Back to all
Webinars

GovLoop | Webinar

How to Reduce Downtime and Improve Operations With AI

Book a 20-minute pilot call

Bring the metric you want to move, like mean time to repair or first-time fix rate. We’ll scope a pilot around it.

Executive Summary

Maddie Wolf, General Manager and Vice President of Operations at Legion Intelligence, joins GovLoop’s Emily Jarvis to discuss how maintenance teams cut equipment downtime with AI. Recorded September 29, 2026 as a GovLoop training sponsored by Legion, the session covers setting reorder points from usage history, troubleshooting against manuals and past work orders, getting fixes right the first time, connecting siloed systems, and tying an AI pilot to one metric. Full summary and transcript below.

Maintenance Is the Supply Line for Readiness

Maddie Wolf, General Manager and VP of Operations at Legion Intelligence, said readiness rests on people, training, and equipment, and maintenance owns the equipment leg outright. Downtime is getting harder to lower because fleets are aging alongside newer systems, and most of it comes down to waiting on parts.

That pressure pushes teams toward controlled exchange, or cannibalization. Wolf compared it to the office fridge: people take what they need, nobody writes it down, and inventory health gets worse.

Three Places AI Cuts Downtime

Wolf walked through three practical steps, each one something maintainers already do without AI:

  • Inventory: set reorder points from historical usage and supplier lead times, and see which suppliers deliver quality parts. One customer found min and max levels that were years old and badly off.
  • Diagnosis: connect a troubleshooting assistant to equipment manuals and past work orders so it can name the likely cause. This helps newer maintainers most as experienced ones retire.
  • First-time fix: show the steps past technicians took and flag where they disagree with the manual, so the repair is done right the first time.

The People Who Do the Job Make the Decisions

IT approves the vendor, but the people who already set reorder points or maintain the manuals should own those decisions, along with the approval gates. Wolf described a first-fix prioritizer Legion built for an aviation maintenance group. It gathers the next sortie time, parts on hand, labor estimate, and which qualified maintainer is on shift. The supervisor still decides which repair comes first.

GovLoop’s first poll found attendees split three ways on their biggest challenge when equipment goes down: finding the right information, getting the right people involved, and figuring out what happened.

Siloed Systems Are a Pit Crew on Different Radio Channels

Wolf compared disconnected maintenance systems to an F1 pit crew where everyone is on a different radio channel. Parts, staffing, and repair steps live in separate places, and teams fill the gaps with calls and messages. Bringing that information into one place means a controlled exchange on one aircraft can prompt a reorder in inventory.

Start With the Data You Have

Connect data only to systems your organization has approved. Wolf noted that Legion is accredited through IL6, so CUI and classified environments are an option. Many teams don’t trust their historical work orders, so she suggested starting with manuals as the source of truth and collecting cleaner data with each new work order.

Common sources include IBM Maximo for facilities, TRAX for aviation, and NALCOMIS in the Navy, reached through an API, MCP, or a straight file upload. On risk, Wolf said free tools often train on what users enter. Access controls and classification markings are usually where teams move from a do-it-yourself chatbot to a paid system built for the job.

Pick a Metric Before You Pick a Tool

Wolf’s main advice was to start from the operational problem and leave AI out of the first question. Pilots that name one metric, such as mean time to repair, carrying costs, or first-time fix rate, and map the process behind it are the ones that get approved. She called technician capacity the sleeper metric as qualified maintainers get harder to hire.

Looking ahead, she wants AI to bring information to maintainers before they ask, such as a depot with the part in stock before someone pulls it from another aircraft.

Legion Packs for Maintenance Teams

Legion Intelligence builds Legion Packs for the maintenance work this session describes. The Facilities Maintenance Pack checks a work order’s symptom against equipment manuals and repair history and proposes a likely cause, technician, and parts for supervisor approval. The Aviation Maintenance Pack gives supervisors a recovery plan for each grounded aircraft, built from its open discrepancies, past fixes, and parts status.

Read the full transcript

This transcript was produced through automated speech recognition and has been lightly edited for readability.

Opening and Introductions

Emily Jarvis: Hello, everyone, and welcome to our session, How to Reduce Downtime and Improve Operations With AI. I'm Emily Jarvis, and I'm the Senior Manager for Events here at GovLoop. Today I have the added pleasure of also being your moderator, so thanks so much for joining me. And a special thank you to our sponsors at Legion Intelligence for helping us put on this session today.

Let's get right to it, because when equipment goes down, maintenance teams need to answer the call quickly. Too often they lose valuable time searching for technical information, locating parts, or working across disconnected systems. AI can help teams find what they need faster, move work forward, and get equipment back in service.

So today we'll talk about how organizations are using governed AI to reduce that downtime and simplify maintenance workflows. We're going to talk through real-world examples from complex maintenance environments, including lessons learned from the Department of War's operations. But before we get to all of that, you all know the drill. We have a few housekeeping items to go over.

First things first, we have some great resources available for you on the console under the Resources button. This is also where you can learn a little bit more about our speaker today, so make sure you check that out. If you need the closed captioning feature, just look for the Show Captions icon along the bottom right-hand side of your panel. That's where you'll find that added accessibility.

