AI wrote this email—and I hated it

Most teams want AI to write faster. Too often, it delivers generic, hallucinated emails that no one would approve.

In this practical and refreshingly honest talk, Aubrey Miller-Schmidt shows how to turn a large language model from an “eager intern” into a reliable email assistant. Drawing on 15 years in the channel and a live demo with a faux brand, she explains what LLMs are (prediction machines) and what they are not, then walks through the workflow, guardrails, and ownership needed to get usable, on-brand copy.

Note: This post is adapted from a transcript of Aubrey’s presentation in July 2025.

This is not just a talk about prompts. It is a playbook for safe, scalable, human-in-the-loop AI email content production.

TL;DR:

• LLMs predict the next word, not read your mind. Garbage in equals garbage out
• Treat AI like a new intern and train it on your voice and goals
• Provide clear prompts with audience, purpose, timing, and context
• Feed it your brand guide and past high-performing examples
• Keep humans in the loop to fact-check, edit, manage technology, and ensure brand fit

Get ready to upgrade your prompts, narrow your model’s world, and ship emails that sound like you.

Aubrey: I am Aubrey Miller-Schmidt. I'm so excited to share information with you today. For those of you who don't know me, I'd like to take a moment to introduce myself.

I currently work at a Fortune 250 insurance carrier, where I manage the email marketing as well as other marketing campaigns and strategy. And over my 15 years of experience in the email space, I have worked on D2C, B2B, nonprofit, and even some political email campaign work. 

All that to say is that email content is not exactly new to me. Like many of you, LLMs, those large language models of AI, have come into my life over the last year and a half or so. I've done a lot of experimentation and learned a lot. 

Another fun fact about me is that I'm a former middle school teacher, and so I know a few things about learning curves and also about tough crowds, but you're not going to be a tough crowd for me today, right? So I'd like to kick it off by asking the crowd here, have you ever been frustrated by the output from your AI? Show of hands? Well, me too. 

So, here are a few things that I've literally told my AI. In preparing for this, I asked it, “can you tell me some things that I've said to you that maybe weren't that nice?” Here’s what it said:

“I hate paragraph two. These aren't specific enough that intro sounds like bull... Not so negative.” and, “That still doesn't sound like a human.”

If you have been experimenting either for personal work or for work work, you've probably seen some places where AI LLMs have been super helpful, and other places where it's been a bit of a dumpster fire. For the purposes of this talk, I created a faux company so that we can enjoy the output that I'm about to show you. So, don't feel bad if you want to bust out laughing and make fun of it. This is why I did this experiment. So what happens when we give AI a generic, average prompt that when you're first starting out, you might give it? 

I asked it, “We’re BlueVine Maps, a clean water access technology B Corp. Write an email about our new water dashboard.” That was a pretty average prompt. Well, here's what it gave me:

"Subject: Introducing the BlueVine Water Dashboard - See the Impact, in Real Time 💧

Hi [First Name],

At BlueVine, our mission is simple: unlock access to clean, safe water-everywhere. As a Certified B Corporation, we believe in transparency, innovation, and measurable impact.

That's why we're excited to introduce the BlueVine Water Dashboard-a new way to track and visualize water access improvements in real time.

What you'll find on the dashboard:

📍Live data on water system performance and uptime

🌎Regional insights into water accessibility and usage 📊Impact metrics showing how we're driving sustainable change

🤝Community-level reporting on system health and support needs

Whether you're a partner, donor, policymaker, or community advocate, this dashboard gives you the visibility you deserve-and the data to take action.

🔗[View the Dashboard Now]

Together, we're transforming water access from a promise into a measurable reality.

Thank you for being part of the movement.

Cleanly,

The BlueVine Team"

So, what is wrong with this email? It’s a faux company – have at it. 

Here’s some answers from the crowd:

• The emojis and the dash give it away.
• There are stock phrases. 
• It's written like a sales letter. 
• No one's going to say “cleanly.” (Yes, I hope not.)

Well, I don't know about you, but if this came across my desk, I would reject it immediately and probably say, “This is written by AI.” It's not what I want to use AI for in my workspace. 

So, it does beg the question, where did it get all this information from — right? We read through my prompt. It also hallucinated all of this information. It didn't know this stuff about this company. It assumed all of these things about my product and my company and how we talk. So why is AI so bad? I think we need a level set on what large language models are for just a second to understand this.

LLMs are statistical models for prediction. What does that mean? It means that they are trained on a bunch of information, and then they predict the next thing to say, word by word by word, based on what they've been fed. When we think about AI, we might be thinking about “Her”. Has anyone seen the movie “Her”? Or we might be thinking about the movie “The Terminator”. We might be thinking about all these other scenarios where we might be talking about artificial general intelligence. 

