#600: AI Agents in Marketing with Raj Rikhy, Microsoft

Welcome to today’s episode where going to talk about AI agents as well as the intersection of artificial intelligence and marketing with Raj Rikhy, Principal Product Manager at Microsoft. We’ll explore the functionalities and strategic uses of AI agents and how marketers can leverage them to enhance their initiatives.

About Raj Rikhy

As a Principal Product Manager at Microsoft, I have over six years of experience in developing and delivering innovative products that utilize Generative AI technologies, including Language Models (LLMs), to enhance data science and engineering capabilities for Microsoft Fabric. My mission is to empower customers and partners with cutting-edge AI solutions that solve complex and high-impact problems across various domains and industries.

I collaborate with engineering teams, data scientists, customers, and partners to identify customer needs and market opportunities, define product requirements and roadmaps, and manage the entire product life cycle, from conception to launch. Previously, I was a Principal Product Manager at Microsoft Project Bonsai, a cloud-based deep reinforcement learning platform for industrial control systems. I also have a strong background in Data Science and Deep Learning, having worked as a Group Technical Product Manager for the Global Chief Data Office at IBM, where I enabled the end-to-end user experience for Data Scientists and Data Engineers, and scaled the adoption of distributed deep learning frameworks and tools.

Resources

Microsoft website: https://www.microsoft.com

Wix Studio is the ultimate web platform for creative, fast-paced teams at agencies and enterprises—with smart design tools, flexible dev capabilities, full-stack business solutions, multi-site management, advanced AI and fully managed infrastructure. https://www.wix.com/studio

Register now for HumanX 2025. This AI-focused event which brings some of the most forward-thinking minds in technology together. Register now with the code “HX25p_tab” for $250 off the regular price.

Connect with Greg on LinkedIn: https://www.linkedin.com/in/gregkihlstrom

Don’t miss a thing: get the latest episodes, sign up for our newsletter and more: https://www.theagilebrand.show

Check out The Agile Brand Guide website with articles, insights, and Martechipedia, the wiki for marketing technology: https://www.agilebrandguide.com

The Agile Brand podcast is brought to you by TEKsystems. Learn more here: https://www.teksystems.com/versionnextnow

The Agile Brand is produced by Missing Link—a Latina-owned strategy-driven, creatively fueled production co-op. From ideation to creation, they craft human connections through intelligent, engaging and informative content. https://www.missinglink.company

Transcript

Greg Kihlstrom:
Today, we’re going to talk about AI agents and the intersection of artificial intelligence and marketing with Raj Ricky, Principal Product Manager at Microsoft. We’re going to explore the functionalities and strategic uses of AI agents and how marketers can leverage them to enhance their initiatives. Raj, welcome to the show.

Raj Rikhy: Hey, Greg, thanks for having me. Wonderful to be here.

Greg Kihlstrom: Yeah, looking forward to talking about this with you. So why don’t we get started first with you giving a little background on yourself and your role at Microsoft.

Raj Rikhy: Yeah, totally. Been at Microsoft for about five years now, currently working on a bunch of really interesting stuff with data, data engineering, data science. Spent several years at Microsoft working on deep reinforcement learning, which has a lot of context around agents, but that’s a story for another time. And yeah, had a smattering of experience at other companies, IBM, Salesforce, and the like, also working on AI.

Greg Kihlstrom: Great, great. So yeah, you’re the right person to talk with about this topic. So yeah, we’re going to touch on a few things here, but I wanted to start with this concept of AI agents and applying it to marketing. So first, why don’t we start with a definition here. Again, lots of Lots of people listening to this are using AI in some capacities, but what exactly are AI agents and how do they differ from some of the other AI tools that might be used in marketing now?

Raj Rikhy: Yeah, totally. I think one of the key things to think about when framing this definition is when we talk about AI agents, we’re talking about it in the context of generative AI. So hopefully folks have had an opportunity to dabble around with ChatGPT and the likes of Copilot and similar tools. One of the interesting aspects of this new revolution with generative AI is the AI’s ability to really provide answers that are meaningful to the user. Whether you’re asking questions to ChatGPT, it’s really fascinating, the output that you get. Most of it actually sounds reasonable. Now let’s take that next leap of logic, right? So if you’re able to get a question and an answer that’s meaningful, what if you were able to ask AI to do something for you and get a meaningful action? And so AI agents are the next evolution of basically extending the capabilities of existing LLM tools and allowing them to really interact with the real world environment, whether that’s through, you know, a software interface, like an API, like, for example, like creating an account, or whether that could even be something as crazy as taking an action in the real world, like maybe even placing an order on behalf of a customer, right? So at its core, AI agents extend the capability of generative AI algorithms to be able to take actions based on a user’s input request.