We hope you don't have any technical difficulties today, but if you do, send a message to the Q&A tab and my colleague will get you all sorted out. You can use that same tab to submit questions to our speaker, and we want your questions, so don't you dare be shy. Put your questions in early and often. You can also chat with your fellow Govies using the chat button along the bottom panel.

Finally, we are recording this training, so check your inboxes tomorrow, because we'll email you a link so you can watch the entire session on demand. All right, with all of that housekeeping out of the way, I want to introduce you to today's speaker. Joining us is Maddie Wolf, who's the General Manager and Vice President of Operations at Legion Intelligence. Maddie, thanks so much for being with us today.

Maddie Wolf: Yeah, thank you for having me, Emily. It's nice to be here.

Why Maintenance Is the Supply Line for Readiness

Emily Jarvis: Of course. All right, so we're going to get into the discussion today. I just want to note for our audience that we're also going to have a few polling questions, so we're going to get you involved as well. And I know I mentioned this at the beginning, but don't you dare wait until the end to ask your questions, because Maddie's here to answer them all. We're excited to make this an interactive discussion.

So, Maddie, let's talk about the DoW a little bit. When you're out there talking to DoW leaders about operational readiness, where does maintenance and equipment uptime fit into the broader mission picture?

Maddie Wolf: Yeah. First off, before I answer your question, I just want to say to the audience: please do put your questions in either the chat or the Q&A. All is fair game. I will answer pretty much whatever you want. If you have a question, if you are curious, I would way rather not just talk at the screen for an hour and actually have a good dialogue. So please go ahead and put them in, and I'll monitor as I talk.

But to start out, leaders talk about readiness, but ultimately it is the maintainers who make readiness happen, right? Anything around readiness or availability rates, that's what I think of as the scoreboard. And ultimately, readiness rests on three principles. It rests on people, it rests on training, and it rests on equipment. Maintenance owns the equipment leg outright, but a lot of it also has to do with the people and the training.

When I think about it, maintenance is sort of the supply line for readiness, right? If you cut it down, everything else downstream goes with it. So I think of it as really the front line.

Aging Fleets, Waiting on Parts, and Controlled Exchange

Emily Jarvis: Gotcha. Okay, I want to take a half step back here, because obviously the overall goal is reducing downtime. That makes sense, and I think we've all experienced this in our own capacity. The Wi-Fi goes out at your house. Currently we have a boil water advisory at my house, because our power plant is actually out of power, shockingly. So I think we've all experienced these instances, and the reduction in downtime is so important. It becomes a critical focus for organizations in keeping mission-critical assets available and operational.

What's shifted in that conversation around downtime, and why is it becoming perhaps an even more influential part of the discussion than it was before?

Maddie Wolf: Yeah, so I think there are a couple of different things. Downtime obviously is just: if you're down, you're not up, right? Pretty simple there. The issue is that our fleets have become more complicated, and by that I mean a lot of them are aging. Not only do we have this new influx of technology coming in, with all these new, more modern fleets, but you still have, whether it's a fleet of aircraft or facilities or whatever it is, stuff that is 10, 20, 30, 40, 50, dare I keep going, 60, 70 years old. That makes things a lot more complex.

And really, downtime is typically waiting on parts. So the aging nature and the ever-expanding nature of the fleet, or the inventory, or whatever it is someone's maintaining, means the more challenging this is going to get. What it comes down to now is that people are having to get creative when it comes to how do I make sure this fleet or this asset is mission ready and that I'm reducing that downtime.

So they're having to do things like cannibalize, bluntly, which you don't always want to do. Controlled exchange, or however you want to call it, is a very real thing that has to happen. You have to start making trade-offs. You have to get creative about talking to different locations, if you're able to share information at all. All of those things are starting to happen.

And it's kind of a self-compounding problem, because if you think about getting really creative to drive down downtime and doing something like controlled exchange or cannibalization, I think of it like the office fridge, right? You go into the office fridge and take something out of it. Hopefully it's yours. You take the sandwich out of it, you take a drink out of it, you take something else. Nobody's writing down what they took. Nobody's replacing it. So it often makes your inventory health even worse.

There's a reason for doing it, and there are times when it's appropriate. But I wouldn't say downtime has become even more of a focus. It's just that it's gotten harder, right? It's gotten harder to lower it, and to lower it well.

Emily Jarvis: That makes sense, and the way you explained it in the context of the environment we're currently operating in makes sense.

Three Places AI Cuts Downtime: Inventory, Diagnosis, and First-Time Fix

Emily Jarvis: All right, I don't want everybody to get too excited, but we already have one question in the Q&A portal. Our friend Dolores is the first one to throw a question in, but she shouldn't be alone, so put your questions in. She asked, "What are some examples of decreasing downtime and increasing operational efficiencies with AI?" She's really nailing the theme, going right into it.

And then she has a follow-up asking about responsibility: who's in charge of monitoring and governing these systems and understanding when a disruption occurs? Who's really on the hook for that? So let's start with any examples.

Maddie Wolf: Yeah, so let's just be super practical here. Decreasing downtime. Let's start with the first thing: why is something usually down? It usually starts with, do you even have the part to fix this? Yes or no? There are a lot of little things you can do to make sure you actually have the inventory you need.

