A Blueprint for Enterprise Agent Adoption

0:00Today on the AI Daily Brief, a blueprint for enterprise AI adoption. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. To join the conversation, follow the Discord link in our show notes.

0:18Hello, friends. Today, we once again have a slightly different type of episode, but one that I'm really excited about. It is undeniable that the biggest theme this year for most enterprises, or at least the most exciting theme to most enterprises, is agents. I have a whole slew of theories around why I think agents have businesses thinking differently even more than perhaps some Gen.AI assistant type tools have. But in this conversation, I'm joined by Swami Chandrasekharan, the head of the USAI Center for Excellence for KPMG. Rather than just a general overview of agents, this conversation comprises a part one of something of a blueprint for thinking about enterprise agent adoption or at least testing.

1:00Swami shares his taco framework, thinking about different types of agents broken down as taskers, automators, collaborators. We discuss the most common challenges that he's seeing among enterprises trying to adopt agents. And ultimately, we try to provide some positive steps that you can take as an enterprise to advance your agent strategy. We certainly don't get through an entire blueprint for an agent strategy. We will have to have Swami back to keep going on that. As you'll see, Swami is definitely not your standard consultant. He has a deep technology background, working previously as an executive architect at IBM's Watson, among other roles.

1:32He holds more than 30 patents and has authored multiple books and articles on applied AI. And so without any further ado, let's dive into this conversation. All right, Swami, welcome to the AI Daily Brief. How are you doing, sir? Doing good, Nathaniel. Thanks for having me over. Big fan of your show. Appreciate it. Yeah, we were just joking. So up until about, I don't know, 24, 48 hours ago, we were talking about the hottest topic in AI. I think for a very brief moment, that's been displaced maybe by DeepSeek and R1. But broadly speaking, I think this conversation about agents still is pretty down the middle of where a lot of people are thinking.

2:09Maybe before we get into it, though, I would love for you to just share a little bit about what you spend your days doing that gives context for this conversation. So I live in Dallas. I'm a partner at KPMG. I lead the AI and data labs for the firm. And what that actually means is as part of the large transformation program we're running called AIQ, that Steve Chase runs, the AI and data labs is a pretty significant part. The way I explain my job to my 13-year-old is I do three things. When I say I, it's me and the team. We do a lot of experimentation. So for lack of a better word, we don't have a full-fledged R &D function at the firm.

2:49So we do a lot of experiments, innovation, R &D around things that don't exist today but will exist tomorrow, whether it is around how do we use language models or how do we build advanced RAG, knowledge assistant techniques or even agentic frameworks or how do we evaluate these models. The second part of what I do is I help establish standardization when it comes to technology, architecture and patterns for AI. across the firm. So we don't do the same thing five times. And the third part of is, given my history and being in the advisory side of KPMG, I work with a lot of folks in co-incubating new things for our clients.

3:30So I get closer to clients and understand problems. So I don't get too disconnected from what I do. So nutshell, I have, I think the best job in the firm and a lot of fun and a lot of responsibilities as well. Awesome. So perfect setup. I think a lot of the conversation today is going to be about the practical, factual kind of where we are with agents and understanding where you're sitting, especially relative to clients, is useful. Let's actually start there with that question. When you think about 2025 as relates to agents, they are obviously a key theme. They're on everyone's mind. But where are we actually when it comes to agent adoption, particularly in the enterprise, right?

4:14What stage are we at? And let's start there. And there's a lot of branching questions that I have from there as well. Yeah. Let me quickly set the context, no pun intended, right, but to get to web agents. So when large language models came out, we started interacting with it with prompts, ungrounded interactions. We loved it. And then we slowly started to bring in more context to longer prompts, p-shot prompting, and so forth. Then, thanks to meta, we have this approach with retrieval augmented generation where we said, look, why don't I intercept the prompts and go to a corpus, bring back the relevant chunks, and give it to the model.

