Paul Morrison

Yeah, right.
Hello and welcome to Retail and Consumer Pulse, the WNS podcast dedicated to transformation in the retail and consumer goods sector. My name is Paul Morrison, and today we are taking another double click into the world of Agentic AI. It's a topic we've discussed a few times this year, and we're going to take a deeper stock of what it means across the sector. To help us grab hold of this rapidly developing subject, we have no less than Vin Kumar with us, AI enablement and digital operations leader at the Hackett group. Vin, great to be talking with you again. Perhaps you could say hi and share a few words about yourself.

Vin Kumar

Hey, thanks, Paul. And thanks for having me on your podcast. I did listen to the previous one on retail. It was very insightful. So again, thank you for having me and looking forward to this conversation.

Yes, Paul, I know AI enablement, digital operations is big. What does it really mean is I focus on helping clients across industry adopt AI and help them take their operations and move it more and more to digital operations. So I'm all about understanding what technologies are working with our clients, helping them to adopt these type of technologies and giving them a framework to do that. So that's the practice I lead and work with clients and looking forward for this conversation.

Paul Morrison

Great stuff then. Well, it's obviously a busy time for you. Lots on the plate. So, let's dive in, and I'll kick out a few numbers just to put the subject in context, just to bring it to life. So, a report from NVIDIA, “the state of AI in retail,” a couple of weeks back, 58% of companies report improved efficiency with AI agents. Salesforce in some of their analysis concluded that 75% of retailers will see AI agents as essential by next year.

And then, I was struck by a Hackett group report earlier this year as well on Gen AI-driven finance, and there was a finding in that – that said agentic AI reduces accounts receivable workload by 50 to 70%. So, there’s some really big potential savings there.

There's a lot of buzz. There’s a lot of ideas flying around about agentic. So, I'm going to give you the difficult hospital pass of trying to define agentic for us and let's take it from there.

Vin Kumar

Sure, Paul. I expected you to throw me straight into the deep end. So, no love lost here. The way I try to explain what is kind of Gen AI, agentic AI and people are looking at agentic AI and saying, is it just RPA, another word for RPA, or is it kind of knows everything? It's magic and can do and execute too. So, it's this range of scale where it's there. But the way we look and define agentic AI is – one, it's a solution is what we call it. We call it either a micro application or AI-based solution.

What does that mean? It is focused on an outcome. The way it works is it’s all tuned with the objective of fulfilling or getting a particular outcome. That's kind of what agentic AI is. But I think the key thing is, it's got three significant capabilities. One, it is able to do reasoning and decision making. This capability is powered by the Gen AI technologies, and there, the reasoning and decision making is a probabilistic characteristic. And that's what the agentic AI leans on the Gen AI to provide that capability.

Two, the second characteristic of agentic AI – it has a couple of agents, and agents could be autonomous or assisted agents. These agents actually go and execute work. They go in, they log in, they download, they process a particular thing, a particular transaction, they actually execute work. From that perspective, it can be taken as a next step of what an RPA used to do, which was all kind of executing the work. So, agentic AI has a reasoning capability. It's got a set of agents. We call it adoptive execution. That means it could change the execution and customize execution to the situation it is.

There is autonomous agents, which means end-to-end, it can be run by the machine, and if there's an exception it comes out to human, versus an assisted agent, where the machine runs steps one to five, waits for a human input for six and seven, and then continues on to do it. And that's what an assisted agent is. But an agentic AI has a set or a suite of autonomous agents and or assisted agents as part of that solution. And the third capability is the ability to orchestrate and plan.

Vin Kumar

As I said, it's got a set of agents, and which agent to be executed in this particular situation, orchestrating that is a capability that an agentic AI solution has. So, agentic AI, just to summarize, is a solution that has three capabilities – reasoning, ability to orchestrate and has either autonomous and or assisted agents, and that's kind of how it works. And we can obviously talk about some examples and bring it more to life.

Paul Morrison

We'll definitely do that. That's a great definition there. Let me call out some of the distinctive flavors and dimensions there. I guess it's key to emphasize that building on the recent wave of generative AI, with LLMS being central to that reasoning. So it's building on recent technologies, it has agents or bots within the functionality that carry out work that execute and then you have that orchestration element as well. So, you can navigate complex goals, complex workflows, and that all comes together. And I guess, the way I see it in my head is that you are putting generative AI to work to achieve a specific goal, like answering a customer query or resolving a queue of payment issues or rolling out some price changes across your infrastructure. You are rather than just creating and content wording ideas or code, you are making it happen. Is that correct?

