Role of AI Agents in Services Design

It says we are live. As soon as live appeared, Koushik gave a smile. Someone said cheese. So we are live and we have our first viewers. So we know that we are live. So let's begin. Hello and welcome to Momentum Office Hours. My name is Yash and I'm joined by the leadership team at Momentum 91, Krishna and Koushik to discuss the topic of the week, which is role of AI agents in services design.

Yeah.

Our goal with these conversations is to provide you with actionable insights and practical strategies that you can apply to your own business. Throughout the session, we encourage you to engage with us by asking questions and sharing your thoughts. This is a fantastic opportunity to learn from each other and gain new insights that can help drive your digital initiatives forward. So let's get started. Krishna, Koushik, how are you doing today?

Good, nice weekend. Ramping up what happened in the week and planning out what should happen for next week.

Do you find? Yeah.

What should happen in the next weeks? We travelling. There's been a lot of travelling that's going on and I know it will, I will remiss if we don't start by like start before by addressing elephant in the room. is Jay? And so to our jealousy, he is surfing. so it's not like he's busy meetings or anything like that. He's literally surfing like right now.

He's taking what like kayaking and all of those lessons and all. So yeah, so that's why he's not here. But Koushik and Krishna and I are here and we can start talking about role of AI agents in services design. But before we talk about role of AI agents, let's talk a little bit about services design. So Koushik, what is services design exactly?

Services design is about designing your organizational functioning. you have, if you are a business, you could be a business or organization company, irrespective of what you are, you have your people, props and processes. So in props, I mean, platforms that you use. So all these things come hand to hand to make sure that you give the best customer experience or client experience for your ideal clients or customers.

So, services design is all about how do we make sure that this is designed in the right way. That's the only thing. So, that is basically services design.

People, props and processes. Essentially on the side of customer facing. So I could have no services design on the side of fulfillment or vendors or procurement and stuff like that. Service design largely deals with people, props. Props is property or props is like just...

So.

So the reason why we name it in that fashion is because very often service designers try to relate it with whenever a non-technical person tries to ask for a service design example, the easiest thing that they give is a stage play. So you have your actors who are like your people. Then at the back, there's always a backstage and a front stage for a stage play.

So the backstage is where you would have props, would have all the people who are continuously sending and organizing.

lights and the music and timing and all of that.

All of that comes under the people who are in backstage and then there are similarly people, props and processes in the front stage also, which is delivering and the backstage is constantly supporting the front stage to deliver a great experience to the audience who are watching it. So it's like the similar metaphor, you could say, but much more a technical way, services design is the way you go about.

are in. Interesting. Krishna.

Yeah, so I want to know that what are the challenges that we face in, you know, building the AI agents for services design?

One biggest challenge is it's very new, so everyone has a clutter about how are you going to go about doing this. Because any new technology that comes up, the emerging place where it starts from is at an enterprise level. And we are also slowly starting to see AI agent adoption at enterprise level in much more large scale. We are starting to see platforms like Salesforce, SAP.

all of them are trying to offer, starting to offer, you know, AI agents within their platform itself. But that is a much more productized version within the platform is what they are trying to have. So that is the first challenge, a lot of clutter around and trying to understand how this works. So I think the best way, the biggest challenge is, one is the knowledge, second is the who are the right, if you have your in-house team,

then it's great. So you have the right people with you to build it. But if you don't have, for decision makers, it's tough to find out who are the right people for them to reach out to and make a choice upon that. that is another challenge that's currently pretty much prevalent with respect to execution point of view. And once it starts coming in, how do you go about identifying pain points within my business or organization?

And what exactly are these AI agents going to replace or aid in this entire process? So that is another challenge that is there. Like for example, there will be certain people or roles, as we just discussed, people, platforms, and processes. There will be certain processes that will get completely replaced through AI agents in place. And at some point, there will be a set of people who will get completely replaced with AI agents in place. So independently,

One needs to identify within their entire business model, which are the areas that could be aided or replaced using these AI agents. So that mapping and that strategy is a completely new way of approaching services design. That level of mapping is still a challenge because that is very new to service design as a professional. Because it was not ready for a complete shift like this.

an entire module or a core module within the basics of service design is practically getting replaced or aided. which is another big challenge.