4:50So we got our arms and ourselves wrapped around, okay, now I understand the concept of RAG or what we call knowledge assistance in KPMG. But still with both of these paradigms, you are sitting and typing prompts. You are away, you are doing it, you may end up doing lang chain type chaining and those kind of things, but you are still typing prompts. There is the action. So agents come agents. The whole concept is, can I have these machines go given a larger goal can these machines go figure out and plan and go take actions so whether it is researching on a topic or whether it is by reconciling a balance sheet against my ERP systems it's now starting to do things so what fundamentally makes agents are how will you define your instructions your goals expressed as instructions long form prompts how well those prompts are reasoned and understood through a planner into tasks that you have to perform.

5:57And to perform the task, what tools I need to do the job. Then there are things like knowledge, memory, and context and all bunch of things. So fundamentally, it is giving the large language models not only additional tools, but the ability to do reasoning in the context of a goal or adjacent set of goals you're trying to achieve. Okay, Swami, you gave a very good theoretical definition. What does it mean? If you look at what is possible today, all the things I've been explaining are possible in a way using frameworks like Lanchain and Lama Index and others where you can deterministically chain those steps.

6:41For example, if I want to reconcile a balance sheet, I may have two brick functions. Each function may have a long form instruction. I make that execution of function one in Python, give that output to section the second function, and I can achieve it. There's nothing truly agentic about it because you are hard coding the steps. The true agentic behavior is going to be where I express, for example, balance sheet reconciliation. What do I do? As an expert, I say a balance sheet will have these following fields. I look for the following parts in the balance sheet input. Then I go to an ERP system and I do certain things.

7:21So you are expressing that as how a human expert will express. The question now comes, can any large language model even reason and understand what you're saying? Probably until like six months ago or maybe a little before that, they were not. It was very hard for them. And over every iteration of the language models that came out from all the big tech, the reasoning capabilities and more importantly, longer instructions, longer prompts, they begin to do pretty well. Even if you go back to three, two, three years ago, these longer instructions were impossible to achieve. Right now you can do it.

7:59So what you have, the ability is better reasoning, better understanding of what you're saying through these long-form instructions that are very critical for reasons that was not possible in the past. So what leaves this? What does that leave with us? So you can understand instructions well. You can break them down into tasks, probably. And now it then comes to, can you rely? Are those tasks that are broken down and the tools that are used for those tasks, are they reliable enough for you? The answer to the jury is out there. The jury is out there in terms of the tools and platforms we have tried and worked with.

8:40It requires a bit of handhold. Well, the language models can reason. The act of turning that into a set of tasks, apply instructions, and to go execute, it's getting there. It's getting better. But long story short, what we can do today is simple agents. I have come up with a simpler definition or a simpler four ways to define the types of agents you can do. Acronyms, TACO, taskers, automators, collaborators, and orchestrators, which is what I agent orchestrate. And one thing about TACO is people differentiate between, I've heard people talk about, oh, certain agents don't get to access all tools.

9:24My thing is, in the TACO framework, all the categories of agents, four types of agents, are going to get access to the same knowledge corpus. It's going to get access to the same breadth and depth of tools that the agents would need to create actions. It will have access to memory. It will have access to the same algorithms. So all of those four are fixed. So what is different? The difference between the four comes down to planning and autostradition. The T in taco taskers, they're singular goals, one goal, but can break down to multiple tasks. They can be chained, easy to manage, easy to test, easy to roll out.

10:06When you go to automators, which is the next, they typically go to cross-system, cross-application. These are end-to-end processes. Order to cash, lead to cash, procure to pay, hire to retire. They touch multiple applications and multiple systems. So the goal may be similar, meaning ensure streamlined order-to-cache process execution, but they break down to sub-goals. Each of the sub-goals may touch different applications in different systems. So it gets a bit complex in terms of the scope of what it does. Planner gets complicated. Orchestration gets complicated. In the orchestration, you have to manage state.

10:50and all these things. The third part is collaborators. This is where I've been pondering over the question. So there is this concept of can AI be used as teammates, agents be used as teammates. They're no longer you're telling the agent to do something, it comes back. You work with it. It's like how you work with your team member on a daily basis. So there's more skewness towards human collaboration, partnering with the machine to get things done. It's there in the other forms of agent but it is even more so in this is just predominantly built in and the last bit oh in the in the taco is uh the multi-agentness where i have agents calling other agents there is interagent collaboration of course the complexity becomes more with all of this so like i said earlier where are we today i think um there have been a lot of experiments prototyping done with the taskers.