Vin Kumar

Exactly. Yes, you're absolutely right. And the best use cases, as I said there, there are three characteristics of where agentic AI really is applicable is – one is what we call structured work. When we know what the inputs are, we know all the steps to process those inputs. We understand what the outcome is or output is going to be. That type of work is where agentic AI is really suitable for. Also is that work involves a little bit of fuzzy decision making or fuzzy logic decision making, which is by humans, which is what we do. It's based on some sort of experience, knowledge of the domain and understanding the data.

So all of those three come together for us to make some level of decision making. It's not extremely prescriptive because if it is a form, extremely formulaic, you can always write a script for it in an RPA and do it. But where there is some level of probabilistic decision making or we call it fuzzy logic. I know customer X has this characteristic. So, I know when they say something, I can interpret it correctly versus a customer Y. There's some level of this fuzzy logic. That's a characteristic of the type of work where agentic AI can solve.

Also, I tell clients that agentic AI doesn't solve an end-to-end process. It's a sub process. It's a series of tasks in a sub process is what gets automated using agentic AI.

Paul Morrison

Thanks. Very well. Let's come in a moment to how that translates in a retailer consumer context to some specific examples. But before we finished sort of framing the patient here, the issue at hand, it strikes me that there are different ways that you can implement agentic and what I observed with clients and with my colleagues is that we deploy and we see agentic AI in different modes or forms. Sometimes, we see it baked into a core application like an ERP. Sometimes, we see it either in or on the roadmap for one of the leading best of breed solutions like HighRadius or ServiceNow. And also, we see it as maybe an open source or in a custom developed solution using a range of the ingredients we're talking around earlier on. So, agentic is a certain way of approaching automation.

You can buy it or deploy it in different ways, which I think makes it slightly more complicated. Do you see that pattern?

Vin Kumar

Oh, absolutely, Paul. I think the reason behind that is a generated AI by itself can't do agentic AI, and for agentic AI, you need the generative AI capability. You need domain knowledge of that process, and then you need to be in your data. You need those three things to actually go and have an agentic AI solution when you start building it. So, what you do is, the ERPs will obviously have some level of domain knowledge there, but typically, it's industry agnostic.

It is one-to-many kind of level of domain knowledge which is there. Then you get your process-specific domain knowledges, be it the ServiceNow, a HighRadius, a BlackLine or a Salesforce service, any of these type of solutions. They have the domain on which they work, so that is important to do it. And data – there, you're getting closer because as you think about, as I said, the agentic AI has to do some amount of reasoning, it's not general knowledge-reasoning, it's reasoning based on that domain in your data to apply. So you see agentic AI and all these flavors, be it with an embedded within the best of breed solutions, and that's where they are all investing in all this agentic and Gen AI infrastructure, because they are building this agentic agents and agentic AI solutions that you can enable in your company when you implement it.

So, there will be that. But that's not going to solve all your agentic AI opportunities. You are going to have these open studios, design studios, where you can custom build your own agentic AI, and you've got all the way from Microsoft Copilot and agentic AI solutions to crew AI and there's N8N, there are various different tools as the UI parts, the automation anyways, all of them have their own kind of platforms to do that custom. And so, you'll see a flavor of all of these in your enterprise.

And that's a challenge. I know we'll be talking a little bit of why is it difficult because you don't know which one to use where. That's something that they have to solve.

Paul Morrison

I think that's an important thing to call out, and looking at this space, it's hard to avoid the agent washing that you see and hear, where a vendor or a solution provider tags on AI and now tags on agentic AI, where in reality, maybe we're talking about workflow or some other solution, perhaps very useful, but not really agentic, so that is a risk.

Let's change gears now. Let's start to sort of move from what is it to what does it mean for retail and consumer goods. And I guess, taking stock of that most recent part of the discussion, we've got agentic AI, which can tackle work that is defined by a goal that is often fuzzy, that requires decision making to happen during it. That is often repetitive or at scale and that covers a lot of work.

What strikes you when you look at the retail consumer goods process function landscape? What jumps out to you as the big buckets where we can start digging for opportunities?

Vin Kumar

So, I think there are a couple of stand out areas where if we were going to look for in a retail CPG client, kind of where the big opportunities could be is – one is definitely in the customer service area, where we are able to provide a personalized, customized ServiceNow at scale that we can do, be it in the B2C or even in the B2B area to do that. That is one big chunk of work where we see a lot of opportunities to provide that.

One of the biggest case studies is a company called Klarna, which is the buy now, pay later solution provider. They have, I think, publicly stated the level of benefits they've got on using in the customer service area. The second is right up front, which is in the order management and order fulfilment. order configuration.

