AI Governance and Mainframe Modernization Update

Start to think about it. There's a lot of anxiety as well in AI. I've talked to my interns that just came in the last couple of weeks and literally talked about, will my job be here in a couple of years, right?

Which is a big thing and kind of scary a little bit, but that's how people view it. So I can talk a little bit about how I'm applying, how we're applying your advisor. As Dave has mentioned a couple of times, into the technology organization, basically the delivery of applications and tools.

But I do want to talk about, a little bit about that anxiety as well and what I think it means and where we're heading in that. And for those of you that don't think, I'm like the all -the -top guy.

So if you have something you want to say, just shout it out. It's not going to bother me in any way, shape, or form. Unless it's a production problem, then keep that to yourself.

I'll have a

hot

side later. Alright, so here's the first way people think about AI, right? Anybody ever see the movie WALL -E? Yeah. Yeah, yeah. Or a classic story, right? Lonely robot sitting out there full of them by himself, finds love, and then there's some human race, wherever they might be.

And he finds the human race, like that has been AI -enabled, and literally all they do is float around on pillows because AI is doing everything for them, right? Machines are there, and they kind of float around, and they lose all of their ability to do anything, basically.

And that's scary. So, but you know what? We might be heading there someday. I don't know. We'll be around to see that. Seriously, other side of the equation where people think that, yeah, Skynet is what I'm aware. Anybody know what Skynet is?

And I am, and thank you to all the people here as well.

I always know my daughter, she was born on August 29th, 1997. At 2 .53pm on August 29th, 1997 is when Skynet became aware.

So, if you don't know what Skynet is, basically it's like, yeah, an AI model,

and

as Arnold Schwarzenegger says it, I won't do the accent, although I could. But, you know, he talks about a geographic learning where Skynet and artificial intelligence takes over, and the humans try to disable it, right? And the defense mechanism is, I'm going to launch a nuclear strike to wipe out the humans, and like, the machines will survive.

And I only point that out because, fascinating enough, if you've ever read this anywhere, from an AI perspective, and it was spoken about in the first presentation a little bit by Nick. I think the idea was, there have been agents running, right, where they're not put around boundaries, there is no governance around it, and that's why I say governance is imperative.

You can control what models do. So, when he says, hey, they go out there, they do whatever they want, not entirely accurate. I mean, maybe it'll come someday, I don't know, I hope not. But the idea of it was, there were models running, and literally the objective that was given to it in the rule set was, you will survive, right, at all costs.

So, when they tried to turn it off, what they found out was the agents were calling other agents, rewriting their own code, right, so that they would survive, and they literally couldn't turn it off, right? So, that's the kind of thing that, like, makes me nervous when you start thinking about, like, commerce, and you start thinking about transactions and going, like, how do you stop it?

How do you make sure you have the right control? So, governance is going to be critical. I'm going to talk a little bit about how to do some of that in, you know, our organization, but it doesn't go away.

And it's important that people understand that, otherwise, there's a couple of factors around that that was spoken. One is, if everyone just has their own use to it, the cost and the expense becomes astronomical. It's not free. AI is not free, no matter what people might think.

So, the talk about it, etc. And, in fact, I see use in certain areas where the annual cost of, say, using some of the models that Dave was going through before is higher than the cost of an engineer.

So, that would be an interesting dynamic as that goes on because they're inherently sometimes not as efficient as they should be, right?

Oh, excuse me. The, so, smart, faster, better, right? AI. This is the construct of everything that goes on here. And I would argue that I've been in financial service technology for over 40 years now. I have seen these, like, dimensions on every shift in technology over the last 40 years, right?

When I started my career, I was in college, one of my first programming classes, I had punch cards. So, they know what punch cards are? And you carry them around, right? Your resiliency was when you drop them on the floor, you had to sit down and hope you would put the numbers on them, you had to rebuild them all that, right?

Took forever. So, now, when I look at AI, I think it's almost the same thing. It's just an enabler. How do I go faster? How do I make things better? How do I get higher quality? That's the objective of what we're trying to achieve all the time.

I put this slide up. I've actually used this at a town hall I did recently. And a couple, everybody actually has referred to it. AI is unavoidable, right? It's not going away. It's not a fad. It's not something that's going to, like, disappear.

