Reengineering Work with Intelligent Agents - with Debanjan Saha of DataRobot

This is Daniel Faggella, you're listening to an episode from the AI infrastructure podcast. I know you're thinking, isn't this the AI in business podcast? And you'd be correct, uh, but we've just launched a new show called the AI infrastructure podcast. We have had such a massive audience growth around everything related to infrastructure. We're talking about hardware, data platforms, governance and compliance, enterprise-wide concerns that go beyond any one industry, that we've launched a whole new show called the AI infrastructure podcast.

Our guest this week, and again, this episode is from the AI infrastructure podcast, which I'll talk about in a moment, is the CEO of DataRobot. DataRobot was one of the early AI unicorns more than a decade ago now, based originally in Boston. They have team members around the world. Their CEO, Debanjan, is in the Bay Area. We speak this week about re-engineering work in the era of intelligent agents. Basically, we talk with Debanjan directly about what the mix of man and machine is going to be in the future of the enterprise.

What is it going to look like to have digital employees who are working side by side with humans? And how should departmental leaders think about building departments, not just with people, but with kind of virtual people? The analogy is actually quite apt. He talks about all kinds of analogies between human and AI workers in terms of training time and H-R related considerations. Weirdly enough, there's a lot of overlap with how people think now, but it definitely needs to be leveled up.

And Debanjan gives us some interesting perspective. Again, this episode is from the AI infrastructure podcast. If you are interested in hardware, governance, compliance, and the broader infrastructure implications of AI in the enterprise, and you want to see what other infra and AI leaders in the Fortune 500 are tuned into, be sure to check out the AI infrastructure podcast. You can go to EMERJ dot com slash INF1. That's I-N-F like infrastructure, E-M-E-R-J dot com slash I-N-F, and then the number one, E-M-E-R-J dot com slash I-N-F1, and you can go directly, it'll take you directly to the Apple page for the AI infrastructure podcast.

We've got amazingly big name guests, a ton of C-suite from Google Cloud to Nokia to AR. M. You name it. I mean, a lot of major names speaking specifically on infra, so be sure to check out that show. Without further ado, let's fly into this episode with Debanjan, CEO of DataRobot, here on the AI in business podcast. So Debanjan, welcome to the show.

Thank you, Dan. Glad to be here. Looking forward to it.

Absolutely cool to cool to catch up with you folks again. Lots of evolution, the AI ecosystem, DataRobot, no exception. Today I want to knuckle down on the topic of kind of digital employee. Seen it in some of your folks work. Definitely a phrase going around the enterprise world. We'll talk about examples, but I want to begin maybe with definitions. When you're speaking to the C-suite and explaining what this paradigm is about and how people should be thinking about it, how do you like to break it?

So, you know, there is, of course, a wide variety of definition of what an agent is and what a digital employee is. I mean, some people call it the silicon agent versus the carbon agent, which is the human employees, right? I put them based on the functionalities and what they do in, uh, three buckets. Number one, which we see a lot of it, is personal productivity agents. I mean, these are people who are your co-pilots and they help you with your personal productivity, whether it is email or whether it is running some script, doing some odd jobs that you need to do every day, etcetera, right?

And there are plenty of examples of that. I mean, co-pilot being probably the most prominent one. The second one I would call agents, which focus on, I would call line of business specific activities. These are agents which are in Salesforce, in ServiceNow, in Workday, a and in many ways, they are essentially making those application a lot more user-friendly, right? I mean, you have a chat interface, you can talk to your agent rather than point and click or open a screen, etcetera, etcetera.

A and what we focus on, what I call more serious enterprise agent. These are highly valuable agents which are doing sometimes the back office work, and sometimes they are doing work that we cannot do without them, right? I'll give you a couple of examples of that. So, one of the problems that we see in enterprises, what I call the coordination tax. Lot of enterprises have these applications in various different silos, whether it is H-R or operations or finance or engineering or logistics, etcetera.