So, the ability to price dynamically at and personalize and at scale. You can see that's the other aspect that we see in the product configuration, order management, order fulfilment, and then it leads to supply chain. These are the three big areas that we see in the retail and in the CPG space.

Definitely, on the G&A functions and finance and HR and procurement, there are opportunities, but we see the bigger opportunities more in the business operations.

Paul Morrison

If we're going to prioritize things in this discussion in terms of the focus of spend, the size in terms of the expense, then that's the logical place to start. So, let's maybe focus then on that customer service part of the equation.

First of all, I'm just thinking about some of the examples that have struck me over recent months. I guess the thing that jumps out in terms of agentic AI in retail and I guess e-commerce as well, is this whole concept of agentic shopping. So, this is something that's done partly by the enterprise. So you can think of Amazon's assistant Rufus. So this takes a goal of a particular product. It's a type of mini TV or a type of garment, whatever it might be, that there's a goal. You have a context from the user. that has a history that may be many years long. And the assistant can then make recommendations, can interact with natural language, and then can complete the purchase and then can support delivery and returns and feedback as required as a whole set of a whole sub process here. That is being automated through that assistant, and I think to the user, it might look somewhat like chatbots in the past, but behind it is much more sophisticated, much more resilient, much more autonomous processing than has been the case before and therefore is more scalable. And I'll just add to that the other side of agentic shopping that I think is interesting is the customer driven side. So, thinking of the Perplexity Pro, for example, an individual customer using their own preferred generative AI-search
functionality to find, to have the same product discovery journey, is now being offered the same set of finding the best price, finding the best location, arranging delivery and so on. So, agentic shopping is an enterprise driven thing and it's also
consumer driven thing as well. I think that's a super interesting space.

Vin Kumar

It is absolutely, Paul. And I think you're so right on. People may perceive it as – this is just the next “evolution of a chatbot.” But before, it was much more prescriptive. You had to explain the behavior of the chatbot, that is very prescriptive. If it is X, you go down this path. If it's Y, you go down this path and then kick it out here. It's much more dynamic that you're truly having an assistant do your work for you. And from an enterprise perspective, it's almost having you as a knowledge worker having an assistant that you can give a task and the task goes and gets run to do that. So, it can be there in your personal life as an end consumer, having an assistant do work for you, be it – “Hey, I am going to be travelling to Houston next week,” and then it knows it for my calendar. It seemed I made some booking then automatically go and do stuff based on my liking. Give me your things, and if I say yes, and it goes and executes it. It will reach out and make reservation for lunch or dinner or anything like that. It's more like taking a lot of my mind space off and allowing me to focus. You will see that in your personal life, and there are multiple players trying to play that on how you get access to it. But if you come from an enterprise perspective, I think there's a lot which you can do even though there is the ability to optimize that scale is where this is really good.

It says, Okay, we know, for example, there is where I'm located, where your customer's located, the closest warehouse. There is going to be an issue. There is a weather-related event coming there. So, it needs to clear some of the stuff which is there, and it can dynamically push to see - Okay, hey, I can provide this customer win a discount, but not Paul, because Paul is in UK. In Southern California, there is an event happening, and we need to clear some of the capacity. So I believe the dynamic complex is really powerful what it can do.

Paul Morrison

Absolutely. If we're talking customer experience, that's just one small part of the customer experience. Maybe the main one to talk about in agentic is how the use of call center functionality of customer reps is automated. You talked about the client or example at the start and there are estimates that suggest that in two to three years’ time, the majority of some of that customer facing work will be either triaged or automated using agentic AI. What's your perspective on that? Any thoughts on that big opportunity?

Vin Kumar

I think that's been a use case where we are seeing a lot of companies investing in it to do because it is one area, it's similar type of activities that's done in that function. Typically, this area has a large FTE and cost base. If you dial back maybe 20 years ago, this was one of the areas where a lot of outsourcing was happening in this. So, they know how to kind of standardize this. They know how to kind of manage it at scale and operationalize that scale, managing multi vendors, the ability to shift work easily to do it. So, they really understand the activities that goes on there and becomes the perfect use case for using agentic AI to give them that significant benefit of converting of the manual labor to digital labor. So, we see that a huge area and what's going to happen is you're going to have, obviously you've got your IVR telephony call center type of technologies, and there will be a lot of customization. And so they are going to provide you those type of studios where you can design your own agentic AI or you may if it's extremely custom because your infrastructure and application that I have to talk to is a significant number and varies across the organizations. Then you can put your own custom agentic AI solutions there to talk to them.

Paul Morrison

Yeah, exactly. That's loads to unpack there. But we have to now look at supply chain because the clock is marching on. So, customer experience into supply chain, there's there is a load in this space as well.