Greg Kihlstrom: Got it, got it. So yeah, so essentially building on that, that idea that I think a lot of people are familiar with is, you know, you type in a question in a, you know, chat interface or something like that, you get an answer. But instead, what you’re saying is, this is a few levels of complexity beyond that of like, I want you to accomplish something. And then the AI agent goes out there and does it right? Is that is that a safe assumption?

Raj Rikhy: Yeah, I mean, you know, the the neat way to think about this is, if you were to do some kind of an integration with another tool, right today, let’s say you weren’t using AI at all, right? You would probably have to scope out the work, figure out, is there a way that you can interact with it? Let’s say, for example, you know, Salesforce and something like Constant Contact or I don’t know, some other type of tool, right? You’d have to figure out like, okay, let’s sit down, figure out, you know, are both tools applicable for our use case? Do we have kind of an intersection of their APIs? You have to find developers, you have to, you know, figure out like, okay, can I And so one of the fascinating things that generative AI is able to do is you actually don’t even need to do any of those things. In fact, you can kind of, and I’m being reductive here, but give the AI a URL and it can kind of figure it out, which is an absolutely fascinating thing. I’m being very reductive, but the integration costs that’s associated with getting an AI to do something with generative AI is beginning to be vanishingly small. And the more you see the capabilities of these models deepen and extend, not only are the richer capabilities that you’re going to see they have, uh, but you’re also going to see more fascinating things like maybe even multi-step operations. I mean, let’s, let’s get real for a second, right? One of the most challenging things that I’ve found, at least in, in marketing is, you know, it’s, dead simple these days to stand up a landing page. I mean, it is, you know, take your tool off of the shelf and you’ve got a landing page, right? But what do you do after that? Right? And if I’m, if I’m thinking about, well, yeah, it’s easy to stand up, I collect a bunch of leads. I mean, how am I going to turn those into anything resembling an MQL other than saying, oh, they’ve filled out a bunch of information? Well, let’s, let’s think about that next step. And I’m going to get crazy for a second, if you don’t mind, Greg.

Raj Rikhy: You know, What if, and just bear with me, I had a landing page and a customer filled out a bunch of information, let’s say for a potential customer, and they gave me a reason as an example for why they wanted to sign up for their interest in my product. And I took that open text box of the things that they were interested in, and I fed that to an agent. And not only could I create out of that text box, as an example, an account on their behalf, on my site, I could even potentially create a custom demo. If I hooked that AI agent up to my website based on that input prompt that that user provided, and I could have that already, you know, inside of about two to three minutes. I mean, there’s fascinating, fascinating capabilities, you know, and, and, and really the, your, your imagination is, is the old limit.

Greg Kihlstrom: Yeah, yeah. And I think one of the one of the amazing things about that part, too, is marketers and, you know, it’s a marketing automation, you know, enables a lot of things, but they also require a marketer to kind of guess what a customer is going to do. Right. So, you know, if this than that, there’s only so many variations that it even makes sense to set up based on how the size of the audience, so on and so forth. But what you’re saying here is Obviously, there needs to be something set up ahead of time or whatever. But let’s not try to second guess what every customer is going to do. Let’s let them guide us. And then as long as the infrastructure is there, kind of create it for them. Is that is that kind of what you’re saying?

Raj Rikhy: Yeah, I mean, I would even take it a step further, you know, like, let’s, let’s start rethinking some stuff. Right? the main thing that we’ve been sort of coaching ourselves to center our brains around is this notion of like a call to action. That’s the most, you know, immediate sort of thing that you can think happens inside of a collateral or an asset. Like what’s, what are the call to actions? Well, why are we limiting those call to actions to, you know, some kind of, of, of customer input in a form, right? Why don’t we have the AI actually take those actions. Forget about the call. Why don’t we actually action the collateral or the material, right? And that’s the big difference. It’s not about, you know, how do we capture that intent in that moment? Forget about capturing the intent. Just let them deliver. Yeah.

