#280 Can We Create A Universal AI Employee? with Surojit Chatterjee, CEO of Ema

Hi, I'm Maggie, VP of Curriculum at datacamp. And I know firsthand how overwhelming it can be to keep up with the rapid changes in data and AI. But staying ahead doesn't have to be hard. That's why our curriculum team has curated the best learning content to help you and your team stay ahead. Covering everything from data and AI literacy for busy executives, to cutting edge generative AI for developers, and everything in between. Whether you're upskilling yourself or your team, we make learning practical, engaging and flexible. It just works. Power your data and AI transformation journey on datacamp.com.

I think this is a transformation moment for the entire industry. It's like when computing came in. One day there'll be thousands. At least today we have like tens of agents. Like 30, 40 agents are not uncommon for a very complex task creating these AI employees that can mimic human roles. Agents not replacing humans, but agents making every employee a human like super employee.

Welcome to Data Framed. This is Richie. Regular listeners will know that AI agents are the most hyped tech for 2025. I'm very keen to learn what agents I should be building and how I should go about building them. But beyond that, there are some deeper questions around the scope of agents. Many of the immediate use cases for agents are for narrowly defined tasks, the repetitive things that humans get bored with. But there's also a vision for broad agents that can do the entire job of a human. This is clearly a very lucrative proposition, but it opens up questions like do they actually work? And which jobs might you want to replace with AI? And what do you do with the humans who were previously doing that job? Our guest is Surajit Chatterjee, the CEO and co founder at ema, an AI agent company with a mission to build a universal AI employee. Previously, Surajit was Chief Product Officer at cryptocurrency exchange platform Coinbase and VP of Product Management for Google Shopping. Well, at Google, he's also been a product manager at Flipkart and Symantec, and he was executive in residence at Andreessen Horowitz. So let's find out about AI agents and the future of work. Hi, Surajit. Welcome to the show.

Thank you. Very excited to be here.

Excellent. So people have been talking about AI agents being the next big thing for a few months now. Can you tell me what's the coolest example of an AI agent you've seen so far?

There are a lot of cool AI agents out there, a lot of claims. I'll give an example from maybe our own customer base we have these AI employees which are actually slightly more complicated than AI agents. They are complex collection of agents. But we have AI employees today working in, say, customer support, helping resolve 80% or over, sometimes 90% tickets for our customers at a proficiency level better than humans. So to me, that's pretty amazing and shows what's the potential here. We are seeing lot of applications in customer support, hr, employee experience, sales and marketing automation, and across the board horizontally in lots of different interesting use cases.

So that's pretty cool. Sort of 80 to 90% of customer support tickets being resolved by an agent. And you mentioned that it's also sort of HR marketing. Are there any other areas of business or any of the particular types of AI agents that you think are common?

Look, at this point, it's kind of proven that customer support is, it actually works. A lot of companies are using it. We are addressing a lot of that demand ourselves. And there are other companies, other competitors as well. We see wherever there is opportunity for eliminating repetitive tasks, tedious tasks, There is a role for AI agents today. And this task doesn't need to be repetitive in the sense of traditional, like RPAs type cask, like robotic automation. They can be repetitive in the sense of they may require some amount of human judgment, but they're the same similar task over and over again. And today these agents can reason, can make decision independently, can make judgments, may require human guidance, sometimes may require human approval, but all that seems to be increasingly manageable and tractable.

That's interesting. You're wanting the agents to have some kind of reasoning power beyond just being a sort of dumb chatbot. So we'll maybe get into like what sort of reasoning capabilities are possible later. But for now, I'm curious why you might want to create agents. So, I mean, it seems like obviously a bot is going to be cheaper than a human for a lot of cases. But is it purely about cost savings or are there other business benefits to making use of agents?

I think the first inspiration was, for many enterprises was cost saving. But increasingly we are seeing customers are coming to us for, hey, what insights can you derive from all my data? What kind of additional innovation your agents can do? So I think it will not be limited to cost saving. It'll be about creativity, innovation, building new things, finding insights that really benefit the business. We are seeing demand for all such applications. Of course, cost saving is somewhat easier to prove than what's the impact of an insight? And so on.

Okay, so I like the idea that you're wanting agents to be creative. Do you have any examples of what sort of creative tasks an agent might be able to solve?