A few of the use cases that I think are striking. One is Walmart, that's been using agentic AI around inventory management and demand forecasting. So, that's nearly 5000 stores we're talking about. And it's been driving towards improving, back to your comment earlier on about outcomes, driving towards 99% plus in stock rates and reducing inventory costs to some very clear outcomes that have been targeted. And I think at the end of the day, the stack of technology to achieve that in an organization like Walmart is obviously large and complex, but the ingredients that you were talking about earlier on the in the call are there. And I note that Walmart has its own bespoke LLM functionality, Wallaby. So that is part of the stack that provides that reasoning and decision making that we were talking about earlier in the call, but together, it's able to take that take that objective and sense triggers around demand, around anomalies, around changes in supply situation, maybe political factors as well, and political risks and basically change the way that inventory is planned and stocked and moved around proactively with human oversight, but without necessary human intervention on a massive scale. So, I think that's a really significant example.

Vin Kumar

It's a huge area to do it, and obviously, a lot of retailers are not at the scale at Walmart where they could have their own LLM and training. So, here we have to look at is how do we get that domain knowledge which is required for agentic AI. It may not be as big scale as Walmart, but other retailers to do it. And so you have to take your general purpose LLMS, augment them with rag solutions, prompt engineering solutions so that provides the reasoning. And then club it with all your other agents to build your agentic AI solution. You can do it in chunks, because I think it is the complexity, speed and almost 24/7 kind of ability to run. That's why agentic AI becomes a really good solution. And we see here is one where it is more and more custom than kind of an industry-specific solution that you would see in in other kind of G&A functions, be it in HR or finance and P2P or source-to-pay or customer-to-cash kind of solutions. Here, it'll be much more customized where you build those agents, agentic AI solutions.

Paul Morrison

Interesting. I think it's a very good point you make around the fact that this technology doesn't just relate to the very largest players in the space, although I think most of the eye-catching stories and use cases at the moment do come from that direction. I think maybe in some of the other areas we can touch on, we can see the ability to tap into agentic with relatively limited spend in that. As you say, in that supply chain space, maybe it is more focused on around some of the big players. The other case studies that caught my eye – one was that Levi's that was using, again in inventory management, but using sensors, smart-shelf technology to get an extremely up-to-date, real-time view of stock, and then linking that automatically to the replenishment. The staffing models as well will be connected to that. So, another complex integrated system.

Another one on the CPG side is Unilever that has been developing a demand sensing platform, as it's called, that operates agentically. So, multiple sources of input around diverse things like media sentiment or weather patterns, crop conditions from some very specific data providers, and again, feeding that into the planning and inventory systems. That's achieved some significant inventory savings and increased that availability above the high 90s again. So, some real value created there.

Vin Kumar

Yeah, it's a combination of all the IoT. Now that you have the data, putting it on top with the Gen AI kind of models to analyze it. Now you have agentic AI to go and execute certain things. So, it's a combination. Now it's all of them coming together, and I think that's where there’s huge opportunities.

Paul Morrison

Exactly. Well put. It's all coming together. So, maybe you could tell us about order management, and what you've seen in that space. You mentioned it earlier.

Vin Kumar

So, we're seeing in order management a lot on pricing fulfilment to do. So, you're getting the orders, you're getting dynamic ability to price, giving pricing guidance to your account sales, your business client, customer focus, staff – all this information to help them price, configure and then execute the order to do, is something that we are seeing.

And usually, what happens is there are the account sales folks focused on the five to six top configurations that they have to that they work on because that's what they're familiar with. But there are other permutation and combination which gets lost. Having this kind of agentic AI to - hey, based on these requirements, what is the best configuration ability to go and recommend configurations which will be the best suited for that delivery of that order. They could be taking real-time information and executing on that. That is what something that we see.

Vin Kumar

The second part of that is on the fulfilment side. Where do we get these custom orders to be? It may not be for the scale of maybe a CPG, but in a B2B case, where orders need to be, a product needs to be designed and constructed based on the configuration. Which plan do we want to do, and where do we see the where we have stock, where we have impact of trade being reduced there, and how do we design it. So, that part of where do we ship this particular order to get manufactured for the product is also something that is being looked at with the with agentic AI solution. And then, all about later on in the dispute management comes the ability to handle dispute from your customers, all comes into an e-mail, it comes in and then how does it get triaged? Is it pricing, is it quantity? Is it chip date? Is it deduction didn't get applied? Is it anything to do with the payment terms ability to take that dispute which is coming in and triage it and go back and fix it?

That cycles and again, we're seeing a lot of opportunities.

