Custom GPTs: Your New AI Colleague!

Hello, this is Janette Roush, and welcome to another one of our Brand USA Agents of Change sessions. This particular session is on Custom GPTs. A year ago, November 2023, on DevDay, OpenAI announced that with a paid OpenAI account, you could now create these things called Custom GPTs, and they introduced a GPT store where you could have these GPTs published for the wider world to access and use. And for many of us, it wasn't clear what those were, what should they be used for.

The power of a Custom GPT is that it allows you to make instructions and knowledge a reusable prompt so that you can go back to the same exact set of instructions over and over again instead of having to paste instructions into a new chat every single time that you want to do a different task. And so the way I like to think about a Custom GPT is it's a very narrow, specific, single-use prompt. But because it's saved as a Custom GPT, you can come back to it every time that you need to do that single-use task.

So this would be a single task like helping you interpret the specific guidelines that your accounting department has published. Or maybe it's the specific ways that your communication team wants to do press releases. These are both individual tasks, each of which would get its own Custom GPT, and then you can use those Custom GPTs over and over and over again.

Here's one that we made here at Brand USA. It's our travel and expense policy. If you're not familiar with that particular corporate policy, what it is, is when you're traveling as part of your job, or you're incurring expenses that your company is going to pay you back for, you can ask questions of this Custom GPT in order to understand, "Can my company pay for this? Or, is this per diem? What is the per diem rate for the particular city that I'm going to?"

And so what we've done is we've uploaded the most up-to-date version of Brand USA's travel and expense policy, and then any new hires or anyone who has a question, instead of having to read through 40 pages of this particular policy, they can just ask the Custom GPT.

So let me show you an example. Let's say that I'm about to book a flight to Orlando to meet some clients and do sales meetings at Destination International, so I'm going to stay for three nights. And so I might ask the Custom GPT, I'm traveling to Orlando for business. I'm going to be there for three nights. What is the per diem rate, and can you break that down for me by lodging, meals and incidentals? And as soon as I hit enter, this Custom GPT is going to look at the instructions that have been given to it—instructions that are not visible in this chat window—and it will also look into the knowledge file, which is the PDF that we uploaded, and it's going to give us a very specific answer.

So here it goes. The per diem rate for Orlando is $179 per day, and then it gives us that breakdown. $112 for lodging, $59 for meals and incidentals, and $8 for incidental expenses. And the reason that this is useful is not because you couldn't just Google it—you could—but because it allows you to follow up with any other questions that you might have that are specific to Brand USA's corporate policies, not just the per diem.

So one of the requirements in order to get to expense your trip is that you've got to be traveling at least 51 miles away from your primary work site. And so you could ask, what is that minimum distance? And it's going to be able to provide that for you. Or you could say, "I have a question about alcohol. Can I put alcohol on my company credit card?" And the Custom GPT can give you a response that is specific to Brand USA's policy around alcohol and whether or not that can be expensed or whether it should be paid for by the employee and not submitted for reimbursement.

So this is just one example of where a Custom GPT can be really, really useful. Now when you go into a Custom GPT, you can edit it. So for this particular Custom GPT, I built this one. If you are the person who built it, or if someone shares the edit link with you, then you can go in and make changes to it. So if I hit edit, it's going to pop open this little flyout window that shows all of the settings that have been configured for this Custom GPT. And there are really two main sections to a Custom GPT.

One is the "Configure" section, where you can set up the name, the description. You can upload a picture or generate a picture. And then down here is the "instructions" section. The instructions are the prompt. This is the set of rules that ChatGPT is going to follow every single time you ask the Custom GPT a question. And then there's a section here called "Knowledge," and the knowledge section is where you are uploading the source of truth.

So in this case, the knowledge that has been provided is the PDF of the Brand USA travel and expense policy, and this right here, this is the internal source of truth. This is the document that was created by our finance team, and so whenever I ask a question of this Custom GPT, it is going to look to that PDF as its source of truth and respond based on this PDF, not based on generally what it thinks that a travel expense policy might look like.

Now one of the really important things to understand is that these models are non-deterministic. They can give you a correct answer for the very first time, and then if you ask the same question three months later, it gives you a wrong answer, because something in the model shifted, something in the way that the prompt was written was slightly different.

And so because these are not like databases where you would ask a query and you're always going to get back the same exact return—because these are non-deterministic models, you've got to write really, really solid prompts for your Custom GPTs and you've got to test them out yourself and have other people beta test them as well. So I'll give you an example. This Custom GPT was one that I created over a year ago, and I actually used it. I traveled to Chicago, I asked this Custom GPT what the per diem rate was for Chicago, I put it on my expense report, and then my expense report came back and said, "This is wrong, this is not the correct rate."

