Every day, I see another headline about AI transforming the way we work. Another startup raises millions to build the “next generation” of productivity tools. Another YouTube influencer promises that AI will 10x your output if you just use the right prompts.
And yet, when I sit down with everyday users, people who aren’t developers, who don’t spend their weekends reading Hacker News, who just want to get their work done, I hear a very different story.
AI tools feel complicated. They feel like they were built for enthusiasts, not for normal people.
This is a problem. And it’s not going away anytime soon.
Stories and Facts
Let me tell you about my friend Ahmed.
Ahmed is a civil engineer. He’s been designing buildings for fifteen years. He can calculate load-bearing capacities in his head. He knows every building code in the region. He manages teams of architects and contractors with the kind of calm authority that comes from decades of experience.
A few months ago, he called me, frustrated.
“I’ve been trying to install and use this AI thing,” he said. “Everyone says it’s going to change everything. My junior engineers are using it. But when I try it, the answers are always wrong, after it took him hours to setup and configure. It gives me specifications that don’t match our local codes. It suggests materials that aren’t available here. I spent an hour trying to get it to understand a simple foundation calculation, and then I just did it myself in five minutes.”
He paused. Then he said something that stuck with me: “Am I doing something wrong, or is this thing just not made for people like me?”
Ahmed isn’t alone.
I have a friend named Sarah who teaches high school biology. She wanted to use AI to help her create lesson plans and quizzes. She watched a few YouTube tutorials, signed up for a subscription, and spent an entire weekend trying to get the tool to generate worksheets that matched her curriculum.
“It keeps giving me college-level content,” she told me. “I ask for 10th-grade biology, and it gives me something that would confuse my students. I tried to tell it what textbook we use, what standards we follow, but it just doesn’t stick. By Sunday night, I had nothing usable and I was more exhausted than if I’d just made the worksheets myself.”
Then there’s my friend Hacer. She’s a family doctor. She heard about AI tools that can help with clinical notes, that can summarize patient histories, that can suggest differential diagnoses. She thought this might finally give her back some of the hours she loses every week to paperwork.
“It’s useless,” she told me over coffee last month. “I tried three different tools. One of them suggested a treatment plan that was contraindicated for my patient’s existing condition. Another one hallucinated a medication interaction that doesn’t actually exist. I can’t trust this stuff. In my job, being wrong isn’t an inconvenience, it’s dangerous.”
Three professionals. Three smart, capable people. And three stories of frustration
The Setup Problem
Let’s start with something simple: getting an AI tool to actually work on your machine.
If you’re a Fedora user—like the kind of person who reads guides about post-installation tweaks, you probably don’t mind opening a terminal and running a few dnf commands. You’re comfortable with RPM Fusion, with Flatpak, with the occasional Copr repository.
But think about the average computer user. They don’t know what a repository is. They don’t want to know. They want to download something, click “install,” and have it work.
Most AI tools today don’t offer that experience.
Want to run a local LLM? You’re looking at Python environments, pip installs, CUDA dependencies, and enough configuration options to make your head spin. Want to set up a RAG system, Retrieval-Augmented Generation, the technology that lets AI answer questions based on your own documents? You’re now in the world of vector databases, embedding models, chunking strategies, and a dozen other concepts that didn’t exist in mainstream computing five years ago.
Even the “easy” options, the hosted services with nice web interfaces—require you to understand what you’re signing up for. Which model should you use? What’s a token? Why does this cost money per query instead of a flat subscription?
This isn’t like installing Firefox addons. It’s not like running sudo dnf install vlc and being done with it. The setup friction alone is enough to stop most people from ever getting started.
I remember once a friend called to ask about what the hell is RAG and another one suddenly ask what is a Vector database, both of them has nothing to do with development or engineering, but they believed those are keys the new AI. When I asked why, they said they tried to install some open-source apps.
The Configuration Maze
Let’s say you get past setup. Now you have an AI tool running. Congratulations. Now you need to configure it.
Here’s where things get truly messy.
With a traditional application, say, VLC or GNOME Tweaks, the configuration is finite. There are a limited number of settings. You can explore them, figure out what they do, and eventually arrive at a setup that works for you.
AI tools don’t work that way.
You’re configuring things like:
- Temperature: Controls randomness. But what does that actually mean for your output? Most users don’t know, and the documentation rarely explains it in practical terms.
- Context window: How much history the AI remembers. Too small, and it loses track of your conversation. Too large, and it gets slower and more expensive.
