AI Tools vs. AI Models — #AIwithUX Series #03

What AI tools and AI models actually are, explained through something we all understand.

Post 1 of AIwithUX series was about talking to AI. Post 2 was about how AI actually works underneath. This one’s about a mix-up I didn’t even realize I was making until recently: using “AI tool” and “AI model” like they were interchangeable.

They’re not. And once that clicked, a lot of confusing AI conversations at work started making sense. These notes come from working through Google’s AI coursework as part of my own learning — I’m just putting it into how I actually think about it as a designer.

The mix-up I kept making

I’d hear someone say “we’re building this with GPT” and someone else say “we’re adding an AI feature to the app,” and in my head, both statements meant roughly the same thing. They don’t.

An AI tool is the AI-powered software you actually interact with; it helps you complete a task. An AI model is the program underneath it, trained on data to recognize patterns and do a specific job. Every AI tool is built on top of one or more AI models. Without the model, the tool simply doesn’t work.

The analogy that finally made this stick for me: think of an AI tool like a restaurant, and an AI model like the kitchen behind it. When you sit down and order, you’re interacting with the menu and the waiter — that’s the tool, the part designed to help you get something done. You never see the kitchen, but nothing happens without it. The kitchen is where the actual work gets done — that’s the model, quietly doing the processing behind the scenes.

AI tools work the same way. Speeko, the AI speech coach, is the restaurant experience - what you interact with. Figma AI is the restaurant. So are Adobe Photoshop’s Neural Filters. What’s actually running the kitchen behind each of them is a model - GPT, DALL·E, Stable Diffusion, Gemini, or something similar. Different kitchens for different menus, same basic relationship.

Once I understood that split, I noticed AI tools themselves come in three different flavors.
1. Stand-alone AI tools work independently with minimal setup. Speeko again is a good example, you just open it and use it.

2. Integrated AI features are built into software you’re already using, like Neural Filters living inside Photoshop - you don’t install anything new, the AI just shows up inside a tool you already know.

3. Custom AI solutions are built for one very specific problem - Johns Hopkins Hospital’s AI system for early sepsis detection is the example that stuck with me, a purpose-built tool solving one very high-stakes task.

There’s one more concept worth knowing here:
AI agents: If the restaurant could run itself - take your order, cook it, and serve it without a single staff member involved - that’s what an AI agent is. It’s an AI-powered tool that performs tasks on its own, with very little human oversight: automatically replying to emails, posting content on social media, monitoring a network. You set the rules once, and the agent just runs. That’s a meaningfully different level of trust than a tool that waits for your input every time, and it’s a distinction I think UX designers especially need to sit with.

Put together, the full loop looks like this: you provide an input or a goal, the AI tool is the interface you actually interact with, the AI model processes that input using patterns it learned from data, and you get an output - a useful result that helps you finish the task. One small note that’s easy to miss: some AI tools aren’t powered by just one model. They combine several, each specialized for a different subtask, working together to handle more complex requests.

How a model actually gets built

This is the part I hadn’t really thought about before — models don’t just appear, they’re trained, step by step.

Take a simple example: building a model to predict rainfall. First, someone defines the actual problem. In this case, helping people know when to expect rain on their commute. Then they collect relevant data, like decades of historical weather records. That data gets prepared; labeling things like temperature, humidity, and air pressure, then splitting it into a training set and a separate validation set. The model is trained on that data, learning to recognize the patterns that tend to show up before it rains. Then it’s evaluated against the validation set to see how well it actually predicts. If something’s off - say the data was biased or incomplete, then the team goes back a step and adjusts. Only once it performs well does it get deployed into an actual tool.

And even after deployment, it doesn’t stop. Once real people start using it, the model runs into new situations it wasn’t trained for, so it keeps getting monitored, fed new data, and refined. It’s not a one-and-done process — it’s a loop.

Why this actually matters for design work

Here’s where I’ve landed:
Understanding this difference changes the questions I ask when an AI feature gets proposed for a product. If someone wants to “add AI” to a feature, the real question is which model is doing the work, and whether it’s a stand-alone tool, something integrated into what we’re already building, or something that needs to be custom-built for the problem at hand, because those are three very different scopes of work. And if what’s being proposed is closer to an agent than a tool, that’s a trust and control conversation, not just a UX polish conversation.

None of this makes me an ML engineer. But it does mean I can sit in a product conversation and actually understand what’s being proposed, instead of nodding along at the word “AI.”

Key Takeaway: The tool is what the user sees. The model is what makes it work. Knowing the difference is what lets me ask better questions before a single screen gets designed.

[ This is post three of the AIwithUX series. Still learning this in public, one concept at a time.]

#AIwithUX #ArtificialIntelligence #UXDesign #UIDesign #MachineLearning #AIDesign #ProductDesign #LearningInPublic #UXwithAI

AI Tools vs. AI Models — #AIwithUX Series #03 was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.