Modernizing Army IT for AI: Networks, Data, and the Tactical Edge

Army IT modernization is what makes Army AI possible. Over the past several years the service has unified a network that had splintered into independent fiefdoms, stood up an Army cloud office drawing on Azure and AWS, and fielded low-code platforms alongside a central data layer. Acting Army CIO Gabe Chiulli describes the current priorities as strategic cybersecurity improvement, making information consumable for both soldiers and industry partners, and rationalizing hundreds of business systems down to a manageable number so that data lives in one authoritative place rather than being duplicated across every system that needs it.

Integrity, context, and authoritative ownership are the hardest problems in the Army data layer. David Markowitz, Deputy CIO and Chief Data Officer, frames the risk as a game of telephone: information passed between applications until what comes back no longer resembles what went in, which the Army cannot tolerate for force structure, status of funds, or anything supporting a congressional audit. Solving it meant bidirectional interfaces between business systems and the central data layer, plus a layer of access control that determines which data an agent may reach. Chiulli notes the consequence. Because agents need broad access to be useful, the Army had to flatten the network and move from firewall rules to data authorization, deciding in advance what each agent can operate against.

Bringing AI to the tactical edge means matching the model, the compute, and the power budget to the actual task. Chiulli warns against what he calls peanut butter AI, spreading a general-purpose model across every problem, and points to Model Armory as the alternative: narrow models tailored to specific mission sets, because detecting a drone in freezing conditions does not require a large language model. Ben Van Roo, whose company works across small edge devices, on-prem data centers, and cloud, describes the same constraint from the vendor side. Once cloud access is removed, as it was during Project Convergence Capstone 6, the problem becomes hardware dependent, and the questions that matter are which models a form factor can actually hold, what power is available, and what software harness ties the models and workflows together.

Transcript

This transcript was produced through automated speech recognition and is provided in rough-draft format. It may not be a verbatim record of the proceedings. Names and technical terms may be rendered inaccurately. Where a system, program, or person's name was clearly mistranscribed, the correction appears in square brackets alongside the original text. Section headings have been added for navigation and were not part of the original proceedings.

Opening and Introductions

Amy Kluber: Hi, I'm Amy Kluber, editor in chief at GovCIO Media & Research. Welcome to today's Gov Focus. We're going to be digging in to a few different areas here with our esteemed panelists. Pentagon has been executing some major technology initiatives to modernize and prepare foundations for AI and other advanced capabilities, and the Army is certainly a big part of that effort. To discuss these priorities and some of the challenges associated with those, I'm joined by Army Acting CIO Gabe Julie [Gabe Chiulli], Army Chief Data Officer and Deputy CIO David Markowitz and Ben Van Roo, CEO and Co founder of Legion Intelligence. Welcome everyone.

Panel: Good to be here. Thanks. Good day. Thanks. Thanks for having us. Yeah, thank you.

Army Modernization Priorities for the Year Ahead

Amy Kluber: We'll start with sort of top level here, digging into some of the modernization priorities with the Army. It's been going through vast modernization over the past few years and certainly still undergoing some of those. So Gabe and David, for both of you, what is top of mind for you? What are your biggest priorities, priority areas over the next year? Give us a rundown. And Gabe, how about we start with you?

Gabe Chiulli: Yeah. So, you know, I picked up this role a couple months ago. And when I first kind of started, I had a pretty good discussion with the Under Secretary of the Army and I kind of said, you know, we've got these bundles of areas that we know are kind of hard things to work through. They're all very interrelated. When you think through cybersecurity, like, you know, everything you do on a daily basis is very, you know, inherently linked.

How I want to bring systems on the network is related to like, do I have the instructions to do that? How I monitor those systems and keep them at a high level of availability is all related to like, did I, you know, have the, the, the left and right limits of how to do that kind of outline to me. So not calling it back to basics, but a little bit back to basics, but they're all things that we're trying to push automation into to help those basic things become more fluid in our day-to-day activities.

I like to bucket them into our strategic cybersecurity just improvement that think every organization says that, but I'll say we're definitely taking taking it to the next level. So just make sure make sure we can see ourselves, which is huge for us. The secondary pieces like consumption of information for our end users are, you know, the war fighter and then just industry and organizations that want to do business with the Army is how they interact with us has a lot to do with like do they do they have the user manual and how to really operate within the Army and can follow like kind of a clear delineated set of steps that say to get on our network do these five things right.