One of the challenges, just to add to what Koushik said is also change management, right? So, you generally, if you need services design, you're at least a mid-market or an enterprise firm. And there are just a lot of cogs in the wheel. And for you to be able to implement or for you to be able to expect a new behavior is by the same people is extremely difficult.

Even though the new behavior will offer them more free time, it will allow them to deliver more, it will allow them to create more value. But implementing and expecting people to sort of adhere and implement a new behavior, that's like change management is a real challenge as well. Having said that, so tell us about like if I'm a business, what should be?

And then we'll not talk a lot about what are AI agents because if you want to know what are AI agents, just the previous episode that we did with Jay and he was leading that session which is where he in detail spoke about AI agents and what are the four categories of AI agents and things like that. So wherever you are, whether you're on our website or YouTube or LinkedIn, you'll be able to find that episode and that's a significantly better...

content for you to sort of consume and understand AI agents. But tell us this, how if I want to, I already have, I may not have the best services design, but I already have some design, right? Something is happening. So by definition, there is some design, it could be bad design, good design, room for improvement, but I, definition, I have some services that are being delivered and there is design for that. How do I go about identifying opportunities of what places

Can AI agents come in and aid, optimize, replace, make it more efficient, whatever the use case may be? How do I go about?

Yeah, so we usually have this approach where stakeholder to stakeholder connection and where all, great, I just said sponsor but it was this thing. So stakeholder to stakeholder, so stakeholder to stakeholder connection which would help you to understand this. Let me show you a diagram that we usually use to make people understand. Could you see my screen?

Give me one second. just, I am able to see it. I'll just bring it on stage.

Yeah, sure.

Now everyone can see your screen.

Yeah, so I think the best, so what you could see here is that if this is your entire business, we have users who are one individual stakeholder.

Just one small one. Just can you make this full screen from the bottom right then? Yeah. Yeah.

Yeah, so let's say if you are a business, right? There are two ends to it. One is you as a service provider and then you have your users on the other end. The current way of agentic development that's happening and the most efficient way of implementation with respect to identifying the pain point or the opportunity places is that, so your user would have a certain set of AI agents

in place and you as a service provider will also have certain set of AI agents in place and there would be an intersection point where both the agents would be talking to each other and sharing the data and trying to make sure that things get done. I'll give you an example. I'll take you through along with an example. Then you can understand much better. Now let's say you are a diagnostic lab and if you want to book a test, right?

The usual current way is that I have to browse through the entire set of available options. Then I have to go and land at one website probably. Then there I would see what all different types of testing services are available for me to take or that is being offered by the diagnostics center. Then I would choose the one that I want.

and then I would go about filling up a form or any sort of inquiry CTA and then your user has made a choice now. So up till this point, your user had the option to explore then arrive at one

Sorry, so just so that we are on the same page, users are my team members? Like if I'm implementing this, are they my team members or are they my customers?

Your customers. My customers, okay. We are coming from there, then we will arrive at your team members. We'll come there. So then you have your user was looking across different options, then he arrived at you, then he came to your website, then he booked an inquiry and he has placed a request. This is his journey. Now at your end, there will be a set of employees in your team.

got it.

who would take that, there will be a platform that would be, know, taking in all those requests, then sending it to, you know, scheduling it. There will be another platform which would probably you'll be using currently to schedule these according to the availability of the technicians as well as the centers availability of different diagnostic labs that is over there. And if it is one diagnostic lab, it will seem pretty simpler.

But if it's like hundreds and many labs that are there, then there are multiple layers of complexities that gets added to this. Now, this is your current flow. Now, how would an agentic implementation would look like for this is that your user would be having only one agent. So you, as a service provider, would have created an agent for your user, which would help him to make much more informed choices about what type of test he wants to take.

