Ep.236-Shadow AI: The Invisible Agents Already Inside Your Customers’ Business

Josh Lupresto (00:00)
So right now, inside your customer’s business, there is almost certainly some AI tools, some AI agents doing some real work. Look at i i we know it’s touching real data, we know it’s making real decisions, and of those, probably nobody in IT approved, nobody inventoried, nobody’s watching. That’s not a maybe though. looking at this latest study shows that two-thirds of office workers admit to using AI tools that they believed weren’t allowed.

They admitted to using the ones that weren’t allowed. Who replies to that on a study? nearly half of all enterprise AI usage is is happening through personal accounts that completely bypass a company’s controls. So that’s gone from what was a side issue to now the main event. this year what we’re gonna call it the unapproved shadow AI shows up in forty-three percent.

Of data breaches, right? Say it again, 43%. That’s more than double the year before. And those breaches cost five million dollars each on average. Okay. Here’s the thing your customer cannot see this. It’s a little invisible to them by definition, right? And I think the person that walks in and helps them turn the lights on, that’s you.

That is one of the most winnable conversations you can have this year. You’re going to be able to carry this one into next year as well. So let’s get into it.

Welcome back, everybody. Josh Lupresto SVP of Sales Engineering here at Telarus And today we are talking about the shadow AI problem. it’s the invisible agents that are already loose in your customers’ building. So this is a security episode, but don’t tune out if security is not your thing. Because this one is really about yes, security, but it’s about trust, visibility.

and it’s a door, I think, if you look at it that way, it’s a door into your customer’s environment. And you know, that door’s kind of wide open right now. So let me define this term for anybody that’s not familiar here with what shadow AI is. It’s about to be everywhere. More AI, more shadow AI. So you you guys remember shadow IT, right? That was the old problem. some employee signs up for a cloud application on a credit card, and now there’s a company, you know.

data set that’s living out in this tool that nobody approved, right? We’ve been managing that for, gosh, it’s crazy, 20 years, roughly. So Shadow AI is kind of the same idea, but with a little bit of turbocharge, right? it’s employees using AI tools that the company never sanctioned. So somebody’s pasting some, you know, something you know, simple, just maybe some, some customer records into a free chat bot and summarizing them. somebody’s running a sales list through their

Personal AI account. maybe somebody even connected in an AI assistant to all of the company’s email and calendar because it’s just easy. And they did it without asking anybody. So here’s the number that should probably stop you a little bit. So two out of three office professionals say that they’ve used AI tools they didn’t think were permitted. That’s so crazy. And nearly half of all AI use in enterprise.

Is going through personal accounts. That’s nuts. Nearly half of all AI use in the enterprise is going through personal accounts. Not the stuff that the company’s got managed, monitored, secured. these are personal ones. Think about all the time that we’ve spent in the last couple of years going through securing the environment, MDR, data loss prevention, data backup, resiliency, training the end user, people are the weakest link, social engineering, all those stuff all those things.

And we’re back. All right. Same, same issue, just looks a little bit different. So if you think about what that, you know, all of that personal account stuff, right? That means the data is going in and out, answers are going, coming out, and and that’s completely invisible to the people whose job it is to protect that. So picture for a second your customer’s business. think about it maybe think about it as a building, right? Shadow IT was somebody leaving a side door unlocked.

Okay. Shadow AI, a little bit different. Now it’s dozens of employees quietly cutting their own doors into the walls, and every one of those doors has company data walking through it. It’s kind of a weird weird analogy, but I think it’s a good comparison. that’s just the starting point, right? But it gets a little more serious, I think, because we’re not just talking about chat bots. That’s kind of what we knew AI to be in the beginning, but that’s not what it is exactly.

Anymore. There’s still some of that. So let’s go, let’s get into part two here. Why agents make this a different animal? Okay, so here’s what changed the game last year. And I and kind of why I wanted to do a whole episode on this. We went from AI that answers questions to AI that takes action. Okay. Agents. A chatbot is just this passive thing. You ask it something, it replies. And if somebody pasted insensitive data, that’s bad, but it’s contained.

Now, an agent is active, right? It doesn’t just answer, it does it it it does whatever it needs to do. It does some different things. So maybe it logs into a system, it reads and sends email, it pulls records from a database, it kicks off workflows, it talks to other software, often with very, very little human watching. Okay. Now, combine that with the shadow problem.