And what that means is that's the pinnacle of AI, where it has creative human ability to make up new things, and it doesn't need to be fed information. That's not what we have today. They're doing a lot of testing and a lot of theories and trying to get better, but we're not there right now. We have something below that level, and so when we're talking to our AI robots, we need to keep in mind that it's only as good as the information that is fed. So it is not as smart as we might want to think. That doesn't mean it's useless; however, it just means we have to understand how to talk to it and how to prompt better. 

So when we give AI better prompts and better context, it can actually produce much better results. And the other thing is, is that we have to think of it like we're training an intern. So if you've ever had an intern in your organization, they come in and they're really excited, just like a lot of the AI is right now.

“Sure, Aubrey, I'd be happy to do that for you.” Did your AI ever say that to you? It's a lot like that with your interns. They're excited. They're probably pretty smart, college-educated, but they don't know anything about your company other than what they've seen on the internet, and they don't understand how you talk. They don't know how to write an email for you. 

And so, you have to take this very eager technology and teach it how to talk like you. And if you do that work and you keep poking it back and saying, “I need to get edits,” it can be taught to do better over time. So then, how do we write a better prompt? Because the one that we started with gave us kind of garbage.

I am a recovering middle school teacher, so if we just come back to our basics of how to write for a second, we can talk about the five W's: who, what, why, when, and where. And if we do that with each prompt, we're more likely to get fewer hallucinations because it has more to work with, and we're more likely to get better content that we don't have to endlessly edit and say that still doesn't sound like a human. I'm not going to read this to you, but this is what it might look like if, instead of the generic prompt that I gave at the beginning, we did that with our blue line maps scenario. We provide a lot more context here. 

We're saying, “Who are we trying to talk to?” Because I didn't in the original one, I just said we're launching a new dashboard. But I had no idea who my customer was. Who am I telling this to? It didn't even understand what day it is. This is something really important. If you've tested AI, it can be kind of dumb when it comes to what day of the week is it? What month or year it is? It's very strange. I always like to say, “Today is this, I'm going to send this email on this date, and the thing launches on this other date.” It's super basic, but you need to provide that context, otherwise, it might make it up for you. 

And then, most importantly, we have to tell it – why? Why are we doing this thing in the first place? Why does this email matter? And you know, we didn't do that in the original part. We just said we're launching a dashboard, but we didn't explain, just like any other marketing goal. Why does it matter? If we're launching a new dashboard, we can't just assume that the AI can read our minds. We have to tell it that information. And then, if you remember, in school, you did your five W's, you also got bonus points when you did the how well, the how is. This is where we have to give it quite a bit more information to be successful. 

So it's like Pac Man, we're going to feed our robot everything we need it to know. So we need competitor examples. It doesn't even understand email, y'all, so it's been indexing the entire internet. But that includes personal email, maybe some email marketing stuff. It doesn't often even understand the difference between the two. So we're saying, “This is what good looks like.” I like to go on Really Good Emails for inspiration from other brands. Let's say I think these few examples that I found are good dashboard emails. As you can see, I fed it some information about that. 

And then I said, “Here's our brand guide.” For the purposes of this presentation, I made a fake one up. It's pretty terrible compared to a real brand guide, but I stuck something in there so it could have some guidelines and boundaries on things like: how do we talk? How do we use our logo? All of those important things that you would provide anyone working on your marketing team. And then we also need to go in and feed it past successful emails. So, we are training this robot specifically for our brand. 

These are ones that not just look like us and talk like us, but actually performed well. So, go find things in your archives that had really high click rates or really great sales, depending on what market you're in, and feed it to your robot. And then after we do all of that, with all of that context, we might get to the ghost — it producing us a good email. 

So you might be saying, “Gee, I thought this was supposed to help me. This is a lot of work. I just want to be able to yell at it and give it me an email. It's supposed to build efficiency on my team. I don't want to do this every single time.” Well, this is where custom GPTs come into play. So you may already have one for your organization, where you purchase some kind of contract with one of the big companies, but what I'm talking about here is one specific to your program. So it might be a sub-custom GPT.

If you're in a small company, you know, that might be really easy to do. We'll talk about sort of the logistics here in a second. But a custom GPT is giving the option to have your AI sort of focus. So you're telling it, don't worry about the entire internet anymore. I need you to narrow down what you're looking at. And here is my brand. Here is the task that you're doing. So in this case, we're asking it to write emails. We don't want it to do anything else. And it can reference past guides, past emails in a space that we sort of have more control, so we're taking that intelligence and asking it to narrow down its focus, and so it will no longer introduce our email with I hope this email finds you well, or whatever other junk that it might produce otherwise, and it also gives us that efficiency that we're looking for To reduce the back and forth number of edits. 

