Macpreneur: Helping Solopreneurs Streamline Their Businesses on a Mac

Why Mac Solopreneurs Should NOT Start With AI Agents (And What to Do Instead)

• Damien Schreurs, Technology Mentor • Season 7 • Episode 172

Share episode

Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.

0:00 | 25:22

🆓 Discover your AI Stage: https://macpreneur.com/aistage

In this episode, I introduce the AI Capability Matrix, a two-axis framework that cuts through AI confusion and shows you exactly which AI tool is right for your business stage. 

You'll discover:

  1. why Large Language Models are simulators, not thinkers
  2. the maturity journey you should follow (and why skipping steps gets you into trouble),  
  3. and why the Claude desktop app is the only all-in-one solution Mac users need (even on older Intel machines).

Show notes and video version: https://macpreneur.com/episode172

Highlights:

Text me topics for future episodes

🎤 Want to be a guest on the show? Fill the application form available at https://macpreneur.com/apply or visit the show profile on Podmatch.

👥 Join the Macpreneur Community!
Simplify your digital life and streamline your business with fellow solopreneurs.
✨ Join the Waiting List

✅ Want to be more efficient on your Mac?
Answer a few questions about how you're currently dealing with unnecessary clicks, repetitive typing and file clutter. It's FREE and takes less than 2 minutes!
🆓 Get personalized time-saving tips today!

Follow me:

MP172 - Why Mac Solopreneurs Should NOT Start With AI Agents (And What to Do Instead)


The AI Confusion and a New Framework (00:00)

Damien Schreurs

Recently, I watched a YouTube video where someone was explaining how to install the Hermes agent with all the security risks that it can bring along.

After 25 minutes switching between the terminal and entering API keys for a bunch of third-party services, the guy asks the agent to do some competitive analysis for a business idea that he had.

Sure, Hermes did it, but based on the output, it was very clear to me that Gemini's deep research would have done a better job and would not have needed to pay extra for tokens.

And on top of that, I would have been able to save the report as a Google Doc just in one click.

And that's when I realized that the confusion around AI is actually leading solopreneurs the wrong path.

Today, my Macpreneur friend, I want to give you the one framework that cuts through all of it, and it's the first piece of a series that I've been building specifically for Mac-loving solopreneurs like you.


Welcome to the AI-Enhanced Macpreneur (01:15)

Nova AI

Welcome to Macpreneur, the show for seasoned solopreneurs looking to streamline their business on a Mac.

Unlock the secrets to saving time and money with your host and technology mentor, Damien Schreurs.

Damien Schreurs

Hello, hello, and welcome back to the Macpreneur podcast.

As a fellow solopreneur, I truly appreciate you dedicating these 15-ish minutes with me today.

If you are new, welcome.

This show is all about helping Mac-loving solopreneurs like you save time and money by working smarter, not harder.

We are deep into season seven, which I dub the AI-Enhanced Macpreneur, and today, I'm kicking off a four-episode series called Your Mac, Your AI Stack.

Each episode stands completely on its own, but they're designed to build on each other.

So if today resonates, make sure to subscribe or follow the podcast so that you automatically get the next episode in your feed.


Introducing the AI Capability Matrix (02:24)

Damien Schreurs

This framework, my Macpreneur friend, is called the AI Capability Matrix.

It's going to help you identify which of the three AI capabilities you may either be already using or be considering.

It will help you understand which one is right for your situation right now.

Recently, I heard an AI expert as a guest that repeatedly labeled a custom GPT or a core project as an AI agent.

This thing does not have autonomy.

It is just a well-configured assistant.

If the people being invited as experts are getting this wrong, it's no wonder that everybody else is confused right now.

There is no shame in that confusion because the industry created it.

Let's get back to the AI Capability Matrix, which uses two dimensions.

On the horizontal axis, we have the autonomy, how independent the AI can act without you directing every step.

On the vertical axis, we have the number of roles that the tool can take on.