And then you see why this is kind of a different animal, right? And one of these unapproved agents is not just a little, you know, it’s not just a leaky chat bot. It is an autonomous worker that nobody hired. essentially, if you think about the roll-up here, it is inheriting that employee’s permissions. So whether that’s one agent, that’s a thousand agents doing different things, the same things, whatever, they’re inheriting the permissions of that that employee. So if that employee is trusted.

want to guess it? The agents, agent agents are trusted, acting, you know, all across the company systems, no oversight and very minimal logging. So a hard new problem that that that people need help tackling. Now, I think the scale on this is a little staggering. So the the analysis on this, they they the they kind of are saying that you know by the end of the year, forty percent

Of enterprise applications will have AI agents built into them. Okay. That’s just what I think it was under maybe five percent last year. And and so I think if you think about it like that, 80% of the biggest companies already have active agents running. No surprise. But I think here’s the gap that matters. only about six percent of organizations have what you would call you know.

mature security strategy for these agents, right? Six percent. So some of these studies even even go a little bit further to say half of those deployed agents are running with no security or oversight or logging at all. So let that land for a second. Agents are flooding into every business. It’s faster than almost any other technology in history, right? We’ve seen the run-ups. You’ve seen

How quickly these guys are getting to a trillion dollar, two trillion dollar valuations. Okay. And then the governance, the oversight, the basic who is watching this, that is that’s it it’s years behind, right? That gap between how fast agents are showing up and how slow the controls are, it’s it’s kind of, you know, it’s catching up. that gap, I think, is the entire opportunity, right?

And and this is why it costs so much when it goes wrong. Okay. When a shadow agent is involved in a breach, the damage is is significantly worse. And it takes longer to even find. Okay. But why is that? that you’ve got so many different agents out there, and that agent is acting with real permissions across a ton of different systems, and nobody is logging it. So so what it did,

right?

You can’t see the blast radius. That’s the problem. That’s the whole shadow factor here. So that’s why these shadow AI breaches have this inordinate five million dollar number, is because they’re the you you just the blast radius is so big and you can’t see it. That’s crazy. I and I, you know, if you think about what we talked about before all the time in security, is it’s it’s meaningfully more than a normal breach. And

You can’t respond to an incident that you can’t see happening, can you? So I think the headline for your customer is simple, and it’s a little scary, right? here’s the headline. So you’ve you’ve probably got some autonomous software acting inside your business right now with real access that nobody is watching. The good news is that is fixable. And fixing it is exactly the kind of project.

That makes you, you, the advisor, indispensable. Okay. let’s let’s get into let’s get into this next part here. Why I think this is the advisors, why I think this is your your conversation to win out. So let’s talk about why this lands so perfectly for you, because I don’t want to make this, I I’m not trying to make a doom episode here. I am not a doomer. I think it’s opportunistic, right? This is

This is scary if you’re the customer and they have absolutely no idea what to do about it, right? And here you are coming into this knight in shining armor. That’s kind of the picture that I want to paint here. Or you could look at it as it’s a, it’s a it’s an opportunity episode wearing a security jacket, right? One of those cool black you know, NVIDIA CEO, one of the Jensen jackets. I and I think like this whole thing is is it’s about.

It’s really about what makes a great advisor, right? You want a problem that is urgent, it’s real, it’s expensive, it’s invisible to the customer, and it’s just something they can’t easily solve alone. Think about that though for a second. Think about all of the times that you’ve gone into the customer’s environment. You’ve had this great, great scenario, great product, great whatever. And if they didn’t have urgency, and if they didn’t have you know, some

pain, then it was really hard to move that along. Okay. So this is, you know, this is all those things wrapped up a little bit in in a bow. Okay. So I I think shadow A is it shadow AI is all those things, right? That we mentioned about it’s expensive, it’s invisible, et cetera. It’s real and the data proves it, right? It’s urgent and I think it’s getting worse every quarter.

we’ve seen that, you’ve seen these costs go up. It’s it’s millions of dollars that we’re talking about per incident, and it’s invincible. It’s literally the the definition. And they can’t fix it. well, they can’t fix it alone, most often, because you can’t govern what you can’t see. We’ve said that, right? So this is about, I think, as perfect of a setup as you will ever get. You’re not walking in trying to create a need. The need’s already there, okay, and and and it’s already growing, and the customer doesn’t even know

the size of it.

Okay. You’re just the one who’s turning on the lights. Okay.

Now, here’s the here’s the part that I love. the first step isn’t even a big and scary sell. It’s a conversation and maybe some visibility. Okay. You go in, you help them answer a question they genuinely cannot answer today, or they’re or it’s really hard for them to answer. What AI is actually running in our environment? Who’s using it? And what can it touch? I’ll say it again.

