Setting the Stage for Agentic AI: A Practical Framework

Struggling to understand what agentic AI actually means? Wondering how to cut through the hype and start implementing agentic AI that truly works for your business?

In this article, you'll discover a practical framework for understanding and implementing agentic AI, from simple prompting techniques to fully autonomous systems.

Why Agentic AI Matters for Modern Marketers

Agentic AI has emerged as one of the most discussed topics in the business technology landscape, yet it remains one of the most misunderstood. The term has become so ubiquitous that companies across every industry are claiming to offer agentic AI capabilities, but the reality is far more nuanced than the marketing suggests.

​The confusion surrounding agentic AI creates significant challenges for marketers and business owners who want to leverage this technology effectively. Without clear definitions and understanding, it becomes nearly impossible to separate genuinely valuable solutions from rebranded existing tools with inflated price tags.

​”There are a lot of people who are using this term, bandying about this term, and it means different things based on who you're talking to,” explains Christopher S. Penn, chief data scientist for Trust Insights.

​Penn has observed that this lack of clarity creates fertile grounds for what he calls “snake oil salesmen” to enter the market. These vendors take advantage of the confusion by labeling everything as an AI agent, from simple workflows to complex autonomous systems, making it difficult for regular business professionals to understand where to start or what they actually need.

​The stakes are particularly high because agentic AI, when properly implemented, represents a fundamental shift in how businesses can operate. The technology promises to move beyond simple automation into systems that can think, adapt, and make decisions with minimal human intervention.

The Real Benefits of Properly Implemented Agentic AI

When agentic AI is implemented correctly and lives up to its actual potential rather than just its marketing promises, the benefits can be transformative for businesses of all sizes. To understand these benefits, Penn uses a helpful analogy about AI engines.

​When we think about a tool like ChatGPT or models like GPT-4, those are engines. Those are engines that are very powerful. The way we've been using them is sort of like directly pulling on the chains and levers of the engine. When you open up ChatGPT and you're typing in your prompt, you are essentially controlling the engine directly. Many people ask why they can't use ChatGPT to check their email and tell them what emails they need to respond to, or why they can't use it to do various other tasks. We've all come up with our various workarounds, copying and pasting and doing this and that and stuff.

​Agentic AI means taking the engine, which is a model, and building the rest of the car around it. If you think about what's in the rest of the car—the seats, the body, the radio, the seatbelts, all this stuff—those are the connections to different systems that you care about, like your CRM, your marketing automation software, your accounting system, your email system. That gives you the ability to scale because you are no longer the one having to manually type everything in.

​If you've built an agent that's really good at making these connections, the upside is dramatic. First, it frees up enormous amounts of human time that can be redirected toward higher-value activities. Instead of spending hours on repetitive tasks or coordination work, your team can focus on strategy, creativity, relationship-building, and other activities that genuinely require human insight and judgment.

​Second, agentic AI can operate at a scale and speed that humans simply cannot match. A properly configured AI agent could potentially monitor hundreds of data sources simultaneously, identify patterns and opportunities in real-time, and take appropriate actions without delay.

​Third, and perhaps most importantly, agentic AI can bring consistency and reliability to processes that might otherwise vary based on who's performing them or how tired or distracted that person might be on a given day. An AI agent will follow its instructions and parameters consistently every single time.

​However, achieving these benefits requires understanding what agentic AI actually is and how to prepare your organization to use it effectively.

#1: Understand the Three Levels of Agentic AI

To understand agentic AI, it helps to first understand what an agent is in the broadest sense, even outside the context of artificial intelligence. Penn uses a relatable example to illustrate this concept.

​When you take your phone to a place like Social Media Marketing World and connect to the Wi-Fi network, you simply enter the password, and your phone connects. What you don't see is that behind the scenes, there are four different agents on your phone working to get you connected to the Wi-Fi. These agents are negotiating your IP address, handling BGP protocols, and managing various other technical processes that you never have to think about. All you experience is either a successful connection or a failure message. That invisible work happening in the background is what an agent does.

​When people discuss agents in the context of AI, the challenge is that they have conflated a massive spectrum of capabilities. On one end of this spectrum, you have simple workflows that might automate a few basic tasks. On the other end, you have truly hands-off, self-driving applications that can operate independently and make complex decisions without human intervention.

