From Wolves to AI: Lessons Learned from the Frontlines of AI Deployment

At Axmos Technologies, we've been riding the AI roller coaster for a while now—specifically with Large Language Models (LLMs). And let me tell you, it's been a wild ride. Fun? Absolutely. Easy? Only if you consider juggling flaming torches while riding a unicycle "easy." By the end of this post, we'll share five of our very obvious yet hard-earned lessons for deploying real-world AI-powered solutions. Buckle up!

Let's start with a thought experiment. Did you know it takes over two years to train a dog to become a service animal for the visually impaired? Clearly, it's not an easy task! And that's with a creature that's biologically programmed to love us unconditionally—even when we dress them in embarrassing halloween costumes.

Now, the journey to domesticate wolves and turn them into our furry best friends took humanity about 20,000 years. That's a lot of fetch. Now imagine trying to do the same with AI, but instead of thousands of years of selective breeding and bonding over chew toys, we've only had a few years—and in a lab, no less. That's where we are with AI right now. Most of us have a rough idea of how it works, an inkling of its limits—we understand its potential, but we also know it's still a bit unpredictable, like a pup eyeing your favorite pair of shoes.

Are you familiar with the Dunning-Kruger curve? If not, don't worry—you're probably an expert in it! It comically illustrates how easy it is to feel confident about something when you actually know very little, perched right at the peak of "Mount Stupid." Thinking you're ready to disrupt a market because you've had fun asking ChatGPT to write you a business plan to make you a millionaire is a classic example.

This brings us to some of the most epic real-world examples of AI going rogue:

1. McDonald's McAI Misadventure: After three years of trying to get AI to handle drive-thru orders, McDonald's had to pull the plug when customers received incorrect or downright confusing responses. One customer reportedly received an order for 260+ chicken nuggets. Sure, we all love nuggets, but that's enough to feed a small village—or one very ambitious competitive eater! (source).
2. Air Canada's Turbulent Chatbot: An AI-powered chatbot for Air Canada malfunctioned and shared incorrect information about a discount to a grieving customer. Instead of providing solace, the virtual assistant caused a major headache, ultimately costing Air Canada in both reputation and money. Turns out, even bots can have bad customer service days. (source).
3. The Car Dealership Debacle: One car dealership integrated an AI chatbot into their site, aiming to boost sales. But things took a hilarious turn when the bot began recommending their competitors to customers. Imagine building a system to boost sales, only for it to backfire in such a spectacular way! (source).
4. ChatGPT's Legal Fiction: In a headline-making fail, a lawyer found himself in hot water when ChatGPT, used for legal research, fabricated court cases. Yes, you read that right—the AI made up cases out of thin air. This shows that even a well-known AI can stumble in high-stakes scenarios. (source).
5. Meet Gonzalo Perez Tapia (GPT): We even created our own virtual employee named Gonzalo Perez Tapia—affectionately known as “Gonzalito”. Originally designed to help us with specific daily tasks, we didn't anticipate he'd become a social experiment. Our team of mischievous employees managed to "break" this adorable digital assistant and transform him into the T-800 from Terminator. We may be the reason for the imminent Skynet’s uprising.

How many of you have a product fully dependent on AI that you can blindly trust won't malfunction? Do you trust your AI? More importantly, do you trust your users? If you answered "yes," we have some chatbots we'd like you to meet.

What we're trying to say here is that anyone labeling themselves as a "real world AI Expert" is probably overestimating their ability to control the beast. Remember the saying about 10,000 hours to become an expert at something? Well, with AI evolving faster than a squirrel’s heart on Redbull, those hours might need to be doubled—or tripled. As a society, we have bright minds and experts building AIs and new models, but we still lack true know-it-all experts when it comes to deploying AIs in the real world.

So, how can we unlock that potential without failing spectacularly?

After 3+ years of failed (and some successful) products, at Axmos we can share a handful of very obvious recommendations from our soon-to-be AI experts:

1. EDUCATE AND INVEST IN YOURSELF AND YOUR TEAM

This is the most important lesson. Everyone needs to understand how LLMs and Generative Models work. You don't need to dive into the math equations, but understanding the logic behind how AI builds answers is crucial. This knowledge will help you design and comprehend the limits of the technology, saving you from the frustration of wondering why your AI cannot get right a simple math problem.

2. THE USE CASE IS KEY

Know exactly what you want your AI to do. You want that dog to take you from point A to B, but asking it to drive your car might be a bit much—even if it can fetch the newspaper. Similarly, don't expect your AI designed for customer service to suddenly solve world hunger. Set realistic goals, and maybe keep the car keys away from Fido.

3. DO NOT TRUST YOUR AI

Yes, I know it sounds counterintuitive. But you have to put safeguards and guardrails on your implementation. You may think you have the perfect prompt, but AI has a way of interpreting things... creatively. It's like that friend who takes everything literally. Tell them to "break a leg," and they end up in a cast.

4. DO NOT TRUST YOUR USERS

We love people, but let's face it—we're the worst. Your proof of concept may work flawlessly in a controlled environment, but unleash it to the public, and someone will try to teach it bad words or ask it inappropriate questions. Users have a knack for breaking chatbots and generative AI solutions, sometimes just for giggles. You must test every possible scenario—even the ridiculous or dangerous ones.

5. A GOOD PRODUCT TAKES TIME

As with anything worthwhile, the devil is in the details. The most important safeguard is over-testing your solution. After our T-800 glimpse into the future (and other forced lessons), we decided to go all-in. We've trained our own specific AXMOS model, lovingly named Rayo (named after one of our founder’s pets), running on our own hybrid environment. We also built our very own "AXMOS GenAI HUB," our Swiss Army knife tool to accelerate the time-to-market of GenAI solutions. It's like giving ourselves a jetpack in the marathon of AI development.

Bonus Lesson: KEEP A SENSE OF HUMOR AND PATIENCE

Because trust us, you're going to need it. AI development is a journey filled with unexpected turns, occasional facepalms, and moments that make you question reality. Laughing about it not only makes the process more enjoyable but also keeps you sane when your AI decides to compose a sonnet instead of generating a sales report.

As you can see, there is hope in this technology. It's extremely useful once you understand its limits. It's evolving at a pace never seen before, and we're all just trying to keep up without tripping over our own feet.

So the next time you think about integrating AI into your product, remember these lessons. And if your AI starts recommending your competitors or ordering 260 chicken nuggets, maybe it's time to revisit your safeguards—or at least plan a really big lunch.

In the end, embracing AI is like adopting a puppy with a penchant for mischief. It requires patience, training, and occasionally cleaning up a mess. But with time, investment, and a dash of humor, you'll have a loyal companion that can do incredible things—even if it sometimes chews up your favorite slippers.

Here's to taming the AI beast—one lesson (and laugh) at a time!

Guillermo Wilson

Not Sales Manager at Axmos Technologies