Let me set the scene properly.
The server arrived and I was not prepared for what came out of the box. I'd specced it out with AI's help — CPU, RAM, storage, all very sensible and practical. What I hadn't accounted for was that modern PC builds apparently come with RGB lighting. Lots of it. Shifting, cycling, colorful LED lighting behind a transparent side panel that lets you see every single component glowing like a nightclub.
In the office.
On my desk. Because I didn't have a separate workstation, so the server was just... sitting there. In front of me. Looking festive. With staff throwing me that weird look.
Thank god the see-through panel is on the right!
That was day one.
What we were actually dealing with
I was managing a few companies, handling thirty to forty staff. Like most small businesses in Singapore, we were held together with a combination of determination, muscle memory, and software that had accumulated over the years rather than been deliberately chosen.
The file server was a good example. It had been running since 2003 — old enough to vote — and had grown organically ever since. Inside it we found documents dating back to 2013, buried under layers of folders that made sense to whoever created them and nobody else. Multiple copies of the same file with no clear indication of which was current, which was stale, or why three versions existed in the first place. Finding something was a coin flip. Reports took days to compile because you had to pull data from three different systems, reconcile them manually, and hope nothing had changed by the time you finished.
And the physical documents. Arch folders in cabinets, cabinets spanning half the office floor. Every time someone needed something from that archive, it was an expedition.
The trap I walked into
Here's something nobody warned me about when I started planning all this with AI's help.
On paper, everything sounds straightforward. You describe your problem to an AI, it maps out a solution, the solution sounds logical and achievable. What the AI doesn't tell you — what I now call the digital fine print — is everything that sits between the clean description and the actual working reality. The legacy software that turned out not to be compatible with Windows Server 2025. The recommended tools that looked great in documentation but felt clunky the moment real users touched them. The race conditions, the configuration gotchas, the tools that had been superseded by something better but the AI's training data hadn't caught up yet.
We went through three AI chat platforms before we found one that stuck. AnythingLLM, then Open WebUI, then LibreChat. Each switch felt like a setback at the time. Looking back, it was just learning.
And then there were the ideas that were genuinely too ambitious. At one point I was seriously researching fine-tuning a language model on my own conversations — the idea being that I'd leave behind not just a document archive but a digital version of me. A legacy AI. I still think that's a fascinating idea. It was also wildly beyond what I should have been attempting at that stage.
The AI's ugly side
This is the part the productivity influencers don't show you.
AI hallucinates A LOT. It confidently tells you to do something that turns out to be wrong, or outdated, or technically correct but catastrophically applied to your specific situation. I lost months of container data once because of a configuration change that seemed reasonable and wasn't. I spent entire evenings untangling problems that an AI had talked me into creating in the first place.
When the RTX 4000 Pro finally arrived, my first reaction was mild disappointment. I'd been watching YouTube videos of people unboxing 4090s — massive, triple-fan cards. The Blackwell professional card is nothing like that. Slim, understated, almost boring-looking.
Not the GPU you imagined it to be
It took me a moment to reconcile "this quiet little thing" with "24GB of serious AI compute." Turns out the ones doing the real work don't need to show off.
What we built that actually worked
Thin clients. When a machine broke down, instead of spending thousands on a new PC, we replaced it with a thin client at a fraction of the cost. One thing I discovered: some modern smartphones have a desktop interface mode — connect them to a monitor and keyboard and they work like a thin client.
The document archive. Painful to build. But once stable — genuinely one of the most satisfying things I've built. Every document searchable, tagged, retrievable in seconds. Information on demand, instead of information on expedition.
The AutoCount integration. Our accounting software now talks to everything else. Sales tracking, procurement analysis, commission calculations that used to require hours of manual spreadsheet work now run in seconds.
Purchase tool built with Claude
The picking list. What used to require three people to manually compile and verify now runs with one. The error rate dropped to near zero.
The results
The SaaS contracts we dropped, combined with ending the IT support retainer, freed up thousands a month. For a company our size, this wasn't incremental improvement.
Sales Dashboard completed
It was a genuine step change.
So why did it wind down anyway?
If you built all this, if the savings were real, if the operations improved — why are you winding down?
Because AI can fix a lot of things. It turned out there were things it couldn't fix.
That's Post 3.
Next: What I learned about the limits of technology when the real problems aren't technical.
Top comments (0)