AI at Scale: Revolutionizing Enterprise Strategy and Growth

When Chetty PT phenomenal first happened at the beginning of 2023, everybody was scratching their head. Oh wow, this is so disruptive, there's no playbook. But I think now, two and a half years later, we have learned a lot, we're still learning. Now it's our job to go to all this client to say, "Okay, let's focus on the how." Hello and welcome to the season finale of Where AI Works, Conversations at the Intersection of AI and Industry, brought to you by Wharton in Collaboration with Accenture. I'm your host, Karthik Osanagar, Professor of Operations, Information and Decisions at Wharton, and co-director of Wharton Human AI Research Initiative. It is our goal to cut through the noise to deliver actionable insights for business leaders by combining cutting-edge research with real-world case studies. Things are changing fast, so without further ado, let's dive in. What we see in the data is that the executive urgency to incorporate AI is at an all-time high. There's been a lot of buzz, maybe even hype, around if you don't adopt this now, you're going to be left behind. Business is a game, and if you want to compete, if you want to win, you have to use AI the right way. In this final episode of our season, we're going to transition from AI strategy and use cases in marketing to implementation at scale. To help me unpack the challenges and opportunities for business leaders, I'm thrilled to introduce someone who has helped lead her company to its fastest growth in its history by booking over $3 billion in generative AI-related business last year. Lan Guan is the Chief AI Officer at Accenture. Lan, welcome to Where AI Works. Thank you for having me, I'm so excited to talk to your audience about AI. There's really a lot to talk about today, but where I want to start off is with your role, your title Chief AI Officer, it's a relatively new role in organizations. It's not too long back, I remember even the CTO and CIO roles where new in organizations until then IT was like a back office enabling function, and then it suddenly became strategic to drive business transformation, so what is the role all about and how did it get formed with an Accenture? So, I've been playing this role for a little bit over two years now, first and foremost, I'm responsible for our company's overall AI strategy in the market, so that means things like what do we want to be famous for, what do we invest, and how do we work with ecosystem partners? I'm also responsible for developing highly differentiated industry solutions, assets, platforms, accelerators for Accenture so that our practitioners can be bringing this kind of tangible capabilities to our cross industry clients. Last but not the least, I'm also responsible for leading 65,000 data and AI practitioners. It's different from the typical IT role because I always talk about AI comparing to data, comparing to cloud is so much closer to the business. Make sense, I think it's so exciting that you get to help define what this role ends up being not only within your organization, but over time other companies will draw inspiration and they'll figure out what that role means for their firms as well. What I want to really take this conversation to is understanding what is the current status of AI implementation today, where are enterprise AI investments currently being directed, what kinds of applications are companies buying and building. I was wondering if you can talk about that both in terms of within Accenture, but you work with so many clients with their AI implementation, so what does it look like across hundreds of thousands of companies? Yeah, sure. Other times we can ask, is this reality or hype, right? Right. So I think I can be very upfront, right? Here in this case, given the results that we have seen, this is quite real, right? You mentioned 3 billion revenue that we have been booking for Gen AI. In fact, it's 5.6 billion over the last 18 months, okay? Oh, amazing. That's amazing. It's all public. Good for you. Well, we have completed 2000 Gen AI, now agent AI projects. The progress that we've been making in the market is just enormous. Let me also give you a couple of data points here. So our latest research shows that 86% of the C-suite clients, they're actually planning to increase their investment in AI. 60% actually expect their Gen AI solutions to be scaled across their organization up from 36% in 2024. So I think the point here is, this is real. They are spending money on this. They're shifting their span right from other areas or into AI, or they are just increasing the net new spend. Another piece of the research we have done, those will indicate it less than 10% of the clients have actually scaled. In fact, the number is 8%. So everybody's asking, okay, what happened to the other 92%, so it's our responsibility to collectively go help them. I think the second part of your question is, what are some common use cases, right? So we actually tell our clients that you should be thinking about your AI investments in two categories. The first category is what we call table stake, right? Think about those are no-brainer areas that every company should be investing in. For example, in marketing space, the second category is what we mean by more strategic bets, right? Yeah. These are the areas that tend to be associated with the client's industry value chain, okay? Like, for example, in life science, one strategic bet that we've been investing in is actually helping a lot of life science companies to speed up their clinical review process, leveraging AI, leveraging data so that they can release the new drug faster, right, into the market. In