That agent is a user-based agent. So that is a personal independent AI agent that you have created for your users. And that could do multiple things. Like, for example, it could be about placing a new request for a test, or it could be about downloading an already previously taken test, or the agent could already have the memory of your medical data and then suggest to you monthly tests based upon that. It could be any of this.

personalize the agent according to make sure that the users are always in contact with our product. So that is your user-based agent that you're building at one level. Then you would have your organizational agent that is doing the tasks. So whatever the tasks about scheduling, receiving, then allotting it to right diagnostic center, according to the nearest location of the user, all these things are done by an other agent.

So now these both agents are connected together at the either end and then they are receiving the data and then they're also informing the right stakeholders who are on the other end of the service provider side that is at your side. So this is called AI to AI implementation. We just spoke about two use cases and two AI agents involved. Now the moment, let me stop sharing. And the moment.

you start identifying as many opportunities, you would need to start having multiple AI agents involved for each different kinds of tasks. And each agent would be cross-functionally talking to each other and receiving data and pulling it. But from the user end, what's happening is that it is more like a chat interface that the user could be having and they are talking to that chat interface.

and the chat interface is guiding them through each step by step. Now, why is this different? Now, why could this be interesting is that, so currently what we have, for example, you have your website, which is currently just showcasing everything, right? Going forward, those kind of interfaces would get replaced. So you would probably, you would start seeing complete open interface of just a chat.

space where you go and ask for details. And if you have already logged in into the chat interface, then it already has your data, probably your medical history. In this example, your medical history, your location, a lot of other details. And then it would suggest automatically the right set of things based on your needs. So this is one example of how you could go about implementing. So you identify one stakeholder, second stakeholder.

and then see what all intermediate processes that you have from both the ends and build AI agents that could just get all these things done as a single set of process that automates automatically one after the other. So this is one way to go about it.

interesting here. Krishna?

Yeah, so when we talk about the clients or our customer, right? So how do AI agents understand customer requirements, their user flow, what they want to do in particular product? So how do AI agents will understand that?

So there are two ways of AI agents we are talking about. One is an AI agent which could, which are programmed based on fixed code paths. We have discussed this in much more detail in our previous podcast also. So in this particular AI agents case, the major pointers are two things. One is any AI agent actually has two primary factors on which it is trained upon. One is retrieval systems and another major component is memories.

The third major component is tools, which it uses to take action upon. But retrieval system and memory are key factors that helps AI agents to do the task. So that is what makes them apart from automations. So this is a very common query and sort of misunderstanding we have with respect to how are AI agents any different from automations, right? So what automations doesn't have and AI agents have,

is one is when I say retrieval system, we are talking about access to intranet or internet. intranet is the entire data set of points that you have collected as a company. It could be your user data, it could be your own company data, it could be your service offering data, it could be multiple data points that it is trained upon. So that is your intranet system. Second important thing is memory. So your user has given something

you might not have logged in in your retrieval system data point. But AI has memory of the fact that it has its own memory data structure where it could go back and it could bring it in. So these both exist to make sure two things. One is to avoid errors. So let's say if you're booking a flight.

And let's say the flight has been delayed by an RR2. If there was any sort of automation systems that exist, currently would show you an error that it has been delayed. What would an AI agent do is that AI agent would have access to internet. In this case, let's say internet. It would also have access to your data because you have previously searched for probably for train tickets or probably for flight tickets. Or it also knows where you are booking the flight from.

So automatically what it does is it will use this data to suggest you flight options nearest for you and the next immediate flight. And it would probably go ahead and do before doing the booking, it would just ask you once for a human confirmation and then do the booking. So this is the differentiation. So this is what aids it for better to avoid errors, better accuracy and to actually get things done.