At the bottom, it's narrow and specialized roles, and at the top, it means that they can have broad or general-purpose types of role.

When you plot today's AI tools on those two axes, there are three clear categories that emerge.


Category 1: AI Chatbots (04:05)

Damien Schreurs

The first one is called the AI chatbot.

You direct every step, but they can take almost any role.

They can become a writer, a strategist, an analyst, a marketer, a coding assistant.

This is where most people leave most of the time, and it's perfectly fine.

And in the next episode of this series, I will show you exactly how to talk to a chatbot in a way that consistently gets you the best possible output using a prompting technique that I teach in my corporate AI training sessions.

And to my knowledge, I've never shared this technique on this podcast before.


Category 2: AI Assistants (04:57)

Damien Schreurs

The second type of AI capability is called the AI assistant.

Those are preconfigured, specialized versions of chatbots.

You build them once, and then you can reuse them repeatedly.

Think of the difference between asking a friend a general question versus briefing a dedicated specialist who already knows your context, your tone, and your preferred output format.

That's the shift from chatbot to assistant.

Examples of those assistants are the Co-CEO, Co-CFO, and Co-CMO that I mentioned back in episode 154.

And those three assistants helped me generate additional revenue ideas, but also saved a lot of money along the way.

On top of that, this episode demonstrated through my guest that staying on the free plan of ChatGPT can actually hurt solopreneurs, because they cannot create those assistants unless they upgrade to the plus plan.

So, if you missed this episode, visit macpreneur.com/episode154.

I will put a link in the show notes.


Category 3: AI Agents (06:21)

Damien Schreurs

Now, back to the AI capability matrix, and the third type of capability is called the AI agent.

Those agents can take actions on your behalf.

They can browse the web, they can execute code, they can interact with other apps without you directing every step.

Back in episode 139, I had the pleasure to interview Anthony Franco, who turned his secondary Mac mini into a digital virtual assistant with the help of OpenAI Operator.

This agent was released back in February 2025, but it was discontinued in August of the same year.

Nowadays, it is in the ChatGPT interface as agent mode.

At the very far end of that agentic spectrum, there are tools like Open Claude and Hermes Agent.

Marketing School podcast co-host, Eric Siu, shared that he spent up to $12,000 a month in API costs to run his Open Claude instances, and purchased two NVIDIA DGX Spark computers at $5,000 each to bring the cost down by running AI models locally.

Impressive?

Yes, sure, but (laughs) it's definitely not the right starting point for a solopreneur.

And because it piqued my curiosity, I ran a deep research project to compare the DGX Spark from NVIDIA with an hypothetical M5 Ultra Mac Studio.

And because the Spark is limited to 128 gigabyte of RAM, a Mac Studio with 256 or 512 gigabyte of shared memory will be a better option, even though the raw speed will be lower than of the Spark.

What makes the Mac Studio attractive for Macpreneurs is the ability to run larger models locally, and you get the power of macOS.

Now, agents are powerful, but in some cases, it's possible to get the same kind of power at a fraction of the cost, and without any of the associated risks.

How?

Well, by building AI-powered automations instead.

And I will cover that in the final episode of the series.


The Truth About LLMs: Simulators, Not Thinkers (09:26)

Damien Schreurs

Now, before going deeper into how to leverage the AI capability matrix, there's one thing that I need you to internalize for the entire series.

Claude, ChatGPT, Gemini, Grok, Mistral, all of those large language models, actually none of them understand what you're telling them, and what they are replying back to you.

At least not the way that another human would understand.

In fact, if you know how they are built and how they work, then you realize that they are actually only extraordinarily sophisticated human dialogue simulators.

They predict what a helpful response would look like based on language patterns from their training data, reinforced by examples of actual human conversations.

Now, don't get me wrong.

It doesn't mean that they are useless.

They are super useful.

Frontier models and even the most recent open weight models that can run on a Mac with at least 64 gigabyte of memory are excellent human dialogue simulators.