That's, I know it sounds simplistic, but it's actually very hard to kind of really detail out in a way that's like consumable by everybody, not people on the inside, not people on the outside. So trying to find that way to make information more consumable. And then we have just different areas of just modernization, right? Like we are consistently trying to find the best ways to look across holistically across our programs and, and in in essence, work through how, how you take hundreds of business systems down to like many, many less than hundreds of systems, right?

Get to a really good number. And then and then say, OK, across these, I'm going to start rationalizing the data. You know, historically in a, in a, you know, the Army has systems that they all had their version of ICAM data that was built into it. Now it's like, OK, use that centralized repository for data. You no longer need to maintain that data. So all these different activities is saying, OK, I want to rationalize our data.

What systems should do what lower the threat profile and the threat surface area of the Army and then help figure out better ways to communicate with industry and our war fighters on on information.

Unifying the Network and Standing Up the Data Layer

David Markowitz: So Army's been a multi year journey to kind of build a lot of infrastructure so we can deliver rapid capability to our war fighters and modernize how the Army kind of does it's fundamental man train, equip business to build that scope so we can go faster. We've done several things. We've unified the Army network. We had a bunch of fiefdoms and that's really come together under a unified network plan. We've had expandable compute in store with the stand up of an Army cloud office that provides kind of best in breed from Azure AWS.

We've fielded a series of kind of low code, no code platforms, including a data layer, which is providing rapid capability to those who want a more kind of lowering the barrier of entry for rapid tool development and application development. We've brought in AI tools working closely with OSW as some of their G and AI portions, but also providing cloud ready tools as well. So that can be easily incorporated and use the data that we've now provided.

The area that's a little, it's our new frontier really kind of goal for this year is kind of like access control, which really just kind of said making sure we kind of have the playbook for all of these things. So our industry partners know our environment as and then the new area we're trying to do is access controls to that data that we talked about. Those low code, no code apps have easy ways to get to the data and can can guarantee integrity of the data.

As we start to stand up apps, you could almost think of it as I would call it micro services 'cause we're not that technically astute per SE, but we're providing the ability to do rapid app development off of a central core so we can give capability faster to our war fighters. We're seeing this at the enterprise level and our major flagship program for a tactical environment, next generation command and control, next Gen.

C2 or now C2 and that is providing capability to our war fighters break. Second major modernization priority, Mr. Lee kind of said it first is cybersecurity. Lot of changes in the external environment for our cybersecurity and the department's taking this very seriously. And that includes modernization of some of our old text stat contact debt. But also ways that we want to look and understand how we automate viewing the network, how we automate defense for the network, how do we automate dealing with an actual threat coming in.

What Separates Organizations That Operationalize AI

Amy Kluber: Lots of priorities there that we'll dig into just a little bit, especially in regards to the data later and sort of pushing AI at the edge there. But Ben, before we get there, you've been working not just with the Army, but also other government agencies on sort of this operationalizing AI priority. From your industry perspective, what are you seeing in organizations that have successfully built that foundation?

Ben Van Roo: Yeah, well, thanks for having me here, Amy. The what it says, the most successful organizations for truly adopting and operationalizing AI have thought deeply about everything that was just presented in front of me is how do you think about data layer? How do you think about access control? How do you create the right constructs to enable AI to actually work and be kind of more meaningful than automating my e-mail message or helping me write a promotion letter?

From our standpoint, we are an agentic software harness, so we sit on top of lots of different models. We've been working with the Department of Energy Special Operations Command Army on everything from very small edge devices to on Prem environments and data centers to cloud. So we think a lot about edge to enterprise. And, and so there's a lot of really interesting complex things that we face in how to tie together the models, the compute form factor that you have access to, how you think about resiliency of networks.

And really just like what are the use cases, the war fighters don't care if they can't write emails, maybe necessarily on the edge using AI. But if we've now changed our processes to start to adopt like, hey, you're going to use AI for your after action reports or your intelligence briefs or operational type of workflows, then it becomes really important. And so, So what we see is really everything that has to be laid is there's so much data and foundational work that I think the Army has spent a lot of time focusing on that.

Now as new tools are available online, we can move from even what we know is today. It's kind of like chat interfaces to much more agentic orchestrated AI workflows that are going to attack everything from the most boring things to the most important thing to the war fighter.

The Hardest Problems in the Army Data Layer

Amy Kluber: It'd be nice to attack, tackle the boring things, which I'm sure there's a lot of that across the board. So I do want to get into that data layer. You kind of set that up really well in terms of some of the things to think about in regard to that data layer. So Dave, back to you. The Army, they certainly has, they have put a major emphasis on getting information into the data layer for AI as you mentioned.