Interesting. So I also have another example of implementing AI agents, which we are actually, as a matter of fact, currently evaluating of implementing it for a client, which is not necessarily in services design, but essentially enabling and completely automating and having an agentic procure to pay process for the complete organization, which means that

a request for code is generated, gets sent out to relevant vendors. Once the codes are received from vendors, it figures out based on the access to information from the intranet or the ERP system, it figures out which vendor delivered on time, which vendor had the least amount of.

returns and sort of recommends which is the right L1 based on the quote and based on their previous performance with our organization, awards them the purchase order and then you essentially get the goods. So there are multiple AI agents in this piece. Each one of those agents actually have actually follow a very very specific part of that process or part of that puzzle.

And yeah, so but this is of course not part of services design, but it is in multiple AI agents in our workflow. Another thing, Koushik, that I wanted to understand as far as AI agents were in services design is concerned is how do you actually, so once you identify that here is a, this is a step or this is a process that can be automated or that can be made more efficient.

Who do you reach out to? What's the process like? Because one of the things that people keep on hearing is that there's loads and loads of data and information that's required for AI to train itself, which sort of acts as an entry barrier or a barrier to just start this. Another thing that folks hear from just outside noise is also that building AI agents is extremely expensive and time consuming.

So can you bust little bits of these myths maybe in this answer and before you answer, we have a comment on YouTube. There's a channel called Tribute to JMW Turner, shout out to Tribute to JMW Turner, says thanks from Paris for sharing insights and wish you all the best. just some nice heartwarming gesture.

coming our way, some love coming our way. So can you go ahead and bust a few myths in terms of the amount of data that's required or how difficult is it to build it?

So currently, the first is about your choice of on what LLM are you going to build it. So a very LLM for dummies way of saying how a thing is that if AI agentic system is about AI agentic system is all about how you it's about you doing things. LLM is the knowledge of that you process to know how to do that.

So that is the right connection between both of them.

No, wait, wait. what they're saying is LLM is like your brain which has memory and has logic and has processing power and the agent itself is my hands and legs for lack of a better analogy. But would that be a fair way to...

Exactly. that's actually a very fair way to put it. now this with that in mind, like, so first is about you choosing what LLM you want to work on. So let's say, for example, I'll give you a use case. Let's say if you are a product company and you want to create an agentic system currently to replace, you just have started, but you want you, you will reach a scale after later.

and you want a particular internal tool to be created or you have a lack of engineers in your team or I could, there is one interesting use case that actually came up while I was just building this use case while I'm trying to talk, it just came to my mind that let's say you have less of a QA team currently and you want to build a QA AI agent for your company. Now, how do you go about doing it? So it's very important to understand which LLM is good at.

That is another important point that many people may sort on is people often that we have seen is that people often... God, okay.

This is a great point that you are making. Can't I just ask an LLM as to which LLM is best for... Okay, no, I cannot. I have to do it myself.

People often forget to think from the LLM shoes on what purpose it was made for. For example, Claude is extremely good at Anthropic as a company, created Claude with core purpose and focus on coding as a purpose. Like even in the recent 3.7 Sonnet release, and since the first release, all the releases they focused

majorly on coding and the preview aspect of how coders can use Anthropic better. Whereas, StadgpD trying to be the general purpose person here. Whereas, if we look at Perplexity on the other hand, Perplexity focuses extremely on research purpose. It was not done for coding. So, like if I'm trying to create a product manager AI assistant, then, or a BA AI assistant, I would probably choose Perplexity rather than

going for Clot. Whereas in this case, it's more of a tech requirement, I would rather go for Anthropic's LLM. So Anthropic provides three things, since we have done those implementations. So Anthropic has this thing where, like I mentioned, for the retrieval systems, the three aspects, retrieval systems, tools, as well as the memory aspect.

they have a combined knowledge-based source called as Context Model Protocol, CMP, which you can think of CMP as like a one-stop where you can get access to all of these. And it would connect to your company internet, and then you can use that to train upon. So it is like a template that you could use to train your data.

