Lee Dickson: Yeah, all that sounds great. And for everyone listening, thanks again for joining us on this episode. If you haven't already, hit that subscribe button to stay in touch. Also, check out our sub stack for more content, which will be linked to this episode to continue this discussion. Until next time, stay curious, stay informed, and we'll see you on the next episode. Take care, everyone. Rich: ⁓ Sorry, do it. ⁓ Lee Dickson: Feeling good? That was the, that was not the vote of confidence I wanted where, ready to rock and roll? Just like, that's what I need. That's the energy. I need, I need, I need you to, yeah, there it is. Exactly. They're not going to know what to do if we're not pumped about it. Rich: Which one are we doing first? Yeah, let's go. Let's go. Let's go. Come on. Come on. Come on. There we go. All right. All right. got a half a million people depending on us. we whatever we say, we're moving. We're moving industries. Lee Dickson: Welcome to the AI Guys podcast where we make artificial intelligence easy to understand for everyone. I'm your host Lee Dixon and as always I'm joined by the CAPTCHA answer king Rich Swire. How you doing today man? Yeah right? Rich: Hey, what's going on? Doing great. I think I calculated, I think I calculated that, ⁓ I, I, I, a conservative estimate is, ⁓ 20 hours a year I spend doing two F a typing, typing in, know, where you got to get a two F a and then you're, you, ⁓ yeah, 20 hours, 20 hours. Yeah. I calculated it. I kind of, that's what that's like two. That's like a whole day of my life every year. Lee Dickson: really yeah yeah going to the authenticator app or you know waiting for the text to come through good lord yeah Yeah, so I have, that's why test 123 is my password for everything, Just, yeah, that's fair. If you don't have to turn it on, but yeah, it is ridiculous how much time we spend on it, but that's actually something we're gonna talk about today for everyone listening. We are gonna touch on the security aspect around. Rich: That's how much I care about security. ⁓ No, it doesn't matter. The 2FA, you still gotta do the 2FA. Lee Dickson: launching and deploying agents, managing them internally and everything like that. So whether it's going down the path of securing the agent themselves through access and permissions, if you're trying to talk about even securing the model around data, training, knowledge integrity, or that full life cycle of what it matters with monitoring and feedback, I figured we can give some quick highlights on what people should be thinking about as this adoption curve continues to only increase. Rich: Yeah, mean, security is a big element that has a very vast and wide, you know, you can talk about it for days if not weeks. You know, we're going to focus primarily on enterprise security because that's where we live. But, you know, there's also obviously with personal security, there's other elements that come into play. But, you know, as far as the enterprise goes, when you're rolling out AI, I think the top three things that we hear when we talk to customers are quite honestly, in my opinion, probably the least concerning things, but they seem to be the ones that are always kind of on top of mind. ⁓ The first one is, ⁓ is your model training on my data, right? Or is this model training on my data, right? Lee Dickson: Mm-hmm. Rich: And I think this is one of those things that was like a boogeyman thing two, three years ago. And, and, and it, and it just like what people don't, if they don't understand AI or what a model is or how it works, they just kind of assume that that's the biggest risk. Quite honestly. ⁓ it is certainly worth noting. ⁓ but in most cases, I think all of the major players that offer models, if you're doing any sort of enterprise, ⁓ integration with those models, like what we do. Lee Dickson: Hmm. Rich: ⁓ they, they, they all basically say right off the bat, you know, we do not train on your data, right? Because they understand that enterprise data is very sensitive. The data that they do train on is obviously consumer engagement data. Right? So that's the, the person who's using chat, GBT for free or for a low, a low amount of money. ⁓ that's talking to the AI and basically doing human feedback, right? You're kind of doing machine learning there. unless you opt out of it, right, which they have that option to do, you know, by default, they're going to train on your data because that's how the AI gets smarter. And they're assuming that, ⁓ you know, this just like, you know, that you're participating in this no different than like, ⁓ how, you use social media, you know, they're selling your data to advertisers, you know, and, and, and that's how they're, that's the revenue model, right? It's like, And these models get better because of that. This is one of the reasons why open AI its leadership for the most part in what I would call the generalized LLM models is because they have so many users using it every day that it can consistently learn and can train on that information. Now training on training on data doesn't necessarily mean that if I put my social security number in there, that it's going to pop up, that it's going to be like, Hey, what's Rich Swier's social security number? ⁓ What it means is that Lee Dickson: That's gonna be stored somewhere, yeah. Rich: OpenAI is collecting