It's going to evolve, right? Everybody's job is going to be touched by it in some way, right? Everybody's going to get more efficient or more productive or, you know, if you're lucky enough, you're one of the lobby people and you just kind of flip around and do other things for you, which is cool.

But the idea of it is, like, everybody from a development perspective, every phase of the life cycle is being improved in terms of the better, faster, more efficient, higher quality, etc. And it's baking into everyone's goals, right? And it's baking into everyone's goals, right?

And anytime we do a massive foundational change like this, you've got to bring people along, right? You start with the hearts and minds. Don't be scared of it. Don't think about Skynet. Don't think about Mali either, right? Think about, like, how it's making your job better.

How you can deliver faster and higher quality. And that's the gist of what we're trying to do. And literally, you know, the models that we have and the tools that we use differ from whatever job family you're in, per se.

But the idea of it would be that you would be able to deliver with less and deliver, obviously, much faster. And that's, frankly, that's really what we're trying to drive. But I just put this up here again because this is important for everybody around AI to understand it.

Again, any transformational change, it doesn't just happen, right? You've got to train people. You've got to teach people. You've got to, like, let them know what governance means. Like, really have to understand that it's not, like, just like, oh, I've sent something through.

It happened. It's a miracle. And, you know, there's no cost. Or there's no overhead. There's something associated with it. There it is. So we have a large organization. Everybody here, you probably have large technology organizations. You want to, like, move people along.

You want them to understand what models I use when I use them, what benefit I get. That's where we are right now. We literally started our journey. This is, like, when I think about David mentioned the office transformation earlier.

The last time I was here, we were literally just, like, taking that project off when I was talking about it. I did not envision, like, the AI impact on that project at all, right? That's how fast it's gone.

Like, AI, I might disagree a little bit with the timeline that they put out in the first presentation. Because AI has been around forever, right? People have

been

talking about it, machine learning, evolution from the 1950s, if you will. But it's only really taken off in the last couple of years because of gaming, right? Video games are the key driver for AI, in case you didn't know that, right?

And when you put up the slide about the CPU, right? That's it. CPUs process billions of transactions or billions of instructions serially. They just go and go and learn. So,

video

games made people think differently about, like, a graphical processing unit as opposed to a central processing unit. And all of a sudden, there was a massively parallel community available. So, all those large data sets, you can disperse across a massive parallel infrastructure.

And that's literally where all the advances started coming into play, combining the tools with data, right? So, I want to make sure that people understand, as they go through the modeling, what their job is, where their role plays in, how they can get more effective and more efficient, and take advantage of those capabilities for that.

So, I got a little – this is a little, like, Terminator kind of slide. I should have, like, made it a little more human, which is, like, not bad of me. But here's the gist of, like, what this slide has to represent.

The most important thing is the agent toolbox, right? People were just talking about the last panel. It was talking about the tools that we use when we do transactions, when we do commerce, et cetera. In the engineering space, the idea of the toolbox can be anything.

Where do I ingest data? Where do I ingest information? How do I bring that in? Am I creating a quality agent? Am I creating a coding agent? A code review? All these things are being built right now. I have a team that's, like, super strong in the AI perspective where things that are very basic that you wouldn't really think of are just a large Excel spreadsheets, right?

How to do a proper code review. How to take in all the cyber vulnerabilities that are out there, feeding them into a toolkit that you can apply back to the code. So, why look at the orchestrator? I was going to put my Bernstein's picture up there, and I missed it, right?

So, when I look at the orchestrator, the idea of it is, like, okay, like, maybe I can do things in a certain order. Maybe I get requirements. Maybe I do things in a certain way, and I hand it off to all of these different roles.

They will all have agents or can be agents. Like, we're doing that right now. I'm going to give a couple of examples. But that's literally the idea. How do I turn requirements into stories? How do I get them engineered?

How do I build that? How do I do that QA? All of these things are accelerating at an enormous pace right now. So, hopefully, we're all going to take advantage of that. And I'll go through a couple of examples of that right now.

So, about this transformation.