What AI is actually running in our environment, who’s using it, and what can it touch? That assessment, j just you know, getting visibility, I think, delivers enormous value before you’ve sold a single, you know, big gigantic solution. You’ve shown them a risk that they couldn’t see. That’s I I think that’s trust instantly. And then I think it opens straight into the real work, right? And this is where our

The Telarus solution map does a lot of that heavy lifting because once you can see the shadow AI, then the customer needs to actually govern it. And I think that’s a whole security engagement. So if you think about on the the the Toleris solution map, that lands right in the cybersecurity pillar. Okay. So what do we have there? we’ve got governance, we’ve got risk, we’ve got compliance.

All of these different suppliers that can come in and build a AI acceptable use policy. Think about zero trust, right? Think about SASE, all the evolutions of that, right? All the the our our expert here that knows SASE better than anybody, Josh Hazelhorst, right? We’ve got this zero trust to make sure an agent can only touch what it’s actually supposed to. You’ve got identity. you know, think Microsoft, intra ID, you’ve got access management because

I think a huge part of this is treating of this agent and these these agents that identify with permissions that you control. That’s the key, right? We’ve got to get a lot of limited granular access. also, then, from a security perspective, what do we have? We’ve got virtual chief you know, CISO, chief information security officer, services for the customer who needs a strategy but doesn’t have the security leader, or maybe doesn’t want to pony up.

half a mil a year for a CISO. Maybe it doesn’t make sense. Outsource that. Get that get that service on an as needed basis, somewhat somewhat regular.

And then of course, you you’ve got to monitor it and you’ve got to detect it. So we’ve got that monitoring and detection, that response layer to actually watch what these agents are doing, whether you you’re thinking about CrowdStrike or Carbon Black or Defender, there’s a lot of tools, lots and lots. I’m not gonna go into all of those, but there’s a lot of tools there. So watch the sequence though. You open with visibility, low friction, high trust, and that reveals.

Some of the exposure. Okay. And the I I I think the fix is just it’s a it’s a full security practice that runs right across our map. Okay. Governance, zero trust, identity, monitoring, virtual CISO. one invisible problem, and it opens the door to the stickiest, highest value category that’s out there. Okay. That’s not a threat to anything that you’ve already placed with that customer. It’s

It’s a brand new reason to go deeper. And I think it wraps a protective layer around everything, right? Every everything else that they’ve already got running. All right. let’s go part four. questions. You know, I can’t let you leave here. you and I always talk about questions, probing questions. Okay. So let’s make let’s make this usable for you Monday morning, right? Or whenever you’re listening to this.

The beauty of this topic, I think, is that the discovery questions kind of sell themselves, unlike the vacuums I used to sell door to door for two grand. They did not. because the customer cannot answer them. And that’s the point, right? Ask hard questions, wait for the pause, don’t fill the air, let them answer them. Okay. So write these down, put them in your phone, however you want to use this. Okay. Start with the inventory gap. Ask this one. Question number one.

Do you have a list of every AI tool and every AI agent being used across your business right now? And are you confident that that list is complete? Ooh, pause. Wait for that one. That’s a tough one. Okay. Nobody can say yes to this. That’s the beauty about it. the moment they hesitate, you’ve demonstrated the problem without saying a word. I think that hesitation is is kind of the whole opening. Okay. This is a new one. We’ve not

We’ve made assumptions around this, but we’ve never really talked about it on this podcast before. It’s this whole thing we’ve talked about here is the the the the personal account issue. Okay. So ask how much of your AI that your your team is you know using is running through company managed accounts versus whatever people signed up for on their own. That is a scary question. the surface I I I think this will surface up the shadow usage.

Directly, right? And and it’s almost always way more than leadership thinks. Whatever they think it is, I could guarantee you it is more. And then let’s go, let’s go to the permissions question. Okay. This is the the you know the big one for agents. Okay. Ask for any AI that is connected to your systems, do you know what it is allowed to access and who approved it, right?

Getting the access and allowing the access, I think, is one thing, but even sometimes their own people will blindly approve it and not even not even think about it, right? Not even read the fine print on the the contract with, you know, this agent software that was just a one pager that says, Hey, we give all your, you know, data back to, you know, the the mothership in China or whatever. it’s pretty crazy some of these. If you actually read I’ve gone through and read

I I think this is where the autonomous agent risk lives, right? Blank stairs here. I think that’s a good it’s usually what you’re gonna get. And I think that means real exposure. And of course, it’s it’s it’s a real project, okay? Then the policy question asks if an employee put customer data into a public AI tool tomorrow, is there a policy that tells them not to? Have they ever even done that? This is this is a document and an operational rollout, and anything.

