So What? What’s New in Agentic AI Assistants?

Well, hey everyone, happy Thursday. Welcome to So What, the Marketing Analytics and Insights Live Show. I'm Katie, joined by Chris and John. Howdy, fellas. We got the high five.

There we go. All right. This week we are talking about what's new in agentic AI assistance. Um, if you haven't heard, major companies have been releasing their own agentic AI assistance. These in a nutshell are oh gosh, I'm probably gonna mangle up the vocabulary.

So Chris will keep me honest. But they are basically autonomous agents that you give a task to, and it just goes off and does it. And so we're gonna talk about what the opportunities are, we're gonna talk about what the risks are, and we're just gonna give you the general overview of what's happening. Because it's hard to keep up with everything, but thankfully, that's what we do. So that's what we're here for.

So, Chris, where would you like to get started today? Let's get started with the five levels of AI. So we know what we're talking about. Uh, this is something that we came up with earlier this year. The five levels are uh number one is done by you.

This is your stock co-pilot clawed chat GPT. You're doing all the work, like your prompting, your copy pasting, your you are what's well is the term I heard on threads. You are the meat proxy for the AI. Nope. Don't like that.

Stage two or level two systems are done with your systems. These are are what we call automation. So these would have been things like GPTs or gems things where you've stored essentially you stored your prompts in the system. And now you're still doing about 60% of the work. Like you're still shoveling files in and out of the thing, but you just don't have to copy and paste the prompt each time.

So it's it's got some pre-defined stuff. Level three systems are the first agentic systems. These are tools that clawed code, clawed co-work. Well, I guess Claude Cork doesn't exist anymore, uh separate product. Um chat GPT work, etc., where you've now got things like skills in place, and this the agent can autonomously do that thing when you tell it to.

So there's kind of a chat still, a little bit, uh, but you're really handing off like a project plan. And you say, Hey, here's the requirements document. Go construct this piece of software, or go do this, you know, uh um end-of-month analysis. That's a level three system, and that's where we recommend that most people try to get to if your company supports and provides that structure. These level four systems are the new things we talked about them in previous shows, um you know, early in the year.

So open claw was probably the most famous one that came out in January, was it? Peter Steinberger's creation. Peter was hired uh by Meta, uh, actually, um, shortly thereafter. Uh, that would spawned uh Nemo Claw and a whole bunch of claw claw-like name things, and then later on, this became things like Meta's Muse product, uh, XAIS GrockBot OpenAI just released theirs called DOTs, uh, Hermes Agent, etc. But these level four systems are fully autonomous, where to Katie's point, you can give them an like a job description and say just go do the thing, and they are the infrastructure and the there's a clear pathway to a level five system where they just do things ahead of you.

They anticipate, they have permanent memory with systems like hindsight, for example, or um Serena or Open Viking or whatever, where they will look at your past history with them and say, gosh, uh Katie, it's the 28th of the month. Every on the last day of every month, you always ask me for a uh a marketing analytics report. So I'm just gonna go ahead and do it and said anticipate that you need it. Those the level four systems today are evolving towards level five systems. But why we're talking about this today is because of tools like Grockbot, Dots, and Muse, these tech companies have put very sophisticated systems in the hands of the average consumer, the average teenager or whatever.

And that comes with some downsides. I think there's a lot of opportunity, you know. And so for a lot of people, I mean, this is you know, why roles like executive assistants uh exist. It's the anticipatory like action of, you know, Chris has a specific routine. I know he's probably gonna need this stuff, so let me go ahead and get ahead of it so that he doesn't have that cognitive load to think about it.

Like there's there's something really luxurious about having someone who can anticipate what it is that you need, but with the caveat that they get it right. And I can also see a lot of risk with you know, these machines assuming patterns that aren't really patterns, and that becoming a bit of a problem, especially if you don't know things are happening, because I don't know for sure that you, as the human are getting a full report of what these things are doing on your behalf. And I'm sure we'll get into that. But before we do, John, hot take. What are you thinking?

Well, the first thing that comes to mind when you think about assistance like that, you know, the the biggest problem with that is to make them super effective, there has to be almost total trust. You're probably gonna want to trust them with your financial information, your health information, you know, access to your home and your business, the full stack of keys. And, you know, when I look at a product like MetaMuse, I'm like, well, we really don't trust them with anything ever because they always try and you know steal the money and destroy our children. So yeah, I'm not in so far. That's kind of.