Tauco's case, using Gen.A.I. and Deep Dura Network to help improve their network operations in banking industry, for example, deepen the bank's ability in the anti-money laundry in all these areas. So I think holistically, we believe this technology is right for enterprise ring mentioned. And we go to clients, right, very strategically help them identify low hanging through area, table-stake areas so that you can reap the benefits faster, but at the same time, also help them develop this kind of a roadmap so that they can see continuous ring mentioned, continuous streams of ROI coming in by also tapping into strategic bets. I think one of the things that companies have to contend with is a problem of planting in this space, meaning that almost every exec, especially CTOs, I talk to talk about how they are being bombarded by lots of new vendors who have new AI-powered substitutes for software that are already used or new workflow automation kinds of solutions. But how does one think about where to prioritize one's attention with adoption decisions? And especially on the strategic side, I'm understanding that on the table-stake side, yes, you have opportunities to increase efficiency, you should certainly go pursue those, but I'm really curious how you advise your clients or how should one think about how to prioritize the attention when you're being bombarded on a daily basis with new AI software. Yeah, people can easily get lost, right? Even just in the year of 2024, we counted more than 200 large language models being developed by all kinds of companies, not to mention now it's all about agent AI. So I think from the see-through perspective, we basically asked them the question, what is your strategy? What are you trying to achieve? Are you looking at AI as a competency enhancing technology or something that is actually competency diminishing? So this is the kind of question that's almost like the light bulb moment to them, because it's very easy for people to get enamored by this kind of a tool conversation and then forget about what is the so what? That's why we lay down five tricks that everybody should be talking about. I think the first one is, like I said, focus on the problem, not the tool. What is that strategic needle that you want to move? What is the competitive advantage that you want to be famous for? The second one not to get carried away by a lot of this technology conversation, so you stay calm, you stay focused, it's actually about taking advantage of this kind of platform based approach as opposed to chasing all kinds of use cases. Right. So let me tell you one example, a very large energy client that I was working with in Middle East. They're basically telling me that within their company, because of the size, because everybody is so interested using Gen AI on the monthly basis, they are getting more than 800 use cases coming from all parts of the company, everybody's submitting this kind of use cases. So they came to us and say, hey, should we be catering to all these use cases? So their approach is actually quite smart. They look under the cover with our help across all these use cases. They realize that there are a lot of the common building blocks, like for example, most of these use cases involving working with some kind of unstructured data. So they had that light bulb moment, hey, in this case, why can I just invest in some kind of vector database? Why can I just go build on the enterprise level, some kind of vectorized embedding pipeline, not just working with the texture data, also working with multimodal data, so that this kind of capabilities become the foundation of the platform. Then every single use case from this repository that they're getting from their employees and different business functions is leveraging this industrialized commonly governed capabilities. So we're seeing this across the board now. Sometimes I say, okay, when chatty PT phenomenal first happened at the beginning of the 2023, everybody was scratching their head, oh, wow, this is so destructive. There's no playbook, but I think now two and a half years later, we have learned a lot, we're still learning. Now, it's our job to go to all this clients to say, okay, let's focus on the how and look at this kind of platform based approach, look at this kind of reference architecture and start investing in this kind of strategic bad. I think that's something that is super, super important. So you mentioned focus on the problem. Don't get carried away by the technology. So that's two. I think you said there were five. I'll quickly go through the rest of them also be pragmatic about the embedded AI. A lot of CIO CTO clients. The first question they tell me is I've been investing in a lot of technology in the past. So how do you make sure that Gen AI or agent AI is working with my existing technology landscape? If they have the enterprise platform, if they are working with Salesforce, if they are working with SAP or ServiceNow, they all have embedded AI capabilities. So they work quite well with the data that is specific to the platform. So I always tell my client, get started, turn on those capabilities. So that's the third one. The fourth one is that you need to be demanding substance from the vendors, okay? Meaning it's not just the buzzword, right? It's focus on asking the tough questions, right? How did it work? How is that trained? What are the differentiators between your agents and the other agents, right? So I think asking these kind of questions from all kinds of vendors is super crucial. The last one is actually about aligning with your data and talent readiness. The old phrase of garbage in garbage out is becoming even more