all of your conversations and messages. It's putting it into a database, right? Because remember, a large language model is a read-only database when you're actually communicating with it. It's not like sucking in your data real time and like consuming it like this massive like black hole. ⁓ What happens is your data goes into another database, right? Just like if you're on Google and you're using your email, your emails are in a Google database. If you're on Instagram, your photos and your chat messages are in a database. And the question is, what that company is doing with that data. ⁓ sharing that data, selling that data or using that data to train their model. This is something that you should be aware of. Now on the enterprise side, for the all intents and purposes, ⁓ I think you're sick, completely safe. And that's no longer a major concern. It's a good question to ask and, and, and confirm, but like open AI enterprise, Google enterprise, I am throw topic. They all follow the same basic playbook, which is we're not going to train on your enterprise data. And they don't really need to, ⁓ because they have, ⁓ the training happening on the, on the, Lee Dickson: Yeah. Yeah, there's plenty of activity that's already taken shape, right, where they don't really need to be thinking of it in that larger plan. Rich: other side. Yeah, no- Yeah. And it also, when you're training on your data, there might be proprietary elements like your code or your documentation or your customer base or something like that, that you're, that you're sending back and forth to the model. But remember it's only the conversations that are being sent, right. ⁓ To the, to the model itself. And then what they do with those conversations, they probably create derivative data from that. So they don't take it wholesale. They probably do a good job of cleaning out PII. They clean out all that other stuff because they don't want, you know, they are, they're as much at risk as you are. at exposing personal information. In fact, OpenAI has done a great job. If you're even in chat GPT and you say, hey, I want you to save my social security number, it won't do it because it already has blockers in place for that because they want to maintain their SOC 2 compliance and so forth. So I think the models have done a great job of knocking that one down, but it's still a good question to ask. The second thing that I hear a lot about is, you know, ⁓ Lee Dickson: Right. Rich: And it's kind of a misnomer is, you know, ⁓ don't want, ⁓ these, the AI to be able to access my bank account or access my information. don't use AI because, you know, I've got my tinfoil hat and I feel like at the minute I start doing something in ChatGPT that AI is going to figure out how to hack into my bank account. And the reality is, that AI is nothing more than just a read only database, a very powerful database. Lee Dickson: Mm-hmm. Rich: On top, there's a layer of software on top of that, which is AKA chat GPT or Microsoft co-pilot or Raia or Anthropic Claude. So there is a layer of software, just like everything else that you engage with on the internet. ⁓ but the LLM itself is not just like dangerous component. It's really there just to service and to be prompted. You ask it a question. gives you an answer. Now, some of these products like chat, GPT and others allow you to connect to third party services, right? You can connect to your. Gmail and you can connect to your Google Drive and that's probably an area where there I would say there is some concern but it's not like just because it can connect to a third-party service through an API that suddenly it's going to drain your bank account because AI is evil ⁓ But that being said when you do connect to these third-party services Whether it be your Gmail or whether it be your Google Drive you are certainly expanding the blast Lee Dickson: Yeah. Rich: radius of what that AI has access to. Now it could be limited to what it could do there. It can't necessarily delete it necessarily unless you give it access at that level. Lee Dickson: The exposure. Yeah. Well, and one thing on that... Yeah. to read and write, yeah. Rich: Right, right. But yeah, you are basically saying you have access to these other tools. So obviously there should be some caution there, but to go to the degree of, ⁓ of, ⁓ of thinking that AI is somehow smart enough to find its way or watch what you're doing on your, on your, ⁓ you know, with your, with your bank and so forth and scanning that as a hacker's tool is a little bit outside of the realm. ⁓ of course, unless you're talking about specifically applications at which we'll talk about in future episodes. Lee Dickson: Yeah. Rich: where you're installing something like, for example, ⁓ OpenClaw or something that has command over your desktop, well then I would say yes, then there are serious security issues that you have to read a red flag for. But if you're working just through the browser, just through the hosted version of a chat jpt or any other application, just be aware that when you're connecting to a service that your AI is accessing that to get more context about what you're prompting it. Lee Dickson: Mm-hmm. Yeah. Rich: But that doesn't necessarily mean that it's going to have like this full control over everything. Lee Dickson: the control. Well, it's interesting too, because