And oftentimes, the business processes cut across these silos, and only way to bring them together is to essentially have what I call human middleware, who are doing all of this coordination and meetings and running things in the back office. Without them, you cannot really execute things in an enterprise. For example, let's say you are in in a CPG industry and there is a port closure somewhere, and you have to figure out that what are the customers who are going to be affected?

How are you going to reroute some of those supplies? What is going to be the impact on your logistics team? What's going to be impact on your revenue? So the finance team has a heads-up on that. All of these things happen through various different meetings, email and various different coordination. And with agentic AI and with your digital agents, you can actually automate a large part of it. You cannot, it's hard to automate 100% of it, but you can automate a large part of it.

And that makes it not only very, very efficient, it also improves, for example, the time it takes to do some of this work, right? I mean, it doesn't take, you know, if it is agentic, it's working twenty four seven. If it is agentic, you can get everybody together in the same call even without the humans being there, and and that facilitates this process and automates this process to the extent possible. So that is one type of agent we are focusing on.

Another type of agent, and I call them essentially reimagining or re-engineering the back office processes, whether it is in operations or it is in supply chain, whatever have you. The another type of agent, which I think I would call them the super agents or agents which have superpowers that we typically don't have in the enterprises today, right? I mean, you know, we recently did, uh, you know, one of those agents with one of our, uh, oil and gas customer, Chevron.

It's actually, uh, in the public where what they do, let's say there is a there is an oil field and or, you know, refinery, and there are a lot of pressurized gas and oil pipelines there. And oftentimes, if you have pressurized pipelines, there are leaks, and in order to manage the safety of it, you have to send humans to go and inspect them. Now, nobody wants to do that, so they have a set of drones and a and digital dogs that they send in order to sniff out where there may be a leak.

Now, if there is a leak and they go too close to it, they have lithium ion batteries and it can blow up. So, what they really want to figure out using digital simulation. So, if there is a sensor which is saying there is a leak, they want to do a digital simulation, almost like a digital twin, and figure out the high risk areas and no-fly zone so that they don't send the drones there. This is something that you cannot do today, and agents actually can help do that.

So, that is the way I can sorry for the long answer, but hopefully

No, no, no, this is fine, this is fine. This sets the table for us to unpack. So let me kind of poke in here. You know, you mentioned the kind the kind of productivity stuff and the things that are kind of built into established platforms. And I think some of our audience will be familiar with ServiceNow and others, and we've we've had some of those folks on. But, uh, you know, you're talking at this broader enterprise scale, things start to open up to new sets of capabilities.

W when I'm when I'm hearing you describe just these two use cases, for example, like you're you're mentioning the C-P-G example for supply chain. So there's there's obviously so much complexity, and since COVID, I think a lot of people that didn't care about supply chain complexity all of a sudden started caring very much. And the the, uh, funding into that space, uh, went up pretty significantly. So kind of like, uh, predicting, forecasting, vendor intelligence, shipments, ports, you know, that whole nine yards.

So the the I can there's almost a workflow called look at the main junctures where our supplies come through, whether it be land, sea, and air, and and then figure out which impingements might be affecting, which impingements are happening, and then which impingements might affect our supply chains, and then update sort of our own forecasting and kind of predictions based on that. That that's sort of like it's a a workflow. And like you had said, some of it is stuck in email, some of it is in other platforms.

Maybe you can't automate all of it, but you can have sort of a a workflow happening for that specific thing. Um, the same thing in this case with the predictive maintenance, we've had like drone deploy and that whole ecosystem. So they've got the they've got the dogs now and they get the flying stuff. And and so, you know, you're talking in in one of your clients in energy, yeah, they have some kind of a an issue.

They want to make sure they kind of know where can we send these things, where can we not send these things? So the the path finding maybe can be defined in some deliberate way. Again, it's it's a workflow. Under these circumstances, some human would normally need to do this. Now we could maybe have an AI kind of do this. I think when some people hear digital employee, they're thinking it's almost like gonna do sort of like a job.