And so then I went back into the Custom GPT, and I asked it the question, why did you give me the wrong per diem rate for Chicago? And I found out that the per diem rate that you'll use is dependent on the county. And so Chicago, the city of Chicago, is in Cook County. And depending on what county you're in, you're going to have a different per diem rate. But the city limits of Chicago extend into DuPage County, and so if you're staying in a hotel that happens to be in DuPage County, you have a different per diem rate. And what this Custom GPT was doing is it was defaulting to the DuPage County rate, not the Cook County rate for the city of Chicago.

So I had to go back to the instructions and I had to rewrite the instructions so that they were really clear. You're going to be given a city. When you are given that city, find the county for that city. Look at the PDF and find the rate that corresponds to that county. And so then once I fixed the instructions, which is this right here—the instructions are the prompt, and so if you've gotten really good at prompting ChatGPT, that exact same skill set is what you use in order to create better Custom GPTs. Now when I ask it what's the per diem rate for Chicago, it returns the correct county and the correct per diem rate.

And so that's one of the things that's really important to understand is that the quality of your Custom GPT is determined by two things: the quality of the knowledge that you upload and the quality of the instructions that you write, which is the quality of the prompting. And both of those things, human beings are responsible for. You've got to upload good sources of truth, and you've got to write really good prompts.

And then the very last part of the Custom GPT configuration is there's a couple of different models that you can choose from. You can decide which is the model that's being used in the back end. So right now it's defaulted to a newer model that came out just in the last month called O1-preview. O1-preview is a reasoning model, which means that it spends a lot more time thinking through the problem, and that's really helpful for writing code or solving complex math problems. But for more conversational, general, faster use cases, I often will default back to GPT-4o, which is their older—I say old, but it's not that old—but it's their more conversational model versus the reasoning model.

And then you can toggle on or off different capabilities. Does this Custom GPT have the ability to search the web in order to help respond to questions? Does it have the ability to generate images using DALL-E? Or does it have the ability to run code, Python code, in order to analyze data that you give it? In this case, we've turned those off because we're really only going to be asking it questions about per diem policy. We don't need it to search the web, we don't need it to generate images, and we don't need it to write code. But in other Custom GPTs, you might turn those different abilities on.

Now I want to show you how you would build a Custom GPT from scratch, so we're going to start over. You close out of that one. And when you're in ChatGPT and you've got a paid account, on the left-hand side, there's going to be a section that says "Explore GPTs," and when you select that, it's going to take you to what they call the GPT store. And inside the GPT store, you can find Custom GPTs that other people have published that do all kinds of different things.

You can select from some of the top GPTs or the trending ones. For example, there's one called Canva that helps you create graphics and presentations using Canva. There's one called Consensus that searches through 200 million academic papers in order to help you do research. And then there is this great one. This is a really fun one called DALL-E, which is just a user-friendly interface for generating images. So you click into any of these in order to actually use them. And then to see all of them, you can select "view all" in order to see what other Custom GPTs people have published.

Now, when you're going to create your own Custom GPT—so let's say that we're in communications or we're on a PR team, and we want to make sure that any time we're writing a press release, we're doing it according to the communications department's guidelines and best practices. Then the way that you would do that is inside the Explore GPTs page, there's a button up here that says "create a GPT," and you select that in order to build your own Custom GPT.

And so here we go. Now the GPT builder is going to pop open. This is using ChatGPT to help you build this Custom GPT. So you're going to have a conversation with ChatGPT, and it's going to ask you, "What would you like to make?" Now you can either tell it what you want to make in the most simple terms or you can give it a lot of detail, kind of like you would in a really good prompt. And so I'm going to tell it, "I'm building a tool to help Brand USA employees draft press releases. It should follow AP style and Brand USA communication style guide. Ask me what the press release is about, and then interview me in order to gather the relevant details for that press release."

So I hit enter. And what ChatGPT is going to do is it's going to take what I've given it, and it's going to draft a better, more detailed version of what I've asked for. So it says, "Got it. I'll create a press release assistant. Here's a preview of the GPT's name and profile. We can always adjust these details." And so it's given it a name called "Brand USA Press Release Assistant." I might change that and just call it "Press Release Drafter." And then it's given it a profile picture. I don't love this profile picture. I might change that so that it was the Brand USA logo or something like that. But for right now, we're just going to leave those.

And then it's given it really detailed instructions, which is the prompt. And so this is the exact same process that I would use to help me write the instructions. I would ask ChatGPT to help me write the instructions instead of having to just make them up on my own. It's good at prompting, so I want to leverage ChatGPT to help me write better prompts.

And so this is what it came back with. It says, "You are a press release assistant for Brand USA employees. Your primary goal is to help draft press releases that follow AP style and Brand USA's communication style guide. Start every conversation by asking the user about the topic of the press release. Then, conduct an interview to gather relevant details, such as key messages, quotes, target audience, event details, and any supporting information." And it goes on. It's got about 300 words' worth of instructions for this Custom GPT. I'm going to give it a thumbs up because I like these instructions.