I feel like AI literacy today has never been more important. Right. So like knowing what you're actually trying to accomplish and knowing what tools or what services you're using to accomplish that is, probably the area that most just general users of AI are probably the weakest because you know, they're like you said, there's that fear that might exist there or it's the unknown or maybe it's like the cutting the corners aspect. Like I'm to use, you know, this ⁓ cool.ai tool that I found and you know, you have no understanding about who built it, what it does, like where it's at. that's kind of the some of the you risks you've seen with the the local hosted stuff is like ⁓ yeah I spun this thing up that does all this cool stuff and then they go to bed and they wake up and you know their entire life is ruined because it's gone through because they didn't understand the permissions they didn't understand what they actually gave it access to so it just went through its system process of everything it's supposed to do and made decisions based on that you know either guide rail or lack thereof and if you don't know what you're actually doing with it please for the love of god just pump the brakes and actually understand what you're about Rich: Yeah. Lee Dickson: put in play because I think that's what matters now more than ever. Rich: Yeah. Yeah. And that brings me to the third part, which is these other tools that you mentioned, right? There's a tremendous amount of open source products out there. There's a tremendous amount of like dot AI products out there. Right. So you, you, could like my wife the other day or before she, when she started getting to chat, GPT, I said, ⁓ she should download chat to you on your phone. And when you go and you go on a, on an iPhone, you search chat, GPT, there's like 30 apps. Lee Dickson: Yeah. Mm-hmm. Yeah. Rich: that all look exactly like open AI use the same logo and she downloaded a ⁓ third party knockoff, right? And she's like, I didn't realize you that you had to pay this much money. And I was like, wait, what, and I was like, what is that? Why does it look so different on an iPhone? I was like, no, you downloaded, you know, ⁓ basically a wrapper and it had ads, you know? So yes, I mean, I think it's extraordinarily important that there is a large amount of, of applications coming in, ⁓ and being built that look Lee Dickson: Yep. A knockoff basically. Yeah. Mm-hmm. Yeah. Mm-hmm. Rich: feel exactly like chat you'll see you you're on Instagram you're scrolling like ⁓ look at this like you can talk to all the models right and there's all these different wrapper ⁓ type technologies I would definitely put put a big caution on that as your point clearly said which is because most of those companies are capturing your data and just because OpenAI, which is what they're probably using as a, or a cloud or Google models, even though they're using those models and the experience is very similar, right? You're still talking to it as if it was like Chachipiti. That third party is grabbing that data and putting it into its own database. And they don't necessarily have to... follow the rules, if you will, they can do whatever they want with that data. They could also insert ads and sell that data and so forth. yeah, I I think in typically what you would want to stray away from is any sort of technology. If you're using another app, ⁓ make sure that you do your due diligence and make sure that app is what it is, what it says it is, and probably dig a little bit deeper into... what ultimately what, what you need that app for. Right. I mean, for the most part, like the things you definitely want to stay away from is like the clones, right. where it's like, they're pretending to be chat GBT and you accidentally download it. And now you're talking to a third party thinking that it's open AI thinking that they're following the rules and they're not right. So those are like the three things that I think got a baseline is kind of been like the core. I, what we hear like right off the bat, right. Lee Dickson: Mm-hmm. Yeah, for sure. Yeah. Rich: And then when you dig deeper though, and you really get into the kind of more scary component of security, you're talking about more on the enterprise side. So one of the things that I think first and foremost probably should be like really clear when you're thinking about AI inside of your business, right? Is that it's a... ⁓ It's a huge multiplier. It's also a massive accelerator, which means that more people use it, the more value they get quicker. And so they're going to want to use more. And so you want to make sure that wherever you start off introducing AI into your business. Lee Dickson: Mm-hmm. Mm-hmm. Rich: You want to make sure that you're prepared and isolating, you know, what tools that they potentially are going to want to use at scale and give them instruction and give them guidance on that on day one. Instead of saying, do not use AI, what's going to happen is when you say do not use AI, they're going to still use AI. They're going to still work around the system, but now you're going have Lee Dickson: Yeah. Yeah. They're just not gonna tell you about it. Yeah. Rich: And they're going to have, and then you have a hundred people using a hundred different tools of which a handful of those hundred tools are definitely not ones