However, this is a big ask for your organization, and I think this is a really important point again to think about. Is often organizations get excited about the use of AI, and that's great, but we're not taking the time to understand that, just like any other marketing technology, this one is also going to need certain things to be successful. And so I would advocate, you need an identified product owner. You're going to say, “We want to write email with AI.” Then who is in charge of that? They need to understand how to sort of train this robot, how to write within the context of your custom GPT. It needs to be in charge of training it so going resourcing those high-quality examples to feed your robot and to update them over time. 

Also, there are aspects of security and safeguards, so things that it can't do, places and data that you don't want it to touch how we leverage the information, so you need to have someone that's deeply familiar with this technology in charge of it to make sure that it's being used well. And then there's that ongoing smoke testing and review. So it's not a set-it-and-forget-it sort of situation, even though it might seem initially like, wow, AI is awesome, I don't need to do anything. It's actually an investment, just like any other marketing technology. 

Then it gets more complicated, but it's important to understand that AI is not all the same. So my initial prompt, I put that into chat, GPT 4o. That's only one single model, and I used that one. But fun fact, when I was preparing, I actually plugged it into Claude Sonnet, which is the default as well, and Gemini 2.5, which was their default. And interestingly, Gemini gave me something super similar. And Claude actually did not give me an email. It said, I need more information and asked me some questions. Now, this is not to say that Claude’s is better. I'm not here to be a proponent for anybody's particular AI brand. 

My point is, is that they're all different, and you're going to get something different depending on what you select. So I frequently use ChatGPT, but I don't use the default. So when I'm talking to it, I often switch models depending on what I'm attempting to do. Ethan Mollick is a wonderful researcher, and you don't follow him. I don't know him personally, but he's a Wharton professor, and he does AI research. I borrowed this image from him, because the thing is, this information isn't well documented by any one of these companies at this point. 

So what is it good for? They don't really say they just give you new models to try them. So if you're in there, you can go play with them. You can also follow some of these researchers to better understand which model to select. And then again, it's even more complicated because there are so many models out there, and there are new ones that are being, you know, shared all the time. And we, again, don't see documentation on what they're good at. But beyond that, for business leaders, what you would need to understand is that there's different speeds, there's different costs, so there's a lot to consider. 

And back to that product ownership. How are you going to decide for your company which models you're going to get your team access to, and which ones are building the efficiency that you're looking for? So maybe the model is cheap, but it's extremely slow or it doesn't give you good output. So it's not a simple formula to come out at the end. However, let's say, okay, we learned how to prompt. We've built a custom GPT, we've selected the right model that works for our company, and we've input all of those things. And let’s come back to our BlueVine Maps example. What do we get? So I did that. 

"Subject: New Water Monitoring Dashboard Live August 1

Preview Text: Save your seat for our launch webinar.

<Full-Width Header Image: Suggestion: A screenshot of the dashboard's live map view and a subtle BlueVine water motif.>

Introducing Your Real-Time Water Monitoring Dashboard

Starting August 1, you'll have one place to see live performance metrics for every water system you manage.

View flow rates, spot issues instantly, and pull reports in seconds-so you can plan operations and share transparent updates with funders and stakeholders.

[Button: See What's New on Aug. 1]

<Two-column section: left showing a screenshot of the dashboard; right a short bulleted list >

Key Benefits:

-Real-time visibility across all sites

-Fast, clear reporting tools

-Full transparency for stakeholders

Thanks for partnering with us!

-The BlueVine Maps team"

Well, how did it do? Did it do better? Is it perfect? No, I still would not approve this email, and that is sort of my point. 

AI is meant to build efficiency, and you can get to efficiency. But it I would not come in and say you should replace your entire copywriting team with AI. This is supposed to be an assistant to human beings. I think there are dangers when you just let it go as-is and hit send. We still have to remember that after everything that we've entered, it can still hallucinate. We need to go and read every single word and make sure that that's true about our brand. 

We need to have that human oversight, but we also might want to consider making adjustments, just to make adjustments, because there are potential copyright issues when it is scraped to the entire Internet. You don't know if it has borrowed copy from somebody else. You can go and do research on that, but it's always one of those things, when we talked about the safeguards and security, that you run into that as a potential issue. However, I will say that this can save your team a ton of time. So now we went from a concept to email, maybe by simply speaking to our custom GPT. And we have an email here that maybe we can use. 

I did create a custom GPT for BlueVine Maps, and you are welcome to give it a test drive. Here’s the link to it. It's free to use. I will say that it's quick and dirty. I did not do this like I would have done for my own company, but you can see if I did this for your own company, what might the experience be.

And I encourage you to test things and learn. That’s how you get better with AI, because there's not a lot out there that's going to tell you how to do this. Thank you.