There is a sentence from early computer science that has aged remarkably well:
Bullshit in, bullshit out.
It always meant the same thing: a system can only produce results as good as the data and rules it receives.
With generative AI, this did not become less true. It became more visible. A common misconception today is that AI can somehow smooth over weak foundations, turning unstructured documents into reliable knowledge, vague requirements into clean software, or random prompts into meaningful creative output.
In practice, the opposite is usually true. AI scales existing quality, or existing chaos.
If you want to see concrete local AI workflows, these posts are a direct continuation:
1. Document search: AI does not replace information architecture
A common AI entry point is internal document search, often implemented as RAG.
The starting point is usually the same: thousands of PDFs, Word files, wiki pages, and slide decks. The expectation sounds reasonable:
“Let’s put AI in front of it and it will answer everything.”
Technically, this often works surprisingly well. Organizationally, this is exactly where the hard part starts.
AI does not automatically know
- which document is currently valid,
- which version is approved,
- whether a draft has already been rejected,
- which policy is only historically relevant,
- which SOP applies to one specific location only.
If the system retrieves multiple sources with conflicting statements, it will often generate a plausible answer. That answer can sound convincing and still be wrong.
Not because the model failed, but because key context was missing.
That is why reliable AI programs rarely start with model switching. They start with classic information management:
- clean metadata,
- explicit versioning,
- clear approval status,
- ownership,
- meaningful tagging.
If you want a transparent technical implementation, the article RAG without magic walks through an auditable pipeline from Markdown to retrieval.
2. Agentic coding: The real code is written before coding starts
The same pattern is even more obvious in agentic software development.
Today, AI agents can generate features, refactor code, and even propose architectural changes quickly. That part is impressive. But the difference between short-term working code and long-term maintainable software is created before generation begins.
An agent needs a clear frame:
- How is the project structured?
- Which architecture is being followed?
- Which principles are non-negotiable?
- Which libraries are allowed?
- What coding guidelines apply?
- What is the test strategy?
- How is documentation handled?
Without that frame, AI will usually produce exactly what was implicitly requested: code that runs.
But running code is not automatically good code.
Without guardrails, teams quickly accumulate inconsistent architecture, duplicate implementations, changing patterns, technical debt, and systems that become difficult to evolve.
AI does not remove engineering. It shifts engineering effort further toward architecture and system design.
3. AI music generation: Turning ideas into actual works
For me, the most personal example came from AI-assisted music production.
When I discovered these tools last spring, I was genuinely impressed. A prompt like
energetic electronic folk drum and bass with emotional vocals
can generate a complete track in seconds.
After the initial wow effect, a second impression followed quickly: many results sounded polished, but emotionally generic. Technically strong, artistically interchangeable.
The turning point came when I changed the order. Instead of prompting first, I started by sketching my own material on the MPC Key 37: chords, melodies, basslines, structure, groove.
From there, my musical fingerprint was already in place. I then used AI selectively for arrangement, instrumentation, timbre decisions, and occasional vocal variations.
The results became far more personal. AI was no longer “composing for me”. It was producing with me.
That is where the real value appears: not human versus machine, but human with machine.
AI makes foundations more valuable, not less
All three examples show the same pattern:
AI is not a substitute for structure. It is an amplifier.
- Clean data becomes more valuable.
- Clean architecture becomes more important.
- Solid creative groundwork has more impact.
This also changes what expertise is worth: less raw output production, more design of robust operating conditions.
- Information architecture matters more.
- Software architecture matters more.
- Craft and preparation matter more.
The more powerful AI gets, the more input quality determines output quality.
So maybe the old line deserves a modern extension:
Garbage in, garbage out.
Structure in, excellence out.
AI is not a replacement for good engineering, it makes good engineering visible.