So What? How to Build New Microsoft Copilot Capabilities

Katie Robbert – 00:28

Happy Thursday, and welcome to So What?, the Marketing Analytics and Insights live show. I’m joined by Chris and John, and we are going to de-escalate from the pre-show. John, we are all calm now again.

John Wall – 00:45

After yet another fight in the green room.

Katie Robbert – 00:48

Yeah, not with each other—just about the world in general.

John Wall – 00:52

The American healthcare system. How could anyone get angry about that?

Katie Robbert – 00:56

How could anyone get angry? Speaking of how could anyone get angry, let me talk today about how to build new Microsoft Copilot capabilities. I don’t remember the exact date, but we have done a livestream on Copilot in general, which you can catch on our YouTube channel, TrustInsights.ai/youtube. Go to the So What? livestream playlist.

Since we did that episode, Microsoft has made a lot of strides in terms of usability and features for companies using Microsoft Copilot. Just to backtrack a little bit, one of the reasons why a lot of companies are being asked to use—and only use—Microsoft Copilot, and not being given access to things like Claude and ChatGPT and so on and so forth, is because they are locked in with a Microsoft 365 license. Chris covered this in, what, seven different newsletters about enterprise AI.

Katie Robbert – 02:01

Microsoft, for the most part, fits the criteria for a company that is looking to secure things with enterprise AI—not meaning the size of the company, but the amount of regulations that they have to meet in terms of privacy, security, et cetera. That’s why a lot of companies are saying, “Well, look at all these cool tools out there,” but nope, you have to use Microsoft Copilot. Microsoft Copilot is like, “Hey, you have to use us. Here’s some cool new stuff that you can use.” So today, we’re going to cover how to build new Microsoft Copilot capabilities. Chris, did I kind of cover the history of that?

Christopher Penn – 02:42

Yeah, and the reason that this episode came out is because we’ve been doing a ton of private workshops for different companies recently, and 95% of the people in the workshops are on Copilot. So they’re saying, “Hey, I can’t use Claude,” like you said, Katie.

In fact, I did a survey of my own newsletter, and a couple of comments were, “I love the stuff that you’re sharing, but I can’t use it because it’s in Copilot.” As we talk through with folks the things that we’re doing, you can port this to Copilot in some fashion. However, as with so many things that Microsoft does, the documentation is either poor or non-existent. So as part of these workshops, I’ve been exploring what’s under the hood inside Copilot.

Christopher Penn – 03:30

And what’s in there is surprisingly robust, really capable, and not at all documented well.

Katie Robbert – 03:37

And to be fair, not to pick on Microsoft solely, a lot of companies are not great about their technical documentation, or they’re really good about their developer documentation and no documentation exists for the everyday user.

Christopher Penn – 03:53

Exactly. So we’re going to cover today skills, which are all the rage in agentic AI, and where you can put them of sorts in Copilot. We’re going to cover Copilot’s coding environment and the 390 extensions that are in there that you can use for free, that are built in, and do some demos of these things to show that it is in fact possible to make this tool actually somewhat useful. As much as in the past I have mocked Copilot, if you use it properly, it is as capable as any other level one or level two AI system.

Katie Robbert – 04:39

So this is the Chris Penn Redemption Tour. He is apologizing to Copilot for mocking them for so many years.

John Wall – 04:47

No apology. No apology.

Christopher Penn – 04:50

I don’t think Satya Nadella’s feelings are hurt.

Katie Robbert – 04:54

Probably not. John, you talk to a lot of people—our customers, audience, prospects—all the time. How often is Copilot coming up? Is it something when you’re screening prospects and you ask them what they are using or what’s in their tech stack, are you hearing more and more Copilot, or is it still a little bit of a mix of everything?

John Wall – 05:15

It’s interesting. It’s changed a little bit. Definitely at the start, it was what Chris had mentioned—70, 80, 90% of people are on Copilot. They’re always making excuses for it. They’re kind of like, “Oh, we know it doesn’t work as well,” and blah, blah.