Yeah. So creative within some bounds, of course you do not want particularly in the area we are in B2B applications enterprise. So you want creativity but within the guardrails of the company. And some examples will be like generating for example blog posts, generating writing responses, like creating a business proposal which is creative again within bounds. You won't manufacture things that don't exist. But it requires kind of creativity to answer specific questions maybe your customer is asking. So today we are Automating, responding to RFPs. Large top like Fortune 500 professional services companies are using us to automatically generate the first draft of the response to RFPs and then iterate and create a final draft with minimal human labor or human effort. Often these RFPs are now generated at F speed, 10 to 20 times faster than they were generated before. So this is pretty game changing. Like, think about it like most many companies make revenue by responding to RFPs and there is like a win rate, 25, 30% win rate. Maybe if you respond to more RFPS at a high quality, you may actually get more revenue. So a lot of these tasks are also the tasks that agents are doing are not just saving cost, but generating more revenue. We see same thing with customer support for example, consistently the support like automation actually gives higher CSAT than what your human agents would do. We have a customer honeyview, they're a large B2C lending company, 60 million users or so. They are seeing that normally their CSATs were like 40, 45% which in BFSI is pretty common. And that's going up to 75% or so. Not just automation. It's also better customer experience.

I love that idea of automating boring tasks by using AI agents. Are there any other particular sort of tasks that you think humans wouldn't want to do that you think are going to be taken away by AI like that?

Yeah, look, in the business today, people spend a lot of time just keeping the lights on. 50, 60% of time of average kind of employees are spent in just keeping the lights on, which is tedious tasks, repetitive tasks that could potentially be delegated to an AI agent or an AI employee. A lot of this task will involve using some kind of unstructured data to derive insights, using unstructured data to understand what actions to be taken. And that's what AI agents are able to do.

I like that idea of just anytime you got some sort of boring task or maybe just like you need that little bit of insight to then having an agent to do that for you. It saves you the trouble. Actually, on the subject of agents replacing humans, I think that's maybe like one of the bigger worries that people have is that the job is going to be at risk. Can you talk about like how this might play out from a career standpoint when you've got agents sort of possibly competing with you for parts of your job?

I think actually agents not replacing humans, but agents making every employee like super employee. It's like, will you have somebody come to work today and say, hey, I don't want to use computers, I'll just do everything on pen and paper. They will just not be very effective as an employee, even if they are very smart compared to other employees. Because, you know, why do you want to multiply two large numbers with pen and paper when a calculator can do it? It's the same thing. Why do you want to spend much of your work time on things that you don't like to do? These are kind of the soul crushing part of your job. You are actually better off that an AI employee colleague of yours does it under your supervision, under your instruction. So I think of this more like humans working with AI hand in hand rather than AI replacing humans. AI, just like every phase of every epoch of technology has done, it has just made humans more efficient, given freed up time for humans to be more creative and do what they want to do. I think we are going to enter that era now where a lot of the office work, particularly the repetitive part of the office work, can get done by employees. So humans are freed up to do more value added, more creative tasks. I mean, if you look at small businesses, the biggest challenge they have is not that they have too many employees and who don't know what to do, they have too few employees, they're overstretched. If you look at healthcare, for example, everybody is stretched and getting automation. Come in and take on the repetitive part of your job is a good thing. It's not replacing anyone. It's freeing up people to do what they want to do.

Definitely a big fan of having AI do things that I don't want to do. And I agree that a lot of teams across many companies, they're overstretched and they just need to be more productive and making use of any technology. It's going to be incredibly helpful. Certainly I think that's the case for like most data teams. There's just never enough people. I'm wondering whether it varies. Is there a tipping point like customer service where no company really wants to spend lots of money on this. And there is just lots of, there can be lots of AI agents for this. Is that going to play out differently do you think?

Look, I think in most of these roles the future will be that it's majority done by AI and with humans managing AI, humans giving instructions to AI. I also don't believe it will be a place where AI just runs over humans and just runs completely autonomously. But AI working AI employees or AI agents working under the guidance of human. And a lot of the jobs will shift to kind of more value added part of the business. For example, building or conceptualizing new products, figuring out strategy, maybe doing more actual selling and actually talking to the customer than filling out spreadsheets after a customer call and so on.