One customer we have started with this building block. They asked: do we have the part, do we have the part in stock, are we going to be able to get the part? And they realized something as simple as their reorder points could be improved. I'm sure everyone here knows this, but in case not, you set min/max levels in your inventory to say, hey, we should order when we get to this certain amount, or we shouldn't carry too much here.

A lot of people, honestly, have kind of made them up, or they were set some years ago. Some of them are good and some of them are way off. I've taken a look at them with people and they're like, "Okay, there's no way we need 50 of these things," or, "We go through 10 of this material a week, so why is our reorder point two?" Something super basic like that.

So one thing you can do is use AI to set reorder points. Use it to look at your historical information and the lead times typically coming in from suppliers, all of that. It seems small, but just allowing yourself to have the inventory you need can help. You can also use AI to understand which of your suppliers are good and which of them kind of suck. Sometimes there are some sneaky ones in there, getting in where the quality isn't there. So one is: start with the inventory. Start with the parts you need.

Okay, let's say you've waved your wand somehow and solved that problem. What's the next part? When you have the equipment, are you able to diagnose the problem very quickly? Sometimes it's easy. If we're talking about something like facilities maintenance, it might be, "The toilet doesn't flush." Okay, I've diagnosed the problem. Sometimes it's significantly harder.

So one thing we've done is use an AI troubleshooting assistant, where you hook it up to the different manuals or guidance for the equipment, as well as historical work orders if you have those. You can hook it up to either or both and just ask, "Why is this thing happening? What's the issue that's probably behind this?" And it uses the historical information or the manual to say, "Hey, eight out of 10 times, it's probably this problem."

That's really helpful, I would say, particularly for less experienced maintainers. Maintainers who have been doing this for 30, 40, 50 years probably know what's up. But especially as the workforce starts to slowly retire and you have new folks coming in, getting them up to speed quicker is a big one.

Then part three: let's say your inventory is looking pretty solid and you know what the problem is. How are you going to fix it correctly, and correctly the first time, so you don't have to go back? You want your first-time fix rate to be up. So that means being really prescriptive with maintenance. A lot of people talk about predictive, but prescriptive means saying, "This is the optimal way to fix it, and this is the stuff you should use to do so."

That's something you can, in some ways, use AI for, to say, "When this technician solved it previously, these are the exact steps they took, and according to the manual, that is generally correct." Or, let's say manuals are sometimes wrong: "The manual says to do this, but the last three technicians all noted that's not right." So it's about making sure you're bridging those gaps. That's how I think about it.

So for concrete examples: look at ways to improve your inventory. Make sure you're quickly diagnosing the problem, and once you diagnose it, that you actually fix it right. All of those, I would say, are pretty basic. They're all things we've done before without AI. It's just that AI can make them a bit better.

I think a lot of times AI can be this spooky, magical tool, and sometimes there's definitely stuff where you go, "Whoa." But in terms of maintenance, there's super practical stuff you can do. And people like me are weirdly interested and involved in it and would love to talk about it anytime. I think I answered the first part of that, but I know there's another part.

Who Owns AI Decisions in Maintenance

Emily Jarvis: There is. But we'll let you take a break for a second while I pull up today's first poll. For our first poll, we're asking: when equipment or systems go down, what's usually the biggest challenge? We'll give you a few seconds to answer. It should be a pop-up on your screen. If for some reason you're having technical difficulty, just throw a note in the Q&A and we'll get you sorted out.

Part two of Dolores's question, Maddie, was around the idea of who's in charge of this. Who's responsible for monitoring and governing the connection of information across systems without disrupting the existing workforce? You've obviously got folks on the ground working to maintain these systems. Who do you see as most responsible, or on the hook, for making sure downtime is as small as it can be?

Maddie Wolf: Yeah, so there are two potential answers here, and I will say I think one is more right than the other. One is, of course, when you bring in AI, you're automatically going to have IT, or whoever it is, say, "Whoa, what are you doing?" And that's understandable. You want to make sure whatever provider you're working with isn't doing super weird, crazy stuff you don't want to know about. That is valid.

But once you have a vendor whose AI has been approved by IT, the person who should own, let's say, setting the reorder points, or hooking up your manuals, is the person who already does that function anyway. You might need to partner them with somebody, whether it's a vendor or someone internal, to make sure they're setting it up. But I should not set reorder points myself. That is a terrible idea. I do not know what any of my customers are doing at any given moment. I don't understand their business as well as they do.

So I think a big thing is, whatever you do, make sure the humans who usually do the job, who have the knowledge, are the ones making the decisions around this type of stuff. You don't want to introduce some new workflow, or some AI agent that can talk about all the jargon and buzzwords, that totally disrupts the way you do things today. The whole point is to do things smarter and enhance the way you're doing things. It doesn't mean you delete everything you've been doing already.

So I'd say stick with the people who are experts. Those are the people who have the knowledge and who will know which gates are appropriate, which people need to approve which things, and how that should all be baked in.

Emily Jarvis: I love that context, too, because so often when we have the conversation around AI, it's about AI coming in to replace people's jobs, or do the jobs for folks, or give them the opportunity to do more high-level work. All of that could potentially be true, but the knowledge base from the human who's actually on the ground and has been doing the work also needs to still exist. You're exactly right: we can't know exactly what should be reordered for these large DoW installations. What we can potentially provide is the tool to help them do that. So I think that context is so critically important.