- System prompts: The instructions that define the AI’s behavior. This is where you tell it to be “helpful,” “concise,” or “act as a technical writer.” Except nobody tells you that a good system prompt is often longer than the actual task you’re trying to accomplish.
- Model selection: Should you use GPT-4o, Claude 3.5, Llama 3, or one of the fifty other models available? Each has different strengths, different pricing, different quirks. There is no “default” that works for everything.
And then there’s the workflow configuration.
If you want AI to actually integrate into how you workflow, not just as a chat window you occasionally open, you need to build something. Maybe you want it to summarize your documents automatically. Maybe you want it to help you write emails. Maybe you want it to act as a research assistant that knows your personal notes.
All of these require you to become something between a power user and a developer. You’re configuring automation tools, setting up integrations, writing prompts that work across different contexts. It’s not impossible, but it’s a long way from “it just works.”
The Prompt Engineering Trap
By now, you’ve probably heard the term “prompt engineering.” It sounds technical. It sounds like a skill you need to learn.
And you do. That’s the problem.
Prompt engineering is the art of writing instructions for AI that actually produce the results you want. It’s not just asking a question. It’s understanding how the model thinks, what kind of phrasing triggers better responses, how to structure your input to avoid common failure modes.
Here’s an example:
If you ask an AI, “Summarize this document,” you’ll get something. It might be good. It might be mediocre. It might completely miss the point.
If you instead prompt: “You are a technical writer with expertise in Linux desktop environments. Summarize the following document in 3-5 bullet points, focusing on actionable recommendations for new users. Ignore marketing language and assume the reader is technically competent but unfamiliar with this specific topic.” You’ll get a much better result.
But who teaches you that? Where is the manual?
Most users don’t have time to become prompt engineers. They have jobs. They have families. They want the tool to work for them, not the other way around.
And yet, every AI tool today essentially outsources the instruction-writing to the user. You’re not just using the tool, you’re also training it, every time, with every query. There is no “default good.” There is only “good enough if you know how to ask.”
The RAG Complexity
Let’s talk about RAG. Retrieval-Augmented Generation.
This is one of the most powerful capabilities in modern AI. It’s what lets you upload your documents and have the AI answer questions based on them. It’s what turns a generic language model into something that actually knows your information.
It’s also incredibly complicated for normal users.
To understand why, let’s look at what RAG actually does:
- It takes your documents and splits them into chunks. How big should those chunks be? Too small, and you lose context. Too big, and you waste tokens.
- It creates embeddings for each chunk—numerical representations that capture meaning. Which embedding model should you use? There are dozens.
- When you ask a question, it searches through these embeddings to find relevant chunks. How many chunks should it retrieve? How do you balance relevance against context length?
- It then combines those chunks with your question and sends everything to the AI model. But the model has a context limit. If you retrieve too much, you exceed it. If you retrieve too little, you miss information.
- Finally, the AI generates an answer based on both your question and the retrieved chunks. But it might still hallucinate. It might ignore the chunks entirely. It might blend information from different sources in ways that don’t make sense.
This is not a “one setting fits all” situation. Different documents need different chunk sizes. Different questions need different retrieval strategies. Different use cases need different trade-offs between accuracy, speed, and cost.
And this is before we even talk about keeping your RAG system updated. What happens when you add new documents? What happens when you modify existing ones? How do you handle deletions?
For an enthusiast, someone who enjoys tinkering with this stuff, it’s a fascinating challenge. For a normal user who just wants to ask questions about their notes? It’s a nightmare.
The AI Agents Problem
If RAG is complicated, AI agents are on another level entirely.
An AI agent is supposed to be the next evolution: an AI that can take actions, not just answer questions. It can send emails, update calendars, browse websites, execute code, and generally act as a digital assistant that actually does things.
In theory, this is what we’ve all been waiting for. In practice, it’s a configuration hellscape.
To make an agent work reliably, you need to:
- Define its permissions. What can it access? What can it modify? What’s off-limits?
- Define its tools. What actions can it take? How does it know when to use each one?
- Define its constraints. When should it ask for confirmation? What are the boundaries of its autonomy?
- Test everything. Agents are notoriously unpredictable. A small change in a prompt can cause them to behave completely differently.
And then there’s the safety problem. If an AI agent has access to your email, your files, your calendar, what happens when it makes a mistake? What happens when it’s tricked by a malicious prompt? What happens when it does exactly what you asked, but that thing turns out to be a bad idea?
The current approach to agents is essentially: here’s a powerful but unpredictable system, and you’re responsible for making sure it doesn’t cause problems. Good luck.