So I guess what are the hardest problems that you're still trying to solve and Gabe for for you as well, jump in when you would like. What are the hardest problems you're trying to solve and making it discoverable, curated, accessible, usable and of course secure.

David Markowitz: Great question. 2 directors from our secretary that really kind of drive us. One was to make sure our central data layer advantage was the primary place that we curate least the readiness data that the Army produces really our business side of making sure it's in one spot so we can have it curated and available for the rest of the Army. It's like some of our lessons learned from industry was, you know, don't have a lot of different places have a central place where people can go to, to get kind of get that larger scale adoption.

And the the second was ensuring that we've got a series of bi directional data interfaces between all of our business systems now in the central data layer, which has been a real activity driven by our our deputy Under Secretary for the last really 3060 days and really a huge amount of progress just recently. So the challenges be this actually is now configuration management of the data. And as we're starting to see, we're finding, you know, deduplication where potion where information is making sure we have the right authoritative person in charge of authoritative data so that it can be available for those who need it.

And then making sure we understand the context of that information. The worst case is kind of like, you know, it's it's your five year old's telephone game where you whisper in someone's ear. What do you think the data is? And think of it as now a series of applications, pass it around and by the time you get it back, it's like something completely different. We can't tolerate that for things like how the Army's organized or force structure.

Our finance folks don't want that for like status of funds. You can't play the telephone game. So we've got to find out what's the right spot. And then and now ensuring data integrity for it so that if you have to do a transaction in that area, we understand fully what it means and we've got the right controls, hence this area of access control to make sure we've got the identity. Army has a large obligation for Congress to get audit underway.

And you got to make sure data integrity is there for audit. And our war fighters, we can't like Miss Misplace or make sure we've got tight control over what's happening in a theatre of operations so we can support, support directly what the Army needs to do.

From Firewall Rules to Data Authorization

Gabe Chiulli: And I'm going to add one piece of that. Dave kind of brought it up in the previous discussion, one of the huge benefits of what we have done over the last 45, I'll call it 60 days, it's been a slow trudge of trying to hook up this many systems to like our, I'll call it our data layer, right? One of the really hard problems in that has been first obviously like we have to really explain to everybody what we're trying to do, right?

When we talk about historically everything has been a firewall change, right? Like I need this port to be open to connect to this system. And that's typically the way a lot of organizations think now when we say, look, we're going to connect everything up and now it becomes a data authorization piece, right? Instead of worried about firewall rules, we flatten the network from that perspective. There's no firewall rules now that we have to trip through and try to figure out how to network things together.

And so this wasn't what we had taken advantage of. This bi directional was really a infrastructure problem more than anything else, which was making sure that there was a clear line of traffic so that systems you can talk to other systems using agents. So like any large frontier company and any even non frontier company would tell you in order to really take advantage of agents, everything needs to be fully accessible by those agents, right.

In order to do that, we had to acknowledge that our network was not set up to do that. And so we took a humongous effort and and as he said, the, the the deputy Under Secretary of the Army, Mr. Fitzgerald pushed us and said, connect all these things up. You have 30 days, Gabe, you've got the con because you you have the authority to make this kind of occur, right? And so like, you know, between pencil whipping enough memos to make sure everybody felt like they had the right cover from us and really working through the various systems as Dave kind of talks about, which is what is the authoritative data that we are trying to protect from an authorization perspective.

We're like, hey, the agent can talk to these three things, but should not be able to always talk to this data element over here. So we were doing a couple different things at once. Once is 1 activity is just a firewall and infrastructure rules, which is really making sure that things are plugged together. And then the secondary pieces are like, hey, before we get access, we need to make sure we have these right authorization pieces in place so that when the agent does go and do its thing, we know exactly what that agent is able to do operate against and what data it's going to kind of access.

So the multi fold of that has been huge because now we can say, hey, to be a modern data oriented organization, we've now got the framework in place to start thinking through how we would operate with agents or with AI in a much more clairvoyant way where we're not fighting firewall rules and things that like typically would hinder an organization. And you're kind of lowering the barrier a lot to like interoperability across the across the Army through the actions that we've kind of really undertaken.

Bringing AI to the Tactical Edge

Amy Kluber: Absolutely. That that visibility is also, I know, exceptionally important there. So now going over to the AI at the edge sort of treatment or opportunities that all this work is unlocking. And, and maybe I'll throw it back to you, Dave. What considerations are there when you're trying to bring AI to the edge potentially on a tactical network, a device out in the field or another sensitive environment?