Maddie Wolf: Yeah.

Finding Information Fast: A First-Fix Prioritizer for Aviation Maintenance

Emily Jarvis: All right, about half of folks have taken the poll, so I'm going to leave it up for a few more seconds. And I want to get your take on this, not to spoil their answers or influence them in any way. Maddie, when you're out there talking to your government clients, where are you seeing the biggest obstacles maintenance personnel face when they're trying to find the information they need to diagnose or resolve an issue quickly?

Maddie Wolf: So I don't know if this is the most, but it's definitely the most complained about, and it's going to be number one. Finding information quickly is perpetually the thing I hear people grumble about, or go a little louder than grumbling. And what that information is differs. It could be as simple as, "Where is this thing?" It could be, "Is X part the same as Y part? Can I use these things interchangeably?" It can also just be bringing together a bunch of different information. People say disparate information, or whatever fancy word they want to use, but basically: how can I find a bunch of stuff that's all over the place and make sense out of it?

So one thing, for example, that we built recently was a first-fix prioritizer. What do I mean by that? This was for an aviation maintenance group. They have a bunch of aircraft coming in, and they're trying to figure out which one to work first. Their whole thing is mission readiness: when I'm told this aircraft needs to fly at seven o'clock tomorrow, how do I make sure it's ready to fly? That's ultimately what they're going for.

But a lot of pieces go into deciding which of these discrepancies or gripes is the first one you knock off, and what order you do it in. So for that, it's: let's take the things we can count. When's the next sortie, or the next flight, for this aircraft? That's part one. Part two: do we actually have all the parts in inventory for this? If we don't, should we do some sort of controlled exchange, or are they somewhere else? How long will this labor take? Do I have the right person on shift right now who's qualified and able to do this?

All of those things are things you don't need AI to figure out, but it can make it a heck of a lot faster. If you can just say, "Hey, Mr. or Mrs. AI," I'm giving it that title, "go get my answers for me and come back," that is a lot easier. And then that person should make the decision themselves. But rather than telling them to go call a bunch of people, or figure this out, or walk over to inventory, it's kind of nice having a little buddy that can bring it to you.

To be clear, I don't think the little buddy should make the decision. Mr. AI shouldn't always make the decision, because sometimes there are random little factors, like, "Hey, we heard from someone over here that this thing is really important," and you should take that into account. But it can actually help people.

So to me, that's the big one. And honestly, I feel like this is kind of cheating, now that I'm thinking about it, because that app we built touches finding the right information, making sure the right people are there, and figuring out what happened. So it ties back to this poll, and I kind of feel like I cheated a little bit. It's kind of all of them.

Poll Results: Information, People, and Root Cause

Emily Jarvis: Well, it feels like the audience is in agreement with you, because we have pretty much a three-way split between finding the right information, getting the right people involved, and figuring out what happened in the first place as our top three answers for the biggest challenge. And I think what you touched on hits all of them.

Maddie Wolf: Except, apparently, not tracking down parts. Apparently this group is very anti-tracking.

Emily Jarvis: They've got that inventory on lockdown. I can't find my keys half the time, so I'm impressed with this group.

Siloed Systems as a Pit Crew on Different Radio Channels

Emily Jarvis: All right, so let's talk a little bit about disconnected systems, siloed data, and legacy processes. All of these contribute to downtime and operational inefficiencies.

Maddie Wolf: Yeah.

Emily Jarvis: These three things are real challenges that exist in agencies, because no one's starting new right now. We're not coming in and creating a new environment; we're dealing with the environments that exist. So these three key areas are obviously going to bring challenges into the mix. How do they work together, or independently, to create these operational inefficiencies?

Maddie Wolf: Yeah, so I think this is probably the best place where that AI glue I was talking about is helpful. I'm not a big F1 fan, but I have been deeply into the show Drive to Survive, so bear with me for a minute here.

In Drive to Survive, in the F1 race, it's all about these pit stops, where they stop and change out their tires. The little guys come running out, sometimes also women, whatever, and they're switching the tires as fast as they can. Everyone's like, "Go, go, go." Then they switch, and the car speeds off, hopefully not too fast or they get a penalty.

I think of siloed systems kind of like this race car pit stop situation, except imagine they're all on different radio channels. You're trying to do something as fast as you can and get this thing done, except they're all on different channels. So they're like, "Oh, is the car here? Is the car not here? Did you fix it?" They haven't done that nice, choreographed F1 repair dance yet. That's kind of how I think of siloed systems.

To take this back a bit, in the example I shared earlier about what to fix first: someone comes in, and they know they need to fix this plane. Their question is, do I have the thing in inventory that I need to fix this plane? Do I have the person I need who's able to do this job? If the thing isn't in inventory, where should I get it from? How do I fix the plane? All of those answers live in different places.

So to me, what you need is some sort of layer that goes and grabs all those different parts and choreographs the Formula 1 car dance of maintenance. This is a very weird analogy, but hopefully it makes sense. You need to make sure the pit crew for maintenance, in this example, all know: "Hey, the car is coming. Hey, this thing was changed. Hey, I need to do this thing." Right now, for most groups, it's a lot of calling, a lot of Teams messages or texts. Just bringing all of that information into one place is going to make this a heck of a lot easier.