And now I could test out this Custom GPT before I fully publish it or share it with anyone. So I'm going to ask it, "I want to write a press release." So I'm already testing it out. And it says, "Great! Let's get started. What is the topic or subject of your press release?" So already it's interviewing me. "I want to announce the Agents of Change program." And so then it says, "Perfect! Agents of Change sounds like an impactful initiative. To help craft an engaging press release, I'd like to gather a bit more detail. Could you share the following?" And then it asks me a number of questions. What are the key messages? Who's the target audience?

Now if I wanted to, I could respond to all of these questions conversationally, and it would draft a press release for me. But what I really want to do is I want to upload the communications style guide so that it's not just following AP style—which it already knows, because its training data includes AP style—but now it's also following Brand USA's communications style.

So in order to do that, I need to jump over into the "Configure" tab, and I'm going to scroll all the way down, and I'm going to upload a file. So when I select "upload files," I'm going to find Brand USA's communication style guide, which is a PDF. And I'm going to upload that file. And once I've uploaded the file, that's been added as the knowledge, and then I'm going to go back in and continue on that preview.

And so now that I've added in that communication style guide, I can test this again. So I'm going to ask it, "I want to write a press release." "Great! Let's get started. What is the topic or subject of your press release?" "I want to announce the Agents of Change program." Perfect! Now, when it gives me this draft, it's going to look at that uploaded communication style guide and compare it to what I've given it and make sure that the press release that it drafts for me uses the specific rules and requirements for a Brand USA press release.

So I'm going to paste in a bunch of text. This is a description of what the Agents of Change program is. I'm going to paste that in. And I'm just going to tell it, "Please go ahead and draft the press release using this information." So now it's going to take a minute. It's looking at the instructions. It's looking at the knowledge file. And it's going to draft a press release for me. And then once it gives me that press release, I can read through the press release and either ask it to make changes or I can decide that it's good to go, I'm going to copy it, and then I can paste it wherever I need to use it.

And the benefit of having this as a Custom GPT is that this is a reusable prompt. I can come back to this same set of instructions and this same uploaded file over and over again, every single time that I need to write a press release, instead of having to upload the communications guide every single time and having to re-paste my instructions every single time I need to write a press release.

Now, once you're comfortable with the Custom GPT, the very last step is that you're going to save it and you're going to select who has access to it. So if I hit the "create" button up here at the top right, this is going to finalize the Custom GPT and allow me to select who can access it. You can choose "only me," which means that only you have access to it. You can choose "anyone with a link," which means that as long as somebody has a ChatGPT account, they can use this Custom GPT if you share the link with them. And then the one that I use the most often, which is "workspace," and what this means is that everyone who has a ChatGPT Team account on Brand USA's team, they're going to be able to see this Custom GPT and use it themselves.

So I'm going to select the "workspace" option because I want everyone at Brand USA who's writing press releases to follow the same communications guide. And then I'm going to hit "save," and now I've added it to the GPT store—not the public GPT store, but my internal, just Brand USA GPT store. And anytime anyone asks me, "Hey, I need to write a press release," I could just say, "Go to the Explore GPTs page inside of ChatGPT. Click on the Brand USA press release drafter, and it will walk you through all of the questions that you need to answer and draft the press release for you."

Now I want to just quickly go through a number of other examples where building Custom GPTs can be really useful for a wide variety of different departments. The one that we've been using at Brand USA is one for our meetings team. And so every time they go to a sales meeting or a tradeshow, they're coming back with handwritten notes or transcripts from that meeting, and they need to put those transcripts or those notes into the CRM—the database that the sales team uses. And before AI, what would happen is they would come back from the tradeshow, they'd have all of these handwritten notes, and then they would have to basically spend hours copying those notes into the CRM.

Now what they've been able to do is build a Custom GPT that takes the handwritten notes or the transcripts from the meetings, reformats it in a way that's specific to the fields that are inside the CRM, and it makes it really easy for them to take those notes and put them into the CRM database.

One of the things that you've got to be really careful about is that in order to do this, if you're using a transcription tool like Otter.ai or Fireflies, these tools can join a Zoom call for you or join a Google Meet for you, and they will take the notes and produce a transcript. And you can upload that transcript to a Custom GPT in order to have it reformat it for you.

But you've got to get consent from the other people on the meeting. Especially if you're at a tradeshow, and you're recording a meeting, and the people on the other side of that meeting are your clients and you have not gotten consent from them to record, you cannot record that meeting. You've got to make sure that you're getting consent either through the tools that are part of Zoom that will say, "This call is being recorded," or you've got to verbally ask people, "I would like to record this meeting so that I can take notes afterwards. Do I have your permission?" And if they say yes, then you record, and if they say no, then you can't. So just make sure that you're getting consent and you're not recording people without their permission.

