Implementing AI In The Real World — With Kyndryl's Antoine Shagoury

Thanks Alex, appreciate the time today.

The first thing to reflect on is it's absolutely amazing at the rate and pace in which the conversation has grown, the opportunity discussions have continued to advance. But one of the things we love to talk about, especially with agents, is you don't start there. It doesn't start with an agent. There's a whole different process that gets us into that set of capabilities. One of the more difficult things to break when you go from what the expectation may be to how you want to approach it really starts to lay the groundwork on what we think is the right pattern and practice that helps make it work. The way I refer to that is almost like that journey message, so you kind of have to take that with a grain of salt. But as we start to unpack the business, demystifying it sounds strange: how do we demystify a business? We start with understanding the information, the data sets available. How do we really understand what we're working with? Often we get into demonstrating what we can do with simple AI introductions, some simple automations to step through: do we have the right data sets based on what the client is looking to do? As we build into that, that's when we start to build the pattern of information. We go from simple automations to predictive capabilities, and that is really the first time there's a sign to put an opportunity to how we can deploy an agent to demonstrate going from a predictive opportunity to something prescriptive. So we know enough about a business process, enough about a system's operation, to really have an agent step in to orchestrate, maybe manage the state of an application, a failover process. That really starts to drive it. As much as it sounds a little bit defeating, it really demonstrates that we can change the framework on how clients approach it. I'll put one last comment to the answer too, which sometimes gets a raised eyebrow: ultimately we're saying, how do you want to become autonomous? So agents are really just another step getting closer to automated systems and automated processes. That's how we really look at and approach it.

It can be both.

Yeah, we can go and just apply it in simple automations, but that's not the benefit. It's really the granularity that we can start to map into to help the business, to help provide the outcome. I'd actually look at it like the impact radius, the impact area. There are so many things that agents can impact, whether it's looking at a business process, looking at human involvement—how many people have to be involved in a business process, a service, or a transaction. So it has the opportunity to get very deep into being the better orchestrator of an event, the better analyst to understand the permutations of a situation, or a better distributor or run agent to distribute when something should run. It has a phenomenal opportunity, and we are seeing use cases where clients are able to get deeper into how business runs. It really exposes legacy systems, human intermediation across different processes, applications, even partners. Agents allow us to start to decompose that in a way in which we can focus on what really matters. There's definitely an opportunity.

Sure. Maybe simple is best sometimes. I'll touch on AI coding or AI-assisted coding. We've been able to use our AI agent framework to deploy not only code-assist agents but also test and deployment agents, to shorten the time in which we see errors in coding to be returned back and refined. We deployed that across a client, for example an international client in the telecom space, on how they improved the code lifecycle, the code deployment lifecycle, and the improvement lifecycle. There were enough specifics in what they wanted to deliver within this code factory that the agents were able to overlay it in a way that improved quality—lower defects, clock-type analogy—and how we can improve how we deploy more effective releases, better functionality, faster to the client base. It's specific, it's not dominating and changing the world day one, but it does provide efficiencies. A lot of the early gains go back to the basics: can I provide more efficiencies back into the business? That's one of the perfect use cases.

We spun out from IBM a little over three years ago. The way I can best represent what we were is we were the implementers, the integrators, and the operators for managed services and the infrastructure and systems as a part of IBM and their client base. When we spun off, it's a phenomenal heritage to base our business on. So when you think about what we do, we are experts in running large, complex, global mission-critical systems. As we left IBM, we went from running their portfolio and their integrations and expanded in our partnerships. We became close partners with Microsoft, AWS, Google, Dell, now Nvidia, SAP, Oracle relationships. We can now really extend our integration and implementation experience to the benefit of our client community. We've really become what I call an ecosystem player, but we are rooted in managed services.

Wow, I've got to be careful on how I answer this one. I'm not sure I'm going to help the numbers or statistics in the conversation. I'll probably reflect first a little indirectly and then directly. Indirectly, it's no different than the rate and pace of solutions hunting for a problem. The market is amazing that way, and the investment is phenomenal, so we have so many things pent up on the opportunity side. The best part is we're starting to see the willingness to adopt, the willingness to try. The POCs have actually gone up exponentially; they haven't died down at all. But we often get into the expectation that AI is going to naturally solve the problem where we haven't really defined it yet. So there's a lot of missed expectations. Although an 80% failure rate is not uncommon, or getting thrown on the shelf—first of a kind is last of a kind type of scenario. The approach we often get into is we worked on the POC, but it's actually trying to find out what's missing. A lot of the effort we bring is in the approach: how do we understand what you're looking for, what's the business problem? We often find many of the POCs don't require complex AI or new model development; they require simple automations, more data, or more programmatic changes in how the application or business is operating. A lot of it is just the gap between the understanding of what technology is and what it can do versus how you need to deliver that value. As much as we talk about it and it sounds almost negative, it is a part of the discovery process. The one thing we always want to encourage is what the business is intending to do: how do I uncover, how do I evolve what's actually happening here? That's usually the best way to keep the focus and recraft the thesis. We are also getting better at what we need to do in the POC requests—how do we help shape it? That's where I think you see a lot of maturity happening, not just with Kyndryl but across the market.