So that is available and it gives you and Anthropic gives you. Similarly, DeepSeq has its own way of doing this. So first, it's about choosing the LLM that you want. So first, choose the brains, going back to the analogy that you just said. Second thing is about.

This actually seems like I'm recruiting people, right? Yeah. So I mean, of course, if I'm looking for a product manager, I'll recruit someone who has some amount of research experience. If I'm recruiting an engineer, I will recruit the person who's coded before. so I'm recruiting, you know, perplexity or as an LLM or I'm recruiting Claude or Anthropic for that matter as an LLM. sounds, yeah, makes sense.

Yeah, in fact, you rightly pointed out like going forward, like we were actually having this discussion and very interesting, you know, like a future came up, like a future thought came up that in future you'd start seeing, you know, AI agents apply for jobs, like an Indeed or something where you would have like 10, 20 AI agents applying for product manager roles and it's on AI to AI compatibility, you would start

checking which one is better and then start hiring them for your company.

Isn't that going to be a recruiter's

So because the rate at which people are trying to build AI agents, you will start seeing them as common as websites. And then you are starting to have which AI agent to have for which use case that you have so specifically. So you'll start seeing that. But coming back to your question, second point is that the team currently that is the most. So you have just cracked five, not even 5 % if you cross the first point that I just said.

remaining 40 % of it is the team because the right talent for building AI agents is currently, one is it's scarce and second thing is that how well equipped they are to learn new upcoming things also because the terrain is changing like every day, I would say because it's completely changed. We had a particular notion until, I mean the very video that we are just recording.

was the three weeks ago the service design notion versus what I'm talking about now is completely different.

Very, very different. Yeah, no, absolutely. So I agree. mean, the people who are actually going to work on it and what's their skill level, what's the amount of things that they know is extremely important.

So I think choosing the, let's say if you are not in the realm and your team is currently not, doesn't have the bandwidth to do that, you always want to reach out to an agency who has the right kind of people who can go about doing this or go about doing these implementations.

Curious to know. Do you know anyone that can do this?

So like we are like, hopefully, you know, like, we have been following this up from the time the entire thing started. We have, think, yes, you would know this very well that I think when the first initial API releases for

token releases happened we started having that implemented within the product that we had. Since the time we always we have a huge FOMO it's us versus the world if we don't catch up we get FOMO and we really try to catch up.

Like the last year and a half or so has been a nice tidy balance between paranoia and sparkle in the eye.

Yeah.

So like you are extremely worried but then you are also very very optimistic as to what does this mean? what does this mean? this happened. no this is this is interesting and so on and so forth. But either way having said that yeah so if you are if you want to do all of these things yourself you can go ahead and do the research you can you know we have videos we've written ebooks you can go to our website you can check out all of those things and we have a knack of trying to simplify

as much as possible and not to appear intimidating. It is not that difficult to do, to be honest, right? So once you sort of are able to put together certain fundamentals in place, you will be able to figure it out. But if you don't have time, you'd rather focus on the business, you'd rather focus on growth and things like that, then please consider booking a call. Happy to get in a call and understand your use case and make some recommendations as well. Having said that, we are...

overtime for today and we can talk about this for hours and hours and then so we have to keep ourselves in check but thank you everyone for joining in whether it is on LinkedIn or YouTube it was a pleasure hope you liked it and got some value out of it as well one of the other things that I also since I got a little bit of push from our marketing team

they basically said you need to get more subscribers. So you need to get more subscribers and so this time I'm not going to joke about it. I'm just going to curse you and I'm going to like if you don't not curse at you I'm going to give you a curse. So if you do not subscribe then the next time that you order a pizza someone will put a pineapple on it. thank god. So if you found some value

Yeah.

Do consider liking, commenting, subscribing. Thank you for being with us and do not let the next pizza have pineapple on it. It's really not good. There are some people who like it, but most people don't and I trust you are not that person. But thank you for taking out the time and joining us and we'll see you until next time.

Bye. Bye.

Okay.

Bye.