Preparing IT for AI Agents: How MCP Shapes the Future of AI

Today, artificial intelligence is everywhere. And if you're working in IT or you're a developer, you've probably either already come into contact with AI, or you may very soon find yourself involved in some projects with AI. And I think it's very important that we take a minute and understand that AI is a launching off of the human brain. And so, if we take a moment to think about where we've come from with AI and look at the architecture and the plan of the human brain, it's going to give us some hints about how we need to evolve our IT architecture to be more AI-ready across our technology. So here's a really little cartoon summary. When we think about AI today, we have the internet, and the internet has tons and tons of data. And so really where the current GPT-driven large language models and the sort of new paradigm of AI that we're in today has real heavy overlap with the internet. Because basically, the GPT models kind of swallow the internet and process it and then create these large models that have a really good understanding of text and images and these new applications. So this first paradigm is sort of AI swallows the internet. And so, when we think about this framing, this is not really useful for us inside an organization. When we're inside an organization, we care a lot about specific things that are applicable to our organization. So the data that we care about inside an organization is very different than all the data that's on the internet. And when we think about our organization, we have important applications, and maybe some of that is SaaS and maybe some of that are things that we've developed internally. But we've got this executive functioning layer of our applications. And then of course, we have networking that connects everything together. But the challenge that we have with this sort of existing IT infrastructure in AI is that we're in a paradigm today where we're in this kind of plus AI paradigm. So it's like, this is what we've got. We've got data and apps and SaaS and networking, and we're trying to have AI swallow the enterprise. So AI swallowed the internet, and now we're trying to jam AI into the existing enterprise. And the problem with this is that we're seeing 90% plus failure of AI initiatives that are happening in this style. And so what we really probably need to get to is a world where our AI is a plus, but we need to get to a world where we've got a little bit more organization and separation of what our data sources are and what our tools and executive capabilities are, and then how that's going to interact with AI. And of course, the goal here is to reverse what's happening and try to get to 80% plus success rates in our AI initiatives. So we've talked a little bit about where we came from and where we're at. Let's take a step back now and talk about what we're trying to mimic. When we're talking about artificial intelligence, what we're really trying to do is mimic human intelligence. And so, all biological animals have a body plan. A body plan is a biologist's way of saying an architecture. An IT person would say an architecture, a biologist says body plan. So here we have a brain and we have this main region in the top where most of the action is, which is the cerebrum. And then we've got this midbrain area. And we've got a little thing back here called the cerebellum, which is Latin for small brain. And then you've got this nerve cord that runs down your backbone. So vertebrates have their nerve cord running down a backbone and invertebrates have their nerve cord not running down a backbone. But this body plan is basically designed to process data. So we're processing data from the environment and we're generating responses. We're getting data from all of our senses: touch, hearing, smell, taste, and sight. You've got a big data center coming in from optical data when you're looking at things. And so, you've got all this data coming in, and the brain is organized into three regions: the lower brain, the midbrain, and the upper brain. In the lower brain, you're processing really primitive data and primitive responses. In the lower and a little bit into the midbrain area, we're monitoring temperature, we know if we're hot or cold, we know if something is hitting us or biting us. As we move up into the midbrain, the lower brain is really primitive. The midbrain starts to be about connectivity. The midbrain is handling a lot of data exchange: what gets ignored, what gets stored in memory, what gets moved around to different parts of the brain. And if you look at this picture from the top, most people know that there's a left and a right hemisphere to the brain. We're bilaterally symmetrical. You've got this left and right brain, and right here in the midbrain, you've got this very important structure that connects and allows for communication across those two hemispheres. So really midbrain is about connectivity, lower brain is primitive functioning. And then if you look at all of this real estate that we have, we've got a huge amount of real estate left over up here. And if you go and look at the literature, this really big region up here in the upper brain is oftentimes described as being in four parts. And this frontal part, the frontal area, is where we're really getting our executive functioning. This is really the pilot. So when you get up and whatever you're deciding to do next right now—if you decide to pause this video—it's this frontal brain area that's integrating everything that you're experiencing and deciding what to do next. We've got these sensorial areas that deal with processing auditory data coming in through our ears. We've got a big optical area towards the back of the upper part of the brain. And we've got these upper regions that allow us to integrate and do really long-term strategic thinking. So one of the great things a brain does is integration. The brain is really amazing at integration. So when you're asked about a memory of being at the beach last summer, you may be remembering salt smells, the sound of the waves, the taste of some great crab cakes, and even stepping on a sharp shell and cutting your foot. But the thing that's amazing about the brain is this ability to integrate and know when to take bits of all these different senses and all of this different data. The last thing I'll say about the brain that I think is foretelling about when we think about artificial intelligence and the enterprise is the brain ignores almost everything that it experiences. We completely ignore about 99.8% of all the data that we come into contact with. So if you think about your drive to