It's often the case to start. Again, the journey is important. I do like the ChatGPT example. That became a great opportunity for us to show how agents can help continue to refine and filter responses, even sources. But everyone thought it would have everything they needed, just trying the next iteration or the next opportunity with the product.

There are some really interesting things that are maturing. If I look at some of the telecom clients we have, a lot of the work in how they're approaching AI, information gathering, and analysis is working very well. How we start to apply gen AI looking at procedural, product capabilities, selling opportunities—how do you go into upsell? There's a huge gain we're seeing in certain industries. They've evolved into cart generation: how do I understand what may be better targeted for you, personalization? They've gone from understanding what they built within the workflow, the efficiencies within the workflow and product lifecycle, and now they're using gen AI to assist within that sales process. Now agents are involved in personalizing baskets and shopping carts for you, reducing the time to actually get you to sign a contract for a new service. Automotive is a little different, but there's a big push on understanding personalization—harvesting surveys, interactions, chats, and directing that through. They've started to leverage models and micro models to specialize and target audiences better, and that's influencing supply chain and quality. I think there may be a really different dichotomy in where the industries are, based on access to information and how targeted they are in addressing a client need or an opportunity to a committed contract lifecycle.

Absolutely, well said.

It's a bit of a double-click through the process itself. Think about it: we're all customers of telcos in some realm. As you start to call in, you get almost immediate analysis or information on our profiles—what services do we have, how long have we had the service. A person would normally talk to you when they're trying to upsell you on a new contract or renewal, but the system can very quickly start to identify opportunities: 'Hey, we can reduce your bill by X because you're not using these services. Can I improve your experience?' All of a sudden suggestions start to come in, and you don't need the same human interaction. As you get into that, you can now give scenarios to show you this can reduce your price. Instead of going through a human process—I'm a customer of one large telco in the United States, they call me every year with different scenarios, and I wind up having to do the analysis myself—they've closed the gap. Many of them are starting to close the gap on how they approach and demonstrate the information of your usage, how they can improve your experience and the economics associated with your service, and how fast they can actually change the service for you. They really shorten the lifecycle of opportunity. The agents in the process are collecting information. The orchestration agent can start to break up or shard the query to pull the information and analysis: where are you, what's your demographic, what's your usage pattern, what's your bill associated with your usage? It brings it to the analysis almost immediately. Then from orchestration, you have the calculations of the opportunity agent: how do I give scenarios that may equate to your demographic, income range, family size? So we can start to put opportunity analysis in play. Then we have presentation agents: how do I represent that to you in a language and presentation so the personalization becomes very apparent? It is an accelerant and an augment to helping improve the service or elongate your willingness to keep the service with them, the annuity they have with you.

Oh man, and you probably shouldn't comment, but it's like the agent war in my head—they're going to negotiate with each other.

I made the comment before about the impact statement. There are so many impacts. There's a statement I probably get in trouble for saying, but we are seeing the greatest evolution of business in our lifetime, fundamentally. It's changing business, how we develop product, how we experience services, it's changing job roles. There are so many things this has implications for, and it's mind-boggling in concept.

Absolutely.

I probably can't underscore it enough—it's paramount. It's one of those scenarios where it's probably the investment area that's the most unspoken, the least spoken about, but it becomes the most pivotal. Data inventory, data harvesting, classification, data access become critical. If you can tell by my hairline, I grew up in highly regulated businesses—you couldn't touch a piece of information without 10 people watching it. But you'd be surprised how many industries don't have that rigor. Understanding data lineage within the process becomes absolutely fundamental: understanding the auditability of how information has been generated, collected, propagated, manifested. That's the foundation element. Securing the information within it—there's been a growth, probably too quiet, in new technologies that allow us to tokenize information and protect data in its construct. So we go from data inventory and data lineage management to understanding how to protect the information or components of it. You may need to protect your name but not your race or gender, your geography but not your usage. That's growing within that construct. The security around that has grown exponentially. The next area we often see is how we build transparency within the data usage for models, analytics, and usage itself. That not only puts visibility on how the data has been used but also puts a magnifying glass on how it's been adopted. A lot of those things are starting to mature. It also lends itself to exposing a lot of issues with legacy tech. I could probably spend 10 years on the data issues, but the other area that's concerning is data duplication, data manipulation. Data has become so mutable in the environments from made business because applications are fragile. They layer application services but don't touch them, and they keep modifying source data. When you start to see that, it's a cascade. When you're looking at hallucinations or errors in calculations, these are often the sources. Models can be adapted, but if you start with a bad ingredient, you're going back to square one in regard to the analysis, the assumption, and the metrics.