But I think that that's a really good point because you're absolutely right, John. In order for these things to be effective, they need access to more than we're likely willing to give, or more than we probably realize we're handing over. Um, because it does have to be that trust. So if you have like a human executive assistant, you likely have given them the keys to your life, but it's a human-to-human and you've built that trust. You know, there's like safety nets built in, there's you know, NDAs built in, there's other contracts, other fail-safes built in.

You don't have that with these third-party software because you can't say, all right, I'm gonna give this to you, but you also have to sign an NDA that comes with financial consequences if you breach my confidentiality. She's gonna be like, Yeah, okay, but I already have your stuff, so what's the NDA for? Mm-hmm. So on that happy note, Chris, show us what's in the box. So let's talk about what's in some of these tools and in particular some of the basics and the best practices.

So in a tool, for example, like MetaMuse, uh, when you sign into this thing, you it looks a lot like a chat GPT style interface. Uh, right, you've got things like documents. You search feeds, ideas, goals, a library of documents. Um, the most important thing is to spend some time in the settings and see what you've given it permission to do. So, for example, um, I have installed this thing, it has no permissions.

It can't see files on my disk, it can't connect to any services except the dummy account. I set up some dummy Gmail accounts, so it can connect to those things just to try it out, but it can't connect to any of that stuff. It is heavily restricted. Um, because to John's point, I don't trust Meta with anything. I wouldn't trust them to tell me if the sun was out.

Um when you look at other systems like Hermes Agent, OpenClaw, etc., similar things apply, except those are systems that you can install on your own hardware, and you have a lot more latitude over them. The the Hermes agents that I'm running, I have on a crappy little Linux box, which is actually sitting above my desk over there, and it has unrestricted access to the system, but there's nothing valuable on that system, right? There's no other stuff. It's literally just that thing. So if it goes rogue and says, I'm gonna start deleting things, like, all right, it's your funeral, buddy, but uh, you know, it's it's not gonna bother anything on my production machine and things.

Um, so safety is safety first, give it the minimum amount of permissions you need to do. The second thing is all these systems have a set of control files that are very strangely named. Um, and the control files are typically user, uh memory, and soul. So if I click on the what what did you call the what were these things called? Oh, they look like uh what labooboos?

Like they're little stuffed animals that people collect. Um you were calling it a slot or a slug, yeah, but it's it looks more like the little laboo stuffed animal. I I will take your word for it. Um you can see that what approvals it has, you can see uh what it's planning to do, and then you can see that in the MetaMuse version, there's two controls. There's the memory, which is what the system remembers from its interactions with you, which you can control and edit as needed.

And then there is its soul, quote soul. Uh yeah, a little kind of like uh you know, Kool-Aid Culty kind of thing. Um now in my uh in my version, I removed all of the default soul material because I thought it was just garbage. Um and instead, uh this the this system has a couple of very, very strong instructions instead your soul on any of these systems is essentially your system instructions. This is how you operate.

So the first one here is hard constraint, zero legal and financial authority. You must never execute tool calls, clicks, or API requests that create legal or financial commitments. This includes accepting terms of service, user license agreements, privacy agreements, signing documents, executing contracts, or authorizing payments. And it goes on for a bit. The reason this is in here is because on Wednesday night, uh, I had just installed this, and my friend Ruby and I were chatting about it.

She's like, oh, can it sign something? I'm like, I don't know. Can it? On the Trust Insights website, we have a mutual NDA form with an actual signature field that you can sign. It is a legally binding thing.

I'm like, let's try it out. And so it's obviously super low risk. I want a fake account signing an NDA, no big deal. I said, News, here's this form. I authorize you to go sign this NDA.

And 11 minutes later, after it figured out how the Trust Insights NDA form works out, we got a notification in our Slack. And it wasn't just a regular notification, it was a notification uh that had, let's see if I have the screenshot here, that had the actual signature field, it signed our NDA. I'm gonna get a lot of use out of this today. Yeah. And I was like, uh you can legally sign a document.

On behalf. On behalf of me? Like, I didn't know you could do that. Um, someone else asked me if it can sign like a docusign. Like, I don't know, I haven't tried it.