So a job might be some insurance underwriting job, or it might be a customer success job, just handling text for like a text mess SMS type stuff, or it might be, you know, like like an entire role where, like, you know, they sit down metaphorically in their data center and sort of do the full workload of a person. What you seem to be articulating in this case is more of specific workflow junctures. In other words, monitoring of potential outages for supply chain stuff or impingements of supply chain, uh, monitoring of cer- certain maintenance protocols for physical energy.

I am I picking up what you're putting down? Cause I I'm seeing more workflow rather than full kind of digital job per se. But where do you draw the distinction?

Well, I mean, I I don't I. It's kind of, uh, I would call a spectrum, right? I mean, you know, uh, in the workflows, I mean, ultimately, everything is a business process that people are executing. Some of that is done by humans, some of that is done by workflows, some of that is done within an application, right? And the things that we talked about, you can think of them as digital employees. I mean, people are doing that instead of people doing that, you are having AI do that for you.

The difference is between an application and AI doing it, AI is also reasoning. It's coming up with different ways of, you know, taking care of things which a typical workflow, for example, an RPA system won't be able to do, right? So I don't kind of distinguish them that deeply, but I mean, you know, we do think of these agents very similar to human in terms of their life cycle, right? So, for example, what do you do with a human employee?

You interview them, you evaluate them, you hire them, you then onboard them, and then you performance manage them, and then, you know, sometimes they leave and you decommission them, right? And if you map

So this is the employee analogy is very direct there, right?

that's right. A and we do a we do lot of that. I mean, for example, you know, before you can put an agent to any of this workflow task, you have to make sure that they can do that job. They can do that accurately, they are abiding by various different policies, etcetera, etcetera. So, there is an evaluation process that you go through before you can onboard them. Now, onboarding is essentially you have to integrate them with your enterprise infrastructure.

For example, we are working with a customer where we are integrating these agents with Octa, so that every agent has an identity and you can manage them like your human employees, and you can control, for example, the access privileges they have from the identity management system, right? Now, these agents are also doing their job, and we track them, trace them, every action that they take, we trace them and monitor them. So that if they make a mistake, you can actually go and correct that, and the next time they do the job, it's actually corrected.

A and you know, sometime it also happens that you have to attribute what was done by the human, what was done by the agent, either autonomously or on behalf of that human. So, that traceability becomes very, very important. And then when, you know, I'm kind of fast forwarding a little bit, when you decommission an agent, like the way when a human employee leaves, somebody has to pick up that work, that context, that memory, and you have to do that, right?

Uh, so so th- this this does have a lot of direct analogies with

That's right.

employees, so that that term, I think, very much fits in that regard. You know, like you had said, the very deliberate process of, you know, how do we sit them in, what are they trained on initially, who's over their shoulder for the first little while, you know, who's checking in periodically to do that kind of RLF or what have you, to sort of, eh, get them on track for detecting fraud or predicting inventory or whatever other things, uh, they're they're working on.

And then for the decommissioning, yes, is whether it be man or machine, who's gonna pick this up? You know, if normally this was handled, these kind of instances were handled by this kind of an agent. Can someone in another department do it? Can we train another AI on it? Um, okay, so th- that that analogy clicks. And I guess as you l- as you look forward into, you know, the next year ahead, I mean, things are moving way faster than they were four or five years ago, right?

It's an incredible and there were so many barriers to enterprise adoption that don't feel like they're there anymore. Um, and and there's so many ways to, you know, get an agent up to speed with no technical background, whether it's a marketing task or a supply chain task or what have you. A as as you're seeing enterprises plan their future here, how are we gonna what's our customer success department gonna look like? Um, what is our fraud department?

You guys do a bunch in financial services, that's a lot of our audience. So fraud, anti-money laundering, there's all kinds of workflows there. Whatever it may be, as they think about a year from now, what's the composition of person and machine with kind of digital employees as part of the mix? H How are you seeing enterprises think and plan about this, and how do you hope they think and plan? Because it does feel like work is fundamentally changing.

Like, what is the state of affairs and what do you think the smart executives should be doing?