This is not a tool for daily users. It’s a tool for people who enjoy complexity.
Why Aren’t These Simplified?
You might be wondering: if these tools are so complicated, why don’t companies simplify them?
Part of the answer is that we’re still early. The technology is moving fast. Standards are being set. Best practices are being discovered. It’s hard to build a simple interface for a technology that’s still figuring out what it wants to be.
But there’s another reason, one that’s more uncomfortable to talk about.
Simplicity requires trade-offs. When you simplify something, you decide what the user doesn’t need to know. You hide complexity. You make assumptions about what they want.
And in AI, those assumptions are risky.
If you hide the temperature setting, you’re deciding that the user doesn’t need to control randomness. But maybe they do. Maybe their use case requires deterministic outputs. Maybe the default setting produces unusable results for them.
If you hide the model selection, you’re deciding that one model works for everything. But it doesn’t. A model that’s great for creative writing is terrible for code generation. A model that’s fast and cheap is not the same as a model that’s slow and accurate.
If you hide the RAG configuration, you’re deciding that one chunking strategy works for all documents. But it doesn’t. A technical manual needs different handling than a collection of personal notes.
The people building these tools know this. They know that simplifying means frustrating advanced users. They know that power users will demand access to the knobs and dials. They know that the current state of the technology doesn’t yet support a “just works” experience for everyone.
So we’re stuck in the middle. The tools exist. They’re powerful. But they require you to become something of an expert just to use them effectively.
What This Means for Daily Users
If you’re a normal person, someone who just wants to get work done, not become an AI enthusiast, where does this leave you?
Right now, it leaves you with a choice.
You can wait. The tools will get simpler. They always do. The first computers required you to understand command lines and memory addresses. Now we have smartphones that a child can use. AI will follow the same path. It might take a few years, but the “just works” versions are coming.
Or you can embrace the complexity. Not all of it, but enough to get value from the tools.
This is where the Fedora mindset actually helps. If you’ve ever worked through a post-installation guide—enabling RPM Fusion, installing Fedy, tweaking GNOME settings—you already understand what it takes to get a complex system working well. You know that the initial investment of time pays off in a system that’s customized to your needs.
AI tools are similar. The setup is annoying. The configuration takes time. You’ll need to learn what temperature means and why it matters. You’ll need to experiment with prompts and figure out what works for your specific tasks.
But once you’ve done that work, you have something genuinely useful. You have a tool that understands how you work. You have a setup that produces consistent results. You have a system that saves you time every single day.
A Practical Approach
If you’re ready to take that approach, here’s what I’d recommend:
Start with one tool. Don’t try to build a complete AI workflow overnight. Pick a single task, maybe summarizing articles, maybe drafting emails—and focus on making that work well.
Use hosted services first. Local AI is great for privacy and control, but it adds another layer of complexity. Start with something like ChatGPT or Claude. Get comfortable with the basics before you worry about running models on your own hardware.
Learn prompt structure. A good prompt has three parts: role, task, and constraints. “You are X. Do Y. Follow these rules.” This pattern works for almost everything.
Don’t worry about RAG yet. If you need to work with your own documents, start with the built-in file upload features in services like ChatGPT or Claude. They handle the complexity for you. Only dive into proper RAG when you hit their limits.
Accept that you’ll iterate. Your first prompts won’t be great. Your first configurations won’t be optimal. That’s fine. Treat it like customizing your desktop, you tweak things over time until they feel right.
Conclusion
AI tools are complicated for daily users. There’s no getting around that right now. The setup is messy. The configuration is deep. The prompt engineering is a skill you need to learn. And the advanced features, RAG, agents, local models, require a level of technical comfort that most people don’t have.
But here’s the thing: complexity isn’t always bad.
Fedora is more complicated than Ubuntu. It requires more setup, more configuration, more willingness to read documentation and run terminal commands. And yet, people choose it. They choose it because the complexity comes with control. It comes with the ability to shape the system to exactly what they need.
AI tools are at that stage. They’re not simple. They’re not polished. They require you to understand what’s happening under the hood.
But if you’re willing to put in the time, the same kind of time you’d spend setting up a new Fedora installation, you get something that most people don’t have. You get a tool that actually works the way you work. You get a system that saves you hours every week. You get the benefits of AI without being limited to what someone else decided was simple enough.
That’s worth the complexity. At least, I think it is.
What are some AI tools or workflows you’ve struggled to set up? Share your experience in the comments.
