterprise use cases, we only do enterprise. We don’t do consumer or other use cases. So that’s been something that our customers really liked. You know, the ability to manage this new generation of software, AI employees, AI agents, from one surface, and being able to trust that this will be highly accurate and at much lower price, much lower price point than operators.

Kate 13:05
That is a huge advantage on the enterprise side. I mean, that’s a big deal, because right now, that’s what you read about. There’s a lot of inefficiencies, trust issues, cost issues.So you’re solving for all of that? Yeah.

Surojit 13:23
I mean, if you look at generative AI, a lot of people are on the fence. (Right.) They’re still kind of doing POCs, testing things out because of these issues, like, Can I trust generative AI to give me accurate results? Does it have bias? What happens when it’s inaccurate? How does it affect my business processes? I think the biggest of all is, how long will it take to set things up and make it useful? So a lot of you know, many of our competitors are what I call selling nuts and bolts. And imagine you want a car. You can go and buy the screws and the nuts and bolts and look at a manual and try to build the engineer yourself, right? So for us, I think you get a car, you basically get the car. It’s very easy to deploy, right? Basically, our platform is all conversational. You can describe a complex business problem, and it will automatically deploy the right agents and orchestrate amongst them to solve the problem. While the alternative would be: you hire engineers, you work with many other vendors, figure out different tools, stitch them together to build something that works for you after many, many quarters and months and only a few enterprises have the right talent pool. It just takes a long time. What we have done is built in unique architecture where we have sort of created building blocks. So if you think of any role, any particular role in an enterprise, consists of few building blocks, like it may require some writing of some software, writing some text, maybe email, may require understanding some text, may require using some tools or other applications, pulling some data or writing into some other some application, SaaS applications. So all of those unique modular actions, we have created agents for them. So now for complex actions, we can bring those agents together to actually help solve them, right? So imagine, like, it’s like each one of our AI employees is like a team of humans somewhere, because it has agents underneath it.

Kate 16:00
Wow, that’s so impressive, the way you’re describing it all, it makes me excited about the possibilities, you know, for an enterprise. But then it also takes away, kind of the fear that’s around there, around AI, which we see a lot of still.

Surojit 16:16
That’s a big, big part of it. I mean, you are absolutely right. Enterprises are worried, Oh, if I deploy these AI agents, what happens? Will they run amok? (Right?) Well, they don’t think, right? There are some famous or infamous, I should say, cases, right? The AI agent based like offering the talents or something right to the client, which is not in their policy, as I understand. For us, I think that’s a very important principle to follow while building such a system. (Okay.) We are looking at ways for humans to always intervene, because we don’t think we are at a place where this AI employee should be completely autonomous. They can be if you trust them. There is a way to do that. You can click a button, but we give opportunity to the human bosses to always be able to supervise AI employee, to ask for clarification before taking the next step. Once you get more and more trust, you can say, okay, you can do this part autonomously. I don’t need to double check your work. But I think a lot of exploration, a lot of good R&D innovation, lot of patterns we filed on this specific topic, how humans interact with this new type of software,

Kate 17:46
Right? Yeah, it’s a big concern. So you have really built that into this, with these checks and balances you’re describing. Until you build that trust.

Surojit 17:57
Absolutely. I always think of it as how a self-driving car kind of works. You know, I drive in a Tesla, where it can drive autonomously, but I can take control whenever I want, right? I can put it back in auto mode when I please. And as I got more and more kind of trust in it, I’m like, okay, I can let it run more time autonomously and just keep my hand on it, but let it make the decisions and power the car. It’s kind of similar, right, build trust over time, and it’s the same way humans work with other humans: you hire a new person, you build trust. You don’t let your intern go and present your CEO the next day.

Kate 18:50
That’s true. That’s true, right? That’s true. You wait. Yes, I see what you’re saying. And do you feel like – this is a broad question, I don’t know if you can answer it – but do you feel like the AI industry as a whole is pretty responsible in this area, like I mean, I think this would help people with adoption, right?

Surojit 19:13
I think they naturally have to be responsible, particularly for enterprise use cases. Otherwise enterprises will not adopt AI, so I think the market forces will push the right outcome. And everybody understands this now. The first question I get from any CTO, CIO, Chief AI officer I talk to is, how will I give feedback to this AI employee. How will I make sure they’re working according to my policies and my rules and not something that you know somebody else has coded in.

Kate 19:52
And do you have a good answer?