11:45It will have been inherently because there are quite a few platforms, open source commercial included, where you can build them quickly. And we can talk about that. But I think those are, in the year of agents, if 25 days, I would see more taskers. That's my prediction. Do you think, it's obviously very dangerous to sort of prescribe one right path without the context of any given organization. but do you think that that toggle framework actually is basically, are they four separate categories only or do they have some sort of linear relationship with one another as you're thinking about adoption, if you're sitting in an enterprise where, you know, it makes sense to start with taskers and then move to the next or, you know, how do you think about that?

12:31Yeah, this is not, I don't want this to be a contrived framework where we retrofit everything into one of these four. The framework is meant for a mental model, mental picture. look how can I break down agents not everything because the reason for this was everybody jumped into multi-agent coordination without even thinking about the basics so that is what second is more than likely when you go talk to clients they're going to talk about scenarios which will not only overlap but will require their focus maybe more than likely starting with okay let me do end-to-end all process automation automators because that's what I need I want to streamline my store performance management or i want to streamline my procure to pay process or when you go to another client may say look i'm more focused on augmenting my human potential so give me an ai agent that can act like a teammate for my ap ar erp finance kind of process domains so yes it is dangerous to put everything's into the bucket but that's not the point the point here is to demystify the whole agentic system and how complexity comes and if you start to amalgamate and combined, that's okay.

13:40But at least you understood the individual's feedbacks. Yeah, it's interesting. I think that one of the things that makes agent adoption fascinating as compared to, for example, sort of broader Gen.AI adoption over the last couple of years, enterprises moved very quickly relative to previous technology changes to grab onto Gen.AI and try to sort of harness it. Now, obviously, there's still tons of organizations that are behind, feel behind. Very few organizations, I think, we tend to find that the organizations who are the farthest ahead also have the greatest awareness of how much more they still have to do when it comes to adoption.

14:17So it's not like they're sort of at the end state or anything. But I do think that because they've been watching agents come down the pipeline for a little while, they maybe have a stronger sense in general of how they want to eventually use agents, the possibilities that have them most excited. And I think that it might be leading to some of what you're seeing around, they're jumping to sort of exactly what they would like out of an ideal state of what an agent can do. They're imagining even ahead of where the technology is, rather than sort of just racing to catch up with what it can do now, which can create challenges just based on what's actually ready for primetime and what's not at this exact moment.

14:55yeah everybody has a an expectation and a notion of what this agent should be for them if you go look in the customer service and marketing function they say my version of an agent is can i put a can i put a digital version of a customer for a software development or sales development representative and it can talk to clients it can negotiate it can ask the questions It can help close a sale and get paid a commission and come on. So they start thinking it like synthetic employees. You go into the enterprise, you go into the mid and back office functions. They think in terms of processes. There is a particular way in which I receive, review, approve, or deny invoices as part of my larger procure-to-pay end-to-end process.

15:48So I have a conception of how agents should be in that particular way. It is not one size fits all, like you said. But at the same time, the key responsibility is when you go talk about them, you're not trying to take an existing technology and retrofit and say, oh, I have agents. So as an example, one belief I have is good old business process engineering, like how you sat and designed business processes for end-to-end processes, it took a particular approach. Process engineering came out where you said decompose your domain, break them down into level 1s through level N, or go up to level 7 and 8, where you kind of have a massive swim lane view of how your process looks like.

16:37That's how we represented processes. That's not how machines think. now with the reasoning capabilities I could express that same thing almost like a long form instruction and you leave it to the machine to say look you go to find the process the steps that are needed to execute it so there is also a change in how we approach designing the agents that is also essential and important the outcome is the same I want a better efficient leaner process. But you're approaching it in a different way. So the point being, the entry point for agents are different. They're all going to converge at some point in time, but given where we are in the stage we are in, the expectations are widely done.