Greg Kihlstrom: Yeah. Yeah. I mean, that’s an interesting, I mean, it definitely, it takes a minute to wrap your head around that in a sense, because there’s so you know, so the paradigm is, is such a like, we’re gonna set up these stages and try to, yeah, kind of corral everyone through, like everyone goes through the same stages. Certainly, we’ve gotten a lot smarter with personalization and segmentation and all those things. But that’s still it’s it’s still using the old paradigm of right, we thought of all these things ahead of time. And now we’re going to try to get someone to do at least one out of three of them. Kind of what you’re saying here is it’s a different, it’s kind of it’s a different paradigm, right?

Raj Rikhy: Absolutely. And I’m using this, I’m using this example just to be illustrative about how I how critical I think it is to Obviously, I think folks are investing in learning about generative AI, but one of the main things that have happened is that with the advent of these new capabilities, in a way we kind of have to reteach ourselves of what capabilities are… What tools are available to us, right? And I think one of the scary things about… I’m saying it’s scary. I’m saying it’s scary. One of the scary things about the new AI wave is that, you know, it’s hard to actually figure out now that there’s been kind of an explosion of models and tools and, you know, things are moving so fast that it’s easy to get kind of swept up in the lingo. Like as an example, you know, for a very, for a hot second there, there was a lot of literature, there was a lot of interest around this thing called prompt engineering, right? And now it’s called God knows what else. And like, for like six months, it was like the only thing anybody could talk about, right? You know, like, Oh, we got to engineer this prompt. And, you know, and there’s going to be careers made out of this all that kind of stuff. And, and, you know, here we are, I want to say, like, six months to a year later, I don’t hear anybody talking about prompt engineering anymore.

Greg Kihlstrom: Right, right. That one person that got that like $350,000 a year job is hopefully they still have it, but yeah. I mean, hopefully they learn fast, right?

Raj Rikhy: Right, right. But the thing is that it’s things are moving so quickly that it’s hard to keep up with not only the pace of innovation, but the pace of how innovation opens up the tools that are available to you. You know, not just necessarily as a marketing professional, but like literally anyone. But this is why I think it’s good to think about, go back to first principles. And I actually use this tool with product managers also when it comes to generative AI. Because applications of generative AI… And by the way, there’s a topic for a whole other time, which is like, what is the overlap between product management and product marketing? And what does that look like in the context of AI? But let’s separate that aside for a second. know, when it comes down to first principles around applications of AI, it’s really about like, you know, what are we really trying to accomplish? And what is the core of this technology that allows us to accomplish this? And the core of AI agents is really about eliminating, and again, I’m being reductive, but eliminating integration cost to zero, right? You know, as much as possible so that you can, you know, take the leap of thought, leap of logic, you know, either with your tools or with, you know, customer assets or collateral or whatever it is, go from zero to whatever your heart desires and figure out, you know, how you close that gap. And if that means creating an account, creating a customized RFQ, creating a, you know, even going so far as to like, setting up somebody CRM, I mean, the world is literally open with possibilities. But yeah, that’s that’s kind of the key of it.

Greg Kihlstrom: So is it multiple agents, AI agents working together? Is that is that already happening? Is that like, kind of like the next phase?

Raj Rikhy: Yeah, great question. So the part of the reason why, in a way, I feel like, I have a unique perspective here is because agents stand alone are not actually a new thing in software per se, right? Like, reinforcement learning is a discipline that’s been around, you know, for 3040, I think, 50 or 60 years. Yeah. And the notion of having, you know, multiple, autonomous, little, you know, I don’t know what you want to call it, like little dudes or gals kind of like running amok in a software environment, right? That’s something that has actually been in the zeitgeist for quite some time. Generative AI has offered the capabilities beyond simply just like predefined routines, right? and allows for discovery. That’s kind of the next sort of thing, right? So I don’t have to, you know, figure out how each of these bots are going to work. If you think about, like, back in the day before we had talked about generative AI, there was, like, a lot of, like, bots that you would have on the page and, you know, these bots would… I don’t know if you’ve ever used one of those things, but, like, the amount of investment that goes into, like, figuring out how I put all of the permutations of questions, responses and answers and linking them together and what the outcomes are. I mean, it is absurd. Okay. And to your point around like agents, right? The maturity of where things are today with agents is that most folks are simply using a single agent because of the complexity involved in stitching together multiple agents. And research is kind of trending at least with general AI, research is trending in the direction of like, okay, well, what does negotiation and communication between agents look like? How do you manage hierarchies? What does a dependency model look like between multiple agents? Honestly speaking, that’s a level of implementation detail that I would avoid. Even working through that inside of a product, I wouldn’t even worry about it. But the reason why it’s useful to think about is Even if you’re not thinking about how many different things like, let’s just call it bots, because it’s just easier. Sure. Even if you’re not thinking about how many bots are doing different things, just think about how many things you want to do. The difference is that you probably will have to think about those steps in a sequence, as opposed to parallelizing them. Right. And that’s kind of the main difference between single agents and multi agents that you can parallelize and do things, you know, not necessarily sequentially, but you can have different threads going at the same time. And that’s okay, right? Like it is a limitation. So as an example, you know, instead of being able to fill out the CRM, while you’re taking the user through a guided interface, you probably have to do the guided interface first, then fill it out. That’s okay. That’s not what’s important. What’s important is being able to kind of figure out how do I dip my toe in the water, right of getting things done on behalf of the customer. And then when I schedule that logic for a single agent, how do I do it in a way where the output is useful and valuable to whoever it is that’s consuming this, you know, stuff. Right. Yeah. That honestly speaking, that’s the level of granularity that I would stay at for now until it gets much easier.