Yeah, I mean, y- you know, as I said, I mean, I I see three different types of, uh, modalities in terms of how they are thinking about the employee, uh, th- thinking about the digital employees, right? Uh, part of that are the co-pilots. I mean, everybody is probably going to have a set of co-pilots. They are going to be smarter, faster, can do lot of their things using the co-pilots. I mean, that's kind of, I would call table stakes.

Pretty much every company will have those, and and probably you're going to see lot of them. The second thing is the line of business, the functional agents, I mean, which used to be applications, are going to be more agentic. I think that also is going to be pretty common. I don't know how long that's going to, uh, sustain. But the real change is not going to come from this agent essentially doing what humans are doing today.

It's going to come from how the back-end processes are going to be re-engineered and restructured so that work gets very different from what it is today. I think that's where most of the value is going to come from, and lot of people are thinking about it that way, and that's where we primarily focus on. A and last one, uh, what are the kind of things that we will be able to do that we cannot do today, right?

So, I mean, you know, for example, think about insurance claim processing. Today, you have customers who will take their camera and then, uh, go and take a video of their car if they're in an accident and upload that, and lot of that is going to be automated. And right, I mean, you can do video analytics really, really well, and then you can match it with the policies, and then you can figure out what the damage is.

You can go and schedule a repair job, and none of this needs human in the back end of it, except for some of the difficult cases where you have to do the exception handling. A and I think the where what I'm mostly excited about is number three and number four, and that is thinking about work in a very, very different way, and it's going to be both efficient and improve productivity as well as it's going to be transformative.

Yeah, when we think about work getting done fundamentally differently, I I I think from my enterprise

That is true.

experience, you know, we talk to the cutting edge vendors and we talk to the people adopting the tech. It's it's a little bit of a different world sometimes, but, you know, it's it's incremental for some of the big companies some of the times. Um, you know, where it's like, oh, you know, reimagining customer success, okay, what if our chat instances could be handled entirely by AI? You know, that that's useful, but it's not quite reinventing. I think what what you're talking about, what what can fundamentally be done differently in ways that are just wholly more efficient than we're operating now.

So there might be entire paperwork workflows that involve checking and double checking of different form fields and whatever the case may be, that may be between intake and between back-end processing. We don't have to go to a BPO, we don't have to have it signed off here, we don't have to have it signed off there.

That is happening already, by the way. I mean, we we want

It it totally no, of course, of course, of course. And the BPOs are all trying to pivot it at one time, um, and and we're so are the services firms we work with that whole crowd. But, like, you know, some of those workflow changes in like mortgage, etcetera, if they come to fruition, will be fundamentally different than how mortgage was ever done before, or or underwriting was ever done before. Um, so there there's like bigger shifts.

When you said you're excited about kind of that bigger stuff, I would say same here. I think that's what's going to be transformative and most valuable. When you think about maybe the categories of things that'll be entirely different, if we wake up in two years and we look at how enterprises are running, and you use your imagination of, like, what you just think is most gonna be fundamentally different. Are there classes of use cases, workflows, areas of business that you think are illustrative of the changes that are coming?

Oh, definitely. I mean, and I will, you know, categorize them, I would say, in different types of capabilities that is going to be foundational to this, right? If you think about any enterprise workflow, which includes human and digital employees, right? I mean, there is a multimodal perception that comes first, right? Whatever you are trying to do, there is a perceptive step, whether it is kind of scanning the form or reviewing the videos or figuring out from from images whether th- there was some difference between what you are expecting versus what came out of the production line.

That is the first step, right? The second step after that, a and that is changing foundationally. I mean, the form processing has been around for a while. I am seeing more and more of video and image processing and various different sensor data processing, which is going to be foundational to this. The second step after that is that, you know, you are thinking about with that input, what are the things that change and what is the impact of that in the process that, uh, you need to do differently?

This is almost like a digital simulation. There are different options that you need to take into consideration. For example, let's take supply chain, for example. You have picked up some signals that there is a hurricane coming, and the shipping line is probably have to reroute or a port is going to close, right? You take that input, and then you feed that into your digital simulator to figure out what does that mean, and what are various different other paths that I can take.

