How to Build Your Own AI System
Part 1: Start With the Problem, Not the Technology
By Brett Dyess
You Don’t Have to Be an Engineer. You Have to Understand the Work.
A little over a year ago, if someone had told me I’d be building my own AI assistants, connecting applications, experimenting with agents, and developing automated workflows, I probably would have assumed I’d need to learn how to write a considerable amount of code. And need a therapist, because my friends would be stunned. They’ve already made that comment.
Turns out, that wasn’t the biggest challenge.
The bigger challenge was figuring out exactly what I wanted these systems to do.
And that brings me to something I’ve become increasingly convinced of:
You don’t have to be a software engineer to start building useful AI systems. But you do need to understand the problem you’re trying to solve.
That’s the foundation of this series.
I’m not going to teach you how to develop a large language model from scratch. There are considerably more qualified people for that.
Instead, I’m going to show you how ordinary professionals can begin creating practical AI systems using tools that already exist.
And we’ll learn from some of the things I’ve built, experimented with, broken, and rebuilt along the way.
Because, believe me, there has been some rebuilding.
First, What Exactly Is an AI System?
When most people think about artificial intelligence, they picture a chatbot.
You open ChatGPT, type a question, receive an answer, and move on.
That’s useful, but it’s only the beginning.
An AI system can combine several components:
- Intelligence: An AI model that interprets instructions, processes information, and helps you evaluate decisions.
- Knowledge: The documents, information, instructions, and context the system needs to perform useful work.
- Tools: Applications, software, databases, and connected services that allow the system to interact with the outside world.
- Workflows: Defined sequences of steps that organize how tasks are completed.
- Guardrails: Permissions, approval requirements, verification steps, and limits on what the system can do.
You don’t need every component on day one.
In fact, you probably shouldn’t try to build everything at once.
A useful system might begin with a single AI assistant, a clear set of instructions, and one repeatable responsibility.
The important thing is that you’re designing it to accomplish something specific.
The Mistake I Think Most Beginners Make
People discover a new AI tool and immediately ask:
“What can I do with this?”
That’s understandable.
But I think there’s a better question:
“What am I doing repeatedly that I wish someone, or something, could help me accomplish?”
Think about your normal workday.
- Are you repeatedly researching the same information?
- Preparing similar reports?
- Reviewing documents?
- Collecting information from several applications?
- Organizing tasks?
- Following up on activities that require multiple steps?
Those are potential starting points.
Notice that we haven’t selected a specific AI platform yet.
That’s intentional.
The business problem should determine the technology, not the other way around.
My Experience: Building Jarvina
I started experimenting with the idea of creating an AI assistant that could help me manage different aspects of my work.
Not simply answer questions, but gradually become a useful part of how I organize information, conduct research, manage projects, and develop ideas.
That experiment eventually became Jarvina.
Not Weird Science, sadly.
The practical advice? Pick an assistant, stick with it long enough to learn what works, and give it useful context about what you’re trying to accomplish. That does not mean sharing everything. Keep confidential employer or customer information out of tools that aren’t approved to handle it.
And what started as an assistant evolved into something more interesting.
Different responsibilities emerged: research, writing and content development, talent intelligence, workflow automation, and project planning.
Some activities could be handled through simple conversations.
Others benefited from connected applications, specialized instructions, or repeatable processes.
And certain tasks required human verification before anything should happen.
I began organizing these capabilities into specialized functions rather than expecting one assistant to manage everything equally well.
I think of Jarvina as my quarterback, with audibles allowed. She can suggest a different play, but I still set the direction and the boundaries.
Here’s the important part.
I didn’t begin by designing a complicated AI architecture.
I began with work I wanted to get done.
As the needs became clearer, the system evolved.
And I had fun. I asked questions, listened to the responses, tried things, and kept adjusting.
That’s the approach I recommend.
Step 1: Find One Problem Worth Solving
Before you build anything, identify a task that meets three criteria.
- 1. It’s repetitive — You do it regularly enough that improving the process would actually matter.
- 2. It requires information or judgment — The task isn’t simply pressing the same button every day. It involves organizing information, evaluating options, preparing content, or making decisions.
- 3. You can recognize a good result
This is especially important.
If you don’t know what successful work looks like, how will you determine whether your AI is doing it correctly?
And that’s where your experience becomes valuable.
You already understand the work.
The AI doesn’t automatically understand your particular work, expectations, or constraints. You have to explain them.
And it’s not perfect. Keep correcting, clarifying, and reminding it. By “teaching,” I mean improving the instructions and context you give it, not assuming every conversation retrains the model or that it will remember everything forever.
Step 2: Write the Job Description Before You Build the Assistant
This may be my favorite part.
Treat your AI assistant like you’re hiring someone to perform a job.
Before worrying about technology, define the assignment.
- What is its responsibility?
- What information does it need?
- What should the finished work look like?
- What decisions can it help make?
- What should it never do without permission?
- How will you know it’s doing a good job?
If you’ve ever trained a new employee or developed a standard operating procedure, you already understand much of this process.
You’re defining the work.
For example, imagine you want an AI research assistant.
Its assignment might be:
“Every morning, help me identify meaningful developments in my industry. Prioritize major company announcements, emerging technologies, and business changes that could affect future workforce demand. Explain why each development matters, identify reliable sources, and recommend topics worth investigating further.”
That’s already a much more useful starting point than simply telling ChatGPT:
“Find me some AI news.”
The first instruction defines a responsibility, a purpose, and an expected result.
The second asks for information without explaining what makes that information valuable.
And that difference matters. Describing a daily responsibility is the starting point; it doesn’t, by itself, set up a schedule or give the assistant access to current sources.