you want them to use. instead of kind of taking this like a parenting approach of do not go outside, right? Because there's dangerous people outside, maybe hyper-focus that messaging on, I know you're going to go outside, but don't talk to anybody who don't know, or, you know, like be more specific. Yeah. You can talk to Mrs. Jones down the street, but, but like my Lee Dickson: Mm-hmm. Mm-hmm. Yeah. Yeah. or be back before the streetlights are on. Not the guy in the van. yeah, okay. Rich: Right, right. No, no white vans kind of thing. So like it's better to have a strategy around, ⁓ Hey, here are the tools that we've authorized that we feel that are good. And for what purpose and what use case you want to use them for. And it might be a limited use case. Like if you're going to build a video or you're building a presentation, or if you need to have a generalized LLM chat agent. Lee Dickson: Mm-hmm. Rich: You know, here are the tools that we've that we, you know, feel comfortable with, right? Because if you don't, what's going to happen is there's going to be sprawl, right? There's going to be people using everything and you won't be able to pull it back. And the last thing you want. Yep. Lee Dickson: Approved. Yeah. Yeah, and it's back, but it's back to that AI literacy piece, right? Like now you're thinking of less on the consumer side or like maybe a small business side and at the enterprise level, it's like somebody inside doesn't necessarily have to know everything about AI, but has to have some type of guideline that's just stating like, look, you know, here's some tools we've already done research on that we believe are good. If you're utilizing a platform, something like a Raya, you know, having your security team, your legal team, whoever like go through the process with whoever's about to use it. So that way there's no there's no questions. Everyone's on the same page about what's gonna actually take place. Rich: Yeah. Yeah. And so, and then once you've kind of outlined your policy of what tools and what what to use, you also want to make sure that if people need ⁓ to use a different tool, you know, they have a way to register it or way to say, hey, I, you know, I need to use Suno to generate a song or something, which is like an outlier example. But, you you say, all let me go check it out, you know, go through the process of doing vendor, you know, and then add it to the list. Right. And and then so that's the easy part. honestly, I would say most companies that we've engaged with have some form of that, but it's good to talk about like what the baseline of what you should start thinking about because you want people to adopt, you want people to use, but you also want to make sure that you you you keep them to the tool set that you trust. And so the natural conclusion of that, ⁓ you know, moving away from tools Lee Dickson: Mm-hmm, absolutely. Rich: You start to get into this what happens in the natural evolution of an enterprise is they go from tools To current and crossing this chasm into this agentic world right like okay now. I want to use agents I want to deploy more agentic solutions I Want to integrate agents into workflows. I want to tie agents and give them access to our third party applications our SAS apps So that's when you start to get into a completely different type of paradigm when it comes to security, because ultimately ⁓ you have to take an approach of, okay, what am I trying to accomplish with this agent? What does it need? What agency do I need to give it? What access do I need to give it to it to, to, to, actually do what I needed to do functionally? Cause if I could get it to do X, Y, and Z, that's a huge time saver. That's a huge productivity gain for me. Lee Dickson: Exactly. Rich: And then limit it, it's, it's capabilities based on that. Right. So it's a functional scope of instead of giving AI access to everything. You say, I need this AI to be able to insert a note into my CRM. Right. And so you have an API access, you give it that limited endpoint and you give it that access and you have a zero trust approach. The concern I have right now is that most companies. who are going from tool-based implementations to kind of graduating to what I would call enterprise-wide rollout of those tools and or agentic. Lee Dickson: Mm-hmm. Rich: approach is that it all starts at the workspace. whether it's Google workspace or Microsoft 365, what these providers have done is they said, Hey guys, you know how we give you email and spreadsheets and we give you all this other cool stuff. Well, we just put AI right in the middle of that. And oh, by the way, know, pretty much anybody can start using AI any way they want within your ecosystem. Lee Dickson: Yeah. Yeah. Yeah. Mm-hmm. Yep. Rich: And that's, ⁓ even though it's somewhat locked into that ecosystem, right? The idea is behind it's behind the firewall. ⁓ that can be extraordinarily overwhelming and dangerous. ⁓ if you don't necessarily, you know, understand what it means to start building out agents that can be shareable. So it's one thing if it's a tool, right? Because if it's a tool, it's a one-to-one relationship, just like your Gmail, right? Like I have Gmail, you have Gmail. Lee Dickson: Yeah. Rich: You log on to your Gmail. I don't see that. I log into my Gmail. So it's the same application. We're isolated user