17:30Today's episode is brought to you by Vanta. Trust isn't just earned, it's demanded. Whether you're a startup founder navigating your first audit or a seasoned security professional scaling your GRC program, proving your commitment to security has never been more critical or more complex. That's where Vanta comes in. Businesses use Vanta to establish trust by automating compliance needs across over 35 frameworks like SOC 2 and ISO 27001. Centralized security workflows complete questionnaires up to 5x faster and proactively manage vendor risk. Vanta can help you start or scale up your security program by connecting you with auditors and experts to conduct your audit and set up your security program quickly.

18:10Plus, with automation and AI throughout the platform, Vanta gives you time back so you can focus on building your company. Join over 9 ,000 global companies like Atlassian, Quora, and Factory who use Vanta to manage risk, improve security in real time. For a limited time, this audience gets$1 ,000 off Vanta at vanta.com slash nlw. That's v-a-n-t-a dot com slash nlw for$1 ,000 off. If there is one thing that's clear about AI in 2025, it's that the agents are coming. Vertical agents by industry, horizontal agent platforms, agents per function. If you are running a large enterprise, you will be experimenting with agents next year.

18:53And given how new this is, all of us are going to be back in pilot mode. That's why Superintelligent is offering a new product for the beginning of this year. It's an agent readiness and opportunity audit. Over the course of a couple quick weeks, we dig in with your team to understand what type of agents make sense for you to test, what type of infrastructure support you need to be ready, and to ultimately come away with a set of actionable recommendations that get you prepared to figure out how agents can transform your business. If you are interested in the agent readiness and opportunity audit, reach out directly to me, nlw at bsuper.ai, put the word agent in the subject line so I know what you're talking about, and let's have you be a leader in the most dynamic part of the AI market.

19:33Hello, AI Daily Brief listeners. Taking a quick break to share some very interesting findings from KPMG's latest AI quarterly pulse survey. Did you know that 67 % of business leaders expect AI to fundamentally transform their businesses within the next two years? And yet it's not all smooth sailing. The biggest challenges that they face include things like data quality, risk management, and employee adoption. KPMG is at the forefront of helping organizations navigate these hurdles. They're not just talking about AI. they're leading the charge with practical solutions and real-world applications.

20:04For instance, over half of the organizations surveyed are exploring AI agents to handle tasks like administrative duties and call center operations. So if you're looking to stay ahead in the AI game, keep an eye on KPMG. They're not just a part of the conversation, they're helping shape it. Learn more about how KPMG is driving AI innovation at kpmg.com slash US. What do you think about as you're advising clients or even just thinking about it broadly and you're thinking about agent readiness in the enterprise, what are some of the pillars of consideration? How much is it about data? How much is it about policy?

20:41How much is it about understanding objectives as you've just articulated? What are some of the key pillars of agent readiness? Yeah, you kind of gave our three out of the things I was going to say anyway. So first of all, why agents? I start with that question.

21:01Well, what is the rationale? What is the motivation for agents? So first, don't go to technology called agents yet. What is the problem you're always trying to solve? So if I'm a client, if they're a retailer, they come and say, you know what? I want better top line growth increase in my stores. in my brick and mortar stores. Okay. What are you doing today? They say, okay, I have these things, but the sales get affected because certain stores don't follow certain kind of policies and procedures. They don't take into account customer satisfaction or customer reviews and all those kind of things.

21:42Okay. Then we go in and say, okay, the goal and objective is to have a better, more tangential approach to how do you do store performance analysis so you can improve the performance and increase your top line and be on. So number one is what I'm trying to do and is agents even the right answer? So let's assume you've gone down the path of saying, look, I want to optimize my processes, reimagine my processes, at the same time optimizing my human resources. Then you talk about, okay, where is the data coming from? Do you have the data? do you have access to all of the data? Have you even, first of all, instrumented the data if it has to be digitized?