David Markowitz: AI, it's like a really big term and some of it we're trying to like bin it and we've not really come down to I think as a department of like where official bins. But let me let me offer some 1 is like autonomy, you know, flying UAS is swarm UAS is the AI involved with that traditional acquisition testing, making sure the network can support those operations that's going kind of full speed.

There's the embedded into a machine, like hovering like a, you know, MV75 in those days, again, kind of kind of classic. Moving forward. There's the assisting someone with a manual task like target recognition having something blip where the human in the loop, but it's a kind of it's with the weapon system and it's close and you make sure your computing store can do it. Then you kind of get, I think where you're going is, is this other category, which is a bit more relationship between enterprise activity and tactical kind of gets blurred.

Like for us would be called mission command. That's a division level, core level of gate level about where you're getting closer to the edge, but you want some form of connectivity back to some large scale compute, some larger scale data and you want some interactions where you're not exactly sure where we put it. And for us, that's a learning journey with our next generation C2 program or C2, now C2, Army C2, where we're experimenting with where to put this and where we don't.

So it's not as if we know the answer to how to do this well, but we are learning to make sure we are as competent as any adversary out there and can deliver capability fast. So that we're putting compute forward, we're putting the tools forward. We're also making sure they've got connectivity in the rear and to the machinery we just talked about to make sure that it's flexible. Now data integrity between forward and rear somewhere still kind of working, permissions forward still kind of going.

But the idea is to automate a lot of those staff functions that drive forward footprint to get those forward footprint less so we can fight in a austere environment and bring more firepower than anyone else can. And that's what we're here to do.

Model Armory and the Problem With Peanut Butter AI

Gabe Chiulli: And I want to add 1 little piece of that. And and this has been a, not to say a challenge. It's a very interesting dynamic, which is how do we make sure that we don't overload a local commander with too much information and too much AI, right? Because at the end of the day, they still have the war fight, right? And so trying to find that fine line medium, like you may have heard before, we have something called Model Armory.

And what that's meant to do is be a place that we have very tailored models to a very specific mission set, right? Like I always use this an analogy. It's terrible, but I'm going to continue to use it is like the war fighter may only be worried about detecting drones in -32° weather in Texas. And we need a model that's very specific to that. I don't need an LLM to do that.

I need a model that's meant to just look for red lights at that time and period of time and space and all those activities, right? And so trying to make sure that we empower that kind of decision framework versus, I'll call it the peanut butter AI, you know, that kind of gets kind of looked at even from the industry side. They, you know, don't do it purposely, but like they're trying to figure out how to best interact AI into like a war fighting activity.

So we're trying to help kind of that accountability side. And then it helps us understand what level of scale of compute AI, you know, GPU's, whatever might be at that tactical edge because there's not like a, you know, a furnace in the background running and giving you unlimited power either. So it's got to be scaled to what that war fighting is. Absolutely.

Matching Models, Compute, and Power to the Task

Amy Kluber: Ben, sort of what are you seeing in industry that you think could help address some of these constraints or sort of fit into this?

Ben Van Roo: Yeah. I mean, I think I think Dave and Gabe touched on some of the more nuanced challenges that exist that definitely get peanut butter spread over, which is like, let's just let the AI solve it. Like give me the the best model from the Frontier lab. And we're just going to have it like do the things and we wanted to do it on the edge. And then you start to unpack that and it's like, OK, well, that actually one, it might be the wrong sets of models, whether it's a computer vision algorithm that you just need to, you know, find that drone in Texas when it's 32 below or whatever, which doesn't probably happen that often, but but also like that's what you need.

Similarly, even more kind of other use cases that have happened in the world. It's like what, what using LLMS potentially for code development. Well, actually in a lot of cases, Co development's really well suited for other types of models. We've, you know, the, OR guys of the world, the optimization models that exist. There's, there's ways to use simulation. And so, so a lot of what I, I think, you know, what I hear and what we see is, well, what we want to do is we, you know, we maybe have access to a big model on IO5 [IL5].

We want that on our tactical network. You know, we did a bunch of edge work for PCC 6 [PC-C6]. The big project Convergence Capstone 6 that the Army just LED maybe two months ago, and it was very interesting for the groups that we were working with. In some cases, the Army assumed that they had access to the cloud. If you started to take away the cloud, then it really became hardware dependent and and there were there were pluses and minuses of some of the things that happened with the hardware that was available there.