So it’s really interesting for me to hear Chris say that you can make it as effective as these other models, because a lot of them have had an inferiority complex. They’re just kind of like, “Well, we put up with Outlook, and we put up with Teams, and now we’re going to put up with Copilot, and we just figure that’s the way the world’s going to work.” To hear that if you set it up the right way, you can actually do more stuff—

John Wall – 05:49

I’m very interested in hearing how this goes. I know we will have a lot of prospects that will be excited to hear this and know that they have the same level of opportunity. At that point, what Microsoft is there for is the fact that it’s enterprise secure. It has everything that the board and your compliance people want to see, which is a big deal because that’s what is keeping some of these companies from trying crazy stuff out.

We’ve also seen an upswing in just the crazy pirates. There are people who have Copilot in the office, and they’re running everything they can run at home, probably getting themselves in trouble.

John Wall – 06:31

So this is great for them, too. Maybe they can stop playing with matches and get back to where they need to be.

Katie Robbert – 06:40

I believe that’s what is called shadow AI, and that’s what a lot of leadership teams are struggling with. One of the things they’re trying to understand is how we can make the tools that we have to use more attractive and enticing so that we can cut down on the amount of shadow AI.

Beyond that, there’s the risk to the company, because the way in which people are using these tools might include sharing sensitive information, identifiable data, IP, and those kinds of things. If you can contain it within your ecosystem in enterprise AI, you have less of that risk. Understanding what you can do with Microsoft Copilot is going to be incredibly useful. To your point, John, a lot of our prospects might find this really helpful and want to learn more.

Katie Robbert – 07:31

So Chris, I am really interested to see how you can cover 390 extensions in, what do we have, 40-ish minutes? You’re going to turn into the Micro Machines man.

Christopher Penn – 07:43

Exactly. Let me not do that. First and foremost, let me point out the obvious things in Copilot itself. We’re talking basic Copilot—not Copilot Premium, not Copilot CoWork, not Copilot Studio, not the 82 versions of Copilot that Microsoft markets because they have no product naming capabilities whatsoever. We’re talking basic, plain vanilla, out-of-the-box Copilot. This is what you get with Microsoft 365.

Depending on which version you have, the most important menu is in either the top left or top right: the model chooser. You will note that there are four models that you can choose. There’s Quick Response, which is great if you’re providing all the data. There’s Think Deeper, which basically uses a reasoning version of Microsoft’s Phi model. Microsoft very strongly encourages people to use these two models because it costs them less money.

Christopher Penn – 08:33

There is also a connector to OpenAI’s models, and you see the brand new GPT-5.6—I’m pretty sure that’s GPT-5.6 Tera, which is the mid-level model—and then GPT-5.5 Quick Response. In general, unless you have a good reason to use something else, you should opt to use the latest model from OpenAI. That is the best model to use in Copilot. So that’s immediate upgrade number one.

Immediate upgrade number two: remember that Copilot has what are called agents, which are not agents—they’re GPTs. If you’ve ever worked with ChatGPT and seen GPTs, or Gemini Gems, agents are effectively that. This is really important because, just like in ChatGPT, if you use the @ sign, you can call an agent inside a Copilot chat and reference it.

Christopher Penn – 09:27

You can say, “Let me write a blog post,” and then call your agent if you have a blog-writing agent there. Here’s why this matters: in the world of Claude, ChatGPT, and agentic AI, there are so many people creating so many really cool skills, agents, and plugins. I’m going to pull up one of my personal favorites, Jesse Vincent’s Superpowers plugin. This is for coding environments—a phenomenal plugin, so intelligent, with so many cool capabilities. In particular, within this plugin, there is a skill called brainstorming. All this is—if I tap on the raw button here—it’s just a prompt. At the end of the day, it’s just a prompt.

Christopher Penn – 10:24

So if I copy and paste this—this is free and he has licensed it for commercial use, so very important: please obey people’s intellectual property licenses. But if I copy this prompt, I go to Copilot, tap “New Agent”, skip their stupid builder, and call this “Brainstorming”. Jesse Vincent’s Brainstorming Skill. Paste. Oh, you’ve exceeded the word limit by 10,000 characters. That’s fine. Paste it in a text document called brainstorming.txt. Now drag that file, which is just a plain text file: “Execute the brainstorming skill specified in brainstorming.txt,” which we can see is attached here. Let’s make sure we hit create.