It's the biggest hurdle. Not even a lift—it's the biggest hurdle to get past or be able to demonstrate there's enough of an area to prove it, because a lot of the areas need to be proven still. Boards and the C-suite have tremendous visibility not just on tech for tech anymore, but can the technology enable a business opportunity or an outcome? That visibility is the first really large speed bump. The next lift is scale. Scale is the real lift: how do I, if I proved it, scale it? How do I break it up in a way that becomes affordable? It's very expensive. Getting deep in analysis, deep in processing, continual real-time model operations becomes very expensive. A lot of what we talk about is what we can do to run that through partners—how do you become more cloud-friendly, more on-demand, more scalable? Even in our Nvidia work, it's about how we can be better at purpose-driven workloads: how do we understand quality of service on a workload? Does it need the highest performance GPUs? Can it run on a different type of memory environment? Can it run on a slow burn? A lot of what we do is size, shape, and estimate how we can optimize or architect the analysis and what data is needed, because cost matters. The tax can be fairly large in the over-process.

Great question. There is a material shift. We're seeing a lot more business leaders involved in the decision process. I think that comes from a couple of environmental changes. The consumption model is evolving too—there's a lot less appetite for build, a lot less appetite for creating solutions into the business. It's a needed-now scenario. We're seeing the majority of investment being directed through business leaders and how we're driving into it. It also shows there's a lot less patience in drift and a lot less patience in a thesis being proven wrong. But it is a shift within that operating environment.

About a year ago we started to look at the growth of AI being mentioned in company announcements. It went exponential. When we went from it being in the news to how many meetings we had with a C-suite that discussed what the strategy benefits could be, it went from heads of infrastructure, data management, analyst teams, to COOs, and then CFOs started asking questions about operating effectiveness—scary stuff for technologists. Then CEOs became engaged. There was a great conversation early last year where a CEO basically said, 'I want to understand the pulse of my business,' and these are the tools they want to deploy. So it's exciting, but it's a very different audience and appetite, and it is turning that directional tide.

I'm going to underwhelm you: it's a little bit of both. You have to balance what the use case is. There's an interesting split in the market. There's embedded tooling—like Salesforce conversations or ServiceNow conversations—a lot of AI that's embedded within the workflows and information being gathered. You're not necessarily going to replace that fundamentally. But infusing different capabilities within your interaction with data, data classification, model development, even in how you're starting to create micro models—small language model capabilities to target your business—that's where the real linkage starts to come. It's not as if you're creating your own model, infrastructure, and analytics from scratch. The question is, is there enough around how to create the benefit to both? The area we see a great amount of energy is in the integration layer. We did some work to compare or arbitrate large language models: let's understand what you're looking for, what parameters they need within the models, and can you now devolve or create a micro model that's very targeted to your business? It was a recipe. How do I look at what's happening within ChatGPT versus Gemini, how do I start looking at Llama differently? It wasn't brought from scratch, but it was a recipe that created that integration layer. That's where I think the good energy and good results are happening.

We learn something new every turn. You can't hide—they are driving material growth in the market. When we look at the partnership opportunity, people first think it's all about revenue and growth, but to our delight, working with them was really about how do we accelerate opportunity, how do we identify and co-create adoption work. I see your smile because even I was like, 'Wait, this sounds too good to be true.' But from a perspective, it is probably one of the best engineering partnerships that demonstrates you can co-learn, develop and target opportunities, and really think about how you co-create. A lot of what we started to do was based on can we natively extend our capabilities to use their AI and agent frameworks, develop it faster, use their interfaces like the NIM interface to speed to market, and shield our customers from some of the complexity. That was at the root of the partnership. There's a learning part, an opportunity to extend into their platform like the NeMo concepts. We even announced a few months back a partnership between Dell and Nvidia on private AI infrastructure—can we actually plan, build, deploy sovereign, controlled environments? It became a natural extension. It's really an enabler type of contract and relationship.

It is kids in a candy shop. The compute capability, the power you start to see—I'm not sure how old we all are, but it's kind of like 'we're not just the president, we're also a client.' A lot of what we do is managed services. We manage exabytes of information across our client base, thousands of customers. We're looking at how to improve operations. When we start to advantage our workloads and models on their environment, it's an exponential improvement in time and turnover velocity. How do we improve and optimize the models and input accuracy? It really becomes interesting—like throwing a lot of octane in your car really fast.

Very real.

We operate in most countries, so we have an ability to directly connect to most operating environments. You can come to kyndryl.com and we can direct clients the right way. We use our own tools to help direct clients—it's kind of fun when you have one in that space. But we are operating in most countries and happy to engage directly with our customers.

Great session, Alex. Really appreciate it. Thanks a lot.