NORDVEST was feature complete in two weeks. On week four we moved Northwest Compressor onto it. It's now been another two weeks and I'm still refining. If the story ended at week two, it'd be the kind of AI-coding flex I see on X every day. But the story didn't end at week two. The part after is what I actually want to talk about.
The Seduction
When building is cheap, I build things I don't need.
AI gave me a filter system that would have made sense for a generic SaaS — every field, every operator, every permutation. It looked impressive. It had nothing to do with how our company actually searches for anything. I spent days refining filters that shouldn't have existed in the first place.
Same story with the workflow. AI handed me a textbook SDLC-shaped pipeline:
open → triage → plan → in progress → done → invoiced
Very tidy. Very wrong. What we actually needed was:
open → done → closed
Three states. That's it. The six-state version wasn't broken, it technically modeled the work. But it modeled the work the way a business school case study models the work, not the way an air compressor service company actually runs a ticket. When I ran it by our parts guy Kobe he said he didn't want to click so many buttons - he just wants to see the tickets and send the techs to work them.
In the ye-olden-days before AI accelerated development, I would have never gotten that far because the cost of building it would have forced the question before we got there. Post-AI, I was a few days deep before I realized I'd built a distraction.
The Sanding
The last four weeks have been sanding.
Catching what AI slipped in. Even with close eyes on every diff, things get through. A helper function that duplicates one I already have. A default that makes sense in isolation but fights the rest of the app. Forgetting to use the tenant's timezone versus the user's timezone. There's two save buttons on some pages because I forgot to delete the one that was there before. These are the kinds of bugs that only come out when using the app - specifically someone who's not me.
QA that only real-world use can find. The best test suite in the world can't tell me that the flow feels wrong when I'm standing in a customer's mechanical room with oil on my hands. That's a different kind of bug. I only find it by using the thing.
And then the harder category — the stuff AI simply can't do. Deciding what our workflow really is. Deciding which filters match how we think. Deciding that six states is five states too many. These aren't engineering decisions. They're judgment. AI doesn't have judgment about our company because it doesn't run our company.
The Complacency Trap
Here's the part I have to be honest about. The problem isn't that AI is bad at building (Opus 4.6 is crazy good). It's that AI is so good at building that it removes the friction that used to make me stop and think.
Friction is annoying, but friction is also a gatekeeper. It makes me ask do I actually need this? before spending three days on it. Remove the friction and I'll happily spend three days on something I never should have started. The cost of a wrong turn used to be a week. Now it's an afternoon — which sounds like good news, until I realize I can now take ten wrong turns in the time it used to take to take one.
A skeptic will say that's a discipline problem. Fair. I agree. The post isn't "AI is dangerous." It's "AI raises the cost of undisciplined judgment, because the judgment is now the whole job."
Subtractive Mixing, Again
I wrote a while back about a mentor who taught me, when I was learning to mix sound, to ask what can I remove to get the same outcome? Instead of turning the keyboard up, figure out what's drowning it out.
AI is extremely good at turning things up. Six-stage workflows. Filter systems with every conceivable operator. Helper functions for problems I don't have. It will happily build you a louder mix.
It is not good at telling you what to mute.
That job — the sanding, the subtracting, the deciding — is still mine. NORDVEST's week two looked done because it had everything. NORDVEST's week eight is starting to look actually done because it has the right things, and nothing else.
The fun part got cheaper. The hard part is exactly as hard as it's always been — and the seduction to skip it is stronger than ever.
The sanding is still the job.