So I think the biggest thing to me is: how do you unite the information and then actually make sense of it? If this thing's happening here, this other thing needs to be happening over here. If someone's doing controlled exchange and taking a part off this plane, is someone ordering the backup in inventory? Those are the types of things that need to be happening.

Emily Jarvis: Yeah. I am also a big fan of Drive to Survive, so I'll have to pick your brain after the webinar today about your thoughts on McLaren. Glad we're getting to all the important things.

But I have heard this example of the pit crew multiple times in relation to pediatric heart surgeons. There was this issue where heart surgeons were performing these miraculous surgeries, but their patients were coding on the way to the ICU. They couldn't figure out for years why this was happening, because they had a successful surgery. They'd done test after test. And they actually brought in an F1 pit crew. I'm not kidding. They brought in an F1 pit crew to look at their systems, and they realized people were going all over each other. Certain surgeons were moving this way, nurses were moving that way. Nothing was coordinated.

So when they tried to move this operation, with all these different connected pieces, they weren't moving in an effective way. The pit crew told them they needed to be able to talk to each other and rethink the systems. They created protocols that hospitals all over the country still follow. I think it's so cool, because you think about these systems being so independent. The heart surgeon is doing the arteries and veins, or whatever. I'm not a heart surgeon. But to understand that they had to bring in somebody else to rethink the processes in place goes to your point. You've got these siloed systems. Each one is working correctly, but together they're making a bit of a mess and producing a result that doesn't work so well.

Maddie Wolf: I feel like I just learned something. I'm going to look that up after this.

Emily Jarvis: It was something I learned on a webinar not too long ago.

Getting Data Into AI Systems Securely

Emily Jarvis: All right, we've got another couple of questions in the Q&A portal, so thanks to everybody for adding those. This question is around data going into these systems. Obviously, in government systems, data is incredibly valuable and important. The question is: how is data entered into a system when it doesn't currently exist in some database? How are you bringing new data into these systems and making sure it's vetted and secured in the proper way?

And following that: what are some common sources of data for maintenance action, and how do maintainers access the tool or the data with AI for support? You talked about AI being the assistant here. How do you retrain that model for your teams?

Maddie Wolf: Okay, so let's start with a couple of things. For the first one, how is data entered into a system that doesn't really exist? You can throw whatever data you want at AI, so long as it's in an environment where you are allowed to do so. There are cases where, let's say, you have manuals that are on paper, or work orders shoved somewhere else. You need to find a way to connect those to AI, and you also need to make sure you're allowed to do so.

There are now increasingly more vendors out there, Legion's one of them, that are authorized up through IL6. We can do anything with CUI or even Secret, all that kind of stuff, and there are more vendors that can do that. So increasingly, you're able to share that kind of information. But once again, make sure whatever organization you're part of has said, "Yes, this vendor is blessed," before you dump stuff into ChatGPT. Please don't. There are increasingly organizations that have agreements with different vendors, even folks like OpenAI, where they've said you're allowed to put this level of information in there.

Now, if the information doesn't exist at all, bluntly, we can't help you. If you don't have any data, you've got to have something. Whether it's a manual from the 1970s about a certain aircraft, or, "These are the people and the shifts they're working," you're going to have something you can put in there.

Starting Without Perfect Historical Data

Maddie Wolf: As I mentioned, we've talked to some people about historical work order information in maintenance, for example, and pretty much everyone we've talked to has said, "We don't trust our historical information. We don't think the data quality is there." There are ways to work around that.

Like I mentioned, if you're doing troubleshooting and you don't trust your historical work order data, or you don't have it, use the manuals first as your source of truth. Start gathering data as you do your work orders, so that one day you have historical data you trust. Then, in the future, make the switch, or combine the two.

I hear a lot of people say, "Oh, we can't do this thing yet because our data isn't ready." No, you can't do the end-all, be-all dream, but you can do something. Think of it like driving a car. You can follow Google Maps. Maybe you can't take the shortcut the local knows about yet, but you can follow the general path. There are always things you can do.

So to sum it up, how is data entered in? There are ways to integrate it, and there are ways to just drop it in. Just make sure that whatever data it is, you're talking to your security or IT organization and you have something that's blessed and that you're allowed to use. And if you're hearing people say, "Oh, this is CUI, we can't do it," that may be true for that system. It is no longer true for all systems. So if you're living in that world, talk to them, because we exist. We're out here.

Common Maintenance Data Sources and How to Connect Them

Maddie Wolf: In terms of common data sources, the most common ones I see are, one, a CMMS, a computerized maintenance management system. In the world of facilities, I hear IBM Maximo all the time. In the world of aircraft, you hear things like TRAX, or the different brands people work with. In government, it might be a Navy system like NALCOMIS, and others like that.

You're either able to integrate directly via an API, and I don't want to get crazy technical, or there are other ways to do it. There are new connection protocols; you may have heard the term MCP. That's another one. And you can always just do a bit of a data dump, where you say, "Hey, I have a bunch of these really wild manuals from the 1960s that are in some shared drive somewhere." That's another way we can do it, too.