If you don't know Office, I think everybody here knows Office, right? It's a massive mainframe platform. I worry less about platform and more about code, right? It's code that's been built over the last 35 years. COBOL, Assembler, all these modern languages that everybody is well aware of.

And one of the things that we started to do, and this comes a little bit into the governance, was how can I use AI to improve this massive rewrite? Like, we looked at it as literally as like a four or five year program overall.

And we started applying AI to saying, well, I don't want to ramp up 800 engineers. I don't want to like think about how to train people constantly as I have, as I have, as I have been below. So, we're really starting to look into the AI tool set like last July, maybe.

So, literally only a year. We're at a point right now where 70 % of our stories, i .e. what comes out of the requirement into the engineering community, 70 % are done by AI. We're not writing down documents, it's all done by AI.

Code generation, over 40%, right? Now, when we build a model that does all of this, this is where a little bit like the governance comes into play. We fed into the models that we use all of what is obvious today.

70 million lines of cobalt code. Every piece of documentation there is. Every user guide, every incident, everything related about this we defined. If you saw the presentations we did on that defined model and where we're going from a target state perspective, we fed all that in.

The governance came in really, the powerful thing was creating a knowledge called a knowledge graph, right? Which is basically just a visual representation of things that don't really naturally hang together. And you can use that, that's your governance, right?

This is where you say, I need to make sure you're doing these things by these rules. And then you feed it into the codex or the different models that we use. That's what we got into the testing. 63 % of the code generated by AI, that's going to go higher.

And literally, for those that know the releases that are coming, we promptly generated upwards of, or saved at least 250, 300 ,000 hours of development already, and accelerated our timeline over 10 months. And I think it can keep growing.

In fact, it will probably keep to the point where you're going to have to ask us to slow down. And I'm sure your shop is trying to do the same thing, right? It's

going

to be, things coming up so quickly, they're going to have to figure out how to make sure you can absorb the change. Another example that we had was, we just did this direct express implementation to the treasury. A similar thing, I took all of those rules from office transformation and brought them into an existing team that had really no expertise.

Built out the agents that they needed, and we delivered this in like four months. It was over 100 ,000 hours of effort. We had like zero engineers on it. We built it in four months. And the big, this is just numbers of all things we've looked through.

But here's the key number, it's in the bottom middle. We had a massive regression test going on for this application that had never been automated. It had never really been changed. And literally, I would say, it was like a, quote, word, a five -week average, right?

All bank law, right? We fended it to the model and spit out all of the test cases that we need for regression. And something that took five hours, five weeks, we now do it in less than two hours.

Right? And that's going to go faster again. So, again, just some critical examples that I think are important to think about how we can advance and how we can do more overall. This is really just trying to wrap it up in like what I said from a governance perspective.

Control the inputs, right? Understand all of those large things that we put into the model, right? And then you get your context layer. That's where the knowledge track comes in. That's where you get the governance and the protection that we need to execute.

And then you keep building out these AI agents. So, I think phenomenal work that the team has done. We'll start to go live across Pfizer. We're already doing governance across different business areas. So, that's going to have to increase as we go as well.

And then for those that I hear this a lot from people, we are building a model around HLEs as well. So, I know it takes a while to get estimates out there. They span different teams. The requirements are very different.

So, again, we can feed all of that data into a model. We're testing it right now. It's learning. It's going quickly. I would expect us to be able to turn around things in hours that take like weeks right now.

So, hopefully it all leads back to the acceleration and driving speed into the overall environment.

So, that's literally like all I have. I can talk a lot about the engineering side and a little bit more on the details. I'll open it up for questions. I'll leave you with one last philosophical thought, right? So, most advanced systems, we're always depending on humans.

Right? No matter what. I hear stuff about the agents and that's certainly the case. But, I will tell you back to my SkyNet example. I have seen when we started rolling out a lot of models early on and I have data references a little bit.

Without the right controls or without the right understanding, I have seen models go into the engineering where you've got to make sure it doesn't understand the delta between production and certification and integration and testing unless you teach it.

Right? And I've seen models go into the right and try to update a file. Couldn't update a file. I failed here in a record and didn't know what to do. So, you know what you would normally do? You would say, oh, it's not what it all.