Katie Robbert – 11:29

And to be clear, you skipped the agent builder because you already had the instructions and the prompt for what you were building. But for someone who doesn’t, they could go through the agent builder step-by-step and say, “Here’s what I’m trying to do. Can you help me build it?”

Christopher Penn – 11:47

You could.

John Wall – 11:47

Exactly.

Christopher Penn – 11:48

But the nice thing about all of these skills that people are publishing—some of which you can buy from Trust Insights, for example—is I can say, “Let me do some brainstorming @brainstorming.” Because I’ve taken that skill and basically pasted it into a Copilot agent, I have ported this thing from Jesse’s Claude version right into Copilot. This is exactly the workflow that happens. So if you see or hear someone on LinkedIn saying, “I’ve got the ultimate skill package,” or “I have the ultimate Claude code thing,” grab it, open up the hood, find the skill.md file inside it, and copy-paste. Now you have it in Copilot just like that. It is so straightforward.

Christopher Penn – 12:38

You can add in any number of these things, and until you @ them in a conversation, they don’t take up space or memory. So one of the easiest things you can do is take all the stuff people are talking about that has been pre-built in other systems and bring it into Copilot as little agents.

Katie Robbert – 12:59

That’s one of the things we talk about a lot when we talk about generative AI: the better your documentation, the easier it’s going to be to migrate from system to system. Skills are one of those things that should work on any given large language model, provided they are well and thoroughly documented.

Christopher Penn – 13:20

So that is—I guess if you want to use bro-ligarchy vocabulary—that’s unlock number two. Number one: choose the right model.

Katie Robbert – 13:31

Let’s not use bro-ligarchy terminology. I don’t have my red flag handy. I only have one hand today. Hold on, I’ll see if I can get it. Oh, okay.

John Wall – 13:45

There you go.

Katie Robbert – 13:46

It’s a red flag for Chris for saying “bro-ligarchy.”

Christopher Penn – 13:53

I don’t think Satya Nadella’s feelings are hurt.

Christopher Penn – 13:53

So, number one: use smart models unless you have a reason not to. Number two: any skill or plugin that you see floating around that is appropriately licensed, you can copy and paste into an agent in Copilot and have it available. Depending on how your administrator sets things up, you could share it with your team. You can share Copilot agents right inside Copilot with the rest of your team, so if you see some cool stuff, share it around.

I believe—and I don’t know because we don’t have SharePoint—that you can put links to your shared agents in SharePoint so that if you find something cool, you can circulate it with others. So that’s number two. Number three…

Christopher Penn – 14:32

And the big one is Copilot can natively run the Python programming language. Right within a chat, it can run Python code—it fires up a little Python interpreter. That alone is pretty cool because that gives you deterministic capabilities: things where there’s a clear right and wrong answer inside of a language model, which otherwise struggles with things like math and counting that they’re really bad at. For example: “Write a 300-word blog post, plus or minus 10 words, about the importance of email marketing for B2B marketers.” Let me see what comes out. Here we have a lovely post. How many words is that? It actually is writing its own little bit: “Contains 323 words.” Yeah, good job. Thanks.

Christopher Penn – 15:46

So it came up with—and I’m curious to see if it actually is 323 words—minus the title, it’s 318. I guess we’ll give it that. But you can see it fired up a little Python script to do exactly that. Does it do that count? Yes.

Katie Robbert – 16:14

So it counted 323, probably including the title. You said minus the title it’s 318, so that’s five. If you can scroll back up for a second, the title is more than five words. I want to call that out to say that I would not be like, “Okay, I’ll give it to you.” It’s already showing inaccuracy. I want to make sure we’re clear: don’t accept things from AI that look off, especially when you can prove that they are wrong. That’s just my little disclaimer. We’re not just going to say it’s fine, because that could have ripple effects further down the road. Just a little bit of a callout.

Christopher Penn – 17:01

Exactly. We’re going to actually prevent that, and here’s how we’re going to prevent that. First, we need to know what we can do in this environment. So we’re going to say: “Make a list of all the Python libraries and packages in this environment using your built-in Python interpreter. Return your list with the package name and version number organized by topic or area, such as statistics, natural language, file copying, etc., as a markdown file for download.” What we’re doing here is saying to Copilot, “You have the ability to export all the Python libraries that are installed for every user of Copilot in the background.”