Okay, yeah, that's definitely true is that the admin after a call to a customer is obviously the tedious part that all the commercial teams forget to do and then you've got terrible data and it has knock on pack and yeah.

Everyone grumbles and then they would rather spend more time with customer. I mean this is just an example. And you can extend this example to other parts of the organization as well. If you're a salesperson, you want to spend more time with customers, you want to solve customers problem. That's what gives you energy, understand their problems, figure out innovative ways to bring a solution to their problems. But then they end up spending all the time in tools, writing out reports, filling out spreadsheets. Imagine if someone could just do that if every salesperson had an intern. Same is true for engineers, same is true for marketers. Customer success people everywhere you look, finance people. I would rather think of, okay, how do I plan the budget for next year? Rather than entering hundreds of numbers in a spreadsheet and spend hours on that, can that be automated in a way that's also highly accurate?

All right, this is going to really annoy that one person who loves data entry and the one person who loves writing reports. But maybe for everyone else it's net positive. All right, cool. So you mentioned the idea of a universal employee and I noticed that this is part of your company mission is to create a universal AI employee. So how broad can an AI employee be? I think a lot of agents are traditionally confined to very narrow tasks.

Yeah. So we think, and this is why we started this company, the vision of the company two years back when we started, right when ChatGPT was coming out is creating these AI employees that can mimic human roles and take on a lot of the tedious tasks that Humans do at workplace today. And I think there is really wide application. You can look at automating very large number of things today. And just to give some context, our AI employees are not just monolithic agents, they are a collection of agents. So some of the AI employees will have 10, 20, 30 different AI agents working together and orchestrated using our proprietary model. Some of those agents may also be procedural agents. Doesn't need to necessarily be AI agents. So it's a combination of different types of agents working together. And if you look at the scope today, we have customers in healthcare using us for automating, say and accelerating a lot of say, prior authorization for treatments and so on. I talked about sales and marketing, right? Writing automatically, writing business proposals, generating sales leads, reaching out to leads, scheduling a meeting. All that can be automated. And we have customers doing that. Customer support largely can be automated. HR and employee experience. A lot of day to day questions that employees will ask. Hr, right. That can all be automated in a way. Some chatbots, et cetera existed before, but now you can create automation that works very much like a human. So it'll have a fluid conversation with you, not just fixed conversation based on keywords and so on. It's not a frustrating conversation With IVR or whatever you have previous generation, we have like voice automation. Same thing you can call and talk to almost like a human agent. On the other side, AI mimics how human will speak in that role. I think the scope is very, very broad. Every week we find newer use cases working with our customers. I think this is a transformation moment for the entire industry. It's like when computing came in for the first time. Why could you use computers in the business? Some people said originally, oh, you can write some documents and maybe some spreadsheets. Those are still the most common applications. But then you suddenly realize, oh, it can be used anywhere, everywhere. And I think that's what will happen. It will also disrupt a lot of the existing stack and existing software that many companies are using.

Okay. It's fascinating stuff. I guess a lot to unpack there. So I like the idea that not all agents need to use AI. You can still have procedural agents, which is doing if some condition, then do that and you can mix these things together. Okay, so you mentioned the idea of using many agents together. I'm not sure what it's called. It's like a team of agents, swarm of agents, something like that.

Mesh of agents. But you can call it mesh of agents.

Yeah, okay, that's a good phrase to know. Mesh of agents. All right. In this mesh, like, how complex do they get? Are we talking like two or three agents together or is this like thousands? What's the deal? How much does it scale?