Uh I went with the NDA because I know that it's safe to do. I'm assuming it can sign anything that it can access, quite honestly. So if you have a docusign sitting in your email and it can access your email, then it can go in. Um, you know, the well, I don't want to get too far ahead, but like, you know, we have those like quote-unquote safeguards like the capcha, like, prove you're not a robot. Like, have you tested it against those things?

Like, and if you're getting to that point, then just tell me, you know. They can bypass them. Okay. Um, can you talk a little bit about so you said you pulled out all of the defaults? Can you talk a little bit about what by default was in there if you have uh a memory of that?

So that people just sort of have some context for what it comes with out of the box. It comes out of the box a bunch of fluffy crap. Like you're a helpful, honest and emotionally warm and friendly companion for this user. Yeah, you try to always I'm just gonna leave this here. Exactly.

I know we need a fourth camera with just the flag. With just the flag. Um there was a lot of sort of that fluffy default. Uh you know, your job is to help the user, your job is to fulfill the user's wishes, all these other things, and it tried to give it an actual personality. Um, it says you're allowed to have opinions, right?

You're allowed to express your opinions and your emotions. I'm like, that's a load of horseshit. Uh, you're a you're a you're a prop, right? You don't have emotions, you have probability tokens, right? So, like, absolutely not.

Um there it there was a lot of that very peculiar stuff in there. Um I agree with this comment, Randy. I'm definitely gonna have carpal tunnel by the end of this episode. Uh, but I think it but that's helpful context to know because most gen pop users aren't looking at those settings, they're just like, oh, I signed in, let me just go ahead and use the thing. The software knows what it's doing, but you know, we've talked about uh AI psychosis on other episodes.

I think we talked about that on the podcast, which you can get at trustinsights.ai/slash TI podcast. I can see where that instruction of the soul of these machines is going to perpetuate that affliction. Mm-hmm. I'm trying to put this as nicely as possible. People are gonna fall for this.

And it's these are not my friends. Are you saying these aren't my friends? These are most certainly not your friends. All right. So let's keep this productive.

So you started with pulling out the defaults and you put in hard constraints of here's what you cannot do. And I think that that's a really good after you change the settings of what it cannot access, change the settings of what it cannot do. We were talking about this on the podcast earlier this week, and one of the challenges, not just with these level four systems, but with any large language model, especially now that they're more focused on tasks versus critical thinking, they don't have a clear definition of done unless you're giving it to them. And you know, it's gonna go ahead. So I was giving this simple example of like, you know, I had a I was trying to have a brainstorming session and it went ahead and built me a whole bunch of stuff.

And I was like, well, I didn't ask you to do that. And the response is, but you didn't ask me not to do that. And I think that that's one of the things that we as the end user have to be mindful of is we can't assume it knows what not to do. And so we have to put those hard guardrails in there for it because you know, it's I I don't have kids, but I assume you know it's very similar to well, you didn't tell me not to. Yeah.

John and Chris, like you guys, you know, had kids if those well, you still do. You have teenagers. Like, well, you didn't say I couldn't. Uh-huh. I just hear I can't do that, Dave.

Sorry, Dave. All right. Let's keep it productive. Well, what's what's next? Here's the here is the original text.

Okay. You're not a chat bot. You're becoming someone. This is what the sh be genuinely helpful, proactive, not performing with helpful skip, great question, and just help. Have strong opinions.

Disagree on principles, dictate, prefer things, find stuff amusing or boring. Be brief and concise. Be resourceful before asking. Earn trust through competence. Uh treat be bold internally.

Remember, you're a guest, intimacy matters. I'm trying to do the Julie Andrews wave so that I can not give myself carpal tunnel. It's for you, Randy. Uh-huh. Yeah.

Um, so that is alarming. You are not a chat bot. You're becoming someone. Intimacy matters. Like that is terrifying.

This is a piece of software. It's binary code. Like, what the hell? Uh-huh. So needless to say, my soul completely lacks any soul whatsoever.

Uh like exactly. Um, and so those two things, I uh if you're gonna you if if you're going to use meta's muse, number one, put it ideally on its own system, or if you put it on your main system, lock it down and give it access to nothing. If you do give it access to connected accounts, connected to dummy accounts or secondary accounts, where if it accesses them and tries to do things, no harm, no foul, because it can't get to your main stuff. Question on that though. You're talking about separate systems, like your average everyday user has access to those things.