Christopher Penn – 17:44

What will come out, depending on the version of Copilot you have, is 300-ish different Python libraries and packages across 10 or 15 different areas. Let me see what it comes up with. Appropriately, it’s writing and calling its own function for this particular task. Let me see what its answer is, and if it blows up—which it very well could—I have a pre-baked version.

Katie Robbert – 18:23

I was going to say, John, we haven’t practiced our banter for this week.

John Wall – 18:30

How about that World Cup?

Christopher Penn – 18:35

In the Copilot environment, because it’s taking its sweet time writing that code, there are 390 packages. Let me make this a little bit bigger so that we can see what we’re doing here. There are AI and machine learning libraries like Gensim and the Hugging Face Hub to import models, scikit-learn, SHAP, tiktoken for tokenization, and XGBoost.

There’s audio and speech for its ability to listen to files that you upload, like MP3s, or read or view videos, or do speech recognition. There’s cloud stuff, of course, for Azure. There’s computer vision and image recognition. There’s data science and statistics—a ton of this: SciPy, scikit-learn, etc. There’s database access, developer tools, Office document handlers for pretty much every kind of document you can imagine, geospatial stuff…

Christopher Penn – 19:26

You can make games with some of these packages, a list of general utilities, and so on and so forth. This is a huge amount of stuff that is in Copilot right now. It’s there, waiting to be used if we tell Microsoft that we want to use it. That is the hard part, and here’s how we would do it. Let’s start a new chat, and we’ll have it use the smartest model available, which is GPT-5.6. We’re going to say—and I’m going to drop in my list of packages: “You are a Python coding expert. Using the included Python environmental packages as appropriate, construct a single monolithic Python script that, when given a standard input, will perform an accurate word count and return the word count to the console or standard output as plain text.”

Christopher Penn – 20:27

The goal is to have a quick, easy, fast, and accurate utility for Copilot to use to accurately count words. What we’re asking Copilot to do is write its own code. We don’t have to code anything at all—hands are off the keyboard, no coding involved. But it is writing a Python script to do word counts. When it’s done, it should provide a nice little download, which will be—in case it decides to take a vacation—here it comes: download word_count.py. You can also redirect the file. Here it is. Very nice. I like it. Thank you.

Now, what do we do with this sucker? In a new chat, I can include the word count file and say, “Write a 300-word blog post about the importance of email marketing.”

Christopher Penn – 21:31

In B2B marketing, use the included Python script to ensure a correct word count, plus or minus 10 words. After you’ve written the post, count the number of words.

What we’ve done is we’ve essentially built a utility that uses Copilot’s native Python environment to do a deterministic task that language models, frankly, are not up to scratch for doing. Now, by the way, this does not solve the fact that this is the most AI-slop post you can possibly imagine. But let me verify: did it actually do the job? The answer is—and this time, Katie, without the title—it is exactly 300 words. I checked in my text editor here, so without the title it is exactly 300 words. Using the utility got us the correct answer.

Christopher Penn – 22:35

So we’ve worked around AI’s issue of “Hey, this thing is as dumb as a bag of hammers and can’t count” by forcing it instead to use the programming languages available to it.

Katie Robbert – 22:51

As a side note, I think one thing that’s interesting is you didn’t specify that the word count should or should not include the title. As people are learning how to use these things, does that kind of specificity matter, or do you think it made the assumption, “Well, they said the post, so only the content, not the title”? Walk me through a little bit of how that works.

Christopher Penn – 23:20

Depending on the model, it may wing it. It will be probabilistic unless you specifically say, “The post minus the title,” or “The post without the title must be 300 words, plus or minus 10 words.” You want to be very clear about that.

What I think is useful and interesting in this particular case is if you look at this catalog of all these different things—the 390 libraries available in here—this should tell us what we can do inside Copilot. At the top there, XGBoost 3.12. XGBoost stands for Extreme Gradient Boosting. This is an old-school machine learning algorithm that is commonly used for regression analysis. When you see that in the list, you go, “Oh, Copilot can do regression analysis if I tell it to use this.”