Greg Kihlstrom: Yeah. Yeah. I know. I’m, I’m envisioning this, like, org chart of AI agents or something at some at some point, which is then we get into a whole whole other level of things. But yeah, I think we’re I think that’s that’s probably good for for most for now. Yeah. I want to talk about, you know, because there’s a lot of people listening and organizations, let’s just say a very varying stages on the adoption curve. You know, I think I think most are using AI to some degree, whether that’s, you know, a few people using, you know, a few different tools like Copilot or chat GPT or Gemini or something to all the way up to some very sophisticated organizations, what are what are you seeing as far as you know, talking about enterprises, because that’s that’s a big part of the audience, you know, where do you think? Where do you see them being on this adoption curve of gen AI in particular? Yeah.

Raj Rikhy: So that’s an excellent question. So I think it really varies. Yeah, I’ve certainly seen Particularly on the AI front, when it comes to regulated industries, as obvious, and this is not unique to generative AI, truly, when it comes to AI applications with regulated industries, and a lot of those enterprises are in verticals that have regulations to that effect, it is a very fast moving space. And there is a lot of concern around, okay, well, does this training data, Will it capture my input? Will it be retrained? What about the inference? And so I think that there’s certainly some uncertainty attached to that, which, look, if you’re a customer who’s rolling it out, the burden of proof is on the output, but it’s also up to can you change as quickly as necessary with potentially the regulations that are coming out? So there’s some risk associated with that, certainly. But I do think that there’s lower hanging fruit that a lot of enterprises can certainly take muster of. And I mentioned one of the use cases at the top of the call, which is instead of capturing emails, just give them a little extra goodness. That’s not something that’s unique to SMBs or enterprises. That’s something that you can take off the shelf and just be like, OK, well, how do I integrate this in my workflow? I think that for the more sophisticated applications, certainly, you know, I’m seeing enterprises follow the typical adoption curve, right? And this is, you know, you have skunkworks teams who are experimenting, they have sort of isolated projects, they’re usually net new, they’re understanding, you know, what the applications look like before they roll it out. And the one thing that I do want to kind of call out here, that is not again, not unique to generative AI, right? Is you always want to be aware of your own data estate, right? You always kind of want to be in lockstep with, particularly on these skunkworks projects, have some level of visibility, right? Just be like, Hey, this is an experiment. This is time bound where, you know, this is minimal cost, et cetera. But the reality is that especially in these enterprises, as I mentioned, since the burden of proof is on you, you have to figure out how to articulate the value to the organization at large. And really, that’s only done through visionary, sort of proof of concept and leadership. That’s the honest truth, right? You have to be able to, like, bring the, you know, I hope you don’t mind me borrowing this. But like, like, bring bring the tablets, right? Like, show the like, show the goods. Right, right. If you can’t showcase in kind of real order, even if it’s something simplistic, right, what the application could unlock for the enterprise, I think it’s going to be difficult to try and gain traction. And the fortunate part with AI agents in particular, okay, is that a lot of the things that you can accomplish can be done with public information that’s off the shelf. Yeah, you don’t have to necessarily do that in the context of your own data, you can find proxy use cases and those proxy use cases with public data can hopefully, I mean, it could even be Greg, it honestly could even be something as simple as you have public documentation for your product, right? Hook it into AI agents that can ask and answer questions and maybe even something as simple as, hey, can you show a user where this is or take a screenshot of this inside of my product from public documentation for this specific question. Yeah, and that’s an easy way to off the shelf. Like if I’m an enterprise like this is all public stuff. This is out there on the web. Right, right, right, but but at least in my opinion, right? The only way that you’re going to be able to kind of like articulate that value back in the house is find something off the shelf. Cheap is cheap. It’s easy to do. It’s not that difficult. and take it back inside the organization.