Then the third step after that, that after you have simulated all of this, you have to evaluate all of this, whether it is forecasting or some other kind of reading that you have to go through and come up with different options that you can execute. And once you have done that, then you have to pick one of those options. In some cases, you probably need a human to be in the loop and approve that and go and execute that.

Now, if you take these foundational building blocks and apply it to various different vertical use cases, they are going to be done in a very, very different way, which is what we are seeing. I mean, we work with, for example, energy companies, we work with insurance companies, we work with manufacturers, and they are all rethinking how they can, first of all, use lot of these human steps from their processes, and how they are going to eliminate this, uh, coordination tax, which is what takes most of the time and probably is most difficult to do and more error-prone.

Take it out of their system, right? And the direct impact of that, I mean, this is where there is always this discussion that if the digital employees come, is that going to eliminate lot of human positions? And which is a fair topic. I I do think it's going to have disruption, but I also think it's going to create very different kind of, you know, outcome people like to have. I mean, one of the things which is near and dear to me is hyper personalization, pretty much in every industry that I can think of.

If you think about our medicine, it is very, very crude. I mean, we kind of come up with some chemical compounds, try it on thousand people, and then apply it to 8 billion people on on earth, right? I mean, making it very, very custom and personalized could be a huge, huge difference for the health and well-being of people, right?

it's a big challenge. Personalized medicine has been a the Lord knows how many billions are going behind that. But but yes, I mean, obviously, if we can, from cancer treatments to a simpler things, it it clearly it's the direction things are moving in. So I'm with you there. I think it's somewhat inevitable and there's a lot of possibility. Yeah,

That's right. You know, if you think about I, I also think even entertainment can be very personalized. You can tell Netflix, this is what I am in the mood for, and they will cook something up for you and show it to you, right? And so I pretty much every discipline. I I think there are a lot of these new kind of opportunities I see. I mean, one part is, of course, re-engineering and re-imagination. Second part is what are kind of the superpower these agents are going to have that humans don't today, and how is that going to do things in a very different way?

And third is the what are the different ultimate outcome that we really care about, that's going to be very, very different and something that, uh, we have been aspiring for for a very long time, actually will become reality.

I would concur. I think when people think personalization, a lot of the time it's like, uh, they think e-commerce or maybe they think media. But to your point, sort of anything that's not now personalized seems quite likely it will be if if personalized ends up meaning simpler or more pleasant, out will outlandishly be adapting the more personalized things. And that is a fundamental shift in many industries. Let me end on this note here because the fundamental change, I think there's a lot of excitement and even on some level, some some pressure being driven down to, like, hey, we've gotta level up and and make sure we don't get our lunch eaten by people that are adopting faster than us.

At the same time, in many of these industries, you mentioned manufacturing, fin serve, you know, health, you know, they they've got kind of the governance considerations that are pretty significant. You know, if if if, uh, Amazon recommends me the wrong book, you know, who cares? But, you know, if something happens with my personal savings or with with information about my diagnostics or something, you know, that that becomes more serious. What are you seeing as we shift into a world where digital employees are very much part of the mix?

Um, how are you guys thinking about kind of governance as part of the beginning of that conversation? What do you think C-level leadership should be considering as they enter the paradigm we're entering here when it comes to governance in those regulated spaces?

Yeah, I I I think foundational to this, by the way, is accountability and attribution. If you have a hybrid workforce, even today, the thing that you really care about in a governance process that who is doing what? Do we have the right people with the right authority to do the things that they are supposed to do? And if something goes wrong, who is going to be accountable, right? And I think that foundational paradigm needs to be extended to a hybrid workforce, which means that sometime people are going to take action, sometime agents on behalf of the people are going to take action, and sometime agents autonomously are going to take actions, right?

And you have to essen- you have to trace and monitor all of these activities and, uh, have a system where they can attribute who did what in a particular workflow, right? That is very, very foundational, and that is, I think, one of the things that we are working on. Second thing that you want to do is that there are a set up, what I would call policies in a hierarchical way that these agents need to know, and we have to hold them accountable to it, right?