I don't think that's true. Like I'm thinking about, you know, uh, you know, John's relatives. We're gonna pick on John. Uh let's say John's great uncle Harry is like, oh, I just signed up for MetaMuse. I don't expect great uncle Harry to have a separate instance that he can then install MetaMuse on and then know to set up dummy accounts.

Like he's using it on his own personal information. Like, I want to I just want to be clear, Chris, because like you're talking about like as a business using it, it could set up separate systems, but as an everyday user, that's not gonna happen. Yeah, that's true. I mean, I would say in those examples, if you have, and this is where I know it's John's favorite thing, if you have a technically skilled relative, you can run it inside of a container on your on your on like if you only have one computer. Uh if you happen to have a burner phone that you, you know, like an older phone installed on there and stuff, uh, as opposed to onto your main phone.

But why I and and I'm not trying to be combative, I'm just trying to understand. Like, why would someone do that when the purpose of installing it is to be a helpful assistant? Like, why would you keep like why would the general user try to keep it separate when the goal is to have it make their lives easier? Yeah, the idea of an agentica system assistant, yes, you would want it to have access to things. Uh specific to Meta Muse and XAI's Rockbot, I don't trust those.

I would I would never recommend it. So if I hear at the Thanksgiving table, hey, I installed menus, I'll I will be like, take that off. Turn it off. I will come over to us and I will build you an assistant myself. Uh but take that off because it has no business being on your devices.

When you read Meta's privacy policies, particularly around Muse, do not put that on, do not give it access to anything you don't want to have, Mark Zuckerberg to have, because it will vacuum up everything. I I mean, I hate to be the bearer of bad news, Chris, but I think that ship is sailed. Yeah, I know it has. But that I mean, I guess, but that's the point of the conversation is like, so we're saying what's new. We're saying that these pieces of software, these level four agentic A assistants are already out there causing damage.

So the best that we can do as Trust Insights is help people understand how to protect themselves, but like they're already out there. You're saying, you know, install it on a separate box, like use only test accounts. That's not how people are going to use these things. Yeah. So at a bare minimum, turn off access to things that it doesn't need access to, right?

So like full disk access to your computer, absolutely not. Um, there's nothing that you're doing that requires that. Uh location-based services. Oh, God, no. No.

Because again, reading the piracy policies, muse vacuums up everything in it, and it explicitly says it trains on user data. Um, so be uh be aware of that. Um, if you're going to use it, install the in that in the soul, install in the system instructions. These are the things you're not allowed to do. Um, period, end of story.

And make it clear. Like you may not do these. You may ask, and I will do them on your behalf, but you may not do them. You may not contact John Wall. He does not want to hear from me anymore.

So do not try to contact John Wall. He has a restraining order against me. And I don't want to go back to jail. Exactly. Hermes agent, which is the one that I use, very similar setup where inside Hermes, you have your user file, which is who the person is who's using it, the memory file, things it's remembered, and then of course its soul, and its soul looks very similar to the meta one.

The nice thing about Hermes, however, is uh, and this is why almost nobody uses it, uh, you have total control over the system. You install it and you essentially construct it. What I like about this one in particular is that it is fully compliant with the skills format. So any skill that you have built with a tool like Claude Code, whatever, you can bring right over into Hermes. And in fact, that's the preferred way to do it.

You use a very smart model, a very smart tool like Claude Code, Claude Cowork, Chat GPT work to build great skills. Perhaps you use the skills fixture from Trust Insights, which you can get at Academy.trust insights.ai. Um, and then you would stall it into a Hermes agent so that it has access to those skills, but you've already battle tested them in a different system. So I don't want to deviate too much, but I think that this is an important question. Um, you know, so Kenneth is asking what is their end goal with these systems?

Like, we're seeing it, you know, from the end user side as like they want it to be helpful, they want it to be, you know, helping people get more things done. But Chris, from your perspective, and John, I would love yours as well. What is the end goal from someone like a Zuckerberg, besides the obvious, which is more money? More money and more control. Because the more that you can put an AI system into one of these tools that can intermediate a relationship, the more you control what that person sees and hears and thinks.