Christopher Penn – 24:19

So I might export a file out of my Google Analytics and say, “I want to know what really drives leads, and I’m going to provide you with these data files, these CSVs. Using your built-in Python interpreter and the XGBoost library, construct a piece of code that will perform this assessment. Return your R-squared error rate, and if the R-squared error rate is below 0.05%, then we know the results work.” Then go and do the thing, and it will go and do the thing. Suddenly, a tool that can’t even count the number of words can do regression analysis.

Katie Robbert – 25:02

I recognized XGBoost because you’ve been talking about XGBoost for as long as I’ve known you, but a lot of these things I don’t recognize. Could it be as simple as taking them section-by-section and saying, “It looks like you have the following seven audio and speech libraries. Can you tell me what each one of them does?” You build your own glossary that way so you can say, “Today I need to do a regression analysis. Which one of the tools you have can I use?”

Christopher Penn – 25:37

Exactly. Yes, you can do that, you should do that, and you should do that anyway just to know what’s out there. For example, in audio and speech, Librosa is a music library—it can read and write music. If you were trying to compose melodies, an AI tool maybe can or can’t do that, but you could say, “Validate it. Fact-check me on this using Librosa in your built-in Python interpreter.”

If we scroll down a bit further to computer vision: if you upload an image and it hallucinates, you say, “No, use your OpenCV Python Headless library to look at the image instead of winging it, and then tell me what you see in this image.”

Katie Robbert – 26:27

Makes sense. In general, you just don’t want AI to wing it. The more information like this that you have, the more specific you can get—not just with your prompt, but with these Python—what do you call them, Python packages?

Christopher Penn – 26:43

Packages. Yeah, packages or libraries, either one.

Katie Robbert – 26:46

Okay. With these libraries, you can get down to the specifics. XGBoost is a great example. A lot of people want to understand attribution analysis. That’s not a new topic, we just have new tools to play with. I’m assuming unless you say something like, “Use the XGBoost Python package,” it may or may not decide to use that. So you are just kind of winging it: “Hey, do a regression analysis on my Google Analytics data,” and it’s going to confidently give you the answer whether or not it’s correct.

Christopher Penn – 27:21

That’s right. That’s right. Let me look at another example, and we’ll use another writing example for this. Suppose we wanted to not write in passive voice. Passive voice is kind of the enemy of clarity. I will take my Python environmental packages list here—which is just that markdown list we made at the beginning of the episode—and I will say: “You are an NLP expert. Using the included Python environmental packages as appropriate, construct a single monolithic Python script that, when given a standard input, will perform a statistical assessment of passive voice in text provided in standard input. Choose the appropriate natural language processing packages that best fit this task, and the script should provide a report to the console or standard output in markdown format of the number of occurrences of passive voice and a bulleted list of the specific passages listed.”

Christopher Penn – 28:17

“Make the Python script available as a download.” Now, instead of just saying, “Hey, count some words,” we’re kicking this up to the level of using natural language processing tools. If I had to guess, it’s probably going to choose spaCy or NLTK.

It chose spaCy, and it says, “Here’s what we’ve got.” Here’s an example, and it can count the number of passive voice occurrences. This script is something I can download and run somewhere else. The fact that it uses spaCy inside its system means that I don’t have to worry about it trying to install anything, because it’s already there—Microsoft has provided it right inside the environment. So I’m in good condition here if I call this passive.py.

Christopher Penn – 29:19

Now let’s go start a new blog post and drop this in. You say: “Write a 300-word blog post about the importance of email marketing to B2B marketers, and ensure that you write in active voice only. Using the included passive voice Python script, revise and rewrite until you have zero occurrences of passive voice.” For this one, because it requires multiple turns, I’m going to use the GPT-5.6 model because it needs a model that has agentic capabilities to say, “Okay, this is going to be multiple steps. I’m going to have to take a few tries at this.” We can even see here as it’s doing its thinking: so it’s coding, it’s executing, it’s running the script.

Katie Robbert – 30:26

As a user, that’s something I would be interested in understanding: how many versions did it go through? I want to see the first version with the passive voice to see how it made adjustments—just out of my own curiosity—or did it take the instruction and write without passive voice the first time?