The problem is people confuse intelligible with intelligent.

They may be capable of being understood like another human, but they do not think.

They do not understand.

They simulate by predicting a likely response.

And that distinction matters, because it applies everywhere in the AI capability matrix.


Your AI Maturity Journey (11:13)

Damien Schreurs

It doesn't matter how large or narrow the role is and how much autonomy they have, whichever large language model you will be interacting with, remember that you are dealing with a human dialogue simulator.

Now, with that in mind, how can this matrix be useful for Macpreneurs?

These three categories lay out a maturity journey, and I advise you not to skip a step.

Start with chatbots for one-off generic queries where context doesn't really need to persist.

Then graduate to assistants.

When you notice that you are entering the same context repeatedly, or you are expecting consistent output for a recurring task.

That's your signal.

When you find yourself typing the same background information for the third time in a row, that's your assistant waiting to be built.

Move to agents only when a task benefits from the AI acting on your behalf, especially when you are not at your computer.

But only (laughs) do and go to that level once you have very refined assistants you know are working reliably, and also, if an AI-powered automation would not be a more viable option.

Uh, an example is the first Open Claude instance that I built in a virtual machine to protect my computer, and what I realized is I had this agent help me build, uh, an AI-powered automation.

At one point, I- I asked the agent, "So why do I need you now?"

And he said something like, "Yeah, you need me to be able to look at the output of the, uh, automation, make sure that it doesn't break and so on."

And I realized, "Well, I don't need you anymore." (laughs)

And so the reason why I thought I needed an agent was, actually, I didn't need an agent.

What I needed was something that was more predictable, and it was actually a kind of recurring task, recurring automation that itself needed, at one point, an AI step.

Another way to look at it is if you would hire a VA, you would not ask that VA to do a very complicated task from the get-go.

You would start with smaller task, so that you would build trust with that VA incrementally.

And it's the same logic here.

A- and so the- the solopreneurs that I see struggling with AI are almost always trying to run before they have learned to walk, not because they are not capable, but because they haven't yet mastered the fundamentals.

Either they are building assistants that under-deliver because the instruction isn't well-crafted and lacks key reference knowledge and documents, or they lose time and spend a fortune building agents for capabilities that already exist in ChatGPT, Claude, or Gemini without realizing it, but also while exposing their solopreneur business to huge security and privacy risk.

So, before I conclude this episode, there's something important I want every Macpreneur to know.


The All-in-One Solution: Claude Desktop (15:36)

Damien Schreurs

There is an application today that covers all three types of AI capability, and it's the Claude desktop application.

The Claude desktop app is the only one that can run still on Intel Macs.

If you look at the ChatGPT desktop app and the Gemini desktop apps, they both require Apple silicon and macOS 15 Sequoia.

So, whether you're still on a 2015 MacBook Pro, a 2020 iMac like I have, or the latest MacBook Neo and anything in between, Claude desktop is your entry point into everything that we will cover in this series.

And so with Claude Chat, you have actually the AI chatbot capability.

But you can also create projects which are equivalent to Custom GPTs or Gemini Gems, and these are examples of assistants.

So, for instance, my co-CEO that lived in ChatGPT and as a Gemini Gem now exists as a Claude project.

You can also use agentic capabilities with the Claude desktop app.

Claude Cowork is this capability.

You can give Cowork any role, so it can cover a lot of roles, but it has some autonomy, not big autonomy, but still, it has autonomy.

With Claude Cowork, you can give it access to a specific folder on your Mac.

The worst that could happen locally on your Mac is on that folder.

If you give it access to different tools, then it's potentially dangerous.

But from a computer security point of view, the attack surface will be only the folder that you give it access to.

Finally, example of agent with a more narrow role is Claude Code.

It has limited, uh, autonomy in the sense that you will get Claude Code to work.... in a specific re- repository, so in a specific set of folders.

However, uh, because it can run shell commands on your computer, the level of risk is slightly higher.