to user. There's privacy, blah, blah, blah. Now you build an agent on that same ecosystem and unknowingly, or maybe you do know, ⁓ train it or give it information and then share that with me. ⁓ now you're starting to talk. Okay. Well that feels okay. Even though we work for the same company, there's no major problem here. Lee Dickson: Yeah, there's exposure risk now. Rich: but it's too easy to share stuff that I may not necessarily have access to from the raw data perspective. ⁓ now it's not, this is not malicious, but yeah, this is not. Yeah. So ultimately when it comes down to is like AI tends to blur the lines quite a bit, ⁓ at the agentic level. Lee Dickson: Yeah. Or maliciously, or not even maliciously thinking that it's gonna be a problem. No, exactly, but. Rich: So unless you isolate your agents and treat them as like their own employee and give them the access they require and only train them on information that they should be trained on. And you have a process in place and a platform like rare to manage that and say, this is the agent we're building. It's isolated. has its own workflow. has its own API access. has its own vector store that we've opted in and trained on that data. And it's gone through that. process, then anybody who, and also by the way, it has its own access control. So maybe Lee has access to this, but rich does not, unless it's pulled out of that kind of, you know, everybody accessing the same 365 or same Google workspace. If it's isolated out as its own kind of tenant, if you will. It's a lot easier to manage and maintain and make sure that you don't do that accidental cross pollination. So, you know, that zero trust framework, that approach that we take is very important because I do think that there's probably areas where somebody goes into their SharePoint. drags and drops 30 documents into an agent, right? Thinking like, oh, maybe it's an HR agent and maybe they work for HR. So they have access to all these cool documents. They accidentally upload a document that's maybe not supposed to be there. Nobody knows that they uploaded it, they? Cause they're just building out this agent. And then they go, you know, I'm going to share this with a few employees to see if they like it. And lo and behold, one of those documents might've been sensitive information. And then now they're like, they're, and the reason why Lee Dickson: Yep. That's not a whoopsie anymore at that point. You know what mean? Like, right. That's a human problem. Yeah. Rich: Yeah. And that's not an LLM problem. Now realize people don't understand some, but it's also like, it doesn't even touch the LLM. So a lot of people are thinking this is, this is the disconnect. People think that the LLM is everything and it's not. So when they hear, the LLM doesn't train on this data. Okay. So if I just drag and drop this data into an agent and share that agent, then it won't train on it. People won't be able to act. That's not true. What happens is those documents don't actually go into the LLM. They go into the context window. Lee Dickson: Yeah. Exactly, it's reference points. Rich: Or a vector, yeah. Or if you're in the 365 environment or into a vector store, if you're in a more advanced environment, when it's in a vector store and or context window, it kind of is invisible to some degree. Like when you're talking to AI and you're giving it, going through a thread, everything you're telling it and its instructions are kind of all go into its memory. Think of it like a kind of a temporary memory hold. And, and that data is then accessible by whoever talks to that same thread and or agent. So you have to be cautious of. Lee Dickson: Mm-hmm. Rich: of what goes into. ⁓ the context window and or vector store and unless it's isolated out and controlled in like more of a you know a process of like deploying an agent for specific purpose and isolate and isolate its access when it's users there's probably a higher likelihood of failure point and this is a concern I think this is probably one of the kind of concerns ⁓ or hidden issues that people may not even know exists and it's hard to trace back to even if it's happening because we've given probably a little bit too much capability to all users all at once ⁓ without really thinking through what happens when there's when there's documents being cross pollinated across the enterprise, right? Lee Dickson: Sure. repercussions of what that is. Yeah. No, absolutely, man. And yeah, I think, you know, we'll continue to do more over the next few episodes around. different solutions as you kind of referenced here, things that are hosted on, like what kind of permissions you're giving, because I do think as tools become more accessible, solutions become more complicated, users' needs become more broad and varied, like the understanding of the security nuances and what permissions should be given is gonna be a big piece of the success that people see with AI. Rich: Yeah. Yeah. And to give you some height, like to kind of boil it down, if you're like a CISO, if you're like a C a security analyst, the, the, the big things I think, ⁓ at least our approach has been and what we would, ⁓ Lee Dickson: Yeah. Rich: definitely recommend would be, you know, number one, if you are to, there is a big difference between tools and agents. And I think the Googles of the world, the Clauds of