I mean, one day there'll be thousands. At least today we have like tens of agents, like 30, 40 agents are not uncommon for a very complex task. And just to give you some illustration, for example, you are writing a very long form document like a response to an rfp. So some sections of the RFL require say, building a chart or putting an image, and that's a separate agent. Building charts from data. Some sections may require going to the web and synthesizing some information and putting it there. That's a different agent. It knows how to go to the web, synthesize data, look for data, and creating concise text conforming to the style of the rest of the document. Now there are those agents, but then there are some meta agents that actually observe other agents, how they work. So there'll be some agent checking that the entire document is consistent and looks like it's been written by a single person. Not different tones and style and so on. Not a hodgepodge. There'll be some agent. We have also seen there's a very common agent design pattern called reflection. So instead of one agent doing your work just standalone, what if you have another agent observing the work of one agent and giving feedback and that dramatically improves accuracy. This is actually intuitively not very hard to grasp. This is how humans work. That's why you have auditors and somebody who checks the audit cross checks, and that increases accuracy. Even though both of these people may have same credentials. You know, in terms of software coding, people often do like peer coding, like two people coding together or somebody codes and somebody does a code review and that increases accuracy of the software. And then we have explored lots of different design patterns. This is one pattern. You can also have a pattern where there's one agent directing like six or seven or ten agents. Right? This is a common, like a manager versus employee design pattern. You can have multiple experts contributing together to build something, to come up with something. So we explore a lot of this agentic design patterns and we have created the engine that explores them kind of autonomously, automatically tries out many design patterns and figures out what will be the best for solving a problem. I think there is a lot of R and D to be done in this space and a lot of cool innovation that will happen. It's almost like, you know, understanding how organizational design patterns work. If in human kind of organizations, enterprises, different types of organization has different types of hierarchy or organizational design and something may work for certain type of organization. Let's say military is differently organized than a innovative, so software company. And there is, there is reasons for that. And we think a lot of that organizational design literature and studies are very useful to get insights on how agentic design patterns will work.

Okay, that's fascinating that you need to think about organizational hierarchy in order to make these meshes of agents work. And yeah, I can certainly see how in the real world like military hierarchy is very different from, I don't know, a year cooperative of artisans sort of thing. So yeah, can you talk me through what are the most common architectures then for these agent meshes?

The most common architecture would be like a chain of agents. Right. So one agent does some work and you know, hands it over to another agent and so on. That'll be the most simplest agent design pattern for almost all our AI employees. We use something lot more complex. Kind of reflection is a very common pattern we use where two agents or three agents are working together and observing each other's work and giving feedback. That improves accuracy quite a bit. There are, as I described in the space of how these agents can be organized is a very interesting space of research that we keep spending a lot of time on.

I'm curious as to what counts as a successful agent. I presume there must be some kind of accuracy measure for like how well does it perform. But yeah, talk me through. How do you know if your agent is any good or not?

It's a great question because in many cases the definition of accuracy is not very well established because we are sometimes doing generative work, not just work that has like a very fixed answer. So we often resort to human evaluation, like a group of subject matter experts will evaluate. It's almost like giving performance feedback to other humans. How do you know your marketing person is doing good work? You ask other people who are working with the marketing person. You have some okrs, et cetera, some performance metrics. But before beyond that, you also look at qualitative aspects, how good is the writing and so on. Similar principles apply. In some cases there is clear cut whether it's correct or not that can be established. In that case it's easier. In some cases, a procedural agent actually can evaluate an AI agent because the procedural agent can just check the final answer. Is it this or that? In most cases it's a human feedback based analysis.

Okay, so I suppose yeah, it depends what your agent's trying to do. But if you've got something where There's a concrete answer, relatively easy to check if it got the right answer. If it's a bit more fuzzy, like it's something writing a report, then you probably have to have a human read it and decide whether it was any good or not. It seems like most agents are going to need to have access to either commercial data or personal data in order to do their job. How can you make sure that this data is kept safe when you give it to the agent?

We use a ton of safeguards today inside our application. But first taking a step back, this is a big concern. Every CIO has, every CTO CIO has and if you are using agentic architecture, you need to make sure that your data is protected. A lot of your private data may be exposed to like public LLMs and you don't know whether they are learning from that data, they are using that data and so on. So you need, you want to really dig deeper. What we do is we actually automatically redact all pii like you know, personally identifiable information or phi personal health information automatically from anything we are sending to any, any public LLM and we send it, we use a, is a combination of models. We also make sure anything we are generating, we always copyright check to make sure our clients don't get into trouble, copyright issues and so forth. Everything that we generate we code sources and explainable and so on. But I think the data privacy is sacrosanct and you have to just understand how many agents, which agents is touching your data, how and what, what's happening with your private data.

Okay, so it sounds like a lot of it's just about being mindful of what data is going where and just being careful not to send it to send this commercial data to public AI.

Service design with privacy in mind. So if you are building an agentic architecture, it cannot be an afterthought. Okay, what will, where will my data go?

