Is AI Still Lacking Systems Thinking?

I've been experimenting with using AI for two different tasks, letting it "manage" everything on its own.

1. Personal Assistant — I configured Google Calendar, gave it a folder named Personal, and told the AI to manage the documents however it saw fit.
2. Political Database — I provided Neo4j (graph database), Garage (S3-compatible storage), and a folder named Politics, then let the AI organize political news and relationships between politicians however it wanted.

The idea was simple: I shouldn't have to care how the information is stored, as long as the AI can keep track of things for me.

Experiment #1: Personal Assistant

This one actually worked pretty well, at least initially.

I used it to keep track of my mom's medical appointments. The AI would create calendar events and maintain a Markdown file (e.g., Mom.md) containing the names of her doctors and their specialties, based on what I asked it to record.

I also used it for TODO lists and meetings.

Everything worked fine until the number of files started growing. Then the AI began giving incorrect or incomplete answers because it wasn't searching through the notes it had previously written.

Experiment #2: Political Database

This one followed a similar pattern.

Initially, I simply told the AI that the primary purpose of the graph database was to discover relationships between individuals, based on events and personal connections. S3 was for storing facts, while the file system didn't have a specific purpose yet, but I made it available anyway.

Whenever I wanted to know something, I'd ask the AI to research it and store the information in those storage systems however it saw fit.

This approach fell apart remarkably quickly. 😅

The first two or three questions seemed to work well. Then, at some point, the AI asked:

"Would you like me to add a field/property to represent relationships between individuals?"

That made me suspicious.

So I opened Neo4j to inspect what it had actually been doing.

And... surprise.

It had modeled the graph database like an RDBMS!

It treated nodes as tables and properties as columns, but hadn't used a single relationship! 😂

A quick Graph Database 101: A graph database consists of nodes and relationships. Unlike an RDBMS, each node represents a record, while a relationship is also a record that connects nodes and describes how they relate to each other.

From an RDBMS perspective, a table is roughly equivalent to a group of nodes sharing the same label.

And that wasn't all.

The database was full of duplicate data. Records were incomplete and inconsistent, and properties seemed to have been created completely arbitrarily.

Basically, a "whatever I feel like" schema.

How Much Do We Actually Need to Tell AI?

Now, people who know how to use AI might say:

"Well, obviously. You didn't give it sufficiently detailed instructions."

But here's the problem: How much do we actually need to tell an AI?

If we believe AI is intelligent, should we really have to spell out every little detail?

And this leads to a broader architectural question:

Should we build specialized systems/applications, or should we simply give AI the necessary tools and let it do the work?

Which direction should we be heading?

If that distinction isn't immediately clear, there's an interesting paper proposing that letting AI search on its own can outperform traditional RAG approaches:

Is Grep All You Need? How Agent Harnesses Reshape Agentic Search

And then there's that famous quote:

"General methods that leverage computation are ultimately the most effective."

Maybe AI just isn't generalized enough yet. 😛

My Takeaway

I've tried this with Claude Fable/Opus, Grok Bot, and GPT-6.1 Sol (though I haven't dared to try Astra yet 😅).

The results have been remarkably similar across all of them.

My conclusion is that AI has been trained to solve problems, so it tends to approach them by following the path of least resistance.

But it hasn't really been trained for systems thinking. (Astra, perhaps?)

What we have today is essentially an automated engine capable of solving almost any problem we throw at it, as long as it doesn't violate the constraints imposed during training.

But what we don't yet have is an engine capable of taking a collection of ad hoc tasks and turning them into a coherent system.

Then again, while I frequently see AI producing code with terrible design, I also see plenty of people happily building applications with AI.

So perhaps...

If the spaghetti code sells, maybe that's good enough. 🙂