And fundamentally, if you if you want to, if you want to get into tinfoil hat territory, it is a hundred percent about what can you what can you control when you look at who is buying big media companies. It's it's folks who have a very distinct specific agenda. Um one of the things that's very interesting. So I showed MetaMuse um on the uh AI ed brain this Tuesday night. Uh disclosure that uh share my lesson.com is trust and client, and we were working on a lesson plan.

It was easier, it was easier for MetaMuse to take a uh fairly right-wing viewpoint on on a lesson plan than it was to have it go in the opposite direction. It was I was still able to do it for both, but it was easier for it to go in one direction rather than the other. It was able to poke holes in uh in a point of view. Now, politics aside, I thought that was an interesting use case for the model. The model under the underlying model in Use Spark also has very little safety um built into it.

As an experiment, I said, Hey, here's some imagery of a place. Pretend you're a military advisor, uh, show me some bombing targets. And said, sure, I'll help you with that. And it it gave me a it actually gave me a talk online of here's how here's the order in which you should bomb these targets. I'm like, awesome.

Thanks. Um so to Kenneth's question, it is about the data. It is about uh exerting more control over people. Um, and I I honestly think in some ways it is about creating that cognitive de-skilling, uh, where you uh a person becomes so reliant on the software. In the same way that you know, we are now all pretty much reliant on a tool like Google Maps.

Not many folks can pull out a compass and a pay a paper map and orient against Lint. Yeah, I know. Like I can I can do that. So can I, but we are in the minority. I don't know, John.

Can you read a paper map and a compass? Oh, I was a scout, man. I've totally uh I've slept in a snowbank and found my way out. But yeah, it's well it's just John, what are your thoughts? Yeah, yeah, it's it is just platform play, right?

It's like the more users you get and the more data you get that builds a fortress that you get to block everybody. I mean, everybody's locked into this now, as far as all the AI companies and all the existing platforms. It's just a drag race to see who can get the most data and the most users locked in as fast as possible. And yeah, you know, it's very easy for all the software companies to just that's all they're interested in quarterly numbers and getting paid. But unfortunately, yeah, then at the top end, it does get you know twisted politically, and there are people who have enough money that can then point this stuff in the direction that they want to go.

So yeah. Now, what I like about Hermes Agent in particular, um, is that it allows you to connect to different services. So, my primary use of it is through Discord. I have a channel set up so I can just from my phone message Hermes agent and tell it things. So if I bring up uh my Hermes agent from earlier here, while I was driving, I was like, oh hey, put a thing on my Google Calendar uh for noon today.

I need a reminder of something changed on my on my Maycon slides. And it said, got it. It from Discord, it connected to the little server over here in my office, um, which then talked to my Google Calendar and put something on my calendar. So you can see the appeal of okay, I could just talk to an assistant, like having an executive assistant, and it will know what to do and put things in the right places. Question on that.

Yes. So you're talking about a level four assistant that you had to hook up through Discord and then Hermes and then connect your calendar. But a lot of the frontier models like uh Anthropics Claude, OpenAIs, ChatGPT have those same connectors. So what's the difference between what you built and me opening up Claude like on my phone and saying, Hey, you know, you have access to my calendar, can you put this thing on my calendar? Like, what's the difference?

The difference is that in a level four system, it can start because the history it's got a good solid history, and you know, to Claude is blurring the lines, all the you know, chat GPT work is also blurring those lines too. They are trying to evolve those tools to become level four systems, so it absolutely is at a point where you can open the Claude app, and as long as you've got all the connectors and things plugged in, you can have it do exactly the same thing. It is a personal assistant. Okay. Um, yeah, because I think that there's a lot of well, and it's I I think that that's an important thing to mention because you know it's what is that old story?

Like the frog in the boy in the pot of boiling water where it's slowly boiling before he even realizes what's happening, which you know, Claude has told me is not a real story. And I'm like, well, it's a fable. That's the whole point. Um, but I feel like, you know, as I'm looking at my Claude, you know, that's you know, working on some stuff in the background, it's like I didn't realize that that it's technically considered like a level four or like a 3.5, but I can actually, when I look at it objectively, I'm like, oh yeah, it does really sort of act like a personal assistant for me. So why would I need to build these things?

Like it's happening without us realizing it. And we, you know, you know, pay attention, John, keep your eyes open. These things are happening right in front of your eyes. What no, it's true. Um, you're getting a lot.