Greg Kihlstrom: Yeah. And I mean, in that case, to your point, it is similar to which I think is a good thing. You know, it brings it back to this is AI has a lot of potential benefits and I do think it’s a little different than some of the technology advancements of recent years but at the end of the day it is still another technology advancement and therefore approaching it in a similar way there’s you know there there’s precedent if if nothing else for the way that it gets adopted and and thus you know proven the model gets proven and you know again even if there’s way more potential in this than some other you know recent trendy things or buzzwords or whatever i think that’s definitely a tried and true way of doing it and and the other part is you know we are seeing, I mean, I, you know, I remember, you know, the early days of GDPR when no one really knew exactly what the best ways were to implement these, you know, brand new regulations and all those kinds of things. I feel like we’re kind of in the early days of that with AI regulations as well. So, you know, to that end, you know, it’s, it seems like it’s probably starting kind of how it did with GDPR with, you know, EU kind of writing some things, but it’s still, it would be difficult to say two years, three years out, what that’s going to look like, whether it’s, you know, in the EU or in the States or wherever. So starting with some some things that are pretty easy to prove and safe, you know, even for those regulated industries seems seems to make a lot of sense.

Raj Rikhy: Yeah. And honestly, I think the main thing here that’s that’s different distinctly from, you know, some of the other technologies, like if you were to tell me How do you apply, as an example, classical ML or classic AI applications to marketing applications? I would say, well, you’ve got to get the pipeline going, then you’ve got to get the data, then you’ve got to get the inference. The main difference separating these things from agents is that agentic capabilities do not require a lot of tailoring and input to be valuable. because of their ability to eliminate that integration cost. And that’s kind of the difference, right? So if I want to showcase the power of agents, I don’t have to train a pipeline of stuff. And now generative AI obviously is pretty similar in that I don’t have to do that. But let’s be real, right? There’s a big difference between answering questions in a chatbot interface and actually doing something that the user asked it to do, right? Like submitting an order in a shopping cart.

Greg Kihlstrom: right right yeah well hey i could talk about this for a lot longer with you today but um we’re almost at time here so one one last question and then i would definitely love to talk about some of the the other things that you mentioned and maybe in another another episode but um you know as as you look in the the months ahead You know, what are you seeing as far as some, some trends that, uh, you know, particularly marketers, but really, you know, people in the enterprise should be looking at in terms of AI agents, you know, how should, how should companies be preparing for, uh, for further use of these things?

Raj Rikhy: Yeah, totally. And I think, you know, just listing a few of these, right. We’ve talked about it at, at large, but start thinking about your workflow differently. Right. If there is an opportunity where a low hanging fruit opportunity where you can provide something to the user of value immediately and not just capture their their, you know, intent. That’s the place to start thinking and pivoting your own mental model around usage of agents. Right. So whether that means like automating customer interactions vis-a-vis maybe outbound reach out things like crafting email campaigns, like pre-tailoring them to individual customers, right? Like massive personalization at scale, you know, individualized copywriting down to potentially even, you know, the team or even the user level, right? Like automated responses. Another one that’s really, really interesting that I’ve been seeing of late is automated creation of A-B tests. So like, you know, if you have like, let’s say a landing page, instead of going through the loop, which everyone kind of knows, which is like, you know, have one landing page compared to compare A to B, and then make a change and then you know, run another cohort. Well, why not hook up all of that stuff to metrics and just have AI create those differences for you and, and figure out the highest optimizations in A, B tests, like These are the kinds of things that are sort of revolutionary applications that I’m beginning to see. And just to take a beat here, Greg, you know, it seems unapproachable at first. I think it’s helpful to learn about it yourself first. I have, you know, a Maven course that is coming up in January where I will go through agentic product development from A to Z. you know, there’s other resources out there as well, that that folks can take advantage of, that are kind of free and online. And they’re sort of low code interfaces for for agentic, you know, education, I guess, in summary, stay abreast of the changes, you know, start pivoting your thinking, and figure out if there are some of these low hanging fruits out there, like some of the ones I just mentioned that you can incorporate in your day to day.

Image