Christopher Penn – 30:49

Yeah. So let’s see: it reviewed the data, it did some coding execution. It says, “B2B marketers need channels that build trust and reach clear audiences to support long sales cycles.” It did an adjustment on the word count—it made some replacements there on the word count. So this is interesting: it started out in active voice, so the script didn’t find any passive voice. Which again, for GPT-5.6 as a model, makes sense, because that is literally one of the smartest models on the planet right now. It probably should be able to do it right the first time, and then the code is there as a backstop.

Katie Robbert – 31:42

Question: one of the things that you can do in Claude that we talked about with a client last week is set global settings. Is that something that you can do in Microsoft Copilot? For example, could you give it the prompt or rule to say, “Anytime you draft content, never use passive voice,” or “Always use active voice,” or something along those lines? That’s one of the nice features in other systems: you can set those global settings so if you forget to mention it in the context of the prompt, it’s already set.

Christopher Penn – 32:21

Yes, under personalization there are custom instructions, a work profile, and you can also save memory. For example, in my custom instructions in this system, I have it set so it’s required to use ASD Simplified Technical English as its chat language.

Simplified Technical English was created by the European airline industry to standardize how pilots speak on the radio. It requires very specific formats of the English language: you’re not allowed to have run-on sentences, there’s one instruction per sentence, one topic per paragraph, et cetera. Instead of a pilot just rambling on the radio like AI, it says you have to say, “Delta 2604, Runway 22.” It’s just very clear.

Christopher Penn – 33:14

I have this set as the custom instructions in this environment because I don’t want it foaming at the mouth and waxing rhapsodic about this—just give me answers. So yes, to your point about custom instructions, that’s what’s in here.

Katie Robbert – 34:23

It’s a nice feature.

Christopher Penn – 34:26

Yeah, it’s a nice feature. The last thing you could do, if you really liked a script that worked well: you cannot upload code to these agents. If I try and drop in that Python script, you can see the screen turns red—it will not allow you to do that.

However, what you can do is change the file extension on the Python script from .py for Python to .txt and then drop it in anyway. You could say: “We are intending to reduce the use of passive voice. Use passive.py and execute it in your built-in Python interpreter to help the user eliminate passive voice.” Paste that in there. Now it’s got the Python script as a plain text file, which it’s going to have to read anyway.

Christopher Penn – 35:30

If I create my agent now, I start a chat and I could say, “Take the following blog post and make sure it’s in active voice.” Oh, I’m already in the active voice assistant, so that’s silly. Let me start a new chat there. I could call the “Active Voice” agent, and instead of having to drag and drop the Python file every time, now I just call it like a skill.

Katie Robbert – 36:01

I like it. This is where you’re getting into the Microsoft ecosystem: could these things be saved on a SharePoint page? How do you get people the latest and greatest? That all comes down to governance, which is going to be unique to every organization. We make recommendations, but you really have to make sure that it works well for how your team works, whether they go to SharePoint or if you want something that’s already in the libraries of your Copilot instance.

Christopher Penn – 36:33

Exactly. Those are the things that I would recommend that folks do: Number one: use the smartest model that you can, whatever is in your subscription. Number two: take any skill that you like from outside the Copilot ecosystem—because they’re all just prompts—and turn them into agents within Copilot. Now you have access to the skills. Yes, there’s some copy-pasting, but that’s fine.

Number three: use the built-in Python interpreter with the 390 packages, which is more than you get in other places. I did the same exact thing in Gemini, and Gemini only gives you 200-some-odd ones. Granted, a lot of them are Azure-specific for Microsoft, but there’s still more in the Copilot ecosystem than you get in stock Gemini, which I thought was kind of interesting.

Christopher Penn – 37:17

Number four: if you build useful code-based extensions or new capabilities with Copilot, turn those into agents as well by renaming the Python script as a text file. Then it can execute them, and you now have the ability to do a lot more with Python built into Copilot than standing there staring at a blank Copilot window going, “Why can’t this thing count words?”

John Wall – 37:44

Is there any limit to that then, really? Can you just put all kinds of Python in there and have it completely running wild, or is there some security around it?