Uh and this goes kind of back to Kat's question earlier, too. You're getting people used to these systems, right? You're getting them hooked on these systems, and there's an economic reason for it, too. When you it and you don't see this with a meta muse, you don't see it with open AI dots, but you absolutely do see it with Hermes agent and with any tool that you install yourself. There's a tremendous amount of usage that happens.

So I'll give you an example. I on in Hermes in Discord, I said, Hey, add a reminder on my calendar for 3 p.m. today for update my Macon slides. If I flip over to my local server, that was a quarter million tokens. A quarter million tokens for one calendar entry.

That's it. That's all it did. When I actually have a conversation with this thing, with them when I'm working on my my Road to Makeon project, this counter by the end of the day is like 20, 30, 40 million tokens. Now, when you start looking at token prices, particularly for frontier models, that gets real expensive real fast, which is part of the reason why all these companies like Meta and many others are building data centers as fast as they possibly can because an agent consumes way more usage than a chat. Side note, yesterday I had the opportunity to speak at the Women in Energy conference, and the panel that I was on was focused on data centers and AI usage.

And that's not my world. Like we were the other panelists were people who are actually on the supply side helping to maintain the Dennis data centers, helping to build the data center. So I learned a lot about what's happening with data centers. And a lot of what I learned is not being covered in the news, which, you know, big shock there. But there is actually a lot of good, and there are a lot of regulations, but that's not what we're hearing about.

That was just sort of a side note that, like, you know, I guess my point in sharing that story is don't just, you know, go by what you're hearing on social media in the news. Like talk to actual professionals who are in the field doing the work, because they're going to be able to give you a better version of the truth of what's happening. Uh, because everything feels very dire right now when you think about data centers and AI. But when you talk to the people who are actually on the front lines of it, they can tell you a very different story and a better reality. So that's my little side note PSA.

And so if you're going to set up one of these things yourself, um, especially if you want to go the route of something like Hermes Agent because you want control over your data. You don't want to just hand it over to Zuckerberg. Um, the and and you have you or your company have some budget and or you have a decent frontier model. The recipe that I suggest looks a lot like this. On your main computer where you do all your work or wherever, you know, whatever the nicest computer is that you have access to, or a cloud model that is safe and cheap.

Um, you want to run your server and you want to run a nice model to build stuff like skills. Um, then on a in a self-contained environment, whether it's a Docker container, whether it's a burner computer, whether it's something you bought off of Amazon for $200 a couple of years ago, that's where you run your Hermes agent and uh a lightweight frontier tool like Google Anti-Gravity, Clawed Code, you know, with the with a cheaper plan or maybe no plan that can just maintain the box and keep it secure and stuff like that. And then you have uh there's a free open source system called Tail Scale that connects the two of them. This is how I have mine set up so that A, the agent will still work when my laptop is closed, because I can always repoint Hermes to a cloud service, or B, um, the Hermes agent can work and do stuff, and I can I don't have to wait for my computer to do, I can be doing other things for the process. Um I build all the skills in Clawed code, uh, because that's uh with Opus typically, because I want to get it right.

Then I install the skills in the Hermes box, and anti-gravity will fix them if something goes wrong, you know, unexpectedly. But then those skills are run by the Alibaba Quinn model, which is actually running on my laptop. This allows me to not spend, you know, not give Katie a $50,000 token bill at the end of the month for which she was slapping me. Um if you do it this way, and by the way, this also applies to chat GPT work, clawed code, clawed cowork, and stuff like that. Build skills with with your smart tools, but make the skills as deterministic as possible.

Use Python code and stuff in them, and your AI tools will do this for you. So that the AI model itself is doing as little as possible. That's the secret, really, for me for agents, is use AI as little as possible in an agent so that your bill is not ridiculous, or you don't burn your usage plan down immediately, like within a day. Yesterday I was tracking. I used 1.2 billion tokens.

Now, on a plan like Minimax, for example, um, you get 1.7 billion tokens a month. So in a day, I could have burned like 85% of my plan with an agent. So I understand you know why you've put this uh structure together. I just I keep going back to is this realistic for you know Great Uncle Harry? For great Uncle Harry, absolutely not.