work this morning, your brain knows to ignore the sequence of colors of different cars or the make and models of different cars. If a Ferrari went zooming past you, you might remember that because it's really out of the ordinary. But the brain is incredibly good at ignoring almost everything that you're coming into contact with, and it's very good at storing things that really stand out and are likely to be important in the future. So now let's talk a minute about where we're at with our enterprise IT architecture. We're going to keep this super simple in three categories: applications, data, and network. Applications are things that do things, like executive functioning applications. Data is all different types of data that we're storing. Network is how we connect things, our communication plan. So when we start looking at our applications, let's put a couple over here: maybe a CRM system, a human resources information system, a financial accounting system, and maybe a legal system that holds contracts. So here are applications. And then typically, we're going to have some type of data layer, like a data lake. Don't be offended by data lakes. Some data lakes are constructed where everything is just dumping into them, and other data lakes have been carefully constructed with nice logical layers and are really useful. But whether your data lake is a swamp or really useful, the idea is that we've got these applications, the applications themselves have data and users and executive functioning, and we typically connect them and dump data into something that might look like a data lake. In the past, when we're trying to do things across these applications, we end up in a bad star structure. So I've got my CRM, my HR, my financial, my CLM. What we typically do is use an API paradigm. We say the HR system needs to go grab some information from that system, and the CLM system needs to push something to the financial system, and the financial system also needs to talk to this other system. So we end up in this connected world where we're relying heavily on APIs and very structured integrations that do very specific things. When we're building this type of architecture, we really can't leave much to chance. If we leave much to chance, things start to break. So the big question is how do we move from this paradigm where AI is being jammed into the picture, into a paradigm where this picture becomes more AI-ready. Let's take a shot at that. The first thing we're going to need to do is scotch this API as the dominant interface, because we're going to need to introduce a new character into our story. We're going to need two new characters: a new orchestration layer, and the orchestration layer is going to spawn armies of AI agents. So we're going to have lots of different AI agents that are going to be able to be spawned out of this orchestration layer. Now, what good is the orchestration layer if it's not interconnected to the rest of this world? And what good is it if we're in a world where everything is already connected in very specific ways? Without interrupting how we're operating today, we can start to do some new things, which is to make sure that each of these applications gets an MCP service. MCP stands for Model Context Protocol. So for each of these applications and for the different areas in our data lake, maybe we partition our data lake a little bit and chop it up and organize it so that these things can start to become MCP. We can have MCP hosts that are providing MCP service to our agents. And so, when we add orchestration and when we transform our applications into being able to be MCP services, what that's doing is exposing these applications as sets of tools and data sources. Tools are what can I do? Data sources are what do I know about? And when we start to have MCP services distributed here and we can communicate and ask agents to go do specific actions across this infrastructure, we're eking in the direction of looking more like this picture. Because the ability to orchestrate, you can almost look at these different applications as different organs in the brain. Maybe the CRM is like my ears and my auditory center. My HRIS system might be like my olfactory, nose and mouth. My financial and accounting system might start to look more like my upper strategic part of my brain. My CLM system might be back here in the rear as part of motor coordination. But ultimately, we start to have organs which remain specialized. But by putting in the MCP and expressing themselves as tools and data, and by breaking into this data lake and creating an AI-ready data layer, when we get to an AI-ready data layer instead of a data lake, and when we get these executive functioning applications able to interact with MCP, then we can send lots of agents out, and these agents almost start to be like synapses. Synapses are the structure that connect neurons in the brain. So what the brain is able to do is when you want to do something super complicated, your forebrain, this executive functioning frontal area, activates this integration. That's going to be this orchestration layer. So when we want to do something complicated, we're going to activate this orchestration. And then we're going to know, are we calling on hearing or on our olfactory senses? Are we calling on some higher-level strategic functioning? Is the task needing to call on motor coordination? And we can start to develop a very strategic approach to doing very complicated things. That's what this architecture starts to look like because you've got your orchestration as your frontal lobe. It can basically look at, in the agent framework, it can say, let's talk about goals we're trying to achieve and then outcomes that are going to be acceptable. So goals and outcomes start to be more the way we work in the human brain. We can use the orchestration layer and light up all these agents, like synapses, to go out across all these different functional executive functioning parts of the application layer. And the agents can also come down and jump into areas of the data layer as well. So the goal of this conversation was to use a little bit of a cartoon and have some fun, but it is to remind everybody that artificial intelligence is literally trying to replicate the incredibly integrated functioning that we have in the human brain with biologic intelligence. So if we take a look at the human brain and we look at how integrated it is, how compartmentalized it is, how organized it is, its compact nature, the fact that it runs on very low power, then we can start to see where we maybe want to go with the enterprise IT so that we can be more AI-ready across the enterprise.