Surojit 19:55
Yeah. I mean, for us, those were the principles from the get-go. So it’s fully, you know, controllable. Our AI employees take human feedback continuously, can rapidly learn. And every AI employee, I always tell our customers, it’s a different type of software product. The traditional software product, you know, whether you buy it, I buy it, we have the same software, right? Unless you have upgraded and paid more or whatever. But it’s very different, say, I had an employee, you had an employee, maybe you have better training in your company, and that employee gets better faster, right? It’s similar, very similar for AI employees, because this is software that evolves, not a static software. It evolves based on the feedback that you will give, the data you will expose it to, the learnings it will have on the job. So the same AI employee in two customers may evolve very differently.

Kate 21:00
Right? So you probably see that, right? (I see that all the time.) I’m sure. That must tell you a lot about those companies. What’s going on there?

Surojit 21:13
There’s a level of education here. The level of education here is, this is a piece of software, a new type of software where depending on how much investment you will do to the software, in terms of your time giving feedback and so on, and aligning, giving it the right information, you can get more ROI. If you ignore it and let it just rot, you won’t get the ROI. And very simply, what we see most of the time, we’ll deploy an AI employee, and they will discover, Oh, I forgot to give a more this SOP, this document, because, just like a human, very smart, let’s say a very smart graduate from the top schools. They’re very smart, but that doesn’t mean they will do a great job in your company, because you have to give it the right information, context, onboarding document, access and so on. If you don’t do that, they won’t learn. And we see that all the time because there’s a lot of tribal knowledge in any enterprise. Sometimes people don’t write, humans don’t write things. Sometimes it’s in the team. Everybody kind of knows it by talking to each other, but the AI employee is not really calling up others and chit chatting and say, Hey, one day it can do that. (Laughing) That may get annoying. So it requires you to proactively give it access to the tools, to the documents, to the knowledge. And sometimes the knowledge is not written down, so then you have to write it down. So a lot of that knowledge discovery happens, which is good for the enterprise, because what happens in enterprises, some expert human that has been there for 10 years, 15 years, has a lot of knowledge in their head. They have not written it down, and one day they decide to leave to a competitor. And then you are screwed.

Kate 23:22
Yes, yeah, you’re describing two common pitfalls, like, there isn’t good onboarding, right, where you’re passing the tools and processes of procedures to new employees, which in this case would be the AI employee, and then also somebody who’s held all of that leaves and it leaves with them absolutely. So you’re showing that it kind of shows up in the same way.

Surojit 23:46
So 100% of the companies where we have deployed AI employees that have came back and gave us feedback, Oh, as a side effect, processes got better and got more documented, and we feel there is less less Enterprise Risk of critical talent leaving, because now all of those SOPs are encoded or written down, not in someone’s head.

Kate 24:10
Yeah, that’s so fascinating. Okay, I just looked at the time, and we’ve spoken the whole time about this, but it’s been fascinating. And I think by taking us through Ema and what you’ve built, which is so impressive, you’ve really given the listeners a lot of helpful information about AI and how it can be more successful and work on the enterprise side. It’s been fascinating talking to you. Two final questions, we always wrap the show up with this, any just general advice out there for other startup founders listening? They love when I ask successful founders, and I’m going to definitely put you in that category, that question.

Surojit 24:53
I have only maybe one advice. Stick to your guns. There is just, particularly in generative AI, too much noise out there. Too many people saying too many things. If you have conviction, I think any, for any startup, this is true. You need to, as a founder, you have to have strong conviction on your vision, unless you see real data points otherwise. But many people kind of have self doubt, and not that I didn’t have self doubt, right? I think some amount of self doubt is useful too, but I think having your inner conviction is very important, because otherwise you will, you’ll kind of keep changing things too quickly, and it does not help building great products. You need to stick with something for a long time. Great products are built with a lot of love and care and hard work. So you have to stick to your conviction. That’s one thing I would say.

Kate 26:06
That is excellent advice. I think, in this day and age, people give up and they listen to all the noise. There’s so much noise out there now. Thank you. Where can we go and learn more about Ema and what you’re doing? It was such a fascinating conversation.

Surojit
Ema.co.

Kate
Great. It was so fun talking to you. I personally learned a lot more about AI. I feel better. You made me feel better. I hope all the other founders in the space are like you and your team. So thank you very much for your time today,

Surojit 26:43
Absolutely. Thank you so much.

Intro 26:45
You’ve been listening to Startup Success to make sure you don’t miss out on future episodes. Subscribe to the show and your favorite podcast player like what you hear. Tap the number of stars you think the show deserves in Apple Podcasts. For more tools and resources for your own startup success, check out burklandassociates.com. Thank you so much for listening. Until next time.