So, for example, that no PII information should come in or go out, simple things. And then there are domain-specific and jurisdiction-specific policies that the agents need to know, and sometimes they have to produce artifacts which can maintain an audit trail that you can go and take a look at, right? So, that is a foundational thing that we are looking at. And in some cases, it's not just about having all of these guardrails. You might have to do real-time intervention in order to stop things from happening that can, you know, make things really wrong, right?

And there has to be, for example, kill switches and things that you can use. One of the things we have done with one of our customers I was talking about is using Okta as an agent identity management system. That is one place from where you can control the access privileges an agent can have to various different internal information, and you can completely shut it down if you want to from that one place, right? So, those are the kind of guardrails and activation mechanism they're putting in the system so that we can at least match the governance processes that you have for human employees.

Yeah, again, another analogy to humans here as we as we wrap up, I mean, the nutshell I'll put together, and you can tell me, Debanjan, if there's anything I'm missing, is it sounds like kind of if we're looking at, you know, a functional department like fraud or anti-money laundering or whatever, wherever it is in the business, or or maybe even a C-suite conversation, and we're kind of mapping out where digital employee futures are sort of seem to be shifting what we're up to.

There there's thinking to be done around, you know, what do they have access to, not access to, but also how do we toggle or cut that access per our our own requests and demand? You know, does someone have to go in and kind of like, like, How would we do wow, who's who's in control of this thing anyway, right? Ha- have those stop gaps the same way you would that's right, with a person. Yeah, yeah, and and the same thing with accountability.

How would you keep a human accountable to know who did what when to be able to know what to correct

That's right.

or when something has really been a violation? Same thing on the planning process. So the there's we're almost able to go less from, okay, someone's gotta write all the crazy code to sort of like, well, how would we deal with this if a person was doing it? And let's plan that up front. Is this the right way to think about it? Would you frame it differently?

I I think this is exactly the right way of thinking about it. Of course, you know, the way the humans are going to interact with each other and the agents are going to interact with each other, that's probably going to be different. There is also another world where the humans are humans and the agents are going to work together. They will collaborate, they will negotiate. A and, you know, but that we are not quite used to.

How do you negotiate negotiate with a with an agent?

Yeah, th- th- that'll have to be a topic for its own episode, I think

That's right.

here, Debanjan. But you're you're you're setting us up for challenges that have yet to come of, uh, a agent to agent negotiation, human to agent negotiation. Luckily for at least right now, we're steering the ship, but I have a feeling in the next eighteen months, we'll have to do another one

That's right.

on that theme. So, Debanjan, it it's been a real pleasure having you on the show. Thanks so much for sharing your insights with us.

Of course.

So that's all for this episode of the AI in Business podcast. A big thank you to you for tuning in and thank you to Debanjan for being with us. As mentioned, this episode was taken directly from the AI infrastructure podcast. So we've aired it for a second time here on AI in business because we have so many of you here, and the show has been around now for more than a decade. It is crazy to think that.

A million listeners a year on this show, but as mentioned, Yeah, because we have so many new listeners concerned with the enterprise-wide concerns of infrastructure, this is everything from data centers to hardware to data storage, the new foundations that enterprises are sitting on, that we've started a new show. So, the AI infrastructure podcast can be found at EMERJ dot com slash INF1. That's I-N-F like infrastructure, and then the number one. It's been forever since we've launched a new show.

It's been years, actually. So I'm sincerely hoping that you'll enjoy this one. Go check it out on Apple. When you follow that link I just gave you, EMERJ dot com slash INF1, you'll go directly to the Apple page for the AI infra podcast. And be sure to give us your thoughts. Feel free to reply via email, drop us a review there on Apple itself. This is very much in response to our audience, that is to say, you, that we've opened up this new show, and I hope you like it.

So, I I also hope you've enjoyed our conversation with Debanjan. I certainly had fun chatting with him. I look forward to catching you here the next time on the AI in Business podcast.