Paul Morrison

Great stuff. So, there's three or four areas we've looked at. Earlier in the call, you were suggesting, the G&A world around finance and HR as well. And I'm working with colleagues in finance around our agentic platform, Track 1F.

Seeing some of this in action, I think we need another call to go through all the different areas. I mean, do you have any sort of headline perspectives on the finance space or the wider?

Vin Kumar

We do see big opportunities in the finance space, especially in finance operations across the three areas. The record-to-report, customer-to-cash, purchase-to-pay and plan the results. Those four areas is where we are seeing agentic AI being used. So, where they were struggling to leverage Gen AI because Gen AI tells you what it is, but it's not actually doing the work. And then they had maybe higher expectations from RPA and realize that we have to be extremely prescriptive. They see agentic AI bridging those two gaps. Now, it understands what are my requirements, understands how do I need to process the transaction, and now I can go and process it and execute it.

This is what they're expecting to give the savings that we are going to see in finance and those numbers that we publish is where we expect that to come in.

Paul Morrison

Yeah, it's those are big numbers, aren't they? When you're talking about more than 50% efficiencies. So, it's an interesting area, and I mean another area that we haven't had time to touch on is around marketing. It's a topic on previous podcasts we've talked about in terms of the automation of marketing operations and the enablement of hyper-personalization, and that's definitely an area. So, I've talked previously around one of our retail clients, where we were able to massively accelerate the personalization of marketing communications.

And to improve click through rate, that was originally a Gen AI use case and since we’re talking about it, it's been connected and re-platformed. So, it is much more autonomous and agentic. I think we might start to see some of the early pilots and use case around Gen AI being plugged in in a way that makes them more efficient and more impactful in that way. So, lots to go on. We're going to have to start to move towards closing the call. I think what one key topic then is, with all this opportunity, is what's to look out for. What is difficult here? Why are there not hundreds of deployed use cases already?

Vin Kumar

I think it's too nascent yet. People say, “OK, why haven't we started using agentic AI?” I try to remind them that this is more of something that enterprise have started looking at and trying to pilot and use it this year. It's not something that has been there for a long time, so it's a nascent. Secondly, this combination of we can't rely just on general purpose knowledge.

We need to have the domain to actually start doing. Otherwise, the use cases today, we see agentic AI is all about translation. It's all about content creation, but because that can be done on general knowledge. But actually, when you have to process you need to have the domain knowledge, and we need to have your data, all of that, and that's where we are going to see. So, this is just beginning of the agentic AI revolution, I would say. It's in a two year kind of sprint is where we'll start seeing more and more coming in. All these vendors are doing, and it's not put one solution and that's going to solve all your agentic AI. That's the other. Also it's going to be a suite of how do you use the solutions for agentic AI. So, that is also making it complex - which one you do and there is a lot of structural change happening, you had one for a agentic AI. I think it was one blanking it out, which then got acquired by ServiceNow, which is really good, and we saw good progress. But before it could take off, it got acquired for billion plus by ServiceNow. So, you want to know which one to invest in which tool sets to have. There's a lot of changes in evolution going on that we may want to get some stability before we can actually wrap up.

Paul Morrison

I think those are really great points. Absolutely right. This is a very new set of possibilities we're talking about. I completely agree with that. I agree around the point on domain. I think there's a knowledge gap or a skills gap. Making the connections and making the functionality we're talking about work needs quite a complicated mix of process change, technology, leadership, communication skills to come together. So, there's a lot that needs to happen for this to be successful, and I completely agree with that perspective about agentic on its own is not the answer.

It's really an embedded thing. Our clients and our listeners are in the markets to solutions to their problems and that are not for agentic AI.

It's not an end in itself and we recognize that. I guess the reason for having this conversation is that, this wave of technology change will shape the possibilities and the economics of all these processes. So, it's something that we recommend everyone keeps their eyes closely focused on. Before we close, any final tip or if there's one thing to think about?

Vin Kumar

I think, one thing is be more open minded to kind of experiment and try this out. You don't need to have too many proofs and pilots and all of them. But be open minded to try and say, “Hey, leverage, what's going to be coming in your best of breed solutions.” They're all investing it, leverage as much as you can from there, but also have a platform or studio where you can custom design your own and for to do it.
If you select, have that generic one where you can custom design and every time you're looking to acquire a new solution for your end-to-end. Or if you already have it, talk to them and because they are all investing and making it available for you to start seeing them.

Paul Morrison

That's great advice, Vin. We have to finish there. Lovely to talk as usual.

Vin Kumar

Hey, thanks a lot, Paul. Great conversation. We shouldn't have so much fun on these podcasts. Take care. Cheers. Bye.