And you move on. It deleted the entire file structure. Because it said, I don't know what to do with it. So, it must not be necessary. So, it's like you've got to be careful. That's why I think the humans are always going to be part of it or I think they are always part of it.

And make sure that you have that proper governance. That's it. That's what I've got. Questions? Comments? Thoughts?

Okay. We're scaling it. Like I said, across the environment. We're training our engineers. And the expectation is that we're going to go faster.

Mike, I have a question. They didn't give me the mic quick enough. So, this is all amazing. And I think we're going to rapidly get to a point where technology may be standing around going, you've got to tell me what to do.

So, like from a technologist point of view, looking up at product and looking up at businesses, like what are the things that you think that should be adopted, right, that will help the development process on your end? Because without good inputs, the outputs are still probably...

Yeah, I agree. And I think,

you know, I go back to the knowledge graph a little bit. Like if you can visualize and connect those things to what a requirement might be, that you can take out the interpretation of those things, build a model that understands how to match the business process to whatever story is being generated.

You know, again, like Tim said, ultimately, the regulators are going to be heavily involved. They have to be because these are financial transactions. We need to make sure that they work appropriately. Some others have to be the right checkoff.

But I think, you know, in the classic example, you'd be able to do more with less. You definitely will be orchestrating and figuring out how to generate less. So, I don't know if I would be people sitting around.

But I do think the idea that efficiency, like any other change in technology, just naturally leads to that.

That's how you get back at it. That's how you get back at it. One question.

Mike, I'm curious, just as someone doing a lot of getting off mainframe modernization, I'm curious how much, as you complete that work, how much more you have capability does that have? Or where do you see that advancing just the capabilities that Fizer can offer?

I think some of the things that we'll have, obviously, the obvious benefits of, like, when you think about the mainframe and the monologues releases, speed to market, like we do these big quarterly releases, et cetera, all that will disappear, right?

And then AI will be in much smaller microservice mode from an obvious transformation perspective. So those things will go faster. Where I see AI as, like, the key driver of all that is finding, and we're talking about it a little bit from a broad perspective, finding logic gaps in code.

This is what Middose is all about that way. It's not specific vulnerabilities, per se, but it's as much as finding patterns in the code that can be exploited, right? So I think the idea of embedding that AI constantly reviewing what's going through the code gives you the ability to do self -healing, right?

Self -regulation. And, you know, we were talking a little bit about the cloud before. There's always going to be a hardware dependency. There's always going to be a failure somewhere. But if you can, like, figure out from an AI perspective how to, like, self -heal those things, that I think is one of the key drivers and benefits that we're going to have as well.

So I think controlling that business process, teaching the code how to, like, literally, you know, heal itself a little bit. I keep coming back to that, like, agent where the code was rewriting itself. Like, that could be a very healthy thing or it could be a really scary thing.

But that's what we'll start embedding. And those things are just really starting to come online and mature now.

So, okay, so, something I've been thinking about as AI becomes more adopted, I feel like we're kind of in a sweet spot, right? That we have a human in the loop who did the coding, who's done the management so that when they look at the output, they go, okay, this makes sense or doesn't.

But as new generations enter the workforce who maybe weren't ever coders and always leveraged AI, that's a concern or risk that I see go forward. And I'm just curious if you all are thinking about that as well or how we approach that.

Yeah, I think

you have to, right? Otherwise, you approach that WALL -E graph

quickly. And hopefully the machines are humane enough to, like, leave us alone. I don't know. But

it's a great point. It's always going to go back to the business process and less on the platform. What are you trying to achieve? How do you put that into work? What is the goal of the tools that you're building just like anything else?

And I think that was one of the things with Office Transformation was, you know, it was built, like, over the last 30 years. It's very much platform out, not really customer journey and, you know, it. And I think, like, looking at it that way is somewhere there's always going to be a reconciliation and a validation that's working correctly.

There has to be. So right now, the phase we're in is more of the data attribute layer. Does this equal this? Was that what it was supposed to be? And then, really, we should be building agents, and this is where I'm thinking about it.

But the next thing is, how do I tie into the business process to use the example that the panel just had? How do I know those transactions took place the way that we're supposed to? How do I know that that's right?