So there are ways to connect to things. Just make sure, once again, that you're allowed to connect and that the environment is capable. But you'd be surprised at the kinds of systems we can connect to, even ones that are 40 years old. It is weirdly doable.

Start With the Operational Problem

Emily Jarvis: Oh, well, we love things that are weirdly doable. All right, I'm going to go ahead and throw up our second poll question. For this one, we're asking: how would you describe your organization's use of AI today? We're just trying to get a pulse check here.

While folks are answering that question, one of the things that is so interesting is that for myself, I sometimes find myself saying, "Well, I really want to explore an AI use case, but I'm too busy right now to get it done," which I know is counterintuitive. But I think it's a challenge a lot of organizations have: carving out the time to pilot AI and demonstrate how it could work, because you're talking about overhauling not only the technology you're using, but the process by which you're doing it. So how can organizations adopt AI without disrupting existing maintenance workflows or requiring all the technology to be transformed?

Maddie Wolf: I'm going to say something, and don't hate me.

Emily Jarvis: Okay.

Maddie Wolf: I think you're thinking about this the wrong way, and I think a lot of people are. A lot of people say, "Someone's come to me and said, use AI," or, "do AI." And I will be honest: when someone comes to me and says that, I'm like, "To do what? What am I supposed to do?" And I work in AI, and I still feel that way.

So the question I always ask people is: forget AI. I know that's ridiculous given the context, but just ignore that. What is the problem you're having right now, or what are the few biggest problems you're having as a business? Is it, for example, that your first-time fix rate is super bad?

Or maybe it's something as pesky as this. I went to one site months ago, and it still sticks with me. They said, "We have security checkpoints in certain areas of our site, and if I don't bring the correct thing with me and have to go back out of that security checkpoint, I just wasted a full hour." So they asked, "How do we make sure people show up to the problem with the stuff they likely need, so they don't have to leave?" That's not an AI problem. That's just a normal operational problem.

So my question to people is always: what are the operational problems you have? And then, can we solve them with AI? Sometimes the answer is no, and sometimes the answer is yes. My whole thing is: don't just use AI to use AI. Fix whatever problems you need to fix, and maybe AI can help. There are people like myself who can say, "Yeah, for sure." There are also times when I see problems and I'm like, "You don't need AI at all. Don't do that."

So rather than thinking, "What can I do with AI?", think, "What problems do I have?" Then someone who knows AI well, or maybe even you, just messing around with something like ChatGPT, can figure out what types of things it could help with. And the answer is most things, but honestly not everything.

Emily Jarvis: Yeah, I think that's good context, too, and I don't hate you for that answer by any means. I do think there's this pressure sometimes of folks saying, "You've got to use AI. It's going to save X amount of time and create all this efficiency." And I sometimes panic and think, "Well, how? I don't know." But thinking through that problem-and-solution lens makes a whole lot of sense.

Risks of Using AI for Troubleshooting

Emily Jarvis: All right, I'm going to give folks a few more seconds to answer the polling question. We did get a couple more questions in the Q&A portal. This one, Maddie, is about risks, and we love to talk about risks. Are there risks of using AI for operational troubleshooting that could result in unintended sharing of trade or operational secrets? They put "secrets" in quotes, so I will as well.

Maddie Wolf: Yes, there are risks. It depends on your environment, but there are a couple of things to know. If I'm thinking of using an AI chat to troubleshoot: if you use a lot of the free solutions out there, and I'm just generally putting them in a bucket, so please don't sue me, whoever you are that doesn't do this, generally, free solutions often train on data, particularly proprietary free solutions. As you move up to their paid tiers, a lot of them have agreements with their paid customers that say, "We won't use your data to train our models." So that's bucket one.

There's also open source in that bucket. If you use an open-source model, it's not necessarily training on your information. That's the whole joy of open source: you can host it yourself. But model training is one risk. The other risk is that data somehow leaking out, or someone getting in. So there are things you need to do about limiting data exposure to the model itself, limiting data exposure to external environments, and even within your organization. Do you have the access controls you need, because this person over here may be allowed to know different things than this person over here? All of that needs to be thought about, and I recognize that's opening a can of worms.

But interestingly, that's actually where I see people making the jump from DIY-type solutions to ending up paying for something. It's usually right there. Bluntly, you can build a pretty decent AI troubleshooting chatbot with tools online, or you can use GPT or something like that, and it does a pretty dang good job.

Where people start saying, "Hey, I might actually need something official," is when they start thinking not just about model training, but about the access control layer. Are all my documents being stored in my cloud? Are they being stored on premises in my data warehouse? Are they going out to someone else? You think about attributes, you think about classified markings and who can see what. That is a huge can of worms, and solving it yourself is really dang hard. That's usually when I see people make that shift. Hopefully that answers it. And I see the poll is up, and I want to get to that.

Poll Results: How Agencies Use AI Today

Emily Jarvis: I think you did answer it. And it looks like the vast majority of our audience, 58%, are exploring what's possible, which we love to see. About 6% said they're actively using it in operations, 26% said pilots or test cases, and 10% said they weren't quite exploring it yet. I'm sure that's going to change in the next year or so. But more than half are in the exploration phase, and I'm not surprised by that number.

Maddie Wolf: That sounds about right.