​The problem is that no one knows what you mean when you say AI agent or agentic AI because the term has been stretched to cover this entire spectrum. This ambiguity, Penn argues, is the biggest misconception of all. There is no clear, universally accepted definition, and the meaning varies dramatically based on who you're talking to and what they're trying to sell.

​This confusion mirrors what happened with AI optimization, AEO, GEO, and other trendy acronyms that emerged in the SEO space. 

Penn observed that many practitioners simply took their existing SEO practices, scratched out SEO, wrote AI optimization instead, and added a zero to their price tag. The same pattern is now repeating itself in agentic AI. Companies are taking their custom GPTs that were popular two years ago, rebranding them as agents, and increasing their prices accordingly.

​To effectively prepare for and implement agentic AI in your business, you need to understand that there are distinct levels of sophistication in how these systems can be built and deployed. Penn uses what he calls the product market fit framework to explain this, drawing on familiar meal analogies that help clarify the often confusing landscape of agentic AI offerings.

Done-By-You

This is the AI equivalent of DIY cooking, where you're starting with raw ingredients and you have to do all the work yourself. If you think about cooking, you go to the store, you buy flour, you buy butter, you buy eggs, you buy chocolate chips. You get home, you get your cookware out, and you have to mix it all together and bake it yourself.

​In the AI context, done-by-you represents tools like ChatGPT in their most basic form. There's the blank window, and you have to do all the work—prompting, conversation, uploading data, this, that, and the other thing. This is where most people are today when they're interacting with AI. You're essentially doing everything manually, typing in prompts, copying and pasting information, and managing the entire interaction yourself.

Done-With-You

Penn explains that this is the AI equivalent of meal kits, where some of the work is done-for-you and some of it is not. In the cooking analogy, think about services like HelloFresh or Blue Apron. A box arrives on your doorstep with ice packs. You unpack it and stick stuff in the fridge, and you follow the directions. It's like, just put this container in the oven for forty-five minutes and your meal is done. Or frozen TV dinners at the grocery store. A lot of it's been done-for-you. Some of it's been done-for-you, some of it you still have to do.

​In the AI world, done-with-you includes things like custom GPTs, Claude Projects, and Google's Gemini Gems. These are things where there's some of it that's pre-baked. If you build a custom GPT or a Claude Project or a Google Gem, you have system instructions that you may have built in the done-by-you stage. You have background knowledge, or what Penn calls knowledge blocks, that's pre-baked into this thing, so that you have a mini app inside your AI tool that you can use for that specific task.

​You might have a writing voice custom GPT where you've loaded examples of how to write like you, here's some samples and stuff, and that becomes a Gem in Gemini. Every time you're on the road, you dash off a voice memo, you get back to the office, just put the voice memo into that custom mini app and say, “Make this, clean this up, but it still has to sound like me.” It will go ahead and build that based on the existing pre-baked instructions and the pre-baked examples.

​Penn notes that if you don't have these mini apps built yourself, there's a marketplace inside of OpenAI's ChatGPT where you can find these things as well. However, those mini apps don't really connect to other systems. Part of what you want to do with agentic AI is connect them to other things.

Done-For-You

Penn describes this as when you go to a restaurant. You sit down and you say, “I want a steak,” and they come out and they bring out this thing. As the product market fit framework goes up, you do less work, but you pay more. There's more work being done by somebody else.

​This done-for-you level represents fully autonomous AI systems that handle complex workflows end-to-end with minimal human intervention. These are the systems that truly embody the promise of agentic AI, where you can define an objective and the system figures out how to achieve it, connecting to multiple services, making decisions, and completing tasks without requiring constant supervision.

​Penn emphasizes that understanding these three levels is critical because it helps you set realistic expectations and make informed decisions about which tools and approaches are right for your specific needs and current capabilities.

#2: Use These Essential Prompting Techniques for Any AI

Before diving into more advanced agentic systems, Penn emphasizes that there are three types of prompts that are bottom-line essential to learn how to do for any use of AI, whether you're just using ChatGPT day to day or whether you're building agentic systems.