At the end of the day, he peeled the onion back and said, well, how do you know that's right? How do you know that's right? How do you know that's right? How do you know that's right? But eventually, it has to come to the point where, you know, you're either accepting what is truth and, like, you've validated it at enough time.

Or you, you know, get a shot that I can go live out in Montana.

That's a good lifestyle, too. I didn't see that. But it's a great question. And, yeah, I think over the next 10, 15 years, it's going to evolve pretty quickly. I do think it's interesting. Like, you know, I'm a big believer in the free market and capitalism.

So, like I said, I see money use already that's more expensive than people. So, like, is it cheaper to just go hire people? Like, that might be, you know, somewhere we drive back the cost and bring the human element back into play from, like, more of the volume of what happens.

So, we'll see. I don't know. I couldn't believe in the future. I place a lot of bets right now. Good question, though. See, that's why you're in terms. Which camp are you in? Are you in the Wally? Or the Skynet?

Skynet.

All right. Take it out. That's my question. Mike,

can you talk a little bit, I've been talking a lot about the act of modernization.

So,

there's big, big sort of movement and modernizing. And then you've got AI coming. They're very converged. Right? And you're doing some of those things. Sounds like the project is a state building. Some of those small things. And they're not small.

Those are actually quite big. So, I was kind of just curious on the timing on some of those things occurring. And then when do they sort of merge? So, that they're moving in the same direction. You might figure out how

to go backwards. So, the convergence, right, is happening now. Right? So, like I said, the idea, like, if I, so what about the project? We were going out, like, into 2031 simply because the volume of, like, understanding the requirements, the volume of understanding the story that we generated, the way we were working, right?

So, when we brought AI to the model, that's what I'm saying. Like, our story creation literally, and that's what the model is doing and referencing back to the model graph is it's taking, like, all of that cobalt code that I referenced, like, 80 million months of pizza into the model.

New requirements for things that we're building right now into the model. And the driver literally is simply telling it, give you the code that does this function as identified here, right, for simple terms. That's happening right now. And that's why I said, like, half of our code is definitely going to be AI created and driven.

At least 25 % of our testing is already there. And I do think very traditionally as an engineer and somebody who's written code for a good part of my life, the idea of AI, the most powerful thing is, this goes back to many of us asking, obviously the business process is more important than the platform.

And one of the great things that's cool to get from the agent perspective is I can, like, look at it from that business need and create an agent that will do the stories, the code, the test plans, and the validation from that view.

And that's the customer journey view in, not the platform out, which is, I think, the way that everything was built in, like, in my career. So that's all converting right now. The last slide I had up before is, like, for us, it just needs to scale.

Like, I'm using that in, like, you know, learning as we go in a few key areas. But it will take us time. It will take us a couple years to scale it across five certain times that, like, we talk about.

But that's what we're doing in programs and the key tracker right now. Like, are you testing the AI code differently?

So I'm starting to ask whether we test the AI code any differently, not really. I mean, the agents are, like, executed at the business case level. So they're all doing business process validation. And then really using it for all.

We can compare a lot more data fields and, like, outcomes than we ever have for the buck possible. Right? Like, so we create 50 ,000 files a day. You know, like, our master files have, like, 55 ,000 attributes.

I can compare all of that at the end. This is the office transformation. I can compare all of that at the end of the run in a matter of hours. So that is different in that we'll be able to do that more frequently and more consistently and rank up environments.

And a good thing, you know, from an office transformation perspective, a rapid is we lose the burden of, like, that test cycle. Like, I can run 10 cycles a day. Right? So we label that a lot more flexible and now we think about testing overall.

What do you think about

the point where you said you're looking at people who are testing, have you look at media testing, what are the different developing? Is it more active on that? Do you have any additional data on that?

Like anything else, I mean, like, like anything else, the first batch of code, when we start doing the record compares, we have, like, maybe a 10, 50 % break rate, meaning, like, which is fine for the first pass.

Like, so if I was comparing 50 ,000 attributes, then, you know, 7 ,000 of them work, whatever I thought they were. And then you go back and you fix the code. So that's another great thing from an agent perspective.

You know what that issue is. You quickly find where the code is generated. This is the self -healing part. And it will spit back. It comes to me and say, this is where you have to go change that.