But you really start to go into more technical challenges. So what are the things that you're trying to accomplish? Are you doing lightweight administrative workflows? Do they require very rich advanced reasoning models? What type of software harness that you use to help facilitate the models and the workflows? And all these become really interesting dynamics. And then the follow on questions is like what compute is available? What power sources are available? So, so, so I do think what what we observe is, is we've been working across these problems for almost 4 years at the, at the company and, and, and it, it feels, what's really exciting right now is it feels like we've, we've left the era of, and I've been in natural English processing and predate LLMS for about about 13 years.

No one really got it. And then right around the time, you know, the early models of project Maven, people kind of got some power of AI and we are very early. The company has that one and doing NLP in that space. And now people are getting it. You know, we the army had launched camo GPT [CamoGPT] to help, you know, introduce this for others. They had some early vendors. There's now Jen AI dot mail [GenAI.mil].

People are really wanting to consume AI and in in, but it's still that broad bucket. We want the AI to do the things and now there's a lot of really interesting sophisticated challenges that are right in front of us. And so I think it's a, it's an extremely exciting time to be in the space and try to understand what can, what models are required at what form factor, with what power consumption, what software is around it, what are we really trying to accomplish and what else do we have to do?

And this is the, this is the opportunity to, to test and engage. NGC 2 [NGC2] is a, is a huge program that's really exciting. You know, OSW still pushing a number of programs as well. The other services are as well. And so a lot of very rich kind of challenges are are are coming right in front of us. But I'm encouraged and enjoying to be a part of trying to solve some of these problems. Absolutely.

Ownership and Accountability as AI Becomes Embedded

Amy Kluber: Very interesting Look at that as well from where you sit. So thank you for sharing. We are down to the last couple areas that I wanted to hit and we didn't even get to ownership and accountability yet, which I know is a huge discussion point right now in data discussions and AI. So a question for all of you. As AI becomes more embedded, how is the Army or sort of the industry thinking about ownership and accountability and, and security for that matter?

Gabe Chiulli: Gabe, when it starts off, we've learned how to operate using machine learning in AI in different ways, right? Things that make the plane, the helicopter auto hover, as Dave was talking about, like that's a machine learning algorithm that kind of helps the, the, the, the helicopter autopilot, right? And, and those different functions haven't always had various levels of oversight and how they want to release that into what we call production, right?

And so I, I've always kind of said like, we don't want to have an AI policy at my level because I'll definitely break something in the Army, right? It's just, it's over 2 specific. We'll never get specific enough or we'll be overly nebulous. It'll just look kind of messy, right? So we have to put kind of guard rails around like which ones you should use and how you should operate, but not necessarily like, hey, in order to go to prod, you must do XY and Z steps because I can guarantee you a weapon systems production from test all the way through production to way more complicated than the than the algorithm I used to create an MFR.

And so we, we acknowledge that. And so looking at how best to empower that decision space is kind of what Ben was talking about is releasing enough AI that people actually know what they have capable now, right? And so they can make these really active determinations on how to use it. And then and making sure that they still continue to follow the same processes as they've used over the long period of time that we've done development of capability to make sure that we're we're really rolling that out.

And so the ownership and accountability side is the same from our perspective of making sure that we have the right policies in place where we need to have them delegating down to the same level that we have always delegated down the authorities that they need to make those decisions and making sure we have enough guidance in place that allows them to make those determinations. When you look at it from a security perspective, what do we need to make sure, you know, whether it's a supply chain of the model, it could be the users that are operating the model, all those different pieces.

They're really no different than we've done across any piece of software or hardware that we've always procured. It's just kind of thinking about the problem. And, and, and like I said, we don't want to be overly dictative or prescriptive on, on this stuff because we may actually break something that's actually well defined process and, and actually cause more harm than good. And so we have been working very closely with these different organizations in the army to say, you know, think about it like this.

They broke down like the different chunks AI has been delivered in various different areas, everything from things that are like an app on a phone where you just go in, you ask your question, get a response down to like an FPGA card that's been developed very specific to do an application on a UAV. And so those various different areas have different users, different developers, different use cases, whether you're vibe coding a website or you're making a very specific FPGA card that's been developed to do a very specific task, right?

Those different organizations are very different operators. And so we acknowledge that and we want to make sure that we empower them and making sure, if nothing else, we're able to say, hey, you have unlimited access to this bucket of things and go do good, do go do good work for the Army using the tools that we put in front of you.

Who Fixes the Algorithm When Conditions Change

David Markowitz: I'm going to put a slightly different spin on a challenge area of ownership. And as gave kind of really put a good history on in the past, the algorithm developers were often close to the data and owned the fielding. Now we're starting to see where people are building algorithms and they're separated from the data. And so responsibility to ensure the algorithm is responding right to changing circumstance is a little less clear.