Okay, it does seem sensible to think about this at the start rather than at the end after you've built it and then you have to change anything. Do you have any other tips around keeping things private beyond just think about it upfront?

Other thing will be and which a lot of our customers are increasingly asking for is making sure everything is in your own cloud prem like air gapped cloud prem solution. So it's not talking to the Internet, not bringing anything from the Internet or not sending anything to the Internet. It's all in your private cloud. We do that all the time. Like you can get a, like a user open source for example model Install, deploy on your cloud or create a private endpoint with Azure, OpenAI or any of the other models. And make sure your deployment is not bringing in data from outside or sending data outside.

So I'm curious, what happens when you have a problem. At some point an agent's gonna make a mistake. And with a human employee, most companies have some sort of procedures to deal with mistakes. What happens when your agent employee does something wrong upfront?

You want to put safeguards like what kind of mistakes will you allow? What kind of mistakes will you not allow? For example, to make it real, many of our clients would say okay for sensitive categories, for customer support, let the AI employee create a draft response, not actually send the response, let it be queued to a human employee. Let the human do a double check and verify, accept and then send. That's very explicitly figured out like what is your tolerance? And that may be like only 10% of the tickets or 5% of the tickets. Rest 85, 90% you are like okay, I'm fine. For AI agent to automatically respond predict checks just like human employee, this is a new type of software. You have to kind of do spot checks. You have to make sure they're not going off the rail and doing the right thing. In the end, humans also make mistakes. So there are ways you will recover from human mistakes and you have to treat it the same way. If there are mistakes made, there is, in course there are ways to recover from the mistake. We are honestly entering a new realm here, a new type of software, new type of automation, new technology and we learn a lot in coming years. I think the nearest example is like self driving car today. I think it's better than most humans, but it's also not 100% probably accurate. It'll once in a while make a mistake and you can take over. And humans can take over very easily. And you have to build your software that way, build your agentic architecture that way so humans can take over easily.

So I like this several step process there. So you're building guardrails into the agent to try and ensure high quality upfront or at least make it fail gracefully if it can't solve the problem. And then you want some kind of observability or monitoring to make sure can assess whether it how often it's getting the right answer. And then you have that sort of fallback of well if it does something wrong then let's swap in a human as fast as possible. I'm not sure whether this is an equivalent of an AI getting fired. I Guess like if a human does something terribly wrong, that's what happens, I suppose. Yeah, the agent just gets turned off, it just doesn't get used anymore. So that's probably about the equivalent. So we're talking about processes of making good AI agents. I'm sure that once you introduce AI agents, it's going to disrupt a lot of your other business processes. So do you have any advice on the change management side of this? Like how do you go about updating your processes to accommodate for AI agents?

What we see today is most companies actually have very poorly written like SOPs or the SOPs they have written like processes they have documented are all sale. Very few companies kind of keep them all updated. This is an opportunity. Now AI employees or AI agents will go by what they can understand and learn from what's there. They cannot go inside someone's head and understand, okay, how do humans do this? It's really important to have good documentation and keep updating the documentation. One good side effect we see of deploying AI agents is suddenly the organization wakes up and says, oh, actually Most of our SOPs are stale and it's all in people's head. And it's like tribal knowledge. They're like these three people who have this knowledge and what if they are hit by a bus or something? Then the organization will be in crisis. Right. So it forces them to write down things, understand the processes. But the other interesting part here is the AI agents are learning your process and, and adapting to it. Just very different than say SaaS applications or previous generation of software where humans had to change their workflow to adapt to a software. This is how software works. So you better change your workflow. So it's not so much about companies changing their processes, more about company documenting their processes and also understanding their processes very clearly. Sometimes they don't have clear understanding or processes have just grown kind of automatically, organically.

That's really interesting that you need to be able to describe your processes in order to be able to teach an agent to do so or to design an agent that mimics them. And it's wild how much like how many things you do on a day to day basis that is just entirely in your head rather than being documented somewhere. Maybe that's just me, maybe I just need to write down what I do more. But I think maybe a common problem. Okay, so actually I read a really nice blog post by Julia Wynne on saying I built this fantastic AI agent, but it took me a whole day to do to replace my existing process to Build this thing. So it seems like there's an upfront cost to creating agents in order to get the long term productivity boost. Do you have any advice on how you go about managing that initial hurdle?