But once you do have their permission and you do have those notes, you can upload those notes as a file to a Custom GPT or you can paste them in as text, and the Custom GPT can reformat them for you. So that's a really great use case for sales teams—for reformatting their notes to be uploaded into a CRM.

Another really great use case is let's say you're the manager of a team of sales professionals, and you want to give them coaching and mentorship on the different sales calls that they're participating in. Maybe you've got junior employees and you want to help give them feedback, but you don't have time to sit in on all of their calls. If those calls are being recorded with consent, then you could upload the transcripts from those calls to a Custom GPT and ask the Custom GPT to act as a coach.

So it would say, "Listen to this sales call. Here are the best practices for sales calls. Compare this transcript to the best practices and give me three pieces of feedback for how I can improve as a sales professional." And then you hand that feedback to your employee and they can learn from the feedback that the Custom GPT has given them. And so it can act as a coach for your team.

Now these are just a couple of examples. There's a wide range of other Custom GPTs that you could be building for your DMO. So let me just quickly throw out a couple. You could have a policy summarizer where you upload new policies, new guidelines, and the Custom GPT summarizes them for your employees so that they can get the high-level understanding of a policy without having to read through 50 pages.

You could have a content repurposing tool where you upload a long-form piece of content, like a case study or a white paper, and then the Custom GPT repurposes it for social media, into a blog post, into a video script, into an email—whatever you need, it can take one source of truth, one primary piece of content, and help you repurpose it for all the various formats that you might need.

You could build a crisis communication simulator where you're training your employees on what to do in case of a PR crisis. The Custom GPT can be loaded with the crisis communication playbook, and then it can generate realistic scenarios and train employees on how to respond in the moment. So it can act as a really great training tool.

You could also build an onboarding assistant where new employees are able to ask questions about the company, about their role, about HR policies, benefits—all of that information can be loaded into one Custom GPT that's specifically for onboarding new employees.

Now, one question that I've been asked at almost every webinar that I've given is, "Can I record a conversation at a tradeshow when I'm just walking the tradeshow floor?" And the answer is yes, you can, as long as you're getting consent from the people you're talking to. So if you're at a tradeshow, you're walking around, and you want to just have your phone in your pocket recording your conversations, you've got to make it clear to the people that you're talking to that you're recording the conversation. The way that I do it is I'll just pull out my phone, I'll hold it, and I'll say, "Hey, I'm taking a recording of this so I can take notes later." And if they're okay with it, then I keep recording. If they're not, then I stop recording.

But as long as you're getting consent, you can use tools like Otter.ai, you can use Apple's voice memo app, you can use Fireflies—any of these tools will produce a transcript for you that you can then upload to a Custom GPT. And so it's a really great way to make sure that after a tradeshow, you don't lose all of the great conversations that you had. You've now got transcripts from all of those conversations that you can put into your CRM.

Now, one of the most important things that you need to understand about Custom GPTs is that in order to use them, you've got to have a paid ChatGPT account. If you have a free ChatGPT account, you cannot build Custom GPTs, and you also cannot use Custom GPTs, even if someone shares a link with you. So you have to upgrade to a paid account. And the reason that this is so important is that if you're using a free ChatGPT account for business use, you are putting your company's proprietary information into a tool that is going to use that information to train future models.

And so the way that you opt out of having your data used to train future models is that you upgrade to a paid account, and then inside of the settings, you turn off the toggle that says, "Improve the model for everyone." When you turn that toggle off, now your data is not being used to train future models, and you have much better security and privacy.

So if you are using ChatGPT for business use, if you're putting any proprietary information into it, you absolutely must have a paid account and you must turn off model training. This is one of the most important settings that you can configure inside of ChatGPT.

Now, the very last thing I want to touch on—because I've been asked this question a lot—is what's the difference between a Custom GPT and a project? Projects are a feature that was released by OpenAI maybe six months ago, and a project allows you to organize your different chats. So let's say that you're working on a big campaign, you can create a project for that campaign and then all of the chats related to that campaign will get organized into the project.

You can also upload files to a project, and those files will be available to all of the chats inside that project. So projects are really helpful for organizing your work. The key difference between a Custom GPT and a project is that a Custom GPT can be shared with your team, and a project cannot. A project is only for you. It's for your own personal organization. A Custom GPT is something that you can share with your entire team, and everyone can use the same instructions and the same knowledge files.

So if you're building something that's just for you, that you want to organize your work, you would use a project. If you're building something that you want your team to use, you would use a Custom GPT. And that's the key difference between the two.

All right, so that is a really quick overview of Custom GPTs. I hope that gives you some ideas for how you can start building Custom GPTs for your team and for your organization. Thank you so much for joining me, and I'll see you at the next one.