For example, I've got an image recognition algorithm that industries developed to find AT90 [T-90] and it snows and it's, you know, snows and it's 50° warm, you know, some weird thing. And I can't see the tank anymore because the algorithms got confused. Who's responsible to fix it? We've got some really good examples within a department before how to do this, where especially like electronic warfare where signatures are done, you bring that algorithm back to the Fort, they re modify it, give it to you and it works fast.

We're opening the aperture and we're getting more vendors because we want more vendors on AI, but they don't live with the data, may not own it. That AI person who helped us with image recognition for the T90 [T-90] doesn't own collection on now all the images of the tank. So how do we give that to that person? What's that relationship with industry now? So we can kind of get a quick turn because we may be the owners of the data that our industry partner may not.

So how do we ensure that that's on? So part of what Matru is talking about delegation is making sure the people who are experts in the area, who own the data and own the algorithm choice have the ability to kind of make these teamworks work. And we don't want to break that relationship, but we finding it's it's evolving rapidly and we need to strengthen it so that when you field something, you can't walk away because the algorithms change too much.

Or if you think it's going to walk away, buyer beware. So that we understand how the government can still take responsibility and provide the capability our soldiers need. So that's a a friction we're working through. I wouldn't call it metro friction, but it's a growing community and we want to make sure we've got the right community who can only support our war fighters. Absolutely, Ben, being a part of that community.

What Industry Owes the Army Beyond the Demo

Amy Kluber: What are your thoughts?

Ben Van Roo: Yeah. So I think the when I think about ownership and and our role, a lot of it is, I mean, Gabe brought up five coded websites [vibe coded websites] and little apps. It, it's this really interesting moment in time where it's very easy to now stand up a kind of nifty demo and maybe even sell it into, you know, you from the military, you know, someone like, hey, I think I got this use case and it looks great.

But it's like everything around that demo is where it all falls down. And, and it's everything from how you think about, you know, Gabe and Dave brought up things like access control, how you manage the data plane. You know, we do things in the more classified environments where our systems have to manage portion markings. So, so from the software standpoint, if we want agents to traverse across classified, you know, potentially in secret or, or higher levels sensitive areas, But you, that's a lot of, there's some really interesting challenges around how you monitor the models, how you monitor what the agents are doing, all this infrastructure around the software.

It's it's kind of boring to some, but if you're going to roll out and operationalize AI and we take a lot of care and feeding like we, we don't vibe code, you know, our back end stack infrastructure because it's just like it's going to fall over and or it really won't work as soon as you go into a sensitive environment or a on premise environment or out at the edge. And so the, the, the ownership on the industry, what I'd say that I observe is, you know, I, I aspire to continue to find people that are asking me those secondary and tertiary questions.

How do you manage access control? How does it fit on different types of form factors? Again, like, like how does it work with open source models versus proprietary models? Those are the things that when I see the, that it's like, OK, there's a level of sophistication. And this isn't, you know, the, we're out of the flash in the pan stage. And now we're getting into there's really sense of information. It's got to have really high resiliency.

How is it going to land? So, so that's what that's what we spend a lot of time thinking about. And then, you know, it's we're, we're not only serving the war fighter when we deploy something, it's the IT administrators that are like, yeah, what are all these agents doing? How much are we spending on stuff? It's, it's some of the, you know, did we violate an access control thing or whatever? That's the type of thing that I think is also really important.

And then the kind of, I think the last thing is, is, is we still have a responsibility in industry where we're in it every day to find the ways to connect with AI, be honest when we're going on site with a bunch of FD ES [FDEs] and you know, we pull a lot of people from the Army. And so come what works, what's kind of baloney and, you know, not oversell because it's still like it doesn't solve all things.

And, and frankly, what we know is AI and LLMS, they're actually often the wrong solutions we talked about before. So the, the, I think there's still responsibility on an industry in what our role is, in what we should be owning is people are going to be coming to us for guidance. People are mean coming to us for hard solutions. And so we have to kind of present what is easy, what's hard, what's what's BS, what's what's great.

Absolutely. So you're the BS detector. No, but I, I also aspire to continue to, you know, the BS detectors on the other side and they're like, OK, let's what's brass tacks? Where's this industry even going? What do we do? So yeah, that's what that's we, that's we should all be seeking to do.

Getting AI Capability Into the Hands of the Workforce

Amy Kluber: Fantastic. jenai.mil [GenAI.mil]. It's helped a large number of defense users, you know, get access to generative AI and other AI tools. So what do you think the next phase looks like, especially when we're trying to get more development capability into the hands of the workforce and the soldier army started off with various and I, I, I joke when I was talking about it.