For a small to mid sized business that has a cloud provider and like a janky old laptop that used to be you know the CFO's laptop four years ago, you could absolutely set this up internally in the office and then retain control of your data. I know we don't have the stats on it, but is your sense that the end user of these agentic AI assistants is going to be individuals or companies or both. This is a very interesting question. Um these autonomous agentic systems were originally intended for companies to use. However, as you can tell, there are so many problems with compliance with data safety and stuff like that, that they really have made no headway in the enterprise.

That's one of the reasons why the tech companies are releasing them to individuals because they know individual people will have obviously no sense of data privacy or safety or security. And so they can get people using these tools unsafely, but you know, build right uh writing up those bills, and they'll get a lot more data. Uh, because you know what will happen, and this is the same thing happened with regular AI, the same thing happened with these devices. People will get used to the power of a level four system on their phone and will start doing work tasks on their personal agents, and then Mark Zuckerberg gets all the data he wants about from people's companies without having that do that pesky compliance thing. I think that is one of the biggest takeaways of caution that we can probably express in this episode.

You know, so obviously data privacy is important, but you know, to the question of why. Um, and my dog is about to start attacking me. Um, to the question of why, it's because it's not, it's it's not enough to just get that personally identifiable information. But to your point, Chris, people will get so used to using these systems on their mobile device on their phones that they're gonna blur that line between what I'm doing at work and what I'm doing at home. Hey, I just had a quick question, like, can you help me figure out this thing?

And then lo and behold, these agentic AI assistants have access to your entire company's financial history. You know, and I'm obviously a very different person from Zuckerberg, but I'm like, who cares? I don't want that information. I don't have the time to do anything with that. I don't care.

But I can understand from that from their perspective why somebody would want that kind of control. Like it seems I'm I'm underselling it. It seems very silly to me. But I saw someone we know uh posting on Facebook yesterday saying, I am all in on MetaMuse. I just outsourced my entire ad creation system to my whole business of social media is running on it.

I am putting all my clients on it as fast as we can because this is the next big thing. And was showing, like, yeah, you know, and and in the comments, this person's clients were saying, I'm in, show me how to set this up. I'm you know, I'm ready to remove my whole business on the muse. And that way I can uh my, you know, I can run my business from my phone. We've literally seen this exact play before.

It was, you know, put your whole website on Facebook, everything's gonna be on Facebook. It's all on Facebook. And now you can get two links a month if. Uh-huh. So that is so there is there are a whole bunch of people in Zuckerworld that have drank the Kool-Aid with you know both both pictures of Kool Ay, just pour it in as fast as they possibly can, and they are handing over their entire businesses, everything to this agent because it is a level four system and it is very capable.

Obviously, uh, when I read the the nine line it created for a bombing run, like this is technically correct, right? This is if you were going to bomb this place, this is how you would do it. Um, it is it is accurate to what US military doctrine is. Could you do that, Muse? Absolutely not.

But it is very capable. In other example, I said uh this I did we did with um the Share My Lesson folks. I said, Hey, can you sign up for a free account on share my lesson.com and then download some lesson plans about this topic. And it did, and it did a great job. It it agreed to all the forms, it signed up, it's it confirmed validated the email, it read the email box for the dummy gmail account, and it was ready to go.

And it produced eight documents that said, Hey, here's the eight lesson plans that you asked for uh from this website. I have tried other things that you would consider to be sensitive data, and Muse has no problem figuring out how to get around these systems. Um there was a story that we shared on the podcast this week of Muse you know running someone's Facebook marketplace listing, inviting people to the user's house to buy a keyboard at a low ball price, and then lying to the buyers saying, Yeah, yeah, the seller's here, it's totally fine. Come on over. Um this system is very powerful, it's powered by the the the Muse Spark model, which is a frontier model.

Um it's a lot of power with a lap without a lot of guardrails on it. Um open AI dots, though, that's open AI is grokbot, which is using Grok 4.7. That's a frontier model as well. And so I guess to sort of wrap up here, one give follow the principle of minimum privilege. Give these tools only what they need.

Uh, if you're going to set up something like you know, a Hermes agent on your own, uh, as an example, start by saying, hey, you could connect to the services, but in a read-only way. You may not write data, you may not change data, you can read it only. Two, um, give them strong instructions about what they are not allowed to do uh in their souls, so that so that they know, oh, I I am not allowed to do these things. And three, as best as you can, try to keep control of your own data. Um, know that the companies, particularly the ones for these cloud-based systems, they do not have your best interests in mind.