the world and the Microsofts of the world are blurring those lines too, that, that, that, that are, that are, that are too much. ⁓ I think tools are awesome and can be deployed within your enterprise because each person has access to their tool and it's isolated from one user to another. And security and compliance and access control can be easily implemented just like it's implemented on your mail, on your SharePoint, on your drive and so forth. An agent... sure should live outside of the firewall. Like it should be thought of as a third party actor. It should not be part of that same ecosystem that tools are. Like you should not let employees be building agents that can be shared amongst others and are also beyond that unless it's an environment that's whole that's held to the same threshold ⁓ of of expectation that a third party SaaS product would be. ⁓ Lee Dickson: Yep, 100%. Rich: And that was, and so if it's accessing internal data, it should be through an API, which guarantees that you can control its access. You can control what it has access to, and you can turn off that access at any point in time, just like you would turn off any sort of other API connection that you might get from a third party software product. If it's being trained on company data, it should go through a process and if curation and collection before it's put into the agent ⁓ through a through a relatively ⁓ straightforward process, not dynamically being done in the shadows where nobody knows what's being dragged and dropped into agents because it's been deployed at scale without any sort of oversight. And every agent should also have the ability for anybody within the organization, especially people at the security compliance level, to see all the logs, to see all the conversations, to have complete oversight over everything that agent is doing and easily download conversations and run it through an audit or, or have certain elements like escalation phrases or an audit agent or an auditor component to the agent. So you can escalate when PII or certain violations of certain policies are created. Like that needs to be part of it. And in the tool world, that's not part of what the tool world, because it's a one-to-one relationship. There's really no reason for those things to exist. And I think the problem is, that When you're trying to kind of make those tools a little bit more like agents, you start to kind of bleed into this new world where it's like, okay, come on guys. Like it's starting to feel a little bit dangerous. And I know that they want to keep up with the Joneses they want to have the coolest, you know, product on the, on the block, but having agents inside of your ecosystem being deployed by employees without any sort of platform that has sophisticated ways of managing and monitoring that. And also being able to have that connect. Lee Dickson: Yeah. Yeah. Yeah. Rich: to their own personal Gmail's or personal 365 Outlook or whatever is extraordinarily dangerous and quite honestly probably not worth, the juice isn't worth the squeeze there. Connecting your agent to your personal Gmail or your personal Outlook email doesn't really, I think, give you. the benefit that you think it does versus what the potential outcome could be ⁓ from a security violation. I think following those kind of like, those are like just general guidelines without getting into the specifics. ⁓ I think is kind of where we typically draw the line and try to follow as far as our protocols. Yeah. Lee Dickson: 100%. Yeah. Yeah, as more business users get empowered to be creating and thinking of things to do, they should have a lot of this in mind. And even if you're not, even if you're a developer that is just seeing the potential of what AI is doing and you, again, not cutting corners, but you're using it to accelerate, that doesn't mean you should steer away from just the general security that you probably know should be in place, right? So yeah, 100%. Yeah. Rich: Mm Yeah. Yeah. And the bottom line, and I'll just make this last statement, which is always true in most cases, is that AI is not the security. AI in general is not a massive, ⁓ you know, new thing that's going to be the problem, right? It's always humans, you know, like it's always humans who find a way. Lee Dickson: Yeah. Yep. Rich: to ⁓ accidentally or not ⁓ make it less secure. So AI is no different than any other software product that you implement. The only difference is that when humans start playing with something new, like a new toy, they want to throw a lot more at it. And if you don't have those controls in place, there's a higher likelihood of something, somebody missing a step. I think, I think, so don't go into this thinking no AI because it's not secure. It's like, you know, it's more about saying, ⁓ Lee Dickson: Yeah. Yep. Something breaking. Yep. Rich: All right, AI is this new technology. It's awesome. People are going to start to use it a lot. What is the outcome we want these users to have and what potentially could they do to make it less secure? And let's train, as you mentioned, let's train these humans on how to better use AI. Lee Dickson: Absolutely. Rich: And that's it. I mean, I think everybody should be embracing, and implementing AI solutions. But definitely make that ⁓ distinguish between what's a tool and what's an agent and act appropriately.