Gabe Chiulli: So Doctor Markowitz is a literally a Max PhD [math PhD]. So it's like we, we actually used to get these kind of arguments discussions because like a lot of this stuff is like kind of what he actually study, which is like theoretical math, right? So it's, it's very interesting that we've come from kind of what was those type of math problems into like, you know, the, the, the Jenna I Don Mills [GenAI.mil] sort of capabilities that are out there and how they're going to evolve over time.

You know, us flattening the network was all about making better access to our data. So like the first thing I think you're going to see is better access to the data from these different generative AI tools. Army is going to continue to evolve when it comes to like making sure we're experimenting with what is the next generation of AI, right? There's, there's AI that's developed and delivered to the masses. The masses will use it for automation tasks and doing a lot of good, good pieces.

And then it's on the services to really get with both the CDAO and then even internally and saying what are those niche use cases at the enterprise delivered AI, right? Like I deliver enterprise IT for the for the Department of the Army. Doesn't mean my desktop works on a tactical edge, doesn't mean it works on the helicopter, doesn't mean that, you know, there's a lot of other use cases out there. So I think what you're going to see is we're going to continue to increase the capabilities on Jenny I down mill [GenAI.mil].

We're definitely talking very actively with CDAO on how to tailor that to be better or more inclusive to different use cases. But what you're going to see, I think is just now the, the now that we have kind of that data connectivity and like the access to the AI itself is like, how do we now ejectically turn on different pieces that are kind of low risk and slowly start turning on turning up the risk factors or turning on the capabilities and lowering the risk in different areas.

And so you're going to see some of that I think is like the step one, Step 2 will be kind of that as Army continues to experiment with AI at different echelons that are not really that that enterprise level is, you'll see us continue to experiment with, OK, these edge use cases that like we were talking about earlier, NGC 2 [NGC2] tactical edge AI. How do you want to actually make that useful for the war fighter?

Because they're they're not operating on nipper [NIPR], right? They're operating on their cell phone or a mobile device or whatever it might be that they need to be able to access and do their day-to-day job while they're war fighting. And it can't be, you know, something that's like, hey, waiting for the server to load, like, Nope, we need to make sure that they can do their day-to-day stuff. So you're going to see kind of, I call it now that we've got kind of general AI kind of built out that allows folks to do their day-to-day stuff a lot better.

It's going to kind of free up time for us to really explore those different areas because we were kind of trying to do everything before. And it's like it's, it takes away time from us like focusing on the harder AI problems to try and get it delivered across the board. And so it'll give us a lot better time to really focus on those more niche things that are kind of outside that enterprise delivered AI.

So I think you're going to see a mixture of those kind of things where we're we're going to continuously try to push the limit on what's next and what's the the next generation of, you know, agentics today's where they'll be another word tomorrow. And then just blatantly put protecting against AIAI is also a counter threat, right? And so we have to make sure that we're consistently monitoring what's out there and how we can make sure our cyber being, you know, the DAO for the army, how are they going to protect us from those AI threats across the board?

And so I say across the board, we're looking forward to seeing as AI develops out and you know, there's there's a lot of goodness that's going to continue to to evolve over time.

Which Deployment Model Wins

Amy Kluber: Dave, how about from your end?

David Markowitz: Yeah, yeah. So so Mr. has got like way more experience in this name. I'm like, you know, so you know, if it can be applied, that's like a miracle. You had like last year or two years ago, Gabe, kind of these three different methods of deployment methods for a, for AI. And let me just recap those and kind of discuss maybe the future. One is, you know, kind of your chat bot sitting on the side.

The other is something kind of integrated into a SAS delivery like a for us a Palantir through Vantage [Palantir Vantage] or MSS or a sales force. It's kind of integrated with the workflow development and then you had this kind of cloud ready tools for us through the joint Wi-Fi or cloud compute [Joint Warfighting Cloud Capability, JWCC], which you know can be integrated for more customized there. So which one of these was going to win? I mean, I bets out.

I mean I'm kind of still in the integration with the tool set like a sales force or Palantir because it gets the masses really quick. There's the more specialized tools that you can build down with this JWCC environment. Those are those two areas are kind of hot where we start to end up in the mix between the two, not sure, but they are both becoming wrapped, I think into this kind of like application level, rapid delivery, all this other infrastructure we start off is making the speed of those things faster.