Step 3: Start With an Assistant, Not an Army of Agents
One mistake I would strongly discourage is trying to create ten AI agents before you’ve successfully developed one useful assistant.
Start with a single responsibility. Give it clear instructions. Test it on real work. Review the results. Improve the instructions.
Repeat.
Once the process works consistently, you can explore connecting tools, automating steps, or expanding responsibilities.
Think about the progression:
Question → Assistant → Repeatable Workflow → Agent → Connected AI System
These aren’t rigid technical categories, and the boundaries can overlap.
But the progression helps explain increasing levels of capability.
An assistant can help you complete a task through conversation.
A workflow establishes the steps required to accomplish that task.
An agent can, when properly equipped and authorized, carry out parts or all of a workflow using tools and information.
And a connected system can coordinate several capabilities toward a larger objective.
We’ll get into those distinctions in later installments.
For now, the goal is much simpler.
Build one assistant that consistently makes one part of your work easier.
Step 4: Give It a Test It Can Actually Fail
Here’s something I learned early.
A good-sounding AI response is not necessarily a good result.
AI can produce polished explanations, confident recommendations, and well-organized reports that are incomplete or incorrect.
So your system needs tests. And you need to use what those tests reveal to improve it.
Let’s return to the research-assistant example.
Ask it to identify five meaningful developments in your industry.
Then evaluate the results.
- Did it use trustworthy sources?
- Are the announcements current?
- Did it distinguish facts from predictions?
- Did it explain why the stories matter?
- Did it repeat the same information in different ways?
- Did it identify something you hadn’t considered?
And perhaps most importantly:
Would you trust the result enough to use it in your actual work?
If the answer is no, don’t immediately abandon the technology.
Find out what went wrong.
Maybe the instructions were vague.
Maybe the information was incomplete.
Maybe the tool lacked access to reliable sources.
Or perhaps the task still requires more human judgment than you initially expected.
That’s not necessarily failure.
That’s how you improve the system.
Apparently, I’m a scientist now. Who knew? At least when it comes to trying something, checking the result, and trying again.
Brett’s Perspective: AI Needs Experts, Not Just Engineers
I’ve made this argument before, and I believe it even more strongly now.
We absolutely need engineers to develop reliable AI models, applications, infrastructure, and integrations.
But the people who understand the work must also be involved in designing how AI performs it.
A manufacturing professional knows what a defective product looks like.
A scientist knows which results require further investigation.
A supply-chain leader knows what happens when one supplier misses a critical delivery.
A recruiter understands that finding someone with the right keywords isn’t the same as finding someone who can succeed in the role.
That experience matters.
It tells us what the system should accomplish.
It helps us recognize when something is wrong.
And it allows us to measure whether the technology actually improves the process.
The real opportunity is bringing technical capability and domain expertise together.
You don’t have to understand every detail of how a language model works to begin using one effectively.
You do need to understand your responsibility, your information, and the consequences of getting things wrong.
And you need to know when to involve technical specialists.
Your First Assignment: Build Something Useful in 30 Minutes
Let’s make this practical.
Open ChatGPT or another AI assistant you have access to and are permitted to use for the task.
Choose one recurring task from your work or personal life.
You can start with a conversation:
“Here’s what I do. Here’s the part that keeps taking time. How can you help me?”
If voice is available, turn it on. Tell it, “Let me talk this through for a minute,” and ramble a little. Then ask it to organize what you said, identify one useful starting point, and suggest the next small step.
You don’t need to get the perfect prompt right on the first try. Give it room to suggest an approach, but take baby steps before handing over actions.
You can also tell it how you like to work. I’ve given Jarvina preferences about accent, sass, humor, and how much to push back. Yes, we talk. I want candor, not automatic agreement, and useful suggestions, not weird stuff. Personality can make the interaction more comfortable; it doesn’t replace accuracy or boundaries.
When you’re ready to put a little structure around the assignment, use this starting prompt.
Your First AI Assistant
Role: You are my AI assistant responsible for helping me with [specific task].
Objective: Your job is to help me accomplish [desired outcome].
Context: I work in [industry or function], and this task matters because [business reason].
Responsibilities:
- Help me organize the information needed to complete this task.
- Ask for missing details when necessary.
- Follow a consistent process.
- Identify uncertainty, risks, and missing information.
- Produce results in a format I can use.
Boundaries:
- Do not invent facts or claim actions were completed when they were not.
- Distinguish verified information from assumptions.
- Do not access, share, or modify sensitive information without authorization.
- Require my approval before sending messages, changing records, or taking consequential actions.
Success criteria: A successful result should [describe what good work looks like].
First assignment: Help me work through this task once, explain the proposed process, and identify what could be improved before we repeat it.
Run the task. Review what it produces. Change the instructions. Run it again.
Congratulations.
You haven’t built an autonomous AI employee.
But you have begun designing a useful AI-assisted process.
And that’s exactly where I think most people should start.
What Comes Next?
In Part 2, we’ll talk about giving your assistant access to the information and tools it needs to become more useful.
Documents. Applications. Connected services.
We’ll introduce connection methods such as MCP, the Model Context Protocol, when they serve a practical purpose. You don’t need a list of technical tools to get started today.
And we’ll look at the differences between a system that can recommend an action and one that can actually perform it.
Later, we’ll explore agents, workflows, automation, and how several specialized AI capabilities can work together.
We’ll also discuss how to maintain appropriate oversight as these systems become more capable.
Because the goal isn’t to build an AI system that does everything.
It’s to build one that reliably does something worth doing.
And you don’t have to become an engineer to begin.
You just have to understand the work.
Brett Dyess | Brett’s Field Notes
How to Build Your Own AI System · Part 1 of 5