Operational Benefits and the Technician Capacity Metric

Emily Jarvis: Yeah. All right, we shockingly only have a few minutes left, so we're going to do these next couple of questions in rapid-fire format. We've talked quite a bit today about reducing downtime and why that's imperative. Are there any other operational benefits you see from the organizations you're working with that are implementing governed AI? Any use cases or best practices they've found?

Maddie Wolf: Yeah, so there are two buckets I put the benefits in. There's the feelings bucket, and there's the operational business bucket. When I think about the benefits of AI, there are certain things that are about how I make people's experience less crappy. How do I make people actually enjoy the work they're doing a bit more?

When I think about maintainers, they signed up to maintain. They did not sign up to go chasing a manual around for an hour. That is not what they want to do. So one bucket of AI benefit is simply: how do I allow people to do the job they were actually hired to do, and not the stuff no one wants to do? That's bucket one.

Bucket two is: what are the measurable operational outcomes? Unfortunately, that's what most leadership actually cares about. Sometimes it's both, but sometimes it's really, is this thing measurable or not? So downtime is one. Mean time to repair is a really big one. Reducing carrying costs, we see that a lot. First-time fix rate is a big one coming up. There are a lot of key metrics that come up pretty consistently.

I would say the sleeper metric I keep hearing about is technician capacity, particularly as finding really qualified maintainers is increasingly difficult, especially as folks retire. How do you make sure the maintainers you have are spending as much time as possible doing the job you hired them for? One, that's what they want to do anyway. No one wants to chase around paperwork. But two, it's actually good for your organization or operation that they're spending their time doing that, and not chasing down stuff. I see both of those happening.

Where AI Takes Maintenance in the Next Few Years

Emily Jarvis: Yeah, that makes total sense. All right, poll number three is going up right now, and this one's more of a fun poll. We're asking: if AI could solve one work headache, what would you pick? Finding the document you need, completing routine paperwork, tracking down the right expert, or sitting through fewer status meetings. I know what I would pick, but I'm curious to see what our audience picks as well.

We've got a couple more questions in the Q&A portal, though. The first one is really around the future and what's on the horizon. AI has been around for a couple of years now, and it's being implemented in organizations in a lot of different ways. What do you see next, especially through the lens of the DoW? Looking ahead, Maddie, where do you see AI changing the way DoW organizations approach maintenance, sustainment, and readiness over the next three to five years?

Maddie Wolf: Yeah, so two things. One, if anyone tells you they know what's going to happen with AI in five to 10 years, they are either lying or they are delusional.

Emily Jarvis: Yeah.

Maddie Wolf: We don't know. It's changing super fast. So I can't tell you that. I can tell you what I'm seeing right now and in the immediate near term. I would say it's starting to take some of the guesswork out of maintenance and letting you be more targeted in the things you're doing.

There was a big rise in predictive maintenance in the last 10 years or so: how do we use sensor information to predict what's going to fail, and so on. That's still a thing, but there are increasingly other parts we can also get data around or figure out. For example, looking into 30 years of historical work orders to tell that the best way to actually fix something is a certain way, which was never written down anywhere. You're just aggregating that information. That's the kind of stuff people can focus on now.

So it's optimizing and tuning things you usually wouldn't be able to pick out, from treasure troves of gross data that no one could really use before. I think that's where it's starting to go: making sure every single technician is more efficient, and leadership is more efficient. That's what I see happening in the next one to five years. What happens after that? You tell me. I'll let you know in five years when we're there. At least that's where I am now.

Where to Start Learning About AI

Emily Jarvis: That makes sense. All right, one of our wonderful attendees, Vanessa, said, "I recently found out that my agency has its own website on AI. Are there other specific places you would recommend folks go to obtain good training, or is it just spending more time learning what your own organization already has?" Where do you go if you're AI curious but need a little direction?

Maddie Wolf: Yeah, I think the biggest thing right now is to just play around with the tools you're allowed to play with. In DoW, GenAI.mil is great. If you're interested in a chat experience or building simple workflow agents, that's a good place to go. We see a lot of people start there as a jumping-off point. So if you're earlier on, feel free to poke around and play with that, assuming you have access to it. That's a great place to start.

Then, as you're messing around and coming up with more complex ideas, or things that require more intensive agents than something like GenAI.mil can provide today, and I know it's beginning to get better, I would encourage you to talk to people. There are a lot of people within the different services and the different groups of DoW who are tinkering, playing, and experimenting with this stuff, and the more you all talk to each other, the better.

I, of course, am here and down to talk to anybody. But I also know I don't have all the context you all have about your work and your role, and I would not be privy to everything. So I would say play around with the tools you can play around with, and then start talking to people who are in a similar boat, and see what you can make of them. There's a lot out there.

AI That Pushes Information to People

Emily Jarvis: All right, I've got two more questions for you. Dolores posted one of these in the Q&A portal. She asked, "What intrigues you most about AI right now?" I thought that was such a unique question. We always stumble into how it's going to help, but what has you fascinated in the world of AI right now?

Maddie Wolf: Okay, this is going to sound simple, and it is, but this really got me a year or two ago. Everything in AI right now is about you going to AI to tell it to do something. And someone, the CTO at Legion, the company I work at, said, "Why are we going to AI? Why doesn't AI come to us?" And I had this Yoda moment of, "Oh my God." It was so simple, and it seems kind of stupid looking back at it, but it totally changed the way I think about things.