And so I think challenge if it's integrated into a SAS where you're now making like a low code, no code app or I've got now some real quick ability to take some of the cloud native stuff and like field it very rapidly either on the enterprise cloud or some edge node. It's coming like very rapid software micro services. So I think some of the challenges us of this employment methods where I think we're seeing a lot of velocity and speed, which is awesome.

Getting capability to our SOLDIER is management and how do we start to manage and making sure everything is out there So we don't have 1000 flowers out there. And some mentally it's like it, is it OK for that to plant all these things and have both these models push out a bunch of tools and kit and they die and they die in a year or two. And it's OK because the data is there.

And as long as you got consistent data, you'll get the best tool and best rapid fielding. Or is this rapid fielding really a long term investment? You got to make sure you got the skills to kind of keep and keep going on the SAS model. It's a little different way of thinking about it because you've got a real tightly coupled industry partner to make sure it gets delivered and coupled. And on the other one, it's kind of whoa, do we have that internal expertise?

So I think we're still kind of wrestling with what's the best mix to make sure we can kind of get, you know, keep speed to give war fighting capability to our soldiers.

Systems of Record and Systems of Work

Amy Kluber: Actually some things to to look out for Ben from your end from I guess the outside or somewhat inside looking in. What are your thoughts?

Ben Van Roo: Yeah, it's, I think they they really covered Dave and Kate have covered a lot of the right points, even going back to kind of the beginning of this where we're talking about, all right, how do you go from thousands of applications to potentially hundreds and what does that data look like? Where does it exist? When when I think at the enterprise level, you with the work that the Army is doing right now in that space.

If I don't have to go and ask everyone to say actually my the agents we're going to run are going to touch these six systems, can you give me permissions to all of them? I've done that and it's, it's awful. It's, it's, it's really, really hard because people don't have the right contacts. This is hard. So I think thinking about the consolidated plane of Dayla Air [data layer] is really advantageous for the Army for agents to like really when we talk about the work and agents, it's like people are trying to do tasks that often sit across lots of systems.

And so right now you have an opportunity to take advantage of Advantage and Palantir and some of the Salesforce ones, they're big systems of record, which means like that's fundamentally. That was what they're doing and they're going to have a really durable enduring role in this. But it doesn't necessarily. And what my observation is, it doesn't necessarily mean that they have all the workflows, right? And people want to touch other things and people want to go across them.

And, and so the model is like, do we push more into that system of record or do we have some of these kind of universal data layers, layers where agents and people can traverse them with these new workflows? I think you're going to see that happen and I aspire, I hope that that happens a lot more because it really starts to unlock how people are doing work. The second thing I think is going to be a really big change in both the enterprise and operationalizing of AI is now people kind of think like is AI ChatGPT or Gemini, that chat interface and the answer like unequivocally no.

You're going to see people wanting to use it on the flight lines. People are going to want to have geospatial reasoning embedded in it so that the user interface explosion I've believe and what we're observing is now happening. And so when you combine that with AI that can sit across many systems in AI that can look and feel a bit different. That's an opportunity where even some nontraditional vendors, people that are kind of the next generation of system of record, I would say our systems of work or system of agents, things that are designed to take advantage of lots of different agents independent of that base layer of data.

And so that I think is a real opportunity to help consolidate and modernize legacy applications and then allow the warfighter to do more advanced workflows within more common interfaces. To the point, and we've talked a little bit about on the edge, that's where there's going to be some of the most interesting experimental iterations. What compute keeps evolving, models keep evolving and then the software harness to make them work really well keeps evolving.

And So what we observe is on the smallest form factors, the think of it as like you can kind of do the least amount of tasks. As you start adding more GPU's and larger sizes, you can start moving into more advanced challenges and models in reasoning type of activities. Help me do a KOA [a COA], help me do a planning more dynamically because we're out in the field and we can't use the cloud that's going to keep shifting up the curve.

And so I think there's going to be some really radical deployments and, and iterations in that space of the operational LED. So, so I think that's what's coming. And, and I think again, kind of kudos to, to the Army and, and, and OSW and the services, because this is, you know, Maven was kind of game changing at that time. But the pace at which we're all moving now is so much different than than what I experienced, you know, 910 years ago. And so it's, so it's a, it's a really exciting time to be here.

Closing

Amy Kluber: And that agility is, is definitely important that all of you sort of touched on. And I am would love to keep asking questions, but we are out of time. So thank you so much for all your perspectives. It's certainly an interesting time to be in the space. And as I love learning new things from Gabe, I, you know, the whole peanut butter analogy, that was the first time I, I heard that.

So thank you so much for just breaking down what the challenges are in the space with the priorities are and you know what there is to look forward to. So thank you all. Thanks.

‍