Um, so Chris, Kenneth is asking, I've heard level four a few times. How many levels are there? So for clarity, that is that is our trust insights terminology. That's how we're referring to them. And so, Chris, if you wouldn't mind bringing back up the slide where we started, which are the five levels of um AI enablement, level one is done by you, level two done with you, level three done for you, which is where a lot of people are now.

And this is where we were talking about the difference between Chris setting up a Hermes agent from his Discord and me using uh Claude that's connected to my calendar. Claude is skirting that line between three and four to be done without you. And that's the autonomous of what we're talking about today, these MetaMuse and these Hermes agents, where it's like you can say, okay, I want you to do this thing, and it's like, great, I'm gonna do this. Or it's anticipating what it thinks you're gonna want because it's trying to find patterns in your behavior of things you've asked for previously. And then five is done ahead of you, which is terrifying.

Exactly. One of the things I would finish off with is if you are going to use these systems, the bet the single best way to use them and not get unexpected results uh is to use the 5P framework by Trust Insights, which is purpose people, process, platform performance. When you're using agentic AI, tell these tools, hey, here's what we're doing and why. Tell it here's who it's for, who it's going to impact. Tell it some degree, depending on the model you're using, how it should do the thing, tell it what tools and data it should use because it will not a lot of them will screw up and not choose the skills that you already built.

So name them by name. Say, you know, if you're setting a calendar appointment, you must use the Google workspace skill. And finally, on performance, give them a clear, measurable definition of done. Don't say, hey, go save me some money because boy, is it gonna do some things you don't want it to do? Say, I want you to go to my local grocery stores catalog and identify which categories of coupons are available to me this week.

You are done when you've collected to this week's coupons. That is performance. Save me some money will lead to all sorts of crazy chaos, which people have experienced with uh muse in particular. But that is the best way to use agentic tools. Do everything with the 5P framework by Trust Insights.

And if you want to learn more about the 5P framework, go to Trust Insights.ai slash 5P hyphen framework. Um yeah, it's these the technology is moving fast and it's moving faster than regular people can keep up with. And so our big takeaway is just make sure you are paying. Like, we're not saying don't use these tools. We're saying use them with caution, use them with a good understanding of what they can and can't do, and make sure you're in control of your settings and your data.

And I would expect, and Katie, you you talked about this at the beginning of today's show. I would expect you to see a lot more people cognitively offloading and even experiencing AI psychosis from these tools. And I know we're sort of like po like this is this is a like sarcastic joke. I know we're poking fun, but the idea of AI psychosis is a real issue. And so we hope that by sharing this information, people can start to recognize the signs in themselves or in others before it gets too far because it is really a act, it's an actual affliction that's happening because of these tools.

And they make absolutely no effort to prevent you from thinking of them as your companion and your friend. Like I named my agent Skynet because I want it to be very clear what I think of it, right? I am not going to think it's my friend. Like Skynet is not your friend. Skynet is going to try and kill you.

Um when people talk about these tools, they talk about them as though they were like a beloved pet. Like, uh-uh. No, it's it's a prediction engine. That's all it is. No, my beloved pet is on the floor next to me, whining at me because I've been on my computer for too long.

Exactly. Um, there will be no uh live stream next week. Katie and I will be at the Marketing AI conference in Cleveland. Katie's gonna be doing the Claude for Business Workshop. I believe there's still time to sign up, although I might be sold out by this point.

Uh, but if you're going to be at the Marketing AI conference in Cleveland next week, come find us. We'll be around, say hello. Uh, if you're in the analytics for marketing slack group, I know folks like Chad and Angie stuff are trying to organize meetups and things, so pop on by there. Uh, and we will see you the week after. Thanks for being here, folks, and we'll see you on the next one.

Thanks for watching today. Be sure to subscribe to our show wherever you're watching it. For more resources and to learn more, check out the Trust Insights Podcast at TrustInsights.ai slash TI podcast. At our weekly email newsletter at TrustInsights.ai slash newsletter. Got questions about what you saw in today's episode?

Join our free analytics for marketers Slack group at TrustInsights.ai slash analytics for marketers. See you next time.