And the good news is, all those capabilities, you have access to them from a single app, even if your computer is a bit older and you're not yet on macOS Sequoia or later, or if you are still on an Intel Mac.


Caution: Managing AI Risks and Tools (18:51)

Damien Schreurs

Now, I still want to give you some words of caution, because the power of Claude, and for me, the reason why I switched from ChatGPT to Claude, is the ability to use tools.

In the Claude parlance, it is called connectors.

Even as a chatbot, you can have your chatbot use connectors.

I hooked Claude desktop with the Note Plan MCP connector.

So I'm using Note Plan for my daily goals, weekly goals, weekly reviews, also, my monthly and quarterly planning and goal setting.

Everything is done in Note Plan.

And because Note Plan has access to my calendar, through the Note Plan connector, I can ask, for instance, Claude, can you give me a heads up, uh, what's on my calendar in the next four weeks?

Do I need to pay attention now so that I'm ready for training sessions or thi- things like that?

It's really powerful.

It's so powerful that it has write capability.

It can overwrite stuff and no later than this week, it actually did that.

It wanted to add some paragraphs of text to an existing note, and it wiped the note.

It replaced the text.

And when I ask it, it says, "Oops, sorry. Yes, I used the wrong tool."

They are human dialogue simulators.

They do not understand what they are doing.

The good news is that with Note Plan, it has a versioning system.

Every time a note is updated, it's a database system, and a new version gets created and the old version is still there behind the scenes.

So, I was able to go in Note Plan, revert back to the previous version, and then tell Claude, "Please, now, add the paragraph of text that I want you to put."

I have created a Note Plan note.

All the knowledge for Claude itself on how to best use the connector, this has been documented.

When I ask it, "Can you please reread that note?" it realized that it was preferable to do an append rather than a full replace.

But then, it added a- an extra caution note to remind itself to, uh, be more cautious in the future.

And so, the practical habit I want to give you is that every time you consider giving an AI tool access to something, could be a file, a calendar, a note, an app, ask yourself two questions.

Do I know exactly what action this tool will take?

And do I have a way to reverse any action taken by the AI tool if something goes wrong?

That's another reason why I really like Claude desktop so much, is because it has granular tool controls.

So, there are, typically, for a connector, two types of actions, non-destructive ones and destructive ones.

For all of them, you have all the sub-actions.

And you can say, "Always ask permission," or, "Do not ask permission."

For the read operations, Claude can access the notes.

It doesn't need my permission to read a note.

But to modify a note or to create a note, it needs to ask me first.

You can say, "Never do these kind of actions," right?

Even, "Do not ask, but..." and, "You are not allowed to do these kind of things."

So you have basically three levels of permissions.


Recap and Your Next Step (23:21)

Damien Schreurs

So, let's recap.

The AI capability matrix has two axes, autonomy level and possible roles.

There are three categories, chatbots, which are low autonomy and many possible roles, assistants, which are low autonomy and a narrow set of roles, and then finally, agents, which have higher autonomy and a wide range of roles.

These are a maturity journey that you move through in the right order.

And regardless of which category you are using, the moment that you give any AI tool access to your data or your apps, you need to apply the risk check.

Know what it will do and make sure that you can undo it.

Now, here's where I want to leave you today.

The matrix tells you what exists, but which category should you be focusing on right now, given where you are in your business and your AI journey?

For that, I built a quiz called Discover Your AI Stage, and it tells you exactly where you stand and what your smartest next move is.

The link is macpreneur.com/aistage, in one word.

And I will put the link in the show notes.


Outro (25:08)

Damien Schreurs

I hope this episode has given you a clear map of the AI landscape and a little more confidence about where you're going next.

And until next time, I'm Damien Schreurs, wishing you a productive day.

Nova AI

(upbeat music) Thank you for listening to The Macpreneur Podcast.

If you've enjoyed the show, please leave a review and share it with a friend right now.