​The first and perhaps most important technique is to end your prompts with this specific instruction:

Ask me one question at a time until you have enough information to successfully complete the task.

Penn explains that AI systems are tuned on three basic imperatives: be harmless (don't tell people how to do very bad things), be helpful (follow the user's instructions), and be truthful if you can be, which is kind of a crapshoot.

​Helpful is the most important imperative. So if you say to a tool, “Hey, help me write a small business strategy for my company,” it just goes right into action like an overeager intern that's had three cups too many coffee. It's like, “Yes, here I go. Here's your business strategy.” This is exactly what you want. And you're thinking, “No, that's super generic. It's not very helpful. And it's not tailored to me.”

​If you say instead, “Write me a small business strategy for my business. Ask me one question at a time until you have enough information to successfully complete the task,” it forces the overeager model to slow down and say, “So what is your business? Who are your customers? What do you sell?” It will lead you through this process of gathering the information until it thinks it knows enough.

​What it's doing behind the scenes is it's saying, “Okay, I know what a small business is. I know what a strategy is. I generally know what pieces should be there. And this user has not given me any of these things.” So it will go through that, and then you get the information out of you. Penn emphasizes that if you do nothing else, that one technique will double the quality of your AI results immediately.

​The second essential prompt technique happens at the end of a conversation. Penn recommends saying:

Recap the entire conversation as a set of system instructions for the next time using your prompt engineering knowledge.

What that's going to do is consolidate the entire conversation you've had and all those questions and answers.

​Here's the thing, Penn explains: AI is better at prompting than you are, than I am. Every model knows how to prompt itself. So if we have it recap the conversation as a prompt for the next time, we dramatically cut down the amount of time it takes to not have to do that cold start the next time.

​At the very least, you copy and paste this recap into a notebook so that you have it for the next time. But you can also start to refine it as one of the building blocks for an agent.

​The third and final prompt engineering basic is to avoid asking for just one answer. Penn explains that these tools are probabilistic—they work in probabilities. If you say, “What's the best performing social media channel for a consulting business?”, it's going to say LinkedIn. It's going to come up with a high probability answer.

​Instead, say:

Give me three to seven different options for my consulting business for social media channels.

You're automatically forcing it to widen its internal knowledge and come up with different answers based on probabilities. This does two really important things. One, it makes the model expand its field of probability, which tends to generate better answers. And two, it prevents you, the human, from cognitively offloading and offloading decision-making skills to a machine that may not reflect you.

​When it comes back and says, “Hey, Mike, here's seven different options and the pros and cons of each for your strategy,” you're forced to actually think about this as a person and figure out what you should be paying attention to.

​Penn also recommends one additional technique: after the AI has provided multiple options, ask:

What have we forgotten?

This question leverages AI's strength at having encyclopedic macro views of anything because it has seen all the text on the internet. If you're working with your biases and assumptions, and we all have them, you may have forgotten some things. The AI can help fill in those gaps.

#3: Done-With-You AI Tools

Once you understand the three levels of agentic AI and have mastered the essential prompting techniques, you're ready to explore the tools that operate at the done-with-you level. These tools provide some pre-built functionality while still requiring your input and configuration.

​Penn explains that the done-with-you stage builds on the custom GPTs, Claude Projects, and Google Gems that provide mini apps within AI systems. However, the limitation of these tools is that they don't really connect too much to other systems. Part of what you want to do with agentic AI is connect them to other things.

​This is where a whole raft of what are called low-code or no-code AI workflow tools comes in. These platforms allow you to build connections between AI and the other business systems you use without requiring extensive programming knowledge.

Opal From Google

Opal is a workflow designer from Google that, unsurprisingly, talks entirely to the Google ecosystem. You can use it to connect to YouTube and your Google Drive and your Gmail and things like that, and work within the Gemini ecosystem and push data out to Google Docs and so on and so forth. It doesn't integrate with much else than that. But if you're a Google Workspace shop, that's perfectly fine. There's no infrastructure to host. It's kind of a nice workflow system.

​Penn demonstrates how Opal works by showing how you can type into a prompt box something like:

I want to build an app that will search for news articles about AI and construct a short video explaining them.