Anything that would catch it, right? If they did that. Okay, if they dumped some of that in there, would anything catch it? Two-thirds, like we talked about earlier, two-thirds of the companies have absolutely nothing there. And that’s your governance, that’s your acceptable use conversation teed right up.

And then I think land it, right? Just land land the plane, as we say. land it on the stakes, right? I ask this one if you i if an AI tool that you didn’t know about caused a data leak, how would you find out? How would you even find out? And how would you land that, right? With your customers and your regulators. That kind of reframes it, right? Takes it takes it a little bit, a little bit of a level up and just you

know.

It really does. It puts it all in different perspective. The risk though, I mean, w when you when you think about things like this from a security perspective, that risk is not theoretical. It is a breach they could not see coming, and it is a breach that they could not explain afterward. And that is not the position that they want to be in. And like the last like the last few episodes here, notice notice this arc. Okay. You’re not selling fear.

And you’re not selling a product. We’re not here to talk about Doom and the end of the whatever, right? You’re helping them just see something that’s invisible and it’s genuinely putting them at a risk, right? Their people are going to use these tools. We are going to use AI for better productivity, more efficiencies, 10xing our people, all of those things. It’s just going to happen. It is a freight train coming at them and they need help. So

When they say, okay, okay, okay. How bad is it? And what do we do? That’s the moment, right? you you you don’t go it alone. So final part of this here. you bring the trust. We are gonna continue to bring those resources, right? Here’s the honest handoff. Of course, you’re not expected to personally run an AI security assessment, build a zero trust framework, roll out whatever, or write this, you know, write all this different governance and compliance things, right? you don’t need to.

we’ll teach you if you want to know it. We absolutely will. but your job is to spot that opening, ask the questions, and own that relationship that you do so well, right? The deep work, we’ll we’ll go in on that. Toleris, field engineers, solution architects. if you haven’t met Jarrett yet, Jarrett is amazing. Jarrett Bowl on our solution architect team. incredible deep depth deep depth in addition.

To the broader field SE team. Okay. So that’s again, that’s governance, policy, identity, access, monitoring, security architecture. That is exactly what our sales engineering team here at Toleris is for. We’ve helped you through these things. We are going to continue to help one. This one is a big, juicy one, I think. and there’s just so many different ways that this one will go. So when the customer says, help us get.

get our arms around this, that is your cue to bring us in. We’ll go as deep as it takes across the pillars, you know, in the Toleris solution map, and we’ll help you turn what maybe started out this conversation as an invisible risk into a managed, governed, monitored environment, right? there’s a lot of things to do in between that. So look, you are the trusted advisor who saw what nobody else did.

We’re just here to help you be that horsepower to lock it down. So together, we’re not just flagging this shadow AI problem. We’re solving it. And you again become the person that the customer calls before they adopt anything new. Because if you’re helping them through these really hard things, in my mind, you’re on the receiving end of all of the easy stuff, all the other things. and that is a relationship that no vendor and no free chat bot.

I think can touch. So let me, let me, let me, let me bring it home here just a little bit further. I think every business your customers run now has a you know, a layer of AI in it that’s growing faster than anybody can govern. Okay. It’s just it is what it is. And most of it’s helpful. Okay. Some of it is invisible, and a real slice of it is quietly creating the most expensive kind of risk there is, the kind you can’t see until it’s a headline.

Right? Nobody wants to be target and the whole HVAC and you know, pick a breach, right? It’s not a reason for anybody to panic. It’s a reason to just get a trusted guide in the room before it becomes a problem. That’s all. And I think that is the whole opportunity. You’re not the person spreading fear about AI. you’re the person who just walks in calm and says, Let’s just see what’s actually running and let’s get it under control. Right? In a world where everybody is.

racing to adopt AI, you’re the one making sure that your customers can do it safely. There’s no more valuable seat, I think, than that. So as we as we look ahead a little bit, right, at this kind of fast moving thing here as we crazily approach the end of twenty twenty six, let’s watch the governance frameworks, let’s watch the regulations firm up potentially, because, you know, we didn’t know it was running.

you know is gonna stop being an excuse for these end customers. So watch agents showing up inside of software that your customers already own, whether they asked for it or not. And I think watch for this you know, honestly, to to to become a standard part of every security review. You’ve got to be the one who got there first. So this week, my challenge to you, go pick one customer and ask them the first question we talked about, right? Just as a reminder.

Question is, do you have a complete list of every AI tool and every agent running in your business? And just sit in the silence. That silence is going to be your next project. So that’s the show for today. do me a favor, wherever you came to, Apple, Spotify, go subscribe, go follow the show. It does us tons and tons of miles. I’m your host, Josh Lupresto

S V P of Sales Engineering at Telarus Till next time.