22:25And is that data made, is clean, and it's all the good things about data availability and readiness and everything. The third one is, I don't think you mentioned in the list, Nathaniel, which is, who is the human expert who can articulate what is happening today and what needs to change? How are we going to elicit that knowledge? whether you pick a domain, you pick any intense domain. Even if it is something as simple as customer service, from the point a customer comes and raises a request for refund, what do you do? What is the process you follow? And what is the way to reimagine from that point onwards using agentic concepts?

23:04So human expertise is still needed to articulate. I mean, there are theories floating around. Can I go do simulation? Can I look at what humans being done and learn from that? From that, yeah, you can, but they're not fully reliable yet. So why agents, data, human expertise, articulating the whole thinking process and how agents have to be built? Then getting into policy-driven things. Okay, how much of autonomy you want to give to these agents? It could be at a very broad stroke principle level saying, look, I don't want any decisions that have a financial implication to be approved without human in the loop.

23:49Maybe I want three stages of human in the loop. So there is a whole strategy around how do you bring in humans? Where do you bring them in? Where is the level of oversight? What does the kill switch equivalent for agents look like? What if you want to stop agents for a day? what is your fallback mechanism in case these don't start to work? So all of those policy, trust, security, reliability aspects is one big bucket. The fourth important bucket is everybody, and this is a very opinionated topic I've seen with clients, is how are you going to build agents? Okay, everything fine. You've got the data, you've got the experts, you've got policies.

24:28You know how to build them. Where are you going to go build them? So today, there are a dime a dozen open source frameworks. The big tech, small tech startups, they all have their platforms. So where do you go standardize and build? Again, my thought process there is until this whole thing settles down, you may have to remain in polyglot and pick a few choices, be very opinionated and go build and try them out. And some are going to work, some are not. So you have to be ready for consolidation and merging. So what is the tool technology infrastructure that you're going to go to? I'm not even using LLMs, I'm assuming they're going to get awesome.

25:05They are awesome already. They're going to continue to get awesome. And the last bit is around skills. Do you have the skills to build this? And one more thing after this. Okay, you have the skills. Building agents is one thing. The day two plus operations is a completely different thing. How are you going to sustain? So we've talked about model drift and data drift. Now comes agent drift. What's the guarantee the agents are not going to drift? It's going to deviate away from what it was built for. How do you keep them, upkeep them? Is the data changing? How good of a feedback are you providing back to it for reinforcement?

25:41Those all come in the day two plus operations. So top of my mind, I think these are the kind of categories of things I would look at. And do you have, I think it's a really useful framework. how much do you think people's how much are you seeing people's first experiences being something that they're kind of rolling their own you know with one of these general frameworks versus trying something that's more off the shelf i mean this is kind of only a question for the last few months as more off the shelf things have been available but you know working with a customer service agent or or is this does this have to do with which category of agent to use your framework they're actually thinking about yeah so if you double click into where are you building agents I think it double clicks into three sub-questions or sub-areas.

26:30Are you going to build your own using open source? Are you going to take a commercial platform like a co-pilot studio or agent space? A third option is, are you going to buy the agent? So you go to Agent Force is going to say, okay, I already have a sales coach agent. You're just going to buy it, configure it, and use it. The experience is changing by the month. what we have today is not what we had six months ago again, there's another one the way I look at the whole agentic tooling spaces there is low-code tools like Copilot Studios and those kind of the work then on the far right, you've got the ProCode tools like the Langer apps and Crew AI and Autogens of the world then in the middle, I call them mid-code you can go back and forth I can write code, I can write in GUI, drag and drop so I can do both.

27:26Initially people tend to go use the pro code options and they realize while it gives them a lot of flexibility they have to end up building a lot of things on their own. So there's a lot of lines of code to write and maintain and manage. Brittleness starts to kick in unless you have a well coordinated engineering team, development team, you may end up recreating the same thing. For example, the same tools to do the same thing may get recreated multiple times. So there is that risk of having, and you need to have a special set of skills and capabilities to do coding by yourself. Now, if you come to low code, I mean, I could get started quickly, very easily, but I've seen roadblocks where you say, oh, I want to do this Excel comparison for one of the steps in my agent, and I cannot do very deep Excel analysis because my Excel has multiple complex cells and rows and headers, as an example.