relevant in this case. So I have seen so many examples, so many clients, they are excited to start AI. They thought that they are on the right track and then they hit the road bump, right? Because their data is not ready. Their talent is not ready. Let me give you one quick example. One of the very large telco clients I was personally working with is exactly what's happening to them, right? They are very anxious to implement contact center agent solution. And then right at the moment of pushing this into production, they realized the model performance degraded, right? Drifted. Then they realized, oh, it was because the training data that they use did not have the good quality. They actually have 37 versions of standard operating procedures for their human employees on the floor. If it's a standard operating procedure, SOP, how can you have 37 versions and supposed to be one single source of truth? What we are dealing with every day is this enterprise mess in it, meaning data is messy. The process is messy because of all kinds of merger acquisition that is happening, right? Talent shortages, right, work with a lot of clients, they basically have been investing in the last two decades, right? Building data science talents, building data engineers. Now I'm going to them say, did you know your data science talents needs to get re-skilled? Because now in order for talents to thrive in the age of AI, they need to learn software engineering. They need to learn design. They need to gain this kind of a full stack capabilities, right? So a lot of times clients say, okay, that is a gap, right? How can you help me? Yeah. It's so interesting you say that it makes sense, but at the same time is something that companies haven't necessarily paid a lot of attention to this point about to get those returns in AI investments. You need first the investments in data and the talent. And in fact, at Wharton in our center, we have several studies. My colleague Sanit Ambe has studies that have looked at AI investments by public companies and then their returns in terms of stock market returns, their revenues, profitability and so on. And it's pretty clear that firms that are getting those returns, their AI investments are preceded by the data infrastructure investments. And they also tend to be companies that are in regions or geographies where there's a large talent base that they can recruit from. And I heard you mention multiple times agents. Now in the previous conversations that we've had on this show, it's come up, but nowhere near as many times as you, which makes sense, you're the chief AI officer. So we should spend some time talking about that. A lot of people believe agents will transform software as a service itself, because now you can replace traditional software applications with agents that are able to take action, listen to a sales call, analyze the call and take some follow up action, like sending a follow up email to the customer. So even the human interaction with software sometimes becomes irrelevant. There's so much hype around this, but it's also not a mature space. It's very early. So I'm really curious to hear what kinds of experiments are you running in the agent space? And what are the big challenges and roadblocks to deploying large scale agent base systems and organizations? So Karthik, we are the early adopter of HTTP AI. So one of the best examples is actually the marketing ring mentioned that we were doing two hours off, because as a large organization, we have lots of marketers, B2B marketing is something that very important to us. Last year, we built AI refinery together with one of our partners in India in this case. And we built 15 agents to basically reinvent the strategic planning, because we believe that's the area that is almost like the crux of the marketing, right? Because it involves a lot of cognitive task, it involves a lot of appending tasks, a lot of coordination, right? All this characteristics that matches the agent's capabilities. So let me give you one example of the agent, right? Utility agent level, like calendar agent. Calendar agent wise, that's important, because in B2B marketing's case, we always have multiple campaigns, right? Coming out multiple campaigns in multiple geographies. So how do you avoid the crashing calendar so that we don't have campaigns with a different objective, different messaging, hitting the same audience? Today, it's usually done by people, right manually, creating marketing campaign calendar. Oh, okay. This sounds like a classic assignment problem, right? And we build operational research, mini agent to solve, resolve a lot of this calendar conflicts to optimize audience impression, optimize the marketing budget to make this much more comprehensive and much more optimal in terms of the challenges, right? The second part of your question, I think there's also three things that I think is slowing us down. The first part is because of this confusion around what is agent, how is that different from RPA, should I be using agent together with large language models, especially thinking about reasoning capabilities? I think we as industry, we need to come together and actually put the state in the ground to say, okay, this is the definition, agent needs to have this kinds of four characteristics, agent needs to be able to perceive the outside world, agent needs to have the cognition reasoning capabilities, do troubleshooting, do diagnostic, breaking down larger problems into smaller chunks. But that's not it, also taking action to using a function calling, giving the tooling part of the agent capability is super critical, right, connecting with your enterprise systems. Last one is