Yeah, and that's really powerful. One of the questions in regards to NECO. Did you guys, I know, pretty late, pretty early, was there any interactions that happened with NECO disorders? What did you do recently when they were on NECO?

I'm sorry. NECO. Yeah. And I know there were these, and these, and they said, did they have a chance to kind of do some free testing, check it out every week with Mike's ability? Well, then, yeah, so there was a question around depots, and I'm sure people have that question, right?

So when that first came out, when that was first announced, that was all happening. You know, we run a very strong, very informal cyber program. I'm sure everybody does here as well. We did stand up, you know, with a daily command center to try to understand, is everything coming down here going to be like a zero -day, you know, availability, that we're going to be like a constant state of patching?

Some things might be, some things might not be, right? So we did get through, we were part of the last -week program, the second tranche of testing, which we just literally went through. So we're interpreting those findings, and then just the normal process figure out our remediation path.

We don't have anything that we would call, like, oh, the zero -day, you know, thing right now. But interestingly enough, to your point, like, we scan, like, every month, right, every release part of the CICD pipeline, and it's still found a whole bunch of stuff that we haven't found, you know, our normal scan.

And I'm sure you guys will find the same with your company and your organizations, right? And, you know, like, the interesting thing, I'm scared of crap, I'm thinking of people, but, you know, we manage vulnerabilities to this standard, right?

Critical high, medium low, et cetera. But one thing about ethos was it would, if you go through your code and string together hundreds of loads that maybe nobody's even looking at and create a critical in seconds. So, like, that was the scary thing.

We didn't have any of those. So, but that's the scary thing, right? Because it's just working at machine speed now, right? And we have to figure that out. Good question.

Yes?

I understand AI is rewriting the code for the transformation, but what's actually going to be processing the code nightly or dataly? Is it AI or is it software?

Well, the production environment of transformation, if you will, will all be running on, not on a mainframe platform, per se. It will be basically a mainframe rack, but really in a physical environment called Linux One, which is, you know, another IBM product as well.

And it's just moving off a high -powered, expensive mainframe platform, into a lower -cost, what's called a, like, bare metal frame. So it just runs in differently for people, the technology people in the room, right? It's just running all your containers at a bare metal level where you lose all of that operating system overhead and things like that.

So I fully intend, you could do anywhere, we could run it anywhere, but I fully intend to keep running it in a non -prem data center. I mean, this is all financial privacy data. I don't have any intention of putting it on the public cloud, unless for some reason it's being like that's the safest place to be.

But we are in the data center business, and, you know, I think the necessity of what we do means we're going to stay in the data center business. I have lots of things that run on the cloud. We do overall, but, like, you know, the old ground, for lack of a better phrase, is something, like, I don't have

any place to put that out in the public cloud.

One more question. Yeah.

Is Vicer worried about, like, clients using, or issuers using our own AI models to replace the services that Vicer gives to us? You know, like, we could now build our own AI model and do what Vicer does for them.

Yeah, I mean, that's more of a vague question than a vague question from a product perspective. But, you know, I mean, you can do that today. You don't need AI to do that. So it's just a matter of, you know, what the strategy is and where do you invest your human capital and issuable capital.

I don't know if you had any other answers. I

missed the question, Mike.

I just walked in. The question was, are we all worried about all of our clients using AI to build their own models, their own applications quickly and not be our best friends

at work? They want. Oh, that they would replace them. The short answer is it's not something that keeps me awake at night. I think that at the end of the day, there is, and I've always said this, I think Vicer provides a very resilient core process of platform that does all of the complex transactions.

And I think where time is best spent from our customers is unifying that user experience out at the end of it, right, and really create that uniqueness. And I think we're already seeing that in the market. I can tell you, in the national space, I'm rebuilding, we have a vision platform that's international where we're rebuilding it from the ground like cloud native.

And it isn't an easy task even with AI.

But look, I welcome, and the reason and I've been pushing this direction and Mike too is that I'm trying to open the platform for easy API access. I'm trying to stream the data out quickly to really get the data into your hands because I think that is going to be the catalyst that drives me.

Excellent. All right. I

didn't really classically came on a little more today for you. What? I'm

not.

I'm not.

I'm

not. I'm

not.