So what excites me about AI is that, rather than building these tools or workflows, like the first-fix prioritizer or the controlled exchange helper I mentioned before, and having people go to them, we should be pushing the information toward people. What if, instead of me triggering some tool, it came to me and said, "Hey, you're about to do controlled exchange and take this part off an aircraft, but actually this depot over here has one and can get it to you later today"? It would keep me from making a mistake, or from missing the insight I otherwise would have had.

We should be having AI tools push information to us so we can make decisions, rather than making people go find a tool and go ask it the question. I shouldn't even have to know to ask the question. It should just come to me. If my reorder points are off, why doesn't it just tell me, rather than me asking it? That's what excites me: the more we're going to see it pushing information toward us, rather than the other way around.

Emily Jarvis: That definitely sounds like it would be very useful, because I forget things all the time. Well, I would never, but theoretically.

Maddie Wolf: Other people might.

Biggest Opportunity: Pick a Metric You Can Measure

Emily Jarvis: Other people might. All right, we have our final polling question up right now, and we're asking about the biggest opportunity. So I'm going to ask you this same question, Maddie, and then pair it with your takeaway for today. Where do you see the biggest opportunity for AI to improve operations, and what do you want attendees listening to today's discussion to walk away with?

Maddie Wolf: I mean, all of these things are true. I'm just looking at this poll: faster access to information, that's true. Better maintenance planning is true. The most important one, I think, depends on who you are. Leadership is going to care about reduced downtime. The techs themselves are going to care about getting faster access to information. Maintenance planning is going to matter to someone in that role. All of those things are true. It just depends on who you are.

I hate not answering questions directly, and it feels like I'm not answering this directly. I think all of those things can happen, and the biggest opportunity depends on who you are and what role you're in. The ones I'm seeing a lot right now are better maintenance planning, faster access to information, and improved collaboration across teams. All of those can lead to reduced downtime and disruptions.

And if you want to have AI at your organization, bluntly, someone's going to say, "Why should I pay for that?" You need to have a metric to point to. So that's the biggest thing: pick a metric you actually want to affect, that you can then measure and point to.

People talk about AI pilots failing. They fail all the time, because people just say, "Let's use AI," the same thing we talked about before, and they don't pick a metric and define a workflow, or something that fits that metric. When you start doing that, you get a very high success rate. At least that's what we find. Pretty much all of our pilots convert. But it took us a while to figure that out. It took us a while to figure out that we need to pick a metric and go all in on it, rather than just generally thinking about what we can do.

Closing Takeaways

Emily Jarvis: Yeah, that makes sense. All right, your takeaway from today. We've talked about a whole bunch of things, and the theme today is obviously reducing downtime and improving operations with AI. As folks log out here in a few minutes, what do you want them to walk away thinking about from today's session?

Maddie Wolf: Heart surgery, apparently. The thing you said earlier, I'm still thinking about that. What would I actually say? I would say the biggest thing, and it's not fun, is thinking about the metric, the thing you want to move the needle on, and actually doing the process mapping that maps up to it. That is the way you're going to get any AI anything approved. That is how you're going to get it done.

If you just go with a chatbot or something like that, it's cool, and people really like it, but it's going to be harder to make that connection. So start with something you can actually get leadership, or people, on board with that makes a difference. Then all of it will trickle from there.

The second thing I would say is: don't be scared to ask or talk to people. I'm on LinkedIn. Yes, I hate myself for it, but I'm on LinkedIn. If you want to shoot me a message and you have a question, please go ahead and do it. I'll absolutely respond.

I think maintenance has historically been overlooked by a lot of people. People think it's not that interesting, or for some reason they don't get into it. It is so ripe for being able to do stuff with AI. It gets me super stoked. There is so much cool stuff we can do. We just need to get the right people in the room: the people who actually have the expertise and the history, and the people who understand the technology. If you can get those people together, we can do some really cool stuff. That gets me stoked.

Emily Jarvis: I agree. Unfortunately, though, we are out of time today. I want to thank Maddie for sharing her insights with us. Just a few quick reminders: there are some additional resources available for you to download, so be sure to check those out before you log off today. We'll also be sending you a link to the on-demand version of this training tomorrow, so look for that in your inboxes.

On behalf of GovLoop and Legion Intelligence, I hope you enjoyed today's session. And as always, we really appreciate and value the work that you do. Thanks for taking the time out of your busy schedules to join us this afternoon. Have a great rest of your day, and we'll see you at our next training. Bye.

Maddie Wolf: Thanks, Emily. Thanks, all.

About Legion Intelligence

Legion Intelligence builds AI for national security. Legion’s software does the specific work a team is responsible for and writes the finished work back into the systems that team already uses. It’s up and running in weeks, and people stay in charge of every step that matters, with approvals and full auditability on every action the AI takes.

Customers include the U.S. Army, U.S. Air Force, U.S. Department of Energy, and Missile Defense Agency. Security credentials include IL2 through IL6 and CMMC 2.0 Level 2, and Legion runs across cloud, on-prem, classified, air-gapped, and edge environments, down to no connection at all.