learning, right? Because ultimately, agent is not perfect. So how do you institute this kind of self-reflection so that agent can continuously learn, be environmental aware? You don't need to cling way, very crystal clear way to describe the power of agents so that a lot of noise will actually be not there. I think that's one challenge. The second challenge I would say is agent interoperability. I still think that there are a lot of gaps that we need to be closing. We don't want this technology super powerful to also create fragmentation, to create silos. Right. To me, that's more tangibly, how do we solve the agent interoperability is a big challenge. The third one is, I would say the agent trust. Yes, we talk about responsible AI, explainable AI, probably for the last two decades, it became another buzzword when LLM came out and everybody started talking about AI safety, but Carter, I can tell you, the importance of AI trust is paramount now, simply because of the proliferation of the AI agents. Not everybody will be creating LLM, not everybody will be creating SLM, right? Auto-customization still seems to be limited to the technical resources, but that's not the case with AI agents. I have business users creating AI agents, our powered users want to create co-pilot agents. Then you have the out-of-box agents, I'm just telling you, the proliferation of the agents make this so critical to actually bring the trust worthiness to everything we do. Because how do I know, right? This agent created by Mary, right, in supply chain, should have the right level access. Is this the authoritative one for research? All these other questions that my client is asking. It brings a lot of questions about governance, I mean, you brought a proliferation of agents, you brought up how agents are being created and used by business teams and so on, but let's talk about what can go wrong, and one, I guess, news story that is very fresh in my mind was, I don't know if you saw the one about cursor AI. They had the customer support agent, Sam, that was automating a bunch of customer support functions and sent people emails saying that you can only use our tool on one device, which was incorrect, caused a bunch of people to unsubscribe when then the company had to then follow up and say, no, that's not our official policy or AI agent did that incorrectly. So the question here is, there's a huge opportunity around AI, but also a lot of things can go wrong when you do this level of automation and AI as access to resources can make decisions and so on. How do you think about that governance, what needs to be in any governance framework for these companies? That's a very good question. I think the example that you gave is very real, right? One of my similar examples I always talk about is a terrible story or a horror story that, okay, now in context, in this case, now you have agent out of the control agent just give refund to every customer. So to me, that's not the kind of things that we want to see happening, AI, it's surely becoming so autonomous, but how do we actually not slow down the progress, but also put a guard rail around this and basically say, okay, how do we balance the risk and reward for this kind of the imperfect technology? So our point of view on AI governance, I think the entire playbook needs to be rewritten simply because of all the things that we've been talking about, a lot of cases, we tell the client, this is the area that you need the professional's help, doing this homegrown, things can get out of control very easily. Professionalism of the gen AI, a gen to AI is quite important, right? Which means asking for outside help, right? Don't try to do all this by yourself. I think the second one is how do you automate a lot of this instrumentation? How do you use the algorithmic way to actually measure agent trustworthyness, right? How do you break down something like this into things like explainability, things like access, control, security, safety, ethics? To me, that is the new way of addressing this kind of AI trust and how do you do that in the highly automated way? So you're just not just talking about theories, not just talking about the guard rail, it's not just talking about the design principle because this is what every single client would need. It's a huge gap area that I think we as industry need to come together to close. Lahn, this is being a great way to wrap up our first season. Thank you so much for joining us and sharing your insights here on Where AI Works. My pleasure, Karthik. Now I would like to highlight a few takeaways from my conversation with Lahn. Lahn brought up the fact that many companies are investing heavily in AI, especially generative AI, but many struggle with execution. The biggest roadblocks include integrating AI with proprietary data, building the right data and talent infrastructure in-house, and scaling beyond small experiments. We also talked about AI governance and trust, and there are many dimensions to it. Companies must prioritize explainability, transparency, and ethical considerations to build trust, especially as AI becomes more autonomous with agent-based systems. Companies will need to rethink their trust models and establish clear Godrids. That's a wrap on season one of the podcast. Thanks so much for listening. It's been a great pleasure taking part in these conversations, and I hope you have enjoyed them as much as I have. Season two is just around the corner, so be sure to follow us so you don't miss an episode. This has been Where AI Works. Conversations at the intersection of AI and industry. Talk to you by Wharton in collaboration with Accenture. I'm Karthikosanagar, bye for now.