Within seconds, Opal starts to figure out what you're trying to do. It recognizes that you need an input section so the user has to tell it what the thing is. It determines it has to do some research with deep Google Deep Research, so it's using Gemini Deep Research. It needs to summarize, and it writes the prompts for you. Then it has to generate the explanatory video with Google Veo to create the output.

​In just one prompt, it created a very straightforward workflow. The system output will be a Google video, a Veo video file. And if you had specified, “and upload it to my YouTube channel,” it would add that step as well. Anything in the Google ecosystem that it can talk to, it can build connectors to.

​Penn emphasizes that this is going back to where we started the show. Agentic AI is all about connecting to the other stuff that isn't AI. Opal, as a simple example, is one of those things that says, “Hey, we know you're in the Google ecosystem. If you're willing to stay in that system, here are all the things you can build within it.”

​The service is free and can be found at opal.google. The limitation right now is that it only works within the Google ecosystem, but for internal work, especially if you're using Sheets and Docs and all that kind of stuff, it can be pretty powerful.

Claude Skills

Claude Agent Skills is a relatively new feature inside of Anthropic's Claude. Claude Skills are similar to Projects and similar to custom GPTs, but they take it one step further. Skills in Claude allow you to not just upload documents and have system instructions, but to actually give Claude the ability to perform specific types of tasks with more sophistication.

​The Skills feature is designed to make it easier to build more capable AI assistants that can handle specialized workflows. While this feature is still evolving, it represents Anthropic's push toward making Claude more capable of acting as a true agent rather than just an interactive chatbot.

OpenAI Actions

GPT Actions allows custom GPTs to connect to external services through APIs. This means your custom GPT can pull data from other sources, push information to other systems, and trigger actions outside of the ChatGPT interface.

​Penn notes that Actions can use various models, including the vision model, the speech-to-text model, and various GPT models. It's still in its infancy—it came out about a month ago as of the time of the recording—and it has a long way to go. But for folks who are in that ecosystem, it does exist and represents OpenAI's move toward enabling more connected, agent-like capabilities.

Microsoft Copilot Studio Flow

Microsoft Copilot Studio Flow is a paid add-on to the paid version of Copilot inside Microsoft. Similar to Opal and other tools, it has its own drag-and-drop agent builder. To no surprise, it connects to pretty much everything in the Microsoft ecosystem. So if you use Azure, if you use their classic stuff, if you use Windows and Office 365, Copilot Studio Flow lets you build those same types of drag-and-drop workflows.

#4: A Hybrid Done-With-You and Done-For-You AI Tool

N8N

Penn is particularly enthusiastic about N8N, which he describes as probably the king of the hill in the workflow automation space. It's been around for a while and was actually built in the age before generative AI because it's really good at just connecting things together.

​Penn compares it to Zapier and Make, noting that Zapier, Make, and N8N are all kind of the same class and family of things. What makes N8N interesting is two things. One, it's an open source project. 

There is the ability for you to run it on your own hardware if you want, which, for companies that are very data privacy sensitive, is a fantastic option because it's one less third party you have to be worried about handing your data to. Penn runs it right on his laptop and just keeps it running all the time because it doesn't consume much in the way of resources.

​The second thing, and this is relatively new, is that like Opal, the latest version of N8N now has a prompt window. One of the big obstacles people have faced with N8N is the fact that it is not particularly user-friendly. It's very powerful, but it is slightly more technical. The new version allows you to say, “I want a workflow that does this, this, this, and this,” and you can prompt it over and over again. It can pull out nodes that it needs and connect them all together. You don't have to do all the plumbing work to tie the pieces together.

​Penn notes that this AI-assisted workflow building is only in the cloud version, and mentions that Make also has similar functionality now, where you can prompt it and it will kind of build the basic template for you and allow you to have a head start.

Real-World N8N Example

Penn shares a detailed example of how he and his CEO, Katie Robbert, use N8N with their podcast. One of the things that's really important to them is that they have more representation of women's voices in the AI space, particularly in leadership roles.

​Katie and Penn record their podcast weekly. They get the diarized transcript from Fireflies (though any service that can diarize is fine; diarized means breaking it out by who's speaking). That transcript goes into a folder on Penn's desktop. Then there's an N8N workflow that says: open the folder, read the PDF, feed the PDF to Gemini, have Gemini clean up the transcript, and produce that for use on their blog.