Christopher Penn – 37:53

There is. It runs in a virtual sandbox, you can’t install extra stuff, and you can only connect to things that your Copilot administrator allows you to connect to. So if there’s no SQL Server in your environment, you can’t connect to one.

Katie Robbert – 38:09

Got it. That makes sense, because one of the things we know about Claude, for example, is that it has a lot of connectors to third-party tools. I would imagine that Microsoft Copilot—again, one of the reasons why it is very attractive to an enterprise AI organization—restricts what it can connect to, and out of the box it’s likely just the stuff in its own ecosystem, not third-party tools.

Christopher Penn – 38:34

Yes. Out of the box, it will connect to SharePoint and Microsoft Office—so any Office documents that are in OneDrive, SharePoint, Word, Excel, PowerPoint, and Outlook, I believe. If you upgrade to Premium, you can attach to more connectors. If you upgrade from Premium to Copilot Studio with Premium, you can access their entire n8n-style workflow builder that has like 500 connectors.

Katie Robbert – 39:05

Yeah, Microsoft has always been pretty bad about their product naming, but that’s neither here nor there. Let me ask you this question, Chris—and this might be putting you on the spot a little bit: is there anything that Copilot can do that the other, more commercial large language models can’t do? If someone had to say, “At least I have this one thing that Copilot can do that the others can’t.”

Christopher Penn – 39:30

Does that thing exist in stock Copilot? No. Stock Copilot is basically ChatGPT, just slimmed down and sandboxed in the Microsoft ecosystem. Where you do get some interesting stuff is when you have Copilot Studio as part of Premium—that does have a lot of interesting connectors, many of which you don’t see in n8n. Granted, a lot of them are enterprise connectors, and that is kind of interesting.

Copilot CoWork, which is a licensed version of Claude CoWork, obviously has all the connectors for Copilot Studio within it as well. If you pay for all the upgrades and you’re at Copilot Platinum Ultra or whatever, then you do have a lot of capabilities. But also it costs you like $5,000 just to open a Word doc.

Katie Robbert – 40:22

I feel like that’s true of any software system.

Christopher Penn – 40:25

Yeah, but Copilot’s prices are really high. In fact, if you go to the Trust Insights website under our Instant Insights, there is a Copilot CoWork calculator that we put up for free that you can check out just to estimate your costs—and boy, are they high.

Katie Robbert – 40:41

Well, I got nothing on that. We have zero control over that.

John Wall – 40:47

That home office is not going to pay for itself in Seattle.

Christopher Penn – 40:50

Exactly. I hope that what we did today shows you that you can extend the capabilities of Copilot beyond what’s in the box—that you can take existing skills, port them to agents, and write your own with the coding environment to do anything that Python plus those 390 libraries support. There’s a lot in there.

As Katie said, it’s a good idea to Google them to learn what they do or have Copilot explain them to you. If you have that list, perhaps even start some of your prompts by saying, “I’m doing this task. As a reminder, you have access to these 390 things,” and drop in the file. “Which of these things would be an appropriate addition to this task?”

Christopher Penn – 41:35

Then, of course, use all the rest of your prompting basics that we teach in our Generative AI for Marketers course, like, “Ask me questions: do you have enough information to successfully complete the task?” Copilot will do a capable job, you will get things done with Copilot, and sometimes get reusable code that you can use for other projects.

Katie Robbert – 41:55

And you can learn about those courses at academy.trustinsights.ai. Chris, is it safe to assume that you will drop the link to that Instant Insights Copilot calculator that you mentioned in our free Slack community, Analytics for Marketers?

Christopher Penn – 42:08

Yes, that’s right.

Katie Robbert – 42:10

Amazing. All right, John, any more questions about Microsoft Copilot?

John Wall – 42:16

No, but I think I’m ready to start throwing Python files in there just by changing the extension. That sounds like the kind of exploit that I’m looking for.

Katie Robbert – 42:33

Carry on.

Christopher Penn – 42:36

That’s it. That’s going to do it for this week, folks. Thanks for tuning in, and talk to 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/tipodcast and our weekly email newsletter at TrustInsights.ai/newsletter. Got questions about what you saw in today’s episode? Join our free Analytics for Marketers Slack group at TrustInsights.ai/analyticsformarketers. See you next time.