Yeah, I think there is upfront cost if you want to code up an agent. I think one day is actually still very good. Many of our customers will say they have been trying for nine months or a year to make the agent because it's really task for experts. I don't think anybody can go and make it all work. And that's where our company, Emma, what we have done is you can build these AI employees just with the conversation. So we built a model from a conversation. We can build the entire workflow, we can recruit the agents, figure out the right agent design pattern automatically generate the orchestration code and so on. So we make it really, really simple. And a business user can use it and can create an agent like a new AI employee within a few minutes. But in most cases it takes a lot longer. Is the investment worth it? I think it's, you know, even if it takes long, it's still worth it because you are going to save tens of thousands of hours of human labor post that.

It seems like there's a big difference in like sometimes you can create something in five minutes, sometimes it's going to take months. So you have to be a bit picky about which use cases you go for then. Do you have a sense of what's simple to build and what the time consuming difficult bits are?

Yeah, I think building is not the most difficult part. Making it work is the difficult part. You can probably build something quickly, but will it work at the level of proficiency that you expect from your humans in that role? And that takes time. Again, using the right technology will cut down the time dramatically. If you're building on your own, it may take for a long time unless you are really an expert, like a PhD in machine learning or something. But if you choose a tool, the right tool, it may be way faster. And that's kind of what we have been engaged in. So democratizing use of AI agents that anybody can use it, anybody can build one and it doesn't take like tens of hours, it's just few minutes. And then making the AI employee actually efficient and useful. Again, the approach we have is you just give feedback. Just like you'll give feedback to a human employee, natural language feedback. It absorbs the feedback and retrains automatically.

So it seems like maybe you're going to have to have some sort of iterative Process where you build something and try and use it and then you get feedback on where it works, where it doesn't, and just keep adding to it. So it's that sort of agile flow of continually improving it.

Absolutely. It's like a human employee. You hire a great person, but they're great on paper. Will they be great in the role in your company? It requires a little bit of work, right? You need to give some feedback. You need to make sure they're, they're mentored. Well, they have all the information and so on. And once they have ramped up, well, they will do great. Okay.

So yeah, I like that. Even agents have to do some training and upskill themselves. It's not just humans. All right, so you mentioned two very different scenarios. So one is you've got a business person creating an agent and another one is like, oh, I've got a PhD in analytics and they're creating agent. So I guess in each of these cases, what sort of skills do you need in order to make agents work?

In the case where you have like a PhD in machine learning or computer science, you basically need to just understand you are building from scratch, maybe using some APIs available somewhere. The skill is actually today not widely available. Not that many engineers are very good at it, but the basic level of prompt engineering. But you also need to understand do you need to build a small model? Do I need to fine tune a model? What kind of agentic design pattern will be useful? You have to experiment a lot and so on. So that's on that bucket. If you want to build on your own, it will require you to from the curating your data, understanding data, lot of data engineering skills, to training models, building small models, maybe training system skills, agentic building agents and making agents work together, write orchestration code and so forth. On the other side of it, if you are using a tool like ours, the skill set required is minimal. It's basically can you communicate with natural language in the business, so you will give feedback to Emma and it's similar to how you will give feedback to a human colleague or another human employee. So really minimal skill set required. In fact, we want you to focus more on the business side of it, so the business knowledge can be incorporated into the AI employee very easily.

Okay, so just to wrap up, tell me what you're most excited about in the world of AI agents.

Look, I think the future will be we'll have personal agents. I mean I'm very excited about kind of next generation of devices that may come in mobile phones may have a bunch of agents that make my life simpler, maybe booking travels, managing my calendar, all kinds of things that will happen in the space we play. I think it's really exciting to see opportunity. There's so much inefficiency and so much opportunity to address that inefficiency with their agents. Next few years, it's really game changing. I think we will see real kind of boost in productivity across the board in economies that adopt this technology faster. This will be like the next Internet or next computing or maybe kind of almost the next Industrial Revolution.

Okay, that's big talk. It's like the next Industrial revolution. So. Yeah, well, yeah, certainly exciting stuff. And that's an interesting idea of incorporating agents into things like mobile phones. So everyone's got them on hand at all times. Super. Thank you so much for your time. Surajit.

Thank you so much.