​Then it takes the timestamps from the transcript and says:

Find the thirty- to sixty-second section where Katie says the most insightful thing that our audience would care about. 

They have a copy of their ideal customer profile that it loads and looks through. It produces that finding, and then it produces a little command for the command line on a Mac.

​There's a command-line tool called FFmpeg, which is a free video editor with no interface. N8N writes the command to say:

Here's the original video, just cut out this section, here's the start time, here's the end time, clip this out.

​So it produces for Penn not just the summary, not just the transcript, not just the part that Katie says, but it also does the work for him and clips the video so that he doesn't have to do the video editing. He can take that and load it to their social media scheduler (they use Agorapulse). It saves him so much work. All he has to do is drop the transcript in and hit go.

​This example illustrates the power of connecting multiple systems together with AI to create an end-to-end workflow that dramatically reduces manual work while ensuring that the content being promoted aligns with the business's strategic priorities.

#5: Done-For-You AI Tools

For businesses that are ready to move beyond workflow automation tools and want to explore more sophisticated implementations, Penn discusses several advanced platforms that represent the current state of the art in this space.

Google Vertex AI

Penn describes Google Vertex AI as Google's AI agent ecosystem. It is incredibly powerful—you can build customer service agents that can interact with customers, create crazy automations, and leverage Google's massive AI infrastructure. However, Penn notes that it is about as user-friendly as a porcupine.

​Vertex is designed for developers, not business users. The documentation is written for people who understand concepts like API endpoints, JSON formatting, authentication protocols, and cloud infrastructure. While you can access hundreds of different models and services through Vertex, including Anthropic's models right from inside Google's platform, the learning curve is steep enough that most marketing and business professionals would struggle to use it effectively without significant technical support.

​Vertex requires a Google Cloud account, which is separate from Google Workspace, and comes with its own billing structure, security considerations, and management overhead. Penn's assessment is that while it's powerful and comprehensive, it's built for developers and enterprises, not for marketers who want to implement AI without a computer science degree.

​Microsoft Azure and Amazon AWS have similar enterprise-focused AI platforms, and Penn's view of them is similar to his assessment of Vertex. They're powerful, comprehensive, and absolutely capable of supporting sophisticated agentic AI implementations, but they're built for developers and enterprises.

Open Router

Penn describes Open Router as a fascinating service that connects to approximately four hundred different AI models. The way Open Router works is that you define your priorities—whether that's lowest cost, highest quality, fastest response time, or some balance of these factors. Then Open Router automatically routes your AI requests to the most appropriate model based on your priorities and the nature of each specific request.

​This kind of intelligent routing is, in Penn's view, an example of an AI agent working in the background to optimize your AI usage without requiring you to understand or make decisions about which specific model to use for each task. You just make your request, and Open Router's system determines whether it should go to GPT-4, Claude, one of Google's models, or any of the hundreds of other options available.

​Penn notes that we're even seeing this kind of multi-model routing within single platforms now. GPT-4, for instance, isn't actually a single model. It's multiple models that work together, with the system automatically routing different parts of your request to whichever model is best suited for that particular type of task. The user experiences this as a single, unified interface, but behind the scenes, there's sophisticated orchestration happening to optimize performance, cost, and quality.

Other Notes From This Episode

• Connect with Michael Stelzner @Stelzner on Facebook and @Mike_Stelzner on X.
• Watch this interview and other exclusive content from Social Media Examiner on YouTube.

Where to subscribe: Apple Podcasts | Spotify | YouTube Music | YouTube | Amazon Music | RSS

✋🏽 If you enjoyed this episode of the AI Explored podcast, please head over to Apple Podcasts, leave a rating, write a review, and subscribe.

---

Stay Up-to-Date: Get New Marketing Articles Delivered to You!

Don't miss out on upcoming social media marketing insights and strategies! Sign up to receive notifications when we publish new articles on Social Media Examiner. Our expertly crafted content will help you stay ahead of the curve and